diff --git "a/data_all_eng_slimpj/shuffled/split2/finalzzsiwd" "b/data_all_eng_slimpj/shuffled/split2/finalzzsiwd"
new file mode 100644--- /dev/null
+++ "b/data_all_eng_slimpj/shuffled/split2/finalzzsiwd"
@@ -0,0 +1,5 @@
+{"text":" \n* * *\n\n# Knowledge, Technology and Law\n\n* * *\n\nThe relationships between knowledge, technologies, and legal processes are central to the constitution of contemporary societies. As such, they have come to provide the focus for a range of academic projects, across interdisciplinary legal studies and the social sciences. The domains of medical law and ethics, intellectual property law, environmental law and criminal law are just some of those within which the pervasive place and 'impact' of technoscience is immediately apparent. At the same time, social scientists investigating the making of technology and expertise \u2013 in particular, scholars working within the tradition of science and technology studies \u2013 frequently interrogate how regulation and legal processes, and the making of knowledge and technologies, are intermingled in complex ways that come to shape and define each other. This book charts the important interface between studies of law, science and society, as explored from the perspectives of socio-legal studies and the increasingly influential field of science and technology studies. It brings together scholars from both areas to interrogate the joint roles of law and science in the construction and stabilization of socio-technical networks, objects, and standards, as well as their place in the production of contemporary social realities and subjectivities.\n\n**Emilie Cloatre** is a Senior Lecturer in Kent Law School at the University of Kent.\n\n**Martyn Pickersgill** is Wellcome Trust Senior Research Fellow in Biomedical Ethics in the Centre for Population Health Sciences at the University of Edinburgh.\n* * *\n\nLaw, Science and Society series\n\nGeneral editors\n\nJohn Paterson | Julian Webb  \n---|---  \n_University of Aberdeen, UK_ | _University of Melbourne, Australia_\n\n* * *\n\nLaw's role has often been understood as one of implementing political decisions concerning the relationship between science and society. Increasingly, however, as our understanding of the complex dynamic between law, science and society deepens, this instrumental characterisation is seen to be inadequate, but as yet we have only a limited conception of what might take its place. If progress is to be made in our legal and scientific understanding of the problems society faces, then there needs to be space for innovative and radical thinking about law and science. _Law, Science and Society_ is intended to provide that space.\n\nThe overarching aim of the series is to support the publication of new and groundbreaking empirical or theoretical contributions that will advance understanding between the disciplines of law, and the social, pure and applied sciences. General topics relevant to the series include studies of:\n\n\u2022 law and the international trade in science and technology;\n\n\u2022 risk and the regulation of science and technology;\n\n\u2022 law, science and the environment;\n\n\u2022 the reception of scientific discourses by law and the legal process;\n\n\u2022 law, chaos and complexity;\n\n\u2022 law and the brain.\n\n**Titles in this series:**\n\n**Absent Environments**\n\nTheorising environmental law and the city\n\n_Andreas Philippopoulos-Mihalopoulos_\n\n**Uncertain Risks Regulated**\n\n_Edited by Ellen Vos and Michelle Everson_\n\n**The Regulation of Animal Health and Welfare**\n\nScience, law and policy\n\n_John McEldowney, Wyn Grant, and Graham Medley_\n\n**Knowledge, Technology and Law**\n\n_Edited by Emilie Cloatre and Martyn Pickersgill_\n* * *\n\nKnowledge, Technology and Law\n\n* * *\n\nEdited by\n\nEmilie Cloatre and Martyn Pickersgill\n\nFirst published 2015  \nby Routledge  \n2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN\n\nand by Routledge  \n711 Third Avenue, New York, NY 10017\n\na GlassHouse Book\n\n_Routledge is an imprint of the Taylor & Francis Group, an informa business_\n\n\u00a9 2015 Emilie Cloatre and Martyn Pickersgill\n\nThe right of Emilie Cloatre and Martyn Pickersgill to be identified as author of this work has been asserted by them in accordance with sections 77 and 78 of the Copyright, Designs and Patents Act 1988.\n\nAll rights reserved. No part of this book may be reprinted or reproduced or utilised in any form or by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying and recording, or in any information storage or retrieval system, without permission in writing from the publishers.\n\n_Trademark notice_ : Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe.\n\n_British Library Cataloguing in Publication Data_  \nA catalogue record for this book is available from the British Library\n\n_Library of Congress Cataloging-in-Publication Data_  \nKnowledge, technology, and law \/ edited by Emilie Cloatre, Martyn   \nPickersgill.   \npages cm. \u2014 (Law, science, and society)   \nIncludes bibliographical references and index.   \nISBN 978-0-415-62862-4 (hardback) \u2014 ISBN 978-0-203-79760-0 (ebk)   \nI.Technology and law. 2. Science and law. I. Cloatre, Emilie, editor of   \ncompilation. II. Pickersgill, Martyn, editor of compilation.   \nK487.T4K59 2014   \n340\u2032.11\u2014dc23   \n2014013342\n\nISBN: 978-0-415-62862-4 (hbk)  \nISBN: 978-0-203-79760-0 (ebk)\n\nTypeset in Galliard  \nby FiSH Books Ltd, Enfield\n* * *\n\n# Contents\n\n* * *\n\n_Notes on contributors_\n\n_Acknowledgements_\n\n**Introduction**\n\nEMILIE CLOATRE AND MARTYN PICKERSGILL\n\n[**SECTION 1  \nLaw, expertise and public participation**](008_9780203797600_section1.html#sect1)\n\n **1** **Science, uncertainty and the normative question of epistemic governance in policymaking**\n\nSUJATHA RAMAN\n\n **2** **Contingent participation: imaginaries of sustainable technoscientific innovation in the European Union**\n\nMARK L. FLEAR AND THOMAS PFISTER\n\n **3** **The voice of silence: UK patients' silent resistance to the assisted reproduction regulations**\n\nILKE TURKMENDAG\n\n[**SECTION 2  \nObjects and epistemologies in criminal law**](011_9780203797600_section2.html#sect2)\n\n **4** **Unchaining research: processes of dis\/empowerment and the social study of criminal law and investigation**\n\nBARBARA PRAINSACK\n\n **5** **Making the colposcope 'forensic': the medico-legal management of a controversial visualisation device**\n\nGETHIN REES\n\n **6** **Telling tales: some episodes from the multiple lives of the polygraph machine**\n\nANDREW BALMER\n\n[**SECTION 3  \nRegulation, ethics and values**](014_9780203797600_section3.html#sect3)\n\n **7** **Through the thicket and across the divide: successfully navigating the regulatory landscape in life sciences research**\n\nGRAEME LAURIE AND SHAWN H. E. HARMON\n\n **8** **Misconduct hunting: research integrity via law, science and technology**\n\nMARIE-ANDR\u00c9E JACOB\n\n **9** **Financial derivatives and the challenge of performation: where contingency meets contestability**\n\nDONATELLA ALESSANDRINI\n\n[**SECTION 4  \nLaw, technoscience and the stabilization of knowledge**](017_9780203797600_section4.html#sect4)\n\n**10** **Epistemic jurisdictions: science and courts in regulatory (de)centralisation**\n\nDAVID E. WINICKOFF\n\n**11** **Un-knowing exposure: toxic emergency housing, strategic inconclusivity and governance in the US Gulf South**\n\nNICHOLAS SHAPIRO\n\n**12** **A likely story: HIV and the definition of disability in UK employment equality law, 1996\u20132005**\n\nEMILY GRABHAM\n\n**13** **Paper prototypes**\n\nALAIN POTTAGE\n\n_Index_\n* * *\n\n# Notes on contributors\n\n* * *\n\n**Donatella Alessandrini** is Reader in Law in Kent Law School at the University of Kent.\n\n**Andrew Balmer** is Simon Research Fellow in Sociology at the University of Manchester.\n\n**Emilie Cloatre** is a Senior Lecturer in Kent Law School at the University of Kent.\n\n**Mark L. Flear** is a Lecturer in the School of Law at Queen's University Belfast.\n\n**Emily Grabham** is a Reader in Law in Kent Law School at the University of Kent.\n\n**Shawn H.E. Harmon** is Lecturer in Regulation and Risk in the School of Law at the University of Edinburgh.\n\n**Marie-Andr\u00e9e Jacob** is a Senior Lecturer in the School of Law at Keele University.\n\n**Graeme Laurie** is Professor of Medical Jurisprudence in the School of Law at the University of Edinburgh.\n\n**Thomas Pfister** is Director of the Energy Cultures Research Group in the Department of Social Science at Zeppelin University.\n\n**Martyn Pickersgill** is Wellcome Trust Senior Research Fellow in Biomedical Ethics in the Centre for Population Health Sciences at the University of Edinburgh.\n\n**Alain Pottage** is Professor of Law in the Law Department at the London School of Economics and Political Science.\n\n**Barbara Prainsack** is Professor of Sociology at the Department of Social Science, Health and Medicine at King's College London.\n\n**Sujatha Raman** is Deputy Director of the Leverhulme Trust 'Making Science Public' Research Programme in the Institute for Science and Society (ISS) at the University of Nottingham.\n\n**Gethin Rees** is Lecturer in Criminology in the Department of Sociology, Social Policy and Criminology at the University of Southampton.\n\n**Nicholas Shapiro** is a Postdoctoral Researcher in the Sociology Department at Goldsmiths, University of London.\n\n**Ilke Turkmendag** is a Research Associate in the Department of Sociological Studies at the Universtiy Sheffield.\n\n**David E. Winickoff** is Associate Professor of Bioethics and Society in the Department of Environmental Science, Policy and Management at the University of California, Berkeley.\n* * *\n\n# Acknowledgements\n\n* * *\n\nWe are grateful, most of all, to our contributors: for their efforts, ideas, excitement, and patience during the process of this book project. We acknowledge too the work of Colin Perrin, Rebekah Jenkins, and the other staff at Routledge who have facilitated its production. We would also like to thank the UK Arts and Humanities Research Council (AHRC) for sponsoring the 'Technoscience, Law and Society' Research Network. This provided a forum for \u2013 and further propelled \u2013 a number of scholarly engagements that generated debate and exchange which informed the contents of this volume. Martyn gratefully acknowledges too the support of the Wellcome Trust. Last, we are indebted to our home institutions who have generously supported us in this endeavour: the University of Kent and the University of Edinburgh.\n* * *\n\n# **Introduction**\n\n_Emilie Cloatre and Martyn Pickersgill_\n\n* * *\n\nThe relationships between knowledge, technologies, and legal processes are central to the constitution of contemporary societies. As such, they have come to be the focus for a range of academic projects, across interdisciplinary legal studies and the social sciences. Legal scholars are commonly drawn to exploring the legal developments that respond to or are shaped by scientific innovation. The domains of medical law and ethics, intellectual property law, environmental law and criminal law are just some of those within which the pervasive place and 'impact' of technoscience is immediately apparent. At the same time, social scientists investigating the making of technology and expertise \u2013 in particular, scholars working within the tradition of science and technology studies (STS) \u2013 frequently interrogate how regulation and legal processes, and the making of knowledge and technologies, are intermingled in complex ways that come to shape and define each other.\n\nThe interrogation of 'law', 'science', and 'technology' has also worked to render problematic any notion that these are tightly bounded things always already out there in the world, waiting patiently for analytic attention. We might, perhaps, broadly agree on what law and technoscience look like, but seeking to provide minute and enduring definitions only further pluralizes their ontologies. Accordingly, we are sceptical about the possibility of definitively specifiying the natures of knowledge, technologies, and legal processes \u2013 even as we necessarily hold on to imaginaries of each of these in order to work and to live.\n\nIt should be clear, then, that our intention with this book is certainly not to produce any kind of schematic about how 'law', 'science', and 'technology' _should_ relate to one another. Our project is a little different, and in some senses it is modest: to bring together a range of perspectives on these entities, in order to begin to apprehend how different epistemic communities might talk about them and their relationships, and hence how they are talked into existence. Such interdisciplinary conversations around technoscience and law can, we believe, enable the issues at stake to be considered in more critical depth, and bring to light some of the complexity of the interactions between what we know, how we do things, and how we think we should live. Our modest agenda, then, enables us to be more ambitious: by eliding responsibility for answering standard normative questions, we seek to energize reflection regarding (and perhaps the transformation of) the foundations upon which normativity is usually enacted.\n\nScholarship that speaks to our concerns has already been undertaken by a range of authors, and we have no wish to reinvent the wheel. We want to acknowledge our, and our contributors', intellectual debts. In particular, the settlement of rules and the making of knowledge in the contexts of law and science have been drawn into conversation (Latour 2002); the patterns that emerge in the interaction between legal institutions, scientific knowledge, and technological innovation have been interrogated (Brownsword and Yeung 2008; Jasanoff 1997; Aronson and Cole 2009); and conceptual insights from STS have become increasingly deployed to explore the making of legal and political norms and practices (Riles 2011; de Goede 2012; Pottage 2012).\n\nThese explorations have been undertaken within a variety of areas, from environmental regulations, to the practices of biomedicine, as well as in the context of information technology and financial regulations. Medical devices, energy saving technologies, computers, lie detectors, stem cells, and other artefacts, both spectacular and mundane, are now increasingly regarded as being at least partly defined by and constituted through their legality. In the messy networks that stabilize these objects, legal processes thus come to be another site of entanglement between materials, techniques, cultures, and institutions, that are inherently contingent on the deployment of knowledge and technologies, and established and emerging forms of expertise. Technoscience and law, rather than leaning on each other as independent spheres of norm-making, come to co-produce each other (Jasanoff, 2004), while remaining defined by particular systemic modes of enunciation that participate in performing their own uniqueness (Luhmann 1995; Pottage 2012).\n\nScholars have devoted much attention to the links between new scientific possibilities and emerging normative conflicts (as the history of the regulation of reproductive technologies has demonstrated, for example; Ragon\u00e9 and Franklin 1998) and to the various sets of social, legal and democratic questions raised by significant events, or disasters (Petryna 2002). At the same time, studies of practices that are, for different communities, everyday, mundane, or 'discrete' have increasingly enriched our understanding of the many locations in which the entanglement of law and technoscience is visible, and of the various surfaces upon which the production of knowledges is scaffolded, and even concretised, by legal events (Jacob 2012; Riles 2011; Cloatre 2013). Historical engagements with law and expertise demonstrate both the transformation of these over the years through the mutations of the other (Jasanoff 1997), and the great extent to which certain types of tensions and opportunities persevere (Leclerc 2005).\n\nAs we intimated above, underlying much of the questioning of the links between law and science are issues pertaining to the nature of each institution. The portrayal of law and science as separate fields of knowledge and activity once dominated \u2013 and, to an extent, continues to be evident within \u2013 the legal literature. This resulted in a set of normative claims (persistent in much public and legal discourse) regarding how legal processes should (better) emulate the methods of science, and its quest for 'truth' rather than 'settlement'. Similarly, we can see that law is consistently called to be more responsive to technoscientific developments, and even to reformulate its own logics with regard to novel (and tentative) knowledge claims (Pickersgill 2011). As Sheila Jasanoff puts it:\n\n> The representation of law and science as fundamentally different enterprises has given rise to two strikingly recurrent themes in legal writing about science: that of the 'culture clash' between lawyers (or legally trained bureaucrats) and scientists, and the corollary that the culture of law should strive as far as possible to assimilate itself to the culture of science when dealing with scientific issues.\n> \n> (Jasanoff 1997: 7)\n\nSTS-inflected studies of law and science have for some time challenged these claims: the possibility, and desirability, of law becoming 'more scientific' or being made more 'authoritative' through incorporating scientific claims or mimicking its epistemic norms becomes increasingly questionable. STS has demonstrated how science is produced through a heady mix of professional dis\/agreement, the management of uncertainty, the 'dance' between the observer and the observed and between the measure and the measured, the silencing of some voices and the amplification of others, and the ongoing intertwinement of the material and the semiotic. Increasing attention to the role of legal discourse within this cocktail has been suggestive of the extent to which law and science exist in a relationship of co-constitution, or entanglement. Accordingly, early normative claims that law should (or even could) 'follow' science in such predetermined ways became less compelling.\n\nThis reimagining of the relationship between law and technoscience also bears significance in relation to our understanding of law and of broader debates within interdisciplinary legal studies. The work of law and society scholarship, legal anthropology, critical legal scholarship, feminism, and race theory, amongst others, has enjoined questions about what we take law to actually be, over the same period of time as STS was deconstructing and redefining the identities of science (e.g. Constable 2007; Silbey and Ewick 1998). Traditional engagements with the 'impact' of law, in turn, came to be challenged by more fluid understandings of its modes of operation, and its co-constitutive entanglement with routinized practices, complicating further the 'and society' aspect of legal studies (e.g. Cloatre 2008, 2013; Pickersgill 2013a; Jacob 2012). For Alain Pottage,\n\n> there is indeed more society in law than there is in the society that is invoked by studies in 'law and society'. Ironically, however, the effect is actually a retrenchment of the premise that law exists as a singular social instance that it is the business of the theorist to explain.\n> \n> (Pottage 2012: 179)\n\nAccordingly, many assumptions have come to be scrutinized about how law functions, about its actual or imagined role, and about how norms and decisions are shaped or made. Law's internal machinery has been interrogated, and the self-constitution of law as a system of knowledge and authority freshly perceived. The externalities of law, and the meanings law adopts in everyday life, have expanded what our understanding of what legal space might encompass (Sibley and Ewick 2000). Just as boundaries between science and society, nature and culture, that were once regarded as bright and sharp were progressively shown by STS to be fuzzy (at the very least), interdisciplinary scholarship around 'the law' has made the possibility of clearly delineating between this and the spaces it inhabits increasingly questionable.\n\nWe follow this position below, and argue that the insights of STS contribute to a radical revisiting of the making and meaning of law \u2013 a project we call, following other 'expansions' of the remit of STS (e.g. MacKenzie _et al._ 2007), 'social studies of law'. Explorations in law and technoscience, and the joint insights of STS and interdisciplinary legal scholarship in unpacking the processes through which knowledge is made and society ordered, have indeed opened up significant possibilities for questioning the nature of legal processes and the production of technoscience. Before turning to this general claim, however, we engage in further detail with three key elements that interrogations of knowledge, technologies, and law invite us to consider.\n\n## **Understandings of progress and regulatory framings**\n\nMuch of the debate surrounding law and science has revolved around how law responds, or ought to respond, to new scientific developments, risks, and opportunities. The conversations here have often been loaded with significant policy pressures, with governments having to balance ethical concerns with the possibilities offered by new \u2013 and at times controversial \u2013 technologies. The truism that law fails to keep up with scientific and technological innovations has long been deconstructed, as the co-production of law and science, and their inherent interdependency, have progressively been proposed as a more adequate way of imagining these interactions (Jasanoff 2004).\n\nA range of studies have explored how balances are made between contrasting interests regarding perceived needs to enable yet control new technologies, and between the rights and duties of citizens \u2013 as well as how new technologies themselves transform or re-embed particular understandings of citizenship (Flear and Pickersgill 2013; Flear _et al._ 2013; Jasanoff 2011). Genetic technologies have been at the centre of many of these explorations, and the past twenty years have seen an inflation of the number of public debates that surround new scientific opportunities, and the normative challenges that they raise. The regulation of science, and regulation _through_ science, is nonetheless at stake in policy choices made far beyond these examples; for example, when choosing which clean energies to facilitate, how to regulate industries in the face of climate change, or how to engage with nuclear energy.\n\nThese questions all contain complex sets of issues relating not only to the role of law, but also to ideas of development and progress, and of the futures that societies accept, imagine, or strive for. Rather than being simply about either constraining or fostering innovation, normative responses to science and technology can come to establish new social understandings of who 'we' are, and how we hope to live. Legal engagements with technoscience therefore entail particular visions of society, and imply exclusions and silencing, as well as the occasional (re)involvement of various groups and interests (cf. Minnow 1991). The regulatory framing of new technology becomes a useful site to explore how citizens contribute to social futures, how these are constituted, and what is gained or lost in different types of regulatory strategies. Additionally, issues of social justice and (re)distribution are inherently entangled in regulatory choices, as is most visible in debates about patents and ownership over biotechnology and pharmaceutical products (e.g. Hayden 2003; Pottage 2006). Each strategy deployed, and the balancing of rights and duties that are attached to technoscientific activities, will presume particular notions of (un)fairness and (in)justice, and either challenge or perpetuate patterns of exclusion and inequality. The interrelationship between law and scientific innovation is therefore complex, multilayered, and entangled within both choices about technological futures and, more discreetly, about citizenship and social justice.\n\n## **Expertise, citizens, and the making of law**\n\nThe questioning of expertise is also central to exploring the deployment of law, and legal bureaucracies, and has been achieved most thoroughly in the context of the courtroom. There, the ways in which legal and scientific knowledges construct and imagine their own spheres of authority, or achieve a particular 'result', have been charted in contexts ranging from criminal to tort to environmental law. STS scholars have demonstrated how the value of expert knowledge, and of particular technologies, in the context of the courtroom, is itself dependent on negotiation and resolutions that happen both within and outside the courts (Jasanoff 1997). Visions of both law and science in which these are, respectively, conceived of as pure constructions of justice, or of truth, have progressively been rendered more complicated than early narratives of trials might have implied.\n\nExplorations of expertise and the roles played by experts in broadly legal settings have also opened up debates on the nature of democracies. The democratic premise of equal participation in the political sphere is often at stake in a context of (real or perceived) increased reliance on experts in decision-making. At one level, the issue here is to reflect on the balancing that must, or can, be operated between democratic expectations and the necessity of grounding decision-making in informed knowledge. STS scholarship has reconstructed and enriched questions pertinent to this debate. Highlighting the contingency of knowledge, and the entanglements between epistemology, power and politics, leads, for example, to much reflection on how expertise is mobilized and framed by governments (local, national, and supranational). The notion of expertise itself appears (perhaps unsurprisingly) more problematic through analytic engagements with instances of knowledge-making and institutional involvement by 'lay-experts' (Gallon _et al._ 2009; Epstein 1996). Finally, the idea of 'the public', and its homogeneity and positionality with regard to new scientific evidence, has been refined. Indeed, it has been pluralized, with the term 'publics' employed within recent scholarship to underscore the intrinsic heterogeneity of communities, subject positions and identities that are present within 'the public'. Overall, the connections between knowledge and law-making appear to break and reform boundaries in ways that are continuous with broader patterns of power and exclusion that are not specific to technoscience, though at times are amplified by the temptations of technocratic decision-making.\n\nInterrogating the interlinks between expert knowledge and the law has moved from imagining legal and scientific expertise as independent spheres to revisiting them as entangled sites of reciprocal production. As neither law nor science are perceived as isolated sites of social action, questions shift from how expert knowledge can best be used, to how the joint works of law and science proceed to the validation and leveraging of knowledges, often restructuring highly political debates into seemingly technical ones. Tensions re-emerge when this mobility of politics is made apparent \u2013 for example by the displacement to either courts or expert committees of highly political issues.\n\n## **Thinking about law and materiality**\n\nEngaging with law and technoscience invites significant attention to social patterns, events, norms, and citizens, as developed above. However, it also requires a shift of attention to material sites of production, and deployment, of both legal and epistemic processes. Indeed, the role of materiality in shaping, conditioning, and mediating the construction and action of law has been rendered visible by many commentators (Pottage and Sherman 2010; Cloatre and Dingwall 2013; Jacob 2012). This is in part a consequence of the progressive permeation of STS within the study of law. As the social study of technoscience has carefully drawn out the many ways in which materiality 'matters', the seizing of this importance by interdisciplinary legal scholars has opened avenues for new interrogations of the law. Examining specific sites of entanglement of law and technoscience therefore bring to the fore the significance of materiality in the production of norms and the deployment of law. This provides forms of questioning that have not traditionally mobilized legal scholars, but which are now increasingly seen as relevant. For Javier Lezaun:\n\n> The effacement of the labour involved in making legal categories material is evident in the lack of attention of legal scholars to the fabrication of the artefacts that incarnate legal entities. Even when the focus is on 'standardization', or 'harmonization', the emphasis tends to be on the production of texts, rules, and criteria \u2013 in other words, on processes of writing that continue and resemble the textual fabrication of the law.\n> \n> (Lezaun 2012: 38)\n\nNevertheless, the claim that materiality has largely been associated with textual materiality in legal scholarship, though certainly once true, now needs to be nuanced slightly. Increasingly, attention is being paid to how law becomes entangled in other kinds of objects. Much of this has been stimulated by wider interest in actor-network theory and posthumanist philosophy (Barad 2007), and an increased attention within political theory to the import of matter (Bennett 2007).\n\nThis critical reflection is important not only as a way of paying attention to 'things', but also because they often suggest a radical reimagining of social ontologies (Barad 2007). Within these broader theoretical movements, the attention paid to matter within legal scholarship has progressively moved beyond the materiality of texts and towards the broader set of materials that conditions the day-to-day activities of lawyering (Latour 2002; Riles 2000). A further shift directed attention to reflecting on how materials themselves come to embed particular legal scripts and histories, or come to be formed, accepted, and transformed by legal processes (Cloatre 2013; Jacob 2012; Lezaun 2012). In turn, the material environment of the law, and the shaping of environments through the law, have been shown to be essential aspects of our understandings of the nature of legal processes (Philippopoulos-Mihalopoulos 2013). The methodological implications of material explorations of law are, in turn, significant, privileging critical and ethnographic approaches to legal processes \u2013 and fostering further links between legal scholarship and the anthropologies of law and science (e.g. Jacob and Riles 2007).\n\nOverall, at stake in these STS-inflected studies of law are the meanings of things in the shaping of legal processes, and in turn the significance of the law in producing forms of materiality \u2013 and the inevitable interrelationships between these. Such explorations reveal how materials become sites that produce, stabilize, and perpetuate particular kinds of power, and help render apparent (potential) displacements of 'traditional' forms of political action. Objects become socially charged both in regards to what they project of the law, and in what they enable (users of) the law to do.\n\n## **Social studies of law**\n\nSome of the core contributions of STS to the wider social sciences and humanities have been the generation of new forms of critical engagement with the institutions of science and the social meanings of technology \u2013 but also, more broadly, the close documentation of how power\/knowledge are made and deployed in processes of ordering the social. In doing so, STS has participated in challenging and deconstructing categories of sometimes entrenched beliefs, norms, and practices, unmaking assumptions as to what constitutes their foundations. When associated with the rich tradition of critical forms of legal scholarship for the constant interrogation of the meaning and nature of law, of what does (or does not) set it aside from other social institutions and 'regimes of normativity' (Pickersgill 2012), STS opens up questions about the boundaries of the legal space, and enjoins us to ask again: what is law?\n\nWe suggest that articulating legal scholarship and STS has necessary consequences for understandings of law and inevitably results in deep challenges to its institutional uniqueness and singularity as an object of study. If the insights of STS are taken seriously, and integrated with the understandings of legality unfolded through conceptual and empirical interdisciplinary legal scholarship, law seems to become a set of fluidly defined associations within shifting networks, even though its language, institutional landscapes, and modes of deployment may appear specific (or perhaps, more appositely, peculiar). In other words, law can be reimagined as co-constituted rather than presupposed, in the way that social relations and objects are more generally, and this has significant implications for its study.\n\nSTS thus allows for a radicalization of the question of law and, as Pottage suggests when reflecting specifically on materiality, pushes us to question whether a deep engagement of law and STS more generally 'might not actually lead to the dissolution of law as a social instance?' (Pottage 2012: 180). Indeed, if our focus of inquiry turns towards the subtle processes of making and ordering that scholars within STS have tended to pay attention to, the vision of law as a predetermined institution becomes difficult to sustain, as the legal space becomes reimagined as being itself constitutive of, and subjected to, broader and fluid material-semiotic interaction. Processes of settlement can be unpacked in order to explore how things fall into place in a particular way, through the joint work of law, science, and other entangled processes, but the idea of law that transpires from each of these instances is to be determined rather than assumed. It is the analytic agenda that falls out from this set of propositions, with its methodological corollaries, that we term 'the Social Studies of Law'.\n\nWork that we may retrospectively characterize as social studies of law already exists, of course, some of which is referred to above. It has also taken the form of what Annelise Riles (2005) calls a return to the 'technical'; enabled through a closer engagement with practices and the 'micro', and a renewed focus on the details of how legal technicalities and the routinized practices of lawyers matter to broader questions about society. There, STS-inflected studies of the law commonly meet the concerns of legal anthropology, as everyday praxis becomes intrinsic to the constitution of law and its social role. The analysis of legal processes thus becomes inherently dependent on the repeated unpacking of a range of assumptions, discourses, practices, and objects. This, in turn, enables the exploration of the making of particular forms of knowledge (both legal and technoscientific), and of the ways norms, institutions, and the deployment of legal rules are inherently entangled in social, political, and material orders. Enjoined here is a conceptual shift to regard law as a fluid notion that acquires its strength and stability by a range of means (verbal, institutional, textual, material...) through space and time. The methodological implications of this approach bear consideration, as (for instance) a microscopic examination of legal processes becomes a valuable tool through which to interrogate the contingent nature of law.\n\nIf law is, or becomes, something different, through comprehension of its entanglement with other sets of cultural or social practices, then this, as we have suggested above, makes it by 'nature' a fluid object, hard to seize and only temporarily stabilized. Nevertheless, the erosion of boundaries that delineate 'law' from other performances of the social does not mean, however, that technical manifestations of either law or science become irrelevant or redundant for the observer. Just because law and its boundaries are performed by societal praxis does not mean that it is irrelevant: it is a thing that is important because it is believed by many to be _a thing that is important_. Both the thingness and the import are worthy of attention in terms of their construction, and the norms and values that they create, propel and challenge.\n\n## **Knowledge, technoscience, and law**\n\nThe overriding aim of this book, therefore, is to map out contemporary themes and issues that emerge from the close interrogation of the links between law, technologies and forms of knowledge, across thematic fields and disciplinary boundaries. As the dynamics between law and technoscience develop as foci of growing academic interest \u2013 and, it should be noted, of policy and civil relevance \u2013 it remains difficult to find overviews of the range of approaches, ideas, methods, and concepts that shape and direct it. This book seeks to display some of the diversity of strategies, preoccupations, and concepts that animate the scholarship in this area, as well as \u2013 we hope \u2013 provide a means of propelling further theoretical and empirical work in what we have referred to as the social studies of law.\n\nThe chapters which follow are divided into four sections, reflecting key questions and domains of analysis in law and science scholarship.\n\nSection 1, _Law, expertise and public participation_ , interrogates the relationships between expert knowledge and public involvement in legal processes and decision-making. In Chapter 1, 'Science, uncertainty and the normative question of epistemic governance in policymaking', Sujatha Raman interrogates the foundation of public participation in the production of scientific knowledge, through insights from STS and political theory. In doing so, she reflects on the complex role publics play in science-based decision-making. This, in turn, opens up important questions on the role of expert knowledge in public policy, and the positionality of publics in democratic decision-making. Mark Flear and Thomas Pfister continue this thematic focus in Chapter 2, 'Contingent participation: imaginaries of sustainable technoscientific innovation in the European Union'. Here, they focus more specifically on processes of inclusion and exclusion that operate in regulatory decision-making. Looking at the example of EU policymaking, they explore how a narrow definition of sustainable development, based on specific technoscientific imaginaries, has narrowed the space for public participation in EU environmental decision-making. In her essay, Chapter 3, 'The voice of silence: UK patients' silent resistance to the assisted reproduction regulations', Ilke Turkmendag interrogates the nature of participation itself. She achieves this through a case study of how would-be parents engaged with the UK removal of sperm donor anonymity. Turkmendag illustrates how discreet forms of resistance and discontent were deployed in response to the new law, offering a rich illustration of the subtle ways in which publics elect to participate in shaping legal processes.\n\nIn Section 2, _Objects and epistemologies in criminal law_ , contributions turn to explorations of the making and stabilization of knowledge, and of technologies of evidence, in the courtroom. Focused specifically on criminal justice settings, a central focus of much scholarship on law and science, this section also raises broader issues about the role of technoscience in legal settlement. Insights from STS are used by Barbara Prainsack in Chapter 4, 'Unchaining research: processes of dis\/empowerment and the social study of criminal law and investigation', to unpack some of the key assumptions and categorizations that underlie commentaries on criminal justice and forensic science. Focusing on two notions \u2013 the nature of convictions and of convicts, on the one hand, and the role of forensic DNA as a technology of individualization, on the other \u2013 she questions how STS tools can be deployed to critically interrogate unexamined assumptions about the nature of criminal justice processes. A focus on contested criminological technologies is continued in Gethin Rees' analysis of the colposcope in the context of rape trials. In Chapter 5, 'Making the colposcope \"forensic\": the medicolegal management of a controversial visualisation device', Rees analyses its usage and unveils some of the controversies and contestations that surround it and which play out amongst particular communities of clinical, scientific and legal practice. He demonstrates how these debates are rooted in deeper, conflicting visions of the role of forensic medical practitioners in rape trials. Contest and controversy are also of import in Chapter 6, 'Telling tales: some episodes from the multiple lives of the polygraph machine'. Here, Andrew Balmer examines the history of the stabilization of another technology of evidence: the polygraph. His critical analysis demonstrates how a story that has often been told as a determined process of technological development is in fact highly messy, complex, and multidirectional. Balmer suggests that acknowledging the multiple lives of technologies is an important part of making sense of their role in legal systems.\n\nSection 3, _Regulation, ethics and. values_ , provides explorations of the entanglement of regulation and governance with other discourses of normativity, and their practical and political implications. In Chapter 7, 'Through the thicket and across the divide: successfully navigating the regulatory landscape in life sciences research', Graeme Laurie and Shawn Harmon document some of the potential discrepancies between what the law dictates, and what scientists feel is required to govern their work. Looking at the particular example of the regulatory separation between tissue and data, they argue for a more responsive approach to the legal governance of science. Marie-Andr\u00e9e Jacob turns to the regulation of research in Chapter 8, 'Misconduct hunting: research integrity via law, science and technology'. Specifically, she is concerned with how integrity is governed. Jacob explores how self-regulation through technologies has become a central aspect of the governance of research integrity. In doing so, she questions the boundary between legal and technological modes of regulation, and between law and science in the making of ethical norms for scientific practice. The analysis of how a set of financial technologies \u2013 financial derivatives \u2013 can be understood as value-making technologies is the core of Chapter 9, 'Financial derivatives and the challenge of performation: where contingency meets contestability', by Donatella Alessandrini. Employing insights from literatures in the social studies of finance and political economy, she challenges the boundary between technologies and practices. In turn, Alessandrini asks, and helps us to answer, questions about how new forms of regulating and formatting markets can be derived from this understanding of derivatives in their complexity.\n\nFinally, contributions in Section 4, _Law, technoscience and the stabilization of knowledge_ , focus on the making of knowledge, and the joint roles of law and technoscience in producing and stabilizing forms of knowledge. David Winickoff, in Chapter 10, 'Epistemic jurisdictions: science and courts in regulatory (de)centralisation', articulates the role of science in complex models of multi-level governance. Looking at examples from the regulation of food and climate change, he demonstrates how regulatory science shapes sites of global governance, and mediates jurisdictional conflicts. In Chapter 11, 'Un-knowing exposure: toxic emergency housing, strategic inconclusivity and governance in the US Gulf South', Nicholas Shapiro interrogates the production of certitude in legal processes surrounding chemical exposure in emergency housing in the US Gulf South. He traces the role of scientific claims in processes of un-knowing in federal governance and court proceedings, and explores the interplay between scientific and legal knowledges in generating (un)certainty. Another uncertain relationship between science and law is set out by Emily Grabham in Chapter 12, 'A likely story: HIV and the definition of disability in UK employment equality law, 1996\u20132005'. Here, by exploring the position of HIV\/AIDs patients in the face of employment equality law, she problematizes in particular the making of legal temporalities in the set of legal and medical discourses, objects and practices that mattered to the legal processes at play. Consequently, she provokes us into thinking more deeply about the dynamics between law and time. Finally, in Chapter 13, Taper prototypes', Alain Pottage analyses the nature of patents, by considering patent texts as prototypes. Looking back at the history of patent jurisprudence, and drawing on STS and related theoretical approaches, he explores how a logic of prototyping and reproduction is a central aspect of the constitution of patents, the implications of this for the position of ideas and techniques in the making of inventions, and what this might suggest for our understanding of patents as a particular legal and technoscientific object.\n\nOverall, the chapters chart some of the contemporary directions taken by scholarship in law and science, and start to articulate the variety of challenges that emerge when seeking to critically engage with legal and scientific institutions, processes, norms, and practices. Although the contributors share similar concerns for the complexity and fluidity of these, questions are approached from a variety of conceptual perspectives, which speak to and complement each other, and together represent a vantage point from which we might better view this still emerging field of intellectual enquiry. As a result, we hope that this collection will enrich conversations both in interdisciplinary legal studies and in STS, and provide a platform for future debates and critical interrogations of the movements exemplified in the various case studies elaborated herein.\n\n## **References**\n\nAronson, J. D. and Cole, S. A. (2009) 'Science and the death penalty: DNA, innocence, and the debate over capital punishment in the United States', _Law & Social Inquiry_, 34(3): 603\u2013633.\n\nBarad, K. (2007) _Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning_ , Durham, NC: Duke University Press.\n\nBennett, J. (2010) _Vibrant Matter: A Political Ecology of Things_ , Durham, NC: Duke University Press.\n\nBrownsword, R. and Yeung, K. (eds) (2008) _Regulating Technologies: Legal Futures, Regulatory Frames and Technological Fixes_ , Oxford: Hart.\n\nCallon, M., Lescoumes, P. and Barthes, Y. (2009) _Acting in an Uncertain World: An Essay on Technical Democracy_ , Cambridge, MA: MIT Press.\n\nCloatre, E. (2008) 'TRIPS and pharmaceutical patents in Djibouti: an ANT analysis of socio-legal objects', _Social and Legal Studies_ , 17(2): 263\u2013287.\n\nCloatre, E. (2013) _Pills for the Poorest: An Exploration of TRIPS and Access to Medication in Sub-Saharan Africa_ , London: Palgrave Macmillan.\n\nCloatre, E. and Dingwall, R. (2013) '\"Embedded regulation\": the migration of objects, scripts and governance', _Regulation and Governance_ , 7(3): 365\u2013386.\n\nConstable, M. (2007) _Just Silences: the Limits and Possibilities of Modern Law_ , Princeton, NJ: Princeton University Press.\n\nDe Goede, M. (2012) _Speculative Security: The Politics of Pursuing Terrorist Monies_ , Minneapolis: University of Minnesota Press.\n\nEpstein, S. (1996) _Impure Science: AIDS, Activism and the Politics of Knowledge_ , Berkeley, CA: University of California Press.\n\nFlear, M. L. and Pickersgill, M. D. (2013) 'Regulatory or regulating publics? The European Unions regulation of emerging health technologies and citizen participation', _Medical Law Review_ , 21(1): 39\u201370.\n\nFlear, M. L., Farrell, A., Hervey, T. K. and Murphy, T. (eds) (2013) _European Law and New Health Technologies_ , Oxford: Oxford University Press.\n\nHayden, C. (2003) _When Nature Goes Public: The Making and Unmaking of Bioprospecting in Mexico_ , Princeton, NJ: Princeton University Press.\n\nJacob, M.-A. (2012) _Matching Organ with Donors: Legality and Kinship in Transplants_ , Philadelphia: University of Pennsylvania Press.\n\nJacob, M.-J. and Riles, A. (2007) 'The new bureaucracies of virtue', _Political and Legal Anthropology Review_ , 30(2): 181\u2013191.\n\nJasanoff, S. (1997) _Science at the Bar: Law, Science and Technology in America_ , Cambridge, MA: Harvard University Press.\n\nJasanoff, S. (ed.) (2004) _States of Knowledge: The Co-production of Science and Social Order_ , London: Routledge.\n\nJasanoff, S. (ed.) (2011) _Reframing Rights: Bioconstitutionalism in the Genetic Age_ , Cambridge, MA: MIT Press.\n\nLatour, B (2002) _La fabrique du droit_ , Paris: La Decouverte.\n\nLeclerc, O. (2005) _Le juge et l'expert. Contribution \u00e0 l'\u00e9tude des rapports entre le droit et la science_ , Paris: LGDJ.\n\nLezaun, J. (2012) 'The pragmatic sanction of materials: notes for an ethnography of legal substances', _Journal of Law and Society_ , 39(1): 20\u201338.\n\nLuhmann, N. (1995) _Social Systems_ , Stanford, CA: Stanford University Press.\n\nMackenzie, D., Muniesa, F. and Siu, L. (2007) _Do Economists Make Markets: On the Performativity of Economics_ , Princeton, NJ: Princeton University Press.\n\nMinnow, M. (1991) _Making All the Difference: Inclusion, Exclusion and American Law_ , Ithaca, NY: Cornell University Press.\n\nPetryna, A. (2002) _Life Exposed: Biological Citizens after Chernobyl_ , Princeton, NJ: Princeton University Press.\n\nPhilippopoulos-Mihalopoulos, A. (2013) 'Atmospheres of law: senses, affects, lawscapes', _Emotion, Space and Society_ , 7(1): 35\u201344.\n\nPickersgill, M. (2011) 'Connecting neuroscience and law: anticipatory discourse and the role of sociotechnical imaginaries', _New Genetics and Society_ , 30(1): 27\u201340.\n\nPickersgill, M. (2012) 'The co-production of science, ethics and emotion', _Science, Technology & Human Values_, 37(6): 579\u2013603.\n\nPickersgill, M. (2013a) 'How personality became treatable: the mutual constitution of clinical knowledge and mental health law', _Social Studies of Science_ , 43(1): 30\u201353.\n\nPickersgill, M. D. (2013b) 'Sociotechnical innovation in mental health: articulating complexity', in Flear, M. L., Farrell, A., Hervey, T. K., and Murphy, T. (eds), _European Law and New Health Technologies_ , Oxford: Oxford University Press, pp. 323\u2013342.\n\nPottage, A. (2006) 'Too much ownership: bioprospecting in the age of synthetic biology', _BioSocieties_ , 1(2): 137\u2013159.\n\nPottage, A. (2012) 'The materiality of what?', _Journal of Law and Society_ , 39(1): 167\u2013183.\n\nPottage, A and Sherman, B. (2010) _Figures of Invention: A History of Modern Patent Law_ , Oxford: Oxford University Press.\n\nRagon\u00e9, H. and Franklin, S. (1998) _Reproducing Reproduction: Kinship, Power, and Technological Innovation_ , Philadelphia: University of Pennsylvania Press.\n\nRiles, A. (2000) _The Network Inside Out_ , Ann Arbor: University of Michigan Press.\n\nRiles, A. (2005) 'A new agenda for the cultural study of law: taking on the technicalities', _Buffalo Law Review_ , 53: 973\u20131033.\n\nRiles, A. (2011) _Collateral Knowledge: Legal Reasoning in the Global Financial Markets_ , Chicago: University of Chicago Press.\n\nSilbey, S. and Ewick, P. (1998) _The Common Place of Law: Stories from Everyday Life_ , Chicago: University of Chicago Press.\nSection I\n\n* * *\n\n# Law, expertise and public participation\n\n* * *\nChapter 1\n\n* * *\n\n# **Science, uncertainty and the normative question of epistemic governance in policymaking**\n\n_Sujatha Raman_\n\n* * *\n\n## **Introduction**\n\nWhat role should science play in the making of law and public policy? The question is relevant for at least three reasons. First, science both identifies (e.g., potential hazards to health and the environment) _and_ generates problems (e.g., potential hazards posed by research and associated technologies), and policymakers are increasingly required to acknowledge and enable their governance. Second, science offers more detailed knowledge on the specific nature of problems (e.g., at what level do some potential hazards become significant?), ways of addressing them, and their impacts (e.g., might we be able to clean up pollution with microbial forms?). Nor is this restricted to hazards alone or to the natural sciences, as the social\/economic\/policy sciences also identify, generate, specify and offer solutions to aspirations around finance, social security, education, health and so on. Third, the question has become resonant in wider public discourse especially in the UK, with government science advisors, scientists and science popularizers, politicians, journalists and academics debating the subject. In this context, I will explore how we might think about the role that science should play in the policy context from an STS and socio-legal studies standpoint.\n\nIn a recent example of the wider debate over science and policymaking, astronomer\/broadcaster Brian Cox and actor\/writer Robin Ince (2012) point out that science derives its status as the best available approach to understanding nature from the fact that its findings are regularly subject to testing and revision in the face of new evidence. Science, it appears, can therefore offer the grounding required for policymaking because of the rigour of its standards for generating knowledge. However, STS studies cast doubt on this type of argument for 'science-based' policymaking (e.g., Irwin 2008; Jasanoff 1991; Sarewitz 2004; Weingart 1999; Wynne 2010). In this chapter, I examine the grounds for this critique and outline an alternative normative perspective on science-in-policymaking that builds on an emerging political theory of science and democracy (e.g., Brown 2009; Guston 2000; Jasanoff 2003a, 2003b, 2013). This perspective is concerned with the need for what I call 'epistemic governance' or the governance of knowledge-for-governance.\n\nI suggest that, contrary to common assumptions that STS seeks to dispense with or downplay science altogether, a key insight from STS research is that the scientific knowledge which is important for policymaking is a distinctively hybrid object straddling both science and politics. In practice, some of this hybridity has been lost in technocratic discourse about 'sound science' which has in turn led to suggestions for a return to the primacy of elected law-makers with a lesser role for science. However, the notion that science-in-policymaking is a hybrid form of expertise highlights the role of delegation as a principle of liberal democracy. I will explore the implications of the notion of expertise as delegated authority (Jasanoff 2003a) for how we think about problems of scientific uncertainty and public involvement around decision-making in technical domains.\n\nFirst, a few clarifications are in order. A variety of knowledge claims may be advanced in policymaking processes \u2013 some of them dubious or openly contested \u2013 without changing the policy decisions they have been used to justify. But if we are interested in the role that knowledge _ought_ to play in policy, then we want to know how this knowledge is produced, what it consists of, how real and potential disagreements are managed, and what forms of evidence are 'acceptable' in pluralistic societies (Jasanoff 1991). This is the domain of 'epistemic governance' or the governance of knowledge-for-governance. While governance in practice involves diverse actors within and beyond the state and this is no less characteristic of knowledge issues (Irwin 2008), I focus on the normative challenges for science advice produced through and for the state.\n\n## **Technocracy, decisionism \u2013 and populism? Normative models of science\/policy**\n\nThe headline of Cox and Ince's (2012) opinion-piece is as follows: _politicians_ must not elevate mere opinion over _science_ (italics mine). The phrasing of their normative position is telling \u2013 where politics is represented as embodied in a specific group of people (politicians), science lacks similar grounding. Aside from mentioning two famous scientists, the opinion-piece pits politicians against 'science' as a rarefied force. But if the production of science for policy is a social activity as STS and socio-legal scholars highlight, any normative model needs to be grounded in a sense of how science is embodied and made material to legislative\/political questions. In this section, I examine why the normative question of how science should relate to policy has become a question about the proper role of _experts_ vis-\u00e0-vis law-makers and the public, and consider three ways in which this relationship has been traditionally conceptualized.\n\nIf science is taken to be the body of knowledge produced by researchers and contained in journals, then as Stehr and Grundmann (2011) argue, it is unlikely that much of this can be directly relevant to policymaking. When science puts new problems on the policy agenda such as climate change or when evidence of global warming is challenged, it is expertise that is centre-stage, or, more specifically, experts who interpret, translate and put knowledge in context for policymaking and policy debate (Stehr and Grundmann 2012). Individual studies cannot be the basis of policymaking since any one study \u2013 of, say, the link between hurricanes and global climate change or the risk of transmission of a new cattle disease to humans \u2013 will need to be interpreted and put in context. In practice, policymakers _might_ attempt to ground their decisions in a single piece of evidence, but from a normative standpoint, they need to consider expert judgement on the credibility of evidence, the standards to which any one study was done and how its findings compares with others, and how to interpret the implications of multiple sources of evidence. In this respect, the notion of 'science-based' or 'evidence-based' policymaking can be misleading if it implies that evidence speaks for itself unmediated by particular agents.\n\nSo, while political actors may appeal to 'the science' or scientific evidence for grounding policy claims, academic debate on this question has come to revolve around how we ought to understand the role of experts who mediate that evidence and how this role should be positioned in relation to the public and policymakers (e.g., Collins _et al._ 2010; Stehr and Grundmann 2012; Jasanoff 2003a; Kitcher 2001; Turner 2003). To make sense of what is at stake here, it is worth considering two classic normative positions on experts-in-policymaking: _technocracy_ and _decisionism_ (Weingart 1999). Although these emerged in relation to the role of bureaucrats, they have been translated to make sense of experts, and more recently, experts providing scientific advice to government. The technocratic model associated with Comte's seventeenth-century vision is loosely defined as rule by a secular priesthood of experts ('sociologists' in his account) while Weber's decisionist model put political leaders 'on top' in that they were expected to first define the goals of policy while experts ('bureaucrats' in his account) were confined to supplying knowledge of the means for achieving them and monitoring implementation (Millstone 2007). Weber's notion of 'value-free' science rested not so much on a fear that reliable knowledge would be corrupted by politics but on a desire to keep science out of important value-based judgements which Weber saw as the proper domain of sovereign rulers (Proctor 1991). Decisionism therefore presumes a clear divide between values (the domain of politics) and facts (the domain of science). By contrast, the original Comtean vision blurred the boundary between expertise as a scientific project and a political one (Fuller 2007). Derived from a belief in the superiority of expert judgement over both the sovereign and the populace, technocracy in this sense would give experts the responsibility for value judgements _as well as_ factual ones.\n\nThe normative case against technocracy \u2013 in the sense of simply leaving policy decisions in the hands of experts without political or public oversight \u2013 is straightforward. Technocracy assumes elite privilege, the notion that a class of people is somehow specially equipped to the powers of reasoning required for decision-making (Brown 2009; Kitcher 2001). In contemporary discourse, it is rare to find articulations of a pure Comtean vision. Rather, the Comtean inclination _co-exists_ with the Weberian in a way that allows technocratic advocates to clearly separate politics as arbiter of value-based choices and science as provider of facts (which may be subject to revision, but which is held to be the most reliable knowledge available at any point in time).\n\nFor example, Cox and Ince (2012) observe that politicians are free to disregard scientific findings and recommendations, but they then argue that politicians must clarify in scientific terms what it is they are disregarding and why. They acknowledge that policy responses to problems identified by science are not just scientific ones since they have moral, geopolitical and economic elements \u2013 but they see policy and wider political debate on value judgements and choices for action coming _after_ science has supplied the (value-neutral) facts. The typical argument here is: politicians can choose what to do about climate change or decide how to regulate GM crops, _but_ they must proceed on the basis of 'sound science' which confirms that the risk of GM crops is low and that of climate change is high. This position accommodates decisionism in one sense (elected politicians have ultimate responsibility for policy decisions in existing democracies), and a version of technocracy based on the virtues of 'sound science' rather than of a group of people who are supposed to embody science. Appealing to elite privilege is no longer needed, it appears; instead, we have technocracy-without-technocrats, reconciled with the demands of a democracy-of-elected-rulers.\n\nOn occasion, the technocratic\/decisionist nexus has translated into campaigns for getting more people with scientific training into Parliament, the logic being that these political agents would also be scientific agents and hence able to make 'science-based' decisions. This has the virtue of at least partly recognizing how science is embodied. But neither version of contemporary technocratic philosophy engages adequately with the implications for policymaking of scientific disagreements and value judgements within science, nor does it consider the limits of decisionism if this means simply leaving decisions to politicians. Before exploring this further, we might ask: are there normative alternatives beyond technocracy or decisionism?\n\nA third model is commonly characterized as _populism_ , usually by way of critiquing the concept of public involvement in policymaking rather than to advance it. STS scholars have been charged from within the discipline of failing to stand up to the dangers of populism (e.g., Collins _et al._ 2010) and from outside for promoting 'relativist\/postmodernist' approaches to knowledge (Kuntz 2012; Taverne 2005). Later I will show why this is a misreading of STS contributions; for the moment, let us consider the normative case for involving the public at all in 'science-based' policymaking.\n\nOver twenty years ago, the American political scientist, Peter May, set out the problem of what he called policies-without-publics (May 1991). May argued that technical matters such as natural and technological hazards, biodiversity, endangered species and innovation policy carried few incentives for the formation of organized public groups to participate actively in debate. Perceived as remote from everyday concerns with high costs associated with participation for limited benefit, these domains were dominated by scientific\/technical communities with policy discussions occurring in the 'backwaters of political institutions' (May 1991: 194). Yet, far from being a technocratic paradise, policy domains without publics lacked adequate mechanisms for identifying key issues, defining problems and designing policies. For May, actively engaged publics were essential for policymaking \u2013 even in domains permeated by scientific and technical matters. His argument was less about concerns over elite privilege and more about the _substantive_ matter of adequacy of knowledge for decision-making. Writing in the STS tradition but drawing from Dewey, Marres (2007) makes a similar argument that public involvement in policymaking is important for issue formation. These are not populist arguments in the sense of just leaving decisions to 'the people' but rather arguments for acknowledging the factual and value-based complexity of policy matters.\n\nIndeed, despite their prevalence, all three models \u2013 technocracy, decisionism, populism \u2013 are easy to attack insofar as they rest on the misleading language of 'leaving decisions' to one group of agents or the other. In the rest of the chapter, I will elaborate why STS offers normative insights for making a place in policymaking for science without technocracy, the public without populism, and the political process without decisionism. 'Science-based' democratic policymaking may well be a good model to aspire to \u2013 so long as we are clear about what we mean by science and by democracy and how these are embodied in practice. To unpack this, the next section considers insights from over thirty years of STS research on science and policymaking.\n\n## **Uncertainty and the hybrid object of science-in-policy**\n\nSTS insights on science-in-policymaking are rooted in controversies over risk assessment in regulatory policy, especially around environmental issues. Indeed, May's account of the absence of publics around environmental regulation is at odds with numerous experiences detailed in this literature of how science has been openly contested rather than been left in institutional backwaters, generating uncertainty about the validity of specific scientific\/evidentiary assertions used to underpin decisions. These studies also highlight why the public legitimacy of science advice and policymaking has been called into question for failing to engage adequately with the basis for these contestations. Appeals to 'sound science' have been problematized in the process. As I detail below, the key insight from these studies is about how uncertainty has been managed rather than the fact that it exists.\n\nSince the 1970s, at least in the USA, public debate was amply evident around processes for regulating potential environmental hazards (Jasanoff 1990). Rather than providing the basis for public agencies to depoliticize decision-making, science became embroiled in politics as citizen-groups challenged the evidence used to make policy on hazards of industrial chemicals, nuclear power plants, construction of airport runways and other ostensibly esoteric matters. Likewise, industry groups who questioned the need for regulation by using rival scientific studies to challenge evidence offered of hazard also politicized science, a pattern that persists today. Discussions over the credibility and policy implications of different scientific studies that might have been expected to remain within closed circles were made public by groups from different sides of the political spectrum.\n\nBy contrast, systems of regulation in Europe at the time seemed to fit the closed, depoliticized policymaking style described by May but these too came to be challenged. For example, Horlick-Jones and de Marchi (1995) warned that high-profile controversies were eroding the credibility of science advice and threatening to transform European-style negotiated decision-making into US-style gridlock. The link acknowledged in 1996 between a fatal disease in cattle (BSE) and its variant in humans (v-CJD) provoked widespread criticism of earlier government attempts to downplay the risk. As well as BSE, controversies involving the role of science, science advisors and organized publics in policies related to pesticides, genetically modified crops, vaccination, climate change, classification of drugs and other issues have opened up the question of the role that science plays and should play in policymaking. In this context, there are concerns that science is being politicized with the evidence 'cherry-picked' to suit specific ideological interests and uncertainties exaggerated even where there is insufficient warrant.\n\nEarly STS work highlighted several examples of such deliberate\/intentional politicization of science-in-policy, but located this outcome in the nature of science itself. Relevance to policy destroyed the mechanisms by which scientists achieve consensus (Collingridge and Reeve 1986), and the simultaneous increase in policy involvement by scientists accompanied by a declining capacity to resolve political conflict became a central paradox (Weingart 1999). The point here is that science requires some level of containment or suspension of critique in order for knowledge claims to be certified; these claims retain authority insofar as their underlying assumptions remain tacit. But when the political stakes are high, the likelihood of these assumptions being scrutinized and contested is also high if only because it is possible to make different assumptions and draw different inferences (Collingridge and Reeve 1986). When stakes are high, the relationship between science and policy becomes 'over-critical', i.e., characterized by endless debate between competing interest-groups over technical points. Sarewitz (2004) however, traces the difficulty of grounding policies in science to an 'excess of objectivity' rather than political differences alone. This arises from the internal complexity of science (and ultimately, nature itself), since, for example, molecular biologists and ecological scientists study the hazards of GMOs differently with different criteria for what counts as a hazard and as evidence that one exists. For Sarewitz too, the difficulties of 'science-based' policymaking arise from a mistaken view that science _can_ provide a firm basis for decisions.\n\nHowever, others sharing the emphasis on tacit assumptions that underpin science (and which when made explicit are potentially open to contestation), highlight the distinctive nature of 'scientific' work taking place at the interface with policymaking. Rather than making a case for limiting the role of science _per se_ , these authors make one for limiting the role of _scientism_ (Wynne 2010). This means paying attention to how uncertainty, scientific disagreement and plural inputs are framed and managed in the process of generating knowledge for policy. Independent advisory committees constituted by the state, research divisions within government departments and trans-national advisory bodies such as the Intergovernmental Panel on Climate Change (IPCC) produce 'regulatory science' which is different from academic science as it is specifically generated to answer policy questions. Informed by such concepts as boundary work (looking empirically at how boundaries between science and non-science are drawn in different instances often using different standards) and framing (asking how a policy question and the role of scientific knowledge in addressing that question are implicitly defined in some ways as opposed to others), this research highlights ways in which regulatory science can be understood as a hybrid of factual and value-based judgements (e.g., Irwin 1995, 2008; Jasanoff 1990; Wynne 1992).\n\nFor example, in the case of climate science and policy, the knowledge claims at stake are different from purely scientific ones (e.g., how has global temperature changed over time? what is the relationship between carbon dioxide and global temperature change?). In science-in-policy, the questions are instead: how significant is the relationship between carbon dioxide and temperature change for policy? What sort of standards and actions are appropriate? Who bears the burden of proof or how is the burden distributed? How should the knowledge claims of different parties be engaged? These questions are hybrids of science\/politics and cannot be addressed entirely by either side (Jasanoff 2013).\n\nIn practice, regulatory science has not always acknowledged or engaged sufficiently with different claims from publics, underlying value-based assumptions in scientific assessments, or with the limits and conditions of expert judgements (Irwin 1995; Jasanoff 1990; Stilgoe _et al._ 2006; Wynne 2010). For example, where probabilistic judgements are involved, differences between using frequency data where these exist and Bayesian judgements (which are based on expert interpretations) have invited scrutiny. Likewise, different value judgements are involved in preferring Type 2 (false negative) versus Type 1 (false positive) errors in statistical hypothesis testing of hazards (Jasanoff 1991). This type of conditionality of claims is often invisible in invocations of 'sound science'. For example, Irwin (1995) showed that for farmworkers, the 'recommended' method of spraying the 2,4,5-T pesticide assumed by the Advisory Committee on Pesticides (ACP) simply bore no resemblance to conditions on the ground. The ACP argued there was no evidence to link miscarriages\/birth deformities highlighted by workers to pesticides, but this judgement was based on laboratory conditions which were radically different from working conditions. STS has therefore highlighted differences between risk (where the parameters are known and estimates can be made with confidence) and more complex situations of uncertainty, ignorance and indeterminacy where these conditions do not hold (Wynne 1992).\n\nImportantly, this line of work also considers how uncertainty can be _produced_ and perpetuated by actors who insist on absolutist standards of 'scientific proof'. Like scientists in the wider debate have argued, decisionmaking requires some closing down of uncertainty as well as due recognition of tacit assumptions of scientific claims that might be open to challenge. But what tends to be missed is the point that uncertainty is only a problem for a technocratic philosophy that claims science can supply value-free facts as the basis for policy. Epistemic doubt is amplified where science bears such a burden, Wynne (2010) argues, at the expense of the broader public judgements required for collective decision-making. Some scientists have, of course, acknowledged the distinctive nature of science produced for policy. Indeed, nuclear physicist, Alvin Weinberg (1972) was one of the first to make the point with his concept of 'trans-science'. In the UK debate on drugs policy triggered by the government sacking its chief advisor, David Nutt, science writer Sue Blackmore acknowledged that there was little in science that was 'pure fact'; the advisory committee in question necessarily engaged with judgements about what to count as harm and how to weigh different harms as well as physiological evidence. Drawing on STS work, climate scientist Mike Hulme (2009) makes a crucial point: scientific knowledge is transformed through its interaction in the social world and what will count as knowledge for policy may be different from what counts as such for research\/academic science. What then are the normative implications of these insights on the nature of scientific knowledge (or more accurately, expertise) produced for policymaking?\n\n## **Back to decisionism? Or, epistemic governance as a problem of delegation**\n\nFrom the above diagnosis, different conclusions have been drawn on the role that science can and should play in decision-making. One response seems perilously close to decisionism; for example, the argument that if science is too complex and plural to be able to underpin policy decisions, policymakers must first sort out what are really value-based conflicts using the normal negotiating procedures of democratic politics (Collingridge and Reeve 1986; Sarewitz 2004). This seems to re-draw the boundary between factual claims and value choices, only giving primacy to the latter and a notion of politics that is stripped of the ability to draw on the former. By contrast, Collins _et al._ (2010) strongly underline the need for scientific expertise in policymaking, but still do so on the basis of a clean separation between the role for experts (provide facts) and the role for politics (adjudicate on values). Pielke (2007) acknowledges the difficulty of this boundary, but his normative vision \u2013 especially in domains where there is strong disagreement on both factual and value questions \u2013 is one of scientists being 'honest brokers' between different policy options with their varying knowledge bases rather than advocating a particular position.\n\nBut the return to decisionism fails to get to the heart of the idea that 'science' \u2013 in the sense of expert advice for policymaking \u2013 is supposed to be a form of reason, a hybrid of scientific and political or value judgement. In addition to transforming the received view of science-in-policymaking, this notion also provides the normative ballast to challenge politics as business-as-usual bargaining. Although bargaining may be the normal mode of governmental politics, the notion of politics assumed in the debate on science advice must allow for aspirations of legitimacy. Legitimacy necessarily involves normative judgements of better or worse decisions (not least from publics who may wish to challenge business-as-usual politics), judgements that rely on knowledge as Ezrahi (1990) writing on the origins of liberal-democratic thought clarifies. Ezrahi highlights a key philosophical tension in the eighteenth century between, on the one hand, the values of liberty, equality and Enlightenment notions of contingency, and on the other the case for order and governance. If absolutist rule was no longer justifiable in the absence of a divinely ordained vision of the world, if people (male, property-owning) are free and equal and capable of disagreeing with each other, how then was order possible or even justifiable? How could a select few be legitimately responsible for governing the rest? Putting aside the exclusions embedded in the definitions of freedom and equality at the time, the answer for classical theorists of liberal democracy came from the Scientific Revolution. The notion of modern, experimental science as producing universal (depersonalized) knowledge through methods that were transparent to all made it possible to develop a case for liberal-democratic rule based on public knowledge accessible to the governed and where rulers' actions were observable by citizens.\n\nSo, the expectation that public knowledge informs the practice of government and that this practice can, in principle, be evaluated by the public is central to the normative idea of liberal democracy. This means the question of the proper role for science-in-policymaking cannot be resolved by giving primacy in Weberian-decisionist fashion to a fact-free politics in response to the tribulations of value-free notions of technocracy\/scientism. How though can the central problem for epistemic governance, namely, the uncertainty entailed by _multiple_ knowledge claims on a policy issue and the potential for tacit assumptions in any of these to be contested, be addressed? The principles I outline below are threefold.\n\nFirst is the normative principle of _transparency_ in mechanisms for handling plurality of knowledge. Advisory institutions already have mechanisms for dealing with epistemic disagreement amongst experts, for better or for worse. These may differ across regulatory cultures as Jasanoff (1991) shows, highlighting the problems of overly formal, quantitative methods for resolving discrepancy between different studies that needed to be aggregated to determine carcinogenic risk (in her case study). Rule-based risk analysis attracted criticism for false precision, the failure to be transparent about value judgements made in statistical assessments, and ironically, for being insufficiently scientific in its inability to be responsive to new knowledge. By contrast, the informal, negotiated methods of British regulatory culture have also been critiqued for the failure to consider public and scientific challenges to established risk assessment based on animal models in toxicology for regulation of industrial chemicals (Diggle 2007). This again underlines the need for reasoned judgement and for this judgement to be more transparent to external scrutiny. Hulme (2009) notes that the use of subjective (Bayesian) assessment to estimate risk of future events in the face of uncertainty is an accepted method in science; likewise, when stakes are high and there is a need for timely decisions to be made, methods to develop consensus amongst conflicting expert assessments may be entirely justified. But the substantive legitimacy of such resolutions of uncertainty rests on making their basis public, even if transparency cannot guarantee the closure of controversy \u2013 outcomes in practice will rest on how different political actors mobilize in response.\n\nSecond, normative principles including transparency are necessarily interpreted in the _context_ of judgement about what is the public interest. There may be good reasons to limit or challenge demands for transparency if these arise from purely private-corporate interests oriented towards the deliberate deconstruction of scientific evidence to ward off regulation (see Jasanoff 2006). In addition, criteria for what counts as good judgement will vary across political systems (Jasanoff 2005), though these can in turn evolve through experience and debate as we see in the UK case. An interest in normative principles does not mean that these differences should be erased.\n\nThird and most important, focusing on experts rather than 'science' in policymaking brings to the forefront the _delegation_ of specialist judgements to experts on par with more widely recognized forms of delegation of decisions by the public to elected representatives. Expertise is a 'form of delegated authority' (Jasanoff 2003a) which means that those involved in producing regulatory science through their role in advisory bodies and science-based administrative agencies are specifically authorized to produce knowledge in the public interest. This delegation of expertise in turn invites a series of political questions similar to those raised around more familiar administrative responsibilities in liberal-democratic societies (Brown 2009; Guston 2000; Jasanoff 2003a). How adequate are the structures through which delegation is institutionalized? How do agents to whom specific roles are delegated fulfil their role or, how _accountable_ are they to the system from which they derive their authority to offer advice?\n\nIf public knowledge for policymaking has been delegated to experts, to whom are experts accountable? Insofar as the debate around science-in-policymaking has been framed in terms of 'science' or 'evidence', the typical response is that experts are ultimately accountable to Nature (thus making it difficult to see where the public fits in), while politicians are authorized by 'the people' to act on their behalf (Cox and Ince, 2012). But if regulatory scientists produce knowledge to address questions of public and policy interest, these experts are indeed accountable to the public. The idea is not that a representative sample of publics must be on advisory panels, but that publics have the right to speak back, to challenge, to articulate different positions, to raise alternative interpretations of the evidence, or to rein in the power of experts. The public do not give up that right -they 'grant to experts a carefully circumscribed power to speak _for them_ on matters requiring specialized judgement' (Jasanoff 2003a: 158, italics mine).\n\nIn sum, expert advisors straddle the science\/politics interface insofar as they are capable of speaking authoritatively on the public good, though they may sometimes fail in this task (in the sense of speaking only for a limited and contestable vision of the public good, e.g., 'what is good for UK plc') or make unwarranted assumptions about what publics want or will accept. In the course of this, they are called upon to 'open up' science and scrutinize the basis of conflicting knowledge claims, but also 'close down' some of this uncertainty in delivering their judgements for policymaking. How they perform and manage this tension lies at the heart of the normative challenge of epistemic governance. Unlike scientists, expert advisors are required to engage with challenges by publics who may emerge around specific issues (but are not expected to be necessarily represented in all expert committees as caricatured in discussions of public participation). Where they have focused on too narrow a range of scientific and public inputs, they have been found wanting; likewise, their role has been problematic when they have failed to adequately bound the problem of scientific disagreement\/uncertainty for policy (in their role _as_ experts, not as academic scientists for whom endless disagreement is not, in principle, a problem). So, if the problem of epistemic governance is one of delegation of specialist knowledge to experts, then the issue on which STS sheds light is the nature and manner of public accountability of experts. Accountability in the deliberative democratic tradition is understood in terms of giving public reasons for judgements which demonstrate how multiple perspectives have been engaged with. Hence, 'to the extent that public challenges to expert claims receive an adequate response, such challenges may actually strengthen expert authority' (Brown 2009: 217).\n\n## **Conclusion**\n\nRobert Dahl, distinguished theorist of democracy, defines delegation as a 'revocable grant of authority, subject to recovery by the demos' (Dahl 1997: 114). Building on an emerging political theory of expertise in STS, I have argued that it is this issue which has been at the heart of recent normative controversy around the role of science-in-policymaking. Although STS research highlights scientific uncertainty and articulates a case for opening up science-in-policy to wider public input, this analysis has been carried out in the context of how 'regulatory science' has worked in practice and how it _ought_ to work in order to fulfil a key role at the interface of science, the public and policymaking. This is fundamentally a question of the necessary delegation of specialist knowledge to expertise, the nature of that knowledge, the question of who counts as an expert, and the need for government science advisors in liberal democracies to recognize and engage with multiple criteria for knowledge (including _within_ science).\n\nIn conclusion, I turn to key dilemmas that arise in response to this account of epistemic governance as a hybrid of science, politics and publics. First, does the notion of delegating knowledge for policymaking to experts imply a return to assumptions of elite privilege? If as I have argued, delegation rests on the public having the right to speak back, elite privilege is not an issue. Rather, delegation rests on the inevitable epistemic division of labour in complex societies and a political structure for governing this system for obtaining specialized knowledge for policymaking.\n\nSecond, the case for public engagement around science-based policymaking has been challenged on very different grounds: one, that in the absence of clear limits, science is devalued and governance degenerates into populism (Collins _et al._ 2010); two, and on the contrary, that it idealizes 'democratic citizens', many of whom are unable to inhabit the role of scrutinizer of science (Kerr 2004) and three, that it depoliticizes governance by seeking to tame genuine conflicts (Thorpe and Gregory 2010). From the perspective set out in this chapter, public engagement involves far more \u2013 and less \u2013 than implied by these very different critiques. The concept is far less radical than implied by a cartoon-vision of participation in which citizens are put in the position (or given the burden) of deciding on matters of scientific fact including those on which they have little knowledge. But it is more radical in the sense that more is demanded of expert committees to whom epistemic judgements are delegated. Expert committees are called upon to become more diverse 'mini-republics' (Brown 2003; Jasanoff 2003a), to consider different ways of framing the issues at stake and interpreting the facts that matter (Jasanoff 2003b), and negotiate collective judgements which clarify how they reconcile different ways of conceptualizing the public interest. None of this means that publics decide or that these mini-republics absorb-and therefore displace \u2013 political conflict. Rather, the notion of public engagement helps articulate the essential nature of _public_ delegation of expertise to authorized groups, the demands arising from this relationship, its limits and its possibilities.\n\nThird, the focus on scientific advisory systems which sift the knowledge that has already been produced by science does not address the fundamental question of how science itself is to be organized. For example, some argue that there needs to be greater political representation in decisions about the kind of scientific knowledge to fund in order to cope with the systemic marginalization of some questions (Brown 2009; Fuller 2007). Fourth, and related to the above is the question of how political institutions shape the independence of expertise. So, is it even possible for experts to be independent in a depleted public sector and in the context of demands for science to contribute to 'UK plc'? If scientists conduct their work in partnership with state and corporate interests as Hind (2012) argues, can expertise be independent or impartial? Delegation only works when there is the possibility to hold to account. Formal accountability requirements only work if there is scrutiny by citizens and a response by officials (Brown 2009). But Hind's point is that the public needs the _means_ for this to happen, i.e., to form a public. In the present era, this is a challenge of tall order but one that further reinforces the need for independent expertise.\n\n## **References**\n\nBeddington, J. (2011) 'We need both scepticism and consensus', _New Scientist_ w\u00adw\u00adw\u00ad.\u00adn\u00ade\u00adw\u00ads\u00adc\u00adi\u00ade\u00adn\u00adt\u00adi\u00ads\u00adt\u00ad.\u00adc\u00ado\u00adm\u00ad\/\u00adb\u00adl\u00ado\u00adg\u00ads\u00ad\/\u00adt\u00adh\u00ade\u00ads\u00adw\u00ado\u00adr\u00add\u00ad\/\u00ad2\u00ad0\u00ad1\u00ad1\u00ad\/\u00ad0\u00ad2\u00ad\/\u00ads\u00adc\u00adi\u00ade\u00adn\u00adc\u00ade\u00ad-\u00ads\u00adc\u00ade\u00adp\u00adt\u00adi\u00adc\u00adi\u00ads\u00adm\u00ad-\u00ada\u00adn\u00add\u00ad-\u00adc\u00ado\u00adn\u00ads\u00ade\u00adn\u00ads\u00ad.\u00adh\u00adt\u00adm\u00adl (accessed 30 August 2013).\n\nBrown, M. B. (2003) _Science in Democracy_ , Cambridge, MA: MIT Press.\n\nCollins, H., Weinel, M. and Evans, R. (2010) 'The politics and policy of the Third Wave: new technologies and society', _Critical Policy Studies_ , 4 (2): 185\u2013201.\n\nCollingridge, D. and Reeve, C. (1986) _Science Speaks to Power: The Role of Experts in Policy Making_ , London: Pinter.\n\nCox, B. and Ince, R. (2012) 'Politicians must not elevate mere opinion over science', _New Statesman_ w\u00adw\u00adw\u00ad.\u00adn\u00ade\u00adw\u00ads\u00adt\u00ada\u00adt\u00ade\u00ads\u00adm\u00ada\u00adn\u00ad.\u00adc\u00ado\u00adm\u00ad\/\u00ads\u00adc\u00adi\u00ad-\u00adt\u00ade\u00adc\u00adh\u00ad\/\u00ads\u00adc\u00adi\u00ad-\u00adt\u00ade\u00adc\u00adh\u00ad\/\u00ad2\u00ad0\u00ad1\u00ad2\u00ad\/\u00ad1\u00ad2\u00ad\/\u00adb\u00adr\u00adi\u00ada\u00adn\u00ad-\u00adc\u00ado\u00adx\u00ad-\u00ada\u00adn\u00add\u00ad-\u00adr\u00ado\u00adb\u00adi\u00adn\u00ad-\u00adi\u00adn\u00adc\u00ade\u00ad-\u00adp\u00ado\u00adl\u00adi\u00adt\u00adi\u00adc\u00adi\u00ada\u00adn\u00ads\u00ad-\u00adm\u00adu\u00ads\u00adt\u00ad-\u00adn\u00ado\u00adt\u00ad-\u00ade\u00adl\u00ade\u00adv\u00ada\u00adt\u00ade\u00ad-\u00adm\u00ade\u00adr\u00ade\u00ad-\u00ado\u00adp\u00adi\u00adn\u00adi\u00ado\u00adn\u00ad-\u00ado\u00adv\u00ade\u00adr\u00ad-\u00ads\u00adc (accessed 30 August 2013).\n\nDahl, R. (1997) 'Procedural democracy', in R. Goodin and P. Pettitt (eds), _Contemporary Political Philosophy: An Anthology_ , Chichester: Wiley.\n\nDiggle, R. (2007) _Regulatory science and uncertainty in the risk assessment of pesticide residues_ , PhD thesis, University of Nottingham h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adt\u00adh\u00ade\u00ads\u00ade\u00ads\u00ad.\u00adn\u00ado\u00adt\u00adt\u00adi\u00adn\u00adg\u00adh\u00ada\u00adm\u00ad.\u00ada\u00adc\u00ad.\u00adu\u00adk\u00ad\/\u00ad1\u00ad4\u00ad5\u00ad1\u00ad\/\n\nEzrahi, Y. (1990) _The Descent of Icarus: science and the Transformation of Contemporary Democracy_ , Cambridge, MA: Harvard University Press.\n\nFuller, S. (2007) _New Frontiers in Science and Technology Studies_ , Cambridge: Polity.\n\nGuston, D. (2000) _Between Politics and Science_ , Cambridge: Cambridge University Press.\n\nHind, D. (2012) _The Return of the Public: Democracy, Power and the Case for Media Reform_ , London: Verso Books.\n\nHorlick-Jones, T. and de Marchi, B. (1995) 'The crisis of scientific expertise in fin de si\u00e8cle Europe', _Science and Public Policy_ , 22(3): 139\u2013145.\n\nHulme, M. (2009) _Why We Disagree about Climate Change: Understanding Controversy, Inaction and Opportunity_ , Cambridge: Cambridge University Press.\n\nIrwin, A. (1995) _Citizen Science: A Study of People, Expertise and Sustainable Development_ , London: Routledge.\n\nIrwin, A. (2008) 'STS perspectives on scientific governance', in E. J. Hackett, O. Amsterdamska, M. Lynchand and J. Wajcman (eds), _The Handbook of Science and Technology Studies_ , Cambridge, MA: MIT Press.\n\nJasanoff, S. (1990) _The Fifth Branch: Science Advisers as Policymakers_ , Cambridge, MA: Harvard University Press.\n\nJasanoff, S. (1991) Acceptable evidence in a pluralistic society', in D. G. Mayo and R. D. Hollander (eds), _Acceptable Evidence: Science and Values in Risk Management_ , New York: Oxford University Press.\n\nJasanoff, S. (2003a) '(No?) Accounting for expertise', _Science and Public Policy_ , 30 (3): 157\u2013162.\n\nJasanoff, S. (2003b) 'Technologies of humility: citizen participation in governing science', _Minerva_ , 41 (3): 223\u2013244.\n\nJasanoff, S. (2005) _Designs on Nature: Science and Democracy in Europe and the United States_ , Princeton, NJ: Princeton University Press.\n\nJasanoff, S. (2006) 'Transparency in public science: purposes, reasons, limits', _Law and Contemporary Problems_ , 69(3): 21\u201345.\n\nJasanoff, S. (2013) 'The science of science advice', in Doubleday R and Wilsdon J. (eds), _Future Directions for Scientific Advice in Whitehall_ , Available at w\u00adw\u00adw\u00ad.\u00adc\u00ads\u00ada\u00adp\u00ad.\u00adc\u00ada\u00adm\u00ad.\u00ada\u00adc\u00ad.\u00adu\u00adk\u00ad\/\u00adm\u00ade\u00add\u00adi\u00ada\u00ad\/\u00adu\u00adp\u00adl\u00ado\u00ada\u00add\u00ads\u00ad\/\u00adf\u00adi\u00adl\u00ade\u00ads\u00ad\/\u00ad1\u00ad\/\u00adf\u00add\u00ads\u00ada\u00adw\u00ad.\u00adp\u00add\u00adf (accessed 21 September 2013).\n\nKerr, A. (2004) _Genetics and Society: A Sociology of Disease_ , London: Routledge.\n\nKitcher, P. (2001) _Science, Truth, and Democracy_ , Oxford: Oxford University Press.\n\nKuntz, M. (2012) 'The postmodern assault on science', _EMBO Reports_ , 13(10): 885\u2013889.\n\nMarres, N. (2007) 'The issues deserve more credit pragmatist contributions to the study of public involvement in controversy', _Social Studies of Science_ , 37 (5): 759\u2013780.\n\nMay, P. J. (1991) 'Reconsidering policy design: policies and publics', _Journal of Public Policy_ , 11(2): 187\u2013206.\n\nMillstone, E. (2007) 'Can food safety policymaking be both scientifically and democratically legitimated? If so, how?', _Journal of Agricultural and Environmental Ethics_ , 20(5): 483\u2013508.\n\nProctor, R. (1991) _Value-free Science?: Purity and Power in Modern Knowledge_ , Cambridge, MA: Harvard University Press.\n\nPielke, R. A. (2007) _The Honest Broker: Making Sense of Science in Policy and Politics_ , Cambridge: Cambridge University Press.\n\nSarewitz, D. (2004) 'How science makes environmental controversies worse', _Environmental Science & Policy_, 7(5): 385\u2013403.\n\nStehr, N. and Grundmann, R. (2011) _Experts: The Knowledge and Power of Expertise_ , London: Routledge.\n\nStilgoe, J., Irwin, A. and Jones, K. (2006) _The Received Wisdom: Opening Up Expert Advice_ , London: Demos.\n\nTaverne, D. (2005) _The March of Unreason: Science, Democracy, and the New Fundamentalism_ , Oxford: Oxford University Press.\n\nThorpe, C. and Gregory, J. (2010) 'Producing the post-Fordist public: the political economy of public engagement with science', _Science as Culture_ , 19(3): 273\u2013301.\n\nTurner, S. (2003) _Liberal Democracy 3.0: Civil society in an Age of Experts_ , London: Sage.\n\nWeinberg, A. M. (1972) 'Science and trans-science', _Minerva_ , 10(2): 209\u2013222.\n\nWeingart, P. (1999) 'Scientific expertise and political accountability: paradoxes of science in politics', _Science and Public Policy_ , 26(3): 151\u2013161.\n\nWynne, B. (1992) 'Uncertainty and environmental learning: reconceiving science and policy in the preventive paradigm', _Global Environmental Change_ , 2(2): 111\u2013127.\n\nWynne, B. (2010) 'When doubt becomes a weapon', _Nature_ , 466(7305): 441\u2013442.\n\n1. For example, the House of Lords Select Committee on Science and Technology published a widely cited report on Science and Society in 2000 in the context of controversy over the role of science advice in the BSE crisis and protests over GM crop trials. The think-tank Demos published several reports on science policy and specific areas of policy in the 2000s as did a number of learned societies including the Royal Society. In the 2010s, blogs on science policy matters appear regularly in the _Guardian_ and elsewhere on the internet.\n\n2. The status of social science as a 'scientific' way of understanding social phenomena is more contested even from within, but the broad acknowledgement that it involves or should involve systematic methods (thus meriting the status of specialist knowledge) is less controversial.\n\n3. For example, some of the evidence used by Coalition government ministers in the UK at the time of writing to justify economic austerity, social security cuts or sweeping changes in education has been challenged but to no policy effect.\n\n4. The term 'science-based' policymaking or 'scientific governance' are commonly used to signal this domain, but since STS scholars take issue with the received view of science and open up the question of how different knowledge claims should be engaged with in policymaking, 'epistemic governance' captures it better.\n\n5. Though their argument is perhaps more complex than suggested by the headline.\n\n6. It should be noted that the term 'technocracy' emerged later in the twentieth century, but its roots are usually traced to Auguste Comte and Henri Saint-Simon.\n\n7. Cox and Ince use the example of climate change, but see Wynne (2010) for a comparison of how science-based arguments for policy are articulated in the case of climate change as opposed to GMOs.\n\n8. For example: http:\/\/societyofbiologyblog.org\/do-we-need-more-scientists-in-parliament\/ (accessed 21 September 2013).\n\n9. In keeping with the pluralist tradition in political science, he defined publics (rather than 'the public') as 'professional associations, producer groups, consumer groups, trade groups, public interest groups, neighborhood groups, or other groups surrounding common issue interests' (May 1991: 190).\n\n10. May cited earthquakes, mass-produced vaccines and household radon as examples of hazards. He also added 'budget policy', an area defined by economics, but one I am bracketing for reasons of space.\n\n11. See Cox and Ince (2012); also the argument set out by former UK Chief Scientific Advisor, John Beddington (2011).\n\n12. www.theguardian.com\/commentisfree\/2009\/nov\/03\/drugs-science-evidence (accessed 21 September 2013).\nChapter 2\n\n* * *\n\n# **Contingent participation**\n\nImaginaries of sustainable technoscientific innovation in the European Union\n\n_Mark L. Flear and Thomas Pfister_\n\n* * *\n\n## **Participation in the governance of uncertain futures**\n\nCitizen or public participation (these tend to be used interchangeably, see: Irwin 2006a: 299) is often valorised in the governance of technoscience, both as a good in itself and as an essential component of legal and regulatory decision-making. In this chapter we consider how citizen participation is contingent on imaginaries. Our case study for the latter investigation is sustainable development as it has been linked to and operationalised in the European Union's (EU's) broader agenda on technoscientific innovation. At least in the global West, democracy and citizenship are so deeply embedded that citizens and publics always matter and 'no participation' is not an option. Indeed, there is increasing awareness of diverse citizenry and publics voicing and often contesting elite claims about public interests and needs, for example, with regard to risks and benefits of new technologies or (their attendant) environmental hazards (Fiorini 1990). A particularly important reason for the focus on participation is the recognition that complex scientific and technological systems have repeatedly and increasingly been unable to prevent \u2013 and can even actually produce \u2013 accidents, disasters and risks (Beck 1986; Beck 2009; Giddens 1998).\n\nThis has prompted a turn towards the integration of techniques and practices into decision-making that together can be understood as constituting 'technologies of participation'. These are a means through which citizens regulate and are themselves regulated into producing (procedurally if not substantively) legitimate legal and regulatory decisions in relation to new and emerging technologies (Flear and Pickersgill 2013). As such participation can be seen as a way of assuaging public fear and anxiety about the risks and ethics of technoscientific development, maintaining faith in progress, and the ability of law and regulation to govern in the face of uncertainty. Participation is also implicated in delineating the boundaries of responsibility, helping to distribute credit for success and, more importantly, accountabilities in the event of failure (Black 2005).\n\nAt the same time, faith in progress through the continued harnessing of technoscience, and of enhanced global development, have also been frustrated by the idea that this might not be _sustainable_ , impelling a turn to the multi-faceted and essentially contested concept of 'sustainable development' (Davidson 2011; Dresner 2002; Irwin 2006b; O'Riordan 2009). Here, we examine how this concept is defined, detailed, and mobilised in the context of the EU. The latter develops its approach from the widely cited and influential definition found in the Brundtland Report (Brundtland), which famously describes sustainable development as being about ensuring that humans can meet 'the needs of the present without compromising the ability of future generations to meet their own needs' (World Commission on Environment and Development 1987: 43). By contrast, the second part of the definition is regularly neglected:\n\n> [sustainable development] contains within it two key concepts: the concept of 'needs', in particular the essential needs of the world's poor, to which overriding priority should be given; and the idea of limitations imposed by the state of technology and social organization on the environment's ability to meet present and future needs.\n> \n> (ibid.)\n\nWithin the EU, sustainable development has been increasingly linked to technoscientific innovation, producing a specific framing: 'sustainable technoscientific development'. With regard to non-technical narratives and visual materials we outline the construction of this frame and how it produces and cements a gradual move away from Brundtland's emphasis on global equity and the limits of economic growth. In doing so, we highlight imaginaries or collectively held symbolic views and architectures that help to manufacture and maintain social purpose and order, and which thereby help to legitimate formally accountable power: the frame itself, which is aimed at producing a _knowledge-based economy_ inhabited and seemingly produced by its _innovator citizens_. Crucially, these imaginaries narrow the conditions of possibility for participation in this area, rendering it as a private productive activity in the marketplace. Innovators are ranked above consumers and 'others', and the image of a united Europe aiming for sustainability paradoxically limits the space for democratic-political debate in governance.\n\n## **Technologies of participation and imaginaries**\n\nThe sheer diversity of the techniques and practices that together constitute technologies of participation prevents more than a limited overview. Participation can occur _ex ante_ or 'upstream' (Mandel 2009; Tait 2009; Wilsdon and Willis 2004), in making the decision itself, or _ex post_ such as when technoscience enters the marketplace and can be purchased by consumers. It includes attempts at public deliberation and involvement that can be more or less extensive, ranging from input into technoscientific development and agenda-setting ahead of the decision to regulate, to public relations exercises that communicate governance efforts either alongside of-or perhaps in formally _making_ \u2013 the decision. Wherever it occurs, participation can be broadly active or passive.\n\nMoreover, the participants comprising 'the public' can be rendered as citizens within different models, such as 'citizenship-as-rights' often favoured in liberal accounts of formal citizenship in which the equal status of citizens as rights-bearers is stressed, or 'citizenship-as-participation' inspired by civic republican ideas of participants who genuinely share in power in order to shape and steer governance (for discussion see: Pfister 2011). Participants can also be rendered as _inter alia_ citizens, stakeholders, laypersons and consumers. These latter roles are not necessarily synonymous or coterminous with citizenship even as they are related to it, at least in terms of being part of 'the public'. Perhaps especially within liberal accounts of democratic decisionmaking where it is delegated to trusted elites wielding scientific and technocratic knowledge and expertise (Flear 2009), the figures constituting 'the public' \u2013 and especially those not explicitly termed 'citizens' \u2013 function to quell contestation, produce public legitimacy and maintain a functioning economy. Overall, the use and deployment of participation is contingent upon its design and configuration in relation to increasingly complex, interlaced, and decentralised governance arrangements (Gottweis 2008; Yeung 2008).\n\nLike participation, imaginaries, including those of science and technology, are one important key to the success and perpetuation of a key repository of formally accountable power, the nation-state, for and by which they have long been established, utilised and normalised (Anderson 1983; Ezrahi 1990; Scott 1998). Imaginaries provide symbolic (in contrast to formal) accounts of social order and political identity as well as of the shared history and common goals that bind a community together and legitimate political action. Latterly, scholars working within science and technology studies (STS) have highlighted the salience of imaginaries in attempts by nation-states to link their scientific and technological projects with their societies (through the sociotechnical) and use this to produce collectively imagined forms of social life and social order, so-called 'sociotechnical imaginaries' (Jasanoff and Kim 2009; 2013). In short, imaginaries are a key part of the symbolic and cultural foundations of political and legal-constitutional orders. Within this literature, there has also been an attempt to decentre nation-states and loosen their link to imaginaries through a focus on micro-social processes and how they emerge in transnational discourse (Pickersgill 2011).\n\nThrough our case study of sustainable development and technoscientific innovation we highlight imaginaries found in textual and non-textual elements, including videos and images (Jasanoff 2001). In doing so, we build on extant literatures in order to expand and augment the study of imaginaries within law and political studies, which have engaged with technoscience, but have hitherto paid little if any attention to imaginaries. We do so by (re)focusing on the relations between imaginaries and legal and regulatory decision-making, and specifically that of the EU (arguably the most successful non-national-state supranational polity), which has also been the focus of studies on participation. We highlight how imaginaries help to render the future legible and governable, enlarging the scope of power and the means of producing legitimacy and identity for 'Europe'. In addition, we seek to open a clear line of enquiry into how imaginaries help to determine who can and should participate in decision-making, including through configurations of the 'citizen' or 'public', how and why they are to participate, including by and through which knowledges, and the outcomes to be achieved.\n\nRegard to the EU is also useful in that it is explicitly founded on the objective of promoting European integration, and the decision-making capacity of its legal-constitutional order has been central to this project. This order draws on the historical and cultural heritage and imaginaries of European classical civilisation and, perhaps paradoxically, the subsequent nation-states that it purports to transcend. The EU also draws on related imagery of the nation-state, including a body of law and institutional practice (the _acquis communautaire_ ) produced by and embedded in institutions (especially the European Council, European Commission (Commission), Court of Justice of the EU, and the European Parliament), presidential roles (especially of the Commission and the European Council), a flag and an anthem. Together these are deeply resonant and potent ways of connecting with the disparate component peoples of Europe, figured \u2013 in another echo of European heritage \u2013 as not just publics, but also (EU) citizens.\n\nAt the same time, especially in the current era of austerity in the name of 'Europe', the EU's stateness, democratic legitimacy, as well as the coherence and togetherness of its people(s) seems to be even more limited. The EU is constantly searching for its political identity, as well as for stable and meaningful relationships with its citizens. Although the EU is struggling to draw on established imaginaries of national political orders and national political communities, it nonetheless continues to make huge investments in the production and mobilisation of (old and new) imaginaries of Europeanness, through its polity, shared future and a 'We' or demos. These are particularly salient with regard to sustainable development, which, like the EU's imaginaries, attempt to transcend traditional national conceptual frames and institutions. Yet, as we go on to show, the EU's imaginaries join sustainable development with technoscientific innovation in the frame of 'sustainable technoscientific development' for the symbolic production of its political and legal-constitutional order. This helps to legitimate the EU's actions and regulatory decisions, distribute responsibilities and the credit for success, limit accountabilities for failures, and as part of that to enrol EU citizens in its political projects, while also configuring them in relation to decisionmaking through those technologies of participation.\n\n## **Sustainable technoscientific development and innovation in the EU**\n\nWe turn now to overview how sustainable development has been gradually linked with technoscientific innovation and the configuration of citizen participation in relation to it. As an initial point, although Brundtland has proven hugely influential, there are different formulations of sustainable development circulating within and between policy domains. For instance, the specific meaning attached to the main elements of economic growth (Ihlen and Roper 2011; Kambites 2012), the environment (Ross 2009), and global social justice (Barkemeyer _et al._ 2011; Deutz 2012), as well as to their relative weight within the concept, can all differ. Sustainable development focuses less on specific uncertainties and risks and more on providing wide-ranging guidance on how to steer away from the current (unsustainable) trajectory. Moreover, sustainable development is nearly always formulated as a cross-cutting idea impacting, at least in principle, on all policy domains, including technoscience.\n\nThe EU included sustainable development among its objectives in 1997 with the Amsterdam Treaty, which amended the then European Community Treaty, (now further amended and renamed the Treaty on the Functioning of the European Union (TFEU)). Moreover, several key policy documents refine the meaning of sustainable development and how it should be promoted by the EU \u2013 mostly in the context of the EU Sustainable Development Strategy (EU SDS), launched in 2001. The EU SDS consists of three main elements. First, it prioritises a list of key unsustainable trends requiring most urgent action. Second, it addresses how the EU could promote sustainable development globally. Finally, the EU SDS promotes a new approach to policy-making in order to mainstream sustainable development in all policy fields and integrate economic, social and environmental policy. This final element is closely tied to the EU's 'Better Regulation' agenda, which promotes the political and administrative technology of (environmental) impact assessment as a way to produce coherent policy-making and facilitate mainstreaming.\n\nHowever, sustainable development has been increasingly narrowed through its link with technoscientific innovation within the overarching direction for EU governance provided by the European Council Lisbon Strategy (2000\u20132010) (Council of the European Union 2000; Armstrong 2008) which was followed by the subsequent 'Europe 2020' strategy (to cover 2010\u20132020). The 2000 Lisbon European Council formulated the much-cited strategic objective to make Europe the ' _most_ dynamic and competitive, _knowledge-based economy_ in the world, capable of sustaining economic growth, employment and social cohesion' (European Council 2000, emphasis added). In fact, the _knowledge-based economy_ is not used just as a descriptive term but is also a powerful imaginary, which is constructed to gather diverse political actors around a vision of a better and achievable future (for academic formulations of this vision, see: Bell 1976; Drucker 1993). In order to promote this transition towards a knowledge-based economy, technoscientific innovation became the central _'driver_ for the _production and exploitation of knowledge_ [making it] above all a _linchpin_ in the implementation of the Lisbon Strategy' (European Commission 2005: 2, emphasis added).\n\nOf particular note are some tensions within the basic political and legal-constitutional architecture, which have important consequences for the EU's notion of sustainable development and its related imaginaries, especially of citizens and their involvement. To begin with, at the legal-constitutional level, sustainable development is a fundamental, long-term objective of the EU as defined by Article 3(3) of the Treaty on European Union (TEU) and Article 11 TFEU. However, its importance as a central value is much less significant when looking at political strategies operationalising and giving life to the concept of sustainable development. To begin with the EU SDS has only been designed as a complementary process adding an environmental dimension to the Lisbon Strategy's economic and employment\/social policy dimensions, as also affirmed by the 2009 review (European Commission 2009: 13\u201315). In other words, sustainable development \u2013 a fundamental long-term objective of the EU \u2013 is subsumed under two subsequent strategies to boost growth and employment, which are more limited in scope as part of a medium-term objective of economic optimisation.\n\nThis narrowing of sustainable development was reinforced by the subsequent refocusing of the Lisbon Strategy on growth and jobs in 2005 (Kok 2004) and the adoption of the EU's Europe 2020 strategy in 2010 \u2013 a new ten-year strategy for economic growth, mainly formulated in response to the economic and financial crisis then (and still) engulfing the EU (European Commission 2010a; European Commission 2010b). The focus going towards 2020 remains on fostering growth, particularly through the exploitation of knowledge and the propagation of 'knowledge workers'. In this manner the EU claims that:\n\n> the Europe 2020 strategy is about delivering growth that is: _smart_ , through more effective investments in education, research and innovation; _sustainable_ , thanks to a decisive move towards a low-carbon economy; and _inclusive_ , with a strong emphasis on job creation and poverty reduction.\n\nThis promise is translated into quantitative targets in the areas of employment, innovation, education, poverty reduction and climate\/energy. One important novel element of Europe 2020 is the seven flagship initiatives formulated by the Commission as concerted EU level actions to realize the proclaimed combined notion of smart, sustainable, and inclusive growth.\n\nThe prioritisation of growth implies further reshaping of sustainable development. For instance, the strategic objectives of a low-carbon economy and ambitious goals with regard to reducing CO2 emissions display strong environmental commitments. Yet, the notion of sustainable growth, especially in the context of ongoing economic crisis, also implies growth that is resistant to future crises rather than the reflexive and equity-oriented notion adopted in Brundtland. In addition, in the EU sustainability regime, environmental protection is to be achieved by _boosting_ economic activities instead of limiting them. For example, the EU's industrial policy agenda \u2013 one of two flagship initiatives to promote sustainable growth under Europe 2020 \u2013 is meant to 'focus explicitly on opportunities for sustainable growth and job creation while making the transition to more resource efficiency across industry as a whole' (European Commission 2010c). For this purpose, a main goal is described as developing 'the EU market for environmental goods and services' (ibid.).\n\nThe other flagship initiative under the heading sustainable growth is 'Resource-efficient Europe'. Acknowledging natural limits, the overall tone of this flagship initiative is about efficiency and growth potential. According to the Commission:\n\n> increasing resource efficiency will be key to securing growth and jobs for Europe. It will bring major economic opportunities, improve productivity, drive down costs and boost competitiveness... This will help stimulate technological innovation, boost employment in the fast developing 'green technology' sector, sustain EU trade, including by opening up new export markets, and benefit consumers through more sustainable products.\n> \n> (European Commission 2011: 2)\n\nAll in all, the political and legal-constitutional architecture comprised of the TEU and TFEU, the EU SDS, the Lisbon Strategy and Europe 2020 suggests a very specific meaning of sustainable development in the EU. Sustainable development is not understood or operationalised as a far-reaching transformation of the economy in order to promote global and intergenerational justice in line with the limits and capacities of ecosystems i.e. the Brundtland formulation. Rather, technoscientific innovation is figured as a necessary requirement for sustainable _growth, wealth_ , and _employment_. In short, the sustainability of Europe is framed and understood as being its persistent economic success, which should guarantee that it is resilient enough to manage large economic crises and the challenges arising from globalisation, demographic change and the transition towards a knowledge-economy. These goals are achieved not on the basis of self-restraint, as implied by Brundtland, but rather through the greater efficiency and market leadership flowing from technoscientific innovation.\n\nBased on how sustainable development is mobilised and embedded in the context of the strategies for growth and employment, a very specific frame and imaginary of sustainability in an integrated Europe comes to the fore: _sustainable technoscientific development_. Indeed, as we go on to outline, this works with the imaginary of the knowledge-based economy to order governance in a particular way and shape the conditions of possibility for, the configuration, and design of participation in relation to it.\n\n## **The limits of participation in EU governance of sustainable development**\n\nCitizen participation has played an important role in most conceptual formulations of sustainable development (Barry 2006). For example, although performing rather than claiming direct participation, Brundtland gave much space to citizens voicing their needs and understandings, which were collected in several public hearings. Similarly, when launching the EU SDS, the Gothenburg European Council underscored the importance of participation, albeit limited to 'consulting widely with all relevant stakeholders' and calling on the Member States 'to establish appropriate national consultative processes' (European Council 2001). In spite of this and the more general idea that sustainable development is only achievable through widespread support and citizen engagement (Agyeman and Evans 2006), at the broader political and legal-constitutional level there is little if any explicit mention of participatory processes understood as citizens sharing in power through deliberation in order to shape and steer governance. This might be explained by the way in which sustainable development amounts to an add-on to the much larger process of facilitating sustainable growth through sustainable technoscientific development.\n\nMoreover, the specific design of the EU SDS (which still exists alongside Europe 2020) does limit opportunities for public involvement since it is mostly based on technocratic and expert-led governance. Most importantly, there is a strong emphasis on environmental impact assessment. Thus, this assessment is 'a process that _prepares evidence for political decision-makers_ on the advantages and disadvantages of possible policy options by assessing their potential impact'. In a move that seeks to underline this closed class of decision-maker (those who are 'political') as responsible and accountable, the focus is on ' _improving the quality_ of new policy proposals... by making those _people who are responsible for policy development_ also responsible for assessing the impact of what they propose'. This is a tool based on expert analyses geared towards effectiveness and efficiency, but which limits the scope for substantive citizen participation through input into governance in advance of decision-making.\n\nIn addition, although there is an apparent attempt at epistemic integration of such inputs \u2013 't]his approach ensures that all relevant _expertise within the Commission_ is used, together with _inputs from stakeholders'_[ 10 \u2013 the consultation exercise limits and prefigures who can participate through a focus on 'stakeholders'. This suggests that what is 'at stake' has already been defined, as have, by implication, those with an interest in governance. Indeed, this abets the delegation of decision-making to trusted decision-makers and the valorisation of those providing scientific knowledge and expertise, and implies a limited form of citizen, one that is less interested in or even capable of sharing in power in order to shape and steer governance. Hence while the resort to technical tools might mean the 'Commission's system is both _accountable and transparent_ ', it makes 'political decision-makers' alone competent to assess supposed value-neutral technical evidence and select the appropriate policy response, while also limiting citizen participation in decision-making to _ex post_ contestation. Indeed, the invocation of supposedly neutral expert knowledge as a support for decisions actually attempts to reduce \u2013 and arguably thwart \u2013 such involvement, which is also limited by the dearth of clearly defined avenues for feeding back into _ex ante_ processes and future decisions.\n\nIn a similar vein, there are technologies to measure and report progress with regard to sustainable development. In this regard, the Commission's progress reports are very much central (European Commission 2007), and they are facilitated by the gathering and production of data by Eurostat in accordance with sustainable development indicators. These in turn provide the basis for bi-annual monitoring reports produced since 2007 (Eurostat 2009a; Eurostat 2011), which provide 'an objective, statistical picture of progress towards the goals and objectives of the EU sustainable development strategy... [and which] underpins the European Commission's progress report on the implementation of the strategy' (Eurostat 2009b: 2). The first monitoring report also provided the basis for the 2009 review of the EU SDS (European Commission 2009). A supplementary form of monitoring is found in the use of National Sustainable Development Strategies. This is a 'process for voluntary peer reviews of National Sustainable Development Strategies, aimed at improving the sharing of good practices.' In other words, this is aimed at strengthening the EU SDS overall, and aims at 'improving the coordination with other levels of governments'.\n\nOverall, these different ways of gathering, producing and disseminating data are not only an important underpinning for governance, but also highlight the limited and particular understanding of participation. According to the Director-General of Eurostat: 'Impartial and objective statistical information is _essential for all decision-making_ and provides transparency and openness. Official statistics play a fundamental role in today's society and are essential for the development, implementation, monitoring and evaluation of EU policies' (Eurostat 2009b: 1). Yet this data, while valuable for demanding and contesting decision-making, highlights the centrality of technical processes and the privileging of experts \u2013 and, by implication, the marginal role of citizens and publics \u2013 in gathering, interpreting and using data and knowledge for decision-making (a longstanding and core element of EU decision-making, as seen in: European Commission 2001). Environmental policy within a sustainable Europe is based on imaginaries of environmental policy as being technical, complex, and not in need of public political deliberation. This also affects how citizen participation is understood and configured in relation to governance, with it again being limited to _ex post_ contestation based on the information published.\n\nIn sum, citizen participation hardly appears in the EU's strategies and policies on sustainable development (European Commission 2012). More specifically, although there are references to certain participants (especially stakeholders), this implies that citizens per se are not provided with a specific formal political arena where they can get involved in sharing power in order to deliberate the meanings of sustainable development or the definition \u2013 even broadly \u2013 of a sustainable integrated Europe. There is a distinct lack of clearly defined avenues for feeding back into _ex ante_ processes and future decisions. Yet, as we shall highlight, this does not mean that participation is completely absent from interventions in creating a sustainable Europe. In the following section, we focus on the Innovation Union, another Europe 2020 flagship initiative, which is subsumed under the priority of smart growth, and which exemplifies the idea of sustainable technoscientific development as a focus of governance. This programme is also exemplary for the third main imaginary in this chapter: the citizens and publics in the context of Europe 2020.\n\n## **Imaginaries of a sustainable Europe, its publics and citizens in the Innovation Union**\n\nDiagnosing an 'innovation emergency', the Innovation Union is aimed at 'refocusing R&D and innovation policy on major challenges for our society like climate change, energy and resource efficiency, health and demographic change' and 'strengthening every link in the innovation chain, from \"blue sky\" research to commercialisation'. European citizens appear on the Innovation Union website, for example, in terms of 'what it means to me', where the benefits of this political programme are explained. In short, all Europeans are promised a higher quality of life in a greener Europe, with work in a smarter economy. Moreover, the Innovation Union is claimed to improve primarily and directly the work life and opportunities of entrepreneurs and knowledge workers.\n\nMuch more interesting, however, are attempts to engage with those perusing the Innovation Union website through the significant amount of playful and entertaining elements which project strong images and messages. For example, the online game 'Innovation Union Lab' invites the public to hunt for 'innovation treasure-spheres in a futuristic eco-city' \u2013 also with 'Augmented Reality' if the users wish. Moreover, the e-book 'Tales from the Future' (Billing 2011) assembles four short stories about a highly appealing future where, for example, skin cancer is cured by gene therapy and surgical 'nanobots'. Each story is followed by a short description of several EU funded projects whose innovations could bring about the world described. Finally, the collection closes with a three-page interview with the Commissioner for Research, Innovation, and Science, who while claiming that 'research and innovation are the key to meeting [the most crucial] challenges' to our societies (Billing 2011: 57) also states that the 'Innovation Union aims to involve everyone in innovation' (Billing 2011: 60) \u2013 not only researchers and entrepreneurs but also public authorities, civil society organisations, and communities.\n\nThis aim of involving citizens in the 'European' task of innovation is also projected by videos on the Innovation Union website. The majority of which consist of short documentary clips presenting European innovators (all working in EU funded projects) and their inventions ranging from a remote controlled fire-fighting tank to carbon nanotubes. In addition, the website features videos of a more abstract and symbolic rather than documentary character. For example, in one clip, ten (mostly young) people \u2013 apparently citizens from different countries of the EU \u2013 present 'their' definition of innovation. The various meanings offered range from innovation as 'change' to the 'power to imagine the future' to the allegoric 'first warm and sunny spring day after a long cold winter'. Apart from the last atmospheric example, all presented definitions share a striking characteristic: they all focus on innovation as a personal characteristic (rather than as an ideational or material product). Projected as a citizen's virtue, innovation obtains ethical qualities and as such helps to demarcate a sphere of self-governance with its attendant responsibilities that is distinct from, albeit interactive with, the broader governance strategy aimed at prosperity, welfare, and environmental protection.\n\nA similar verbal and symbolic language is employed in two videos named _Innovation Union Act 1_ and _Innovation Union Act 2_ , although they are even free of speaking agents. The first video follows three people \u2013 one could be a planner or architect, one seems to be a natural scientist in the lab, the third one could be a researcher working at his laptop computer \u2013 who are literally 'drawn' into their technoscientific innovations. For example, the architect enters a futuristic city, which looks very green, very clean, and technically advanced. In the second, _Innovation Union Act 2_ , the same three people appear only briefly and are then followed by a succession of flashing words ranging from 'zero emissions' and 'greener cities' to 'creativity', 'expertise', and 'jobs' to 'improved health', 'better ageing', and 'better life'. Importantly, all terms are positive but their scope and positive connotations increase towards the end until the already comprehensive 'better life' is followed by an ultimate, catch-all concept: 'innovation'. In the process, based on the visual and spoken language used in these videos, innovation itself becomes an iconic label and a synonym for all positive qualities of a sustainable and technoscientific Europe.\n\nIndeed, a main category and component of the public in the context of the Innovation Union is that of what we call _innovator citizens_ : individuals who do not necessarily need to be in possession of formal (EU) citizenship but who have technoscientific knowledge and expertise. They are responsibilised into using it \u2013 to innovate! \u2013 in coming to understand _themselves_ as innovators who work with a wider community of innovators and knowledge workers, and produce _for others_ , in order to create a sustainable European economy. By contrast, those who are not given or are incapable of fulfilling such specific roles \u2013 including those who are unable, unwilling or failing to produce innovations \u2013 are effectively not rendered as full citizens in this specific imaginary. These components of the public are to be educated into technoscientific literacy and consumption; while those who are incapable of participating as such are rendered even more marginal. The Europe inhabited by innovator citizens is the private market sphere of the knowledge-based economy where they build a sustainable Europe through their technoscientific innovations. Political power, deliberation, and contestation as central aspects of social order are, however, entirely absent in this imagined 'society'.\n\n## **Discussion**\n\nIn this chapter, the area of sustainable technoscientific development provided our case study for highlighting how imaginaries help to provide the conditions of possibility for and shape the design of technologies of participation and their relation to decision-making. We highlighted three imaginaries found in textual and non-textual elements of the EU. The imaginary of the _knowledge-based economy_ fuses sustainable development to technoscientific innovation, thereby limiting it to being about economic optimisation in the imaginary of a Europe of _sustainable technological development_. Within this frame, those publics in possession of knowledge and expertise are produced and enrolled in another imaginary as private _innovator citizens_ who become complicit in the EU's project of rule. Consumers are implicit in that they literally consume the ideational and material products of technoscience produced by innovators.\n\nImportantly, this public \u2013 especially innovator citizens, stakeholders (noted above) and consumers \u2013 does not explicitly encompass citizens understood as those who are fully capable of sharing in power in order to shape and steer governance. The limited nature of those explicitly figured as participants eschews this fuller notion of citizenship, implying that not all citizens can equally contribute knowledge to decision-making. Those who are not innovator citizens or stakeholders are limited to _ex post_ and (arguably) passive participation as consumers of the ideational and material products of decision-making. Moreover, the contribution of stakeholders is limited to participation in consultations, while the activity of innovator citizens occurs in the private, self-governing depoliticised knowledge-based economy \u2013 it does not extend to participation _in_ decision-making itself or, apparently, _ex ante_ processes.\n\nThere is little formal institutionalisation of _democratic-political_ participation in this area, but attention to imaginaries helps understanding of this situation and reveals the privileging of experts in the gathering, production and interpretation of technical data. This is used to extrapolate imaginaries of an innovative and (technoscientifically, economically, and socially) optimised future facilitated by the EU. Moreover, in a neoliberal vein, this Europe is a depoliticised space of (private) action, creation, production, and consumption, where the agents imagined are exhorted to self-governance and responsibilised as such towards the production and harnessing of technoscientific knowledge and expertise or consumption of its ideational and material products for economic optimisation. This is deemed essential to sustainability, but understood as being about transcending current limits and propelling further, deeper and ongoing economic growth by and through innovation.\n\nSince the knowledge and expertise available to innovator citizens is defined as being of use merely to their self-governance and, through that, the economy, its and their importance and potential contribution to decisionmaking is frustrated \u2013 and that of consumers, and those who cannot participate as such, even more so. Indeed, this implies Europe's citizenry as a whole is stymied from engaging in public questioning of governance options, and the potential transformation of an innovation-centred Europe. In short, the collective imaginary of a united and sustainable Europe is not up for debate among its citizens.\n\nNevertheless, or actually because of this, understood as providing the conditions of possibility for and constituting technologies of participation, the role of imaginaries for the production and perpetuation of the EU's order becomes clearer. Configured in relation to decision-making, imaginaries not only shape participation, they also play a key role in distributing and sharing responsibilities, the credit for success, the accountability and blame for failure, and legitimating and publicly sanctioning the predetermined imagined future of Europe, while undermining or even precluding the possibility of others. There is, therefore, a need for further attention to the mutual constitution and design of imaginaries, participation and the broader technologies of decision-making.\n\nOverall, our analysis highlights imaginaries and the struggle about meaning that they represent as a crucial site for legal and regulatory design, the politics of new and emerging technoscientific innovations, and the future(s) the EU has in mind for Europe. Making imaginaries visible is not just an analytical exercise: exposure is a precondition for disruption. As such, this creates both an avenue for participation and the possibility of formulating alternative imaginaries, knowledge domains and sustainable futures within and outside of formal governance.\n\n## **Acknowledgements**\n\nWe are grateful to the editors of this special issue for their encouragement, comments and suggestions, as well as to Th\u00e9r\u00e8se Murphy. Thomas Pfister also would like to thank Sheila Jasanoff for hosting him as a fellow in the Program on Science, Technology and Society at the Harvard Kennedy School (2010\/2011) as well as for providing the intellectual environment where many important thoughts for this research took shape. The chapter was written with the support of an ESRC Seminar Series grant (RES-451-26-0764) and a BA Small Grant (SG-48186) to Mark L. Flear, and a DAAD scholarship to Thomas Pfister.\n\n In the EU context see especially Article 9 TEU and Articles 20\u201324 TFEU, which establish EU citizenship and its rights.\n\n 'European Year of Citizens 2013': h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.\u00ade\u00adu\u00ad\/\u00adc\u00adi\u00adt\u00adi\u00adz\u00ade\u00adn\u00ads\u00ad-\u00ad2\u00ad0\u00ad1\u00ad3\u00ad\/\n\n These areas of priority are: climate change and energy; sustainable transport; sustainable consumption and production; conservation and management of natural resources; public health; social inclusion, demography, and migration; global poverty and sustainable development changes.\n\n 'Impact assessment': h\u00adt\u00adt\u00adp\u00ad:\/\u00ad\/e\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/g\u00ado\u00adv\u00ade\u00adr\u00adn\u00ada\u00adn\u00adc\u00ade\u00ad\/i\u00adm\u00adp\u00ada\u00adc\u00adt\u00ad\/i\u00adn\u00add\u00ade\u00adx\u00ad_e\u00adn\u00ad.h\u00adt\u00adm\u00ad\n\n 'Commission launches consultation on EU 2020: a new strategy to make the EU a smarter, greener social market': h\u00adt\u00adt\u00adp\u00ad:\/\u00ad\/\u00ade\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00adr\u00ada\u00adp\u00adi\u00add\u00ad\/\u00adp\u00adr\u00ade\u00ads\u00ads\u00adR\u00ade\u00adl\u00ade\u00ada\u00ads\u00ade\u00ads\u00adA\u00adc\u00adt\u00adi\u00ado\u00adn\u00ad.d\u00ado\u00ad?;\u00adr\u00ade\u00adf\u00ade\u00adr\u00ade\u00adn\u00adc\u00ade\u00ad=\u00adI\u00adP\u00ad\/\u00ad0\u00ad9\u00ad\/\u00ad1\u00ad8\u00ad0\u00ad7. Emphasis added.\n\n 'Europe 2020': h\u00adt\u00adt\u00adp\u00ad:\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00ade\u00adu\u00adr\u00ado\u00adp\u00ade\u00ad2\u00ad0\u00ad2\u00ad0\u00ad\/\u00ade\u00adu\u00adr\u00ado\u00adp\u00ade\u00ad-\u00ad2\u00ad0\u00ad2\u00ad0\u00ad-\u00adi\u00adn\u00ad-a\u00ad-n\u00adu\u00adt\u00ads\u00adh\u00ade\u00adl\u00adl\u00ad\/p\u00adr\u00adi\u00ado\u00adr\u00adi\u00adt\u00adi\u00ade\u00ads\u00ad\/\u00adi\u00adn\u00add\u00ade\u00adx\u00ad_\u00ade\u00adn\u00ad.h\u00adt\u00adm\u00ad\n\n In respect of the environment aspect of sustainable development, and therefore worthy of mention, is the Aarhus Convention (1998), which provides for participation, and has been operationalised in EU governance through secondary legislation. Given its focus on the environment it is, however, rather limited, even as it provides a 'way in' for participation in sustainable development more broadly.\n\n 'Impact assessment': h\u00adt\u00adt\u00adp\u00ad:\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00adg\u00ado\u00adv\u00ade\u00adr\u00adn\u00ada\u00adn\u00adc\u00ade\u00ad\/\u00adi\u00adm\u00adp\u00ada\u00adc\u00adt\u00ad\/\u00adi\u00adn\u00add\u00ade\u00adx\u00ad_e\u00adn\u00ad.h\u00adt\u00adm\u00ad. Original emphasis.\n\n Ibid. Original emphasis.\n\n Ibid. Original emphasis.\n\n Such as 'High Level Group of National Regulatory Experts': h\u00adt\u00adt\u00adp\u00ad:\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/g\u00ado\u00adv\u00ade\u00adr\u00adn\u00ada\u00adn\u00adc\u00ade\u00ad\/\u00adi\u00adm\u00adp\u00ada\u00adc\u00adt\u00ad\/h\u00adi\u00adg\u00adh\u00ad_\u00adl\u00ade\u00adv\u00ade\u00adl\u00ad_\u00adg\u00adr\u00ado\u00adu\u00adp\u00ad\/i\u00adn\u00add\u00ade\u00adx\u00ad_\u00ade\u00adn\u00ad.h\u00adt\u00adm\u00ad\n\n Ibid. Original emphasis. For online publication see 'Impact assessments': h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00adg\u00ado\u00adv\u00ade\u00adr\u00adn\u00ada\u00adn\u00adc\u00ade\u00ad\/\u00adi\u00adm\u00adp\u00ada\u00adc\u00adt\u00ad\/i\u00ada\u00ad_\u00adc\u00ada\u00adr\u00adr\u00adi\u00ade\u00add\u00ad_o\u00adu\u00adt\u00ad\/i\u00ada\u00ad_c\u00ada\u00adr\u00adr\u00adi\u00ade\u00add\u00ad_\u00ado\u00adu\u00adt\u00ad_\u00ade\u00adn\u00ad.h\u00adt\u00adm\u00ad\n\n 'Sustainable development indicators': h\u00adt\u00adt\u00adp\u00ad:\/\u00ad\/\u00ade\u00adp\u00adp\u00ad.e\u00adu\u00adr\u00ado\u00ads\u00adt\u00ada\u00adt\u00ad.e\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00adp\u00ado\u00adr\u00adt\u00ada\u00adl\u00ad\/\u00adp\u00ada\u00adg\u00ade\u00ad\/\u00adp\u00ado\u00adr\u00adt\u00ada\u00adl\u00ad\/\u00ads\u00add\u00adi\u00ad\/\u00adi\u00adn\u00add\u00adi\u00adc\u00ada\u00adt\u00ado\u00adr\u00ads\u00ad\n\n 'National Sustainable Development Strategies': h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00ade\u00adn\u00adv\u00adi\u00adr\u00ado\u00adn\u00adm\u00ade\u00adn\u00adt\u00ad\/\u00adp\u00add\u00adf\u00ad\/\u00adn\u00ads\u00add\u00ads\u00ad.p\u00add\u00adf\u00ad\n\n 'Sustainable Development': h\u00adt\u00adt\u00adp\u00ad:\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/e\u00adn\u00adv\u00adi\u00adr\u00ado\u00adn\u00adm\u00ade\u00adn\u00adt\u00ad\/\u00ade\u00adu\u00ads\u00ads\u00add\u00ad\/\n\n Ibid.\n\n Beyond the EU SDS, participatory elements can also be found in single elements of the Lisbon-oriented economic reform agendas, especially as part of (national-level) consultations on National Reform Programmes or in the context of specific open methods of coordination. However, these consultations always focus on specific strategies and measures to implement common guidelines, for example, in relation to employment policy, but in which sustainable development is not an explicit topic, even as it is supposed to contribute to the objective of sustainable development.\n\n 'Innovation Union':  h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/e\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/r\u00ade\u00ads\u00ade\u00ada\u00adr\u00adc\u00adh\u00ad\/\u00adi\u00adn\u00adn\u00ado\u00adv\u00ada\u00adt\u00adi\u00ado\u00adn\u00ad-\u00adu\u00adn\u00adi\u00ado\u00adn\u00ad\/\u00adi\u00adn\u00add\u00ade\u00adx\u00ad_e\u00adn\u00ad.c\u00adf\u00adm\u00ad.?;\u00adp\u00adg\u00ad=\u00adw\u00adh\u00ady\u00ad\n\n 'Europe 2020': h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00ade\u00adu\u00adr\u00ado\u00adp\u00ade\u00ad2\u00ad0\u00ad2\u00ad0\u00ad\/\u00ade\u00adu\u00adr\u00ado\u00adp\u00ade\u00ad-\u00ad2\u00ad0\u00ad2\u00ad0\u00ad-\u00adi\u00adn\u00ad-\u00ada\u00ad-\u00adn\u00adu\u00adt\u00ads\u00adh\u00ade\u00adl\u00adl\u00ad\/p\u00adr\u00adi\u00ado\u00adr\u00adi\u00adt\u00adi\u00ade\u00ads\u00ad\/s\u00adm\u00ada\u00adr\u00adt\u00ad-\u00adg\u00adr\u00ado\u00adw\u00adt\u00adh\u00ad\/\u00adi\u00adn\u00add\u00ade\u00adx\u00ad_e\u00adn\u00ad.h\u00adt\u00adm\u00ad\n\n 'Innovation Union \u2013 what it means to you': h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/r\u00ade\u00ads\u00ade\u00ada\u00adr\u00adc\u00adh\u00ad\/\u00adi\u00adn\u00adn\u00ado\u00adv\u00ada\u00adt\u00adi\u00ado\u00adn\u00ad-u\u00adn\u00adi\u00ado\u00adn\u00ad\/i\u00adn\u00add\u00ade\u00adx\u00ad_\u00ade\u00adn\u00ad.c\u00adf\u00adm\u00ad?p\u00adg\u00ad=\u00adw\u00adh\u00ada\u00adt\u00adi\u00adt\u00adm\u00ade\u00ada\u00adn\u00ads\u00ad\n\n 'Innovation Union lab': h\u00adt\u00adt\u00adp\u00ad:\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00adr\u00ade\u00ads\u00ade\u00ada\u00adr\u00adc\u00adh\u00ad\/\u00adi\u00adn\u00adn\u00ado\u00adv\u00ada\u00adt\u00adi\u00ado\u00adn\u00ad-u\u00adn\u00adi\u00ado\u00adn\u00ad\/\u00adi\u00adn\u00add\u00ade\u00adx\u00ad_\u00ade\u00adn\u00ad.c\u00adf\u00adm\u00ad?p\u00adg\u00ad=\u00adl\u00ada\u00adb\u00ad\n\n 'Augmented reality': h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/\u00adr\u00ade\u00ads\u00ade\u00ada\u00adr\u00adc\u00adh\u00ad\/i\u00adn\u00adn\u00ado\u00adv\u00ada\u00adt\u00adi\u00ado\u00adn\u00ad-u\u00adn\u00adi\u00ado\u00adn\u00ad\/i\u00adn\u00add\u00ade\u00adx\u00ad_\u00ade\u00adn\u00ad.c\u00adf\u00adm\u00ad?p\u00adg\u00ad=\u00ada\u00adr\u00ad\n\n For all videos, see 'Videos': h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/r\u00ade\u00ads\u00ade\u00ada\u00adr\u00adc\u00adh\u00ad\/\u00adi\u00adn\u00adn\u00ado\u00adv\u00ada\u00adt\u00adi\u00ado\u00adn\u00ad-\u00adu\u00adn\u00adi\u00ado\u00adn\u00ad\/\u00adi\u00adn\u00add\u00ade\u00adx\u00ad_e\u00adn\u00ad.c\u00adf\u00adm\u00ad?p\u00adg\u00ad=\u00adv\u00adi\u00add\u00ade\u00ado\u00ads\u00ad\n\n## **References**\n\nAgyeman, J. and Evans, B. (2006) 'Justice, governance, and sustainability: perspectives on environmental citizenship from North America and Europe', in A. Dobson and D. Bell (eds), _Environmental Citizenship_ , London: MIT Press.\n\nAnderson, B. (1983) _Imagined Communities_ , London: Verso.\n\nArmstrong, K. (2008) 'Governance and constitutionalism after Lisbon', in JCMS Symposium: EU Governance After Lisbon _Journal of Common Market Studies_ , 46: 413.\n\nBarkemeyer, R., Holt, D., Preuss, L. and Tsang, S. (2011) 'What happened to the development in sustainable development? Business guidelines two decades after Brundtland', _Sustainable Development_ , DOI: 10.1002\/sd.521.\n\nBarry, J. (2006) 'Resistance is futile: from environmental to sustainability citizenship', in A. Dobson and U. Beck (1986) _Risk Society: Towards a New Modernity_ , London: Sage.\n\nBeck, U. (2009) _World at Risk_ , Cambridge: Polity Press.\n\nBell, D. (1976) _The Coming of Post-Industrial Society: A Venture in Social Forecasting_ , Harmondsworth: Basic Books.\n\nBilling, R (2011) _Tales from the Future_ , Brussels: European Union, available at: h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adc\u00ad.e\u00adu\u00adr\u00ado\u00adp\u00ada\u00ad.e\u00adu\u00ad\/r\u00ade\u00ads\u00ade\u00ada\u00adr\u00adc\u00adh\u00ad\/\u00adi\u00adn\u00adn\u00ado\u00adv\u00ada\u00adt\u00adi\u00ado\u00adn\u00ad-u\u00adn\u00adi\u00ado\u00adn\u00ad\/\u00adi\u00adn\u00add\u00ade\u00adx\u00ad_\u00ade\u00adn\u00ad.c\u00adf\u00adm\u00ad?p\u00adg\u00ad=\u00adt\u00ada\u00adl\u00ade\u00ads\u00ad\n\nBlack, J. (2007) 'Tensions in the regulatory state', _Publie Law_ , 58: 58\u201373.\n\nDavidson, K. (2011) 'A typology to categorize the ideologies of actors in the sustainable development debate', _Sustainable Development_ , DOI: 10.1002\/sd.520.\n\nDeutz, P. (2012) 'A class-based analysis of sustainable development: developing a radical perspective on environmental justice', _Sustainable Development_ , DOI: 10.1002\/sd.l528.\n\nDobson, A. and Bell, D. (2006) _Environmental Citizenship_ , London: MIT Press.\n\nDresner, S. (2002) _The Principles of Sustainability_ , London: Earfhscan.\n\nDrucker P. (1993) _Post-Capitalist Society_ , New York: Harper Business.\n\nEuropean Commission (2001) _European Governance: A White Paper_ , COM(2001) 428 final.\n\n\u2014\u2014 (2005) _Building the ERA of Knowledge for Growth_ , COM (2005) 118 final.\n\n\u2014\u2014 (2007) _Progress Report on the Sustainable Development Strategy 2007_ , COM (2007) 642 final.\n\n\u2014\u2014 (2009) _Mainstreaming Sustainable Development into EU Policies: 2009 Review of the European Union Strategy for Sustainable Development_ , COM (2009) 400 final.\n\n\u2014\u2014 (2010a) _Smart Regulation in the European Union_ , COM (2010) 543 final.\n\n\u2014\u2014 (2010b) _Europe 2020 Flagship Initiative Innovation Union_ , COM (2010) 546 final.\n\n\u2014\u2014 (2010c) _An Integrated Industrial Policy for the Globalisation Era. Putting Competitiveness and Sustainability at Centre Stage_ , COM (2010) 614 final.\n\n\u2014\u2014 (2011) _A Resource-Efficient Europe \u2013 Flagship Initiative Under the Europe 2020 Strategy_ , COM, (2011)21 final.\n\n\u2014\u2014 (2012) _Action for Stability and Jobs_ , COM (2012) 299 final.\n\nEuropean Council (2000) _Lisbon European Council_ , 23 and 24 March2000. Presidency Conclusions, no. 5.\n\n\u2014\u2014 (2001) _Gothenburg European Council_ , 15 and 16 June2001. Presidency Conclusions.\n\nEurostat (2009a) _Sustainable Development in the European Union. 2009 Monitoring Report of the EU Sustainable Development Strategy_ , Luxembourg: Office for Official Publications of the European Communities.\n\n\u2014\u2014 (2009b) _Sustainable development in the European Union_ , Luxembourg: Office for Official Publications of the European Communities.\n\n\u2014\u2014 (2011) _Sustainable Development in the European Union. 2011 Monitoring Report of the EU Sustainable Development Strategy_ , Luxembourg: Office for Official Publications of the European Communities.\n\nEzrahi, Y. (1990) _The Descent of Icarus_ , Cambridge, MA: Harvard University Press.\n\nFiorini, D.J. (1990) 'Citizen participation and environmental risk: a survey of institutional mechanisms', _Science, Technology & Human Values_, 15: 226.\n\nFlear, M.L. (2009) 'Editorial: an introduction to \"New Technologies, European Law and Citizens\"', _Maastricht Journal of European and Comparative Law_ , 16(1): 3.\n\nFlear, M.L. and Pickersgill, M.D. (2013) 'Regulatory or regulating publics? The European Unions regulation of emerging health technologies and citizen participation', _Medical Law Review_ , 21: 39.\n\nGiddens, A. (1998) 'Bisk society: the context of British politics', in J. Franklin (ed.), _The Politics of Risk Society_ , Cambridge: Polity Press.\n\nGottweis, H. (2008) 'Participation and the new governance of life', _Bio Societies_ , 3: 265.\n\nIhlen, \u00d8. and Roper, J. (2011) ' _Corporate reports on sustainability and sustainable development: \"we have arrived\"_ ', Sustainable Development, DOI: 10.1002\/ sd.524.\n\nIrwin, A. (2006a) 'The politics of talk: coming to terms with the \"new\" scientific governance', _Social Studies of Science_ , 36: 299.\n\nIrwin, A. (2006b) _Sociology and the Environment_ , Cambridge: Polity Press.\n\nJasanoff, S. (2001) 'Image and imagination: the formation of global environmental consciousness', in P. Edwards and C. Miller (eds), _Changing the Atmosphere: Expert Knowledge and Environmental Governance_ , Cambridge, MA: MIT Press.\n\nJasanoff, S. and Kim, S.-H. (2009) 'Containing the atom: sociotechnical imaginaries and nuclear power in the United States and South Korea', _Minerva_ , 47: 119.\n\nJasanoff, S. and Kim, S.-H. (2013). 'Sociotechnical imaginaries and national energy policies', _Science as Culture_ , 22: 189.\n\nKambites, C.J. (2012) '\"Sustainable development\": the \"unsustainable\" development of a concept in political discourse', _Sustainable Development_ , DOI: 10.1002\/sd.l552.\n\nKok, W. (2004) _Facing the Challenge: The Lisbon Strategy for Growth and Employment. Report from the High Level Group Chaired by Wim Kok_ , Brussels: European Commission.\n\nMandel, G.N. (2009) 'Regulating emerging technologies', _Law, Innovation and Technology_ , 1: 75.\n\nO'Riordan, T. (2009) 'Reflections on the pathways to sustainability', in N. Adger and A. Jordan (eds), _Governing Sustainability_ , Cambridge: Cambridge University Press.\n\nPfister, T. (2011) _The activation of citizenship in Europe_ , Manchester: Manchester University Press.\n\nPickersgill, M.D. (2011) 'Connecting neuroscience and law: anticipatory discourse and the role of sociotechnical imaginaries', _New Genetics and Society_ , 30(1): 27.\n\nRoss, A. (2009) 'Modern interpretations of sustainable development', _Journal of Law and Society_ , 36(1): 32.\n\nScott, J.C. (1998) _Seeing Like a State_ , New Haven, CT: Yale University Press.\n\nTait, J. (2009) 'Upstream engagement and the governance of science' _EMBO Reports_ 10: S18\u2013S22.\n\nWilsdon, J. and Willis, R (2004) _See-Through Science: Why Public Engagement Needs to Move Upstream_ , London: Demos.\n\nWorld Commission on Environment and Development (1987) _Our Common Future_ , Oxford: Oxford University Press.\n\nYeung, K. (2008) 'Towards an understanding of regulation by design', in R Brownsword and K. Yeung (eds), _Regulating Technologies: Legal Futures, Regulatory Frames and Technological Fixes_ , Oxford: Hart Publishing.\nChapter 3\n\n* * *\n\n# **The voice of silence**\n\nUK patients' silent resistance to the assisted reproduction regulations\n\n_Ilke Turkmendag_\n\n* * *\n\n## **Introduction**\n\nSince the 1990s, one of the most controversial changes to the UK's Human Fertilisation and Embryology Act (HFE Act) concerned the information that is kept about the egg, sperm (gametes) and embryo donors in donor registers. This information is held by the Human Fertilisation and Embryology Authority (HFEA), which is a quasi-autonomous non-governmental organisation (quango) set by the HFE Act in 1990 to license and regulate clinics that carry out _in vitro_ fertilisation (IVF), other assisted conception techniques using human sperm, eggs and embryos, and embryo research.\n\nOriginally, section 31(3) of the HFE Act 1990 allowed offspring conceived through gamete donation the right to apply for information about the donor when reaching the age of 18. In practice, donors were asked to provide some non-identifying information, which could be passed to potential recipients. However, on 1 April 2005, with the implementation of the Human Fertilisation and Embryology Authority (Disclosure of Donor Information) Regulations 2004, UK law was changed to allow children born through gamete donation to access identifying details of the donor (The Human Fertilisation and Embryology Authority, 2004). Drawing on trends in adoption law, the decision to abolish donor anonymity was strongly influenced by a discourse that asserted the 'child's right-to-personal identity' (Turkmendag, 2012), encouraged by the European Convention on Human Rights (ECHR) (European Convention for the Protection of Human Rights and Fundamental Freedoms, 1950).\n\nThe removal of donor anonymity has aggravated the outgoing donor shortage since individuals were less likely to donate if their relatedness could one day be established (Turkmendag _et al._ , 2008). Moreover, on 5 July 2007, the HFE Act was amended in order to bring the EU Tissues and Cells Directive (EUTCD) into UK law. The implementation of EUTCD's voluntary and unpaid donations policy for tissues and cells has made it even more difficult to recover the donor shortage as donations cannot be promoted by financial incentives (Turkmendag, 2013). As a result, fewer people now have access to the treatment in the UK (Turkmendag, 2013).\n\nPerhaps surprisingly, neither during nor after the removal of donor anonymity did the government face organised resistance from would-be parents through classic avenues of opposition to new legislation; i.e., both those who cannot receive treatment due to the donor shortage, and those who might carry the burden of informing children of their origins if treatment were available and successful (Turkmendag _et al._ , 2008) Those who prefer anonymous donation are reluctant to press their claims because confidentiality matters to them (Turkmendag, 2013). Even those who are pro-openness may be reluctant to face publicity as undergoing fertility treatment, and using donated gametes are private matters. Presumably because they did not articulate their concerns in formal ways of communication (such as lodging complaints for a policy change), there has been little enquiry into the reactions of would-be parents to the new law. But, despite their reticence during the donor anonymity debate, would-be parents articulated _anonymous_ protests against the new law on the discussion forums of online fertility support groups. In this chapter, I examine such discussion forums in order to cast fresh light into the discourse of would-be parents that was propelled by the change to the HFE Act. I examine how strategies of legal avoidance and resistance to the new law were facilitated by the growth of an online community of would-be parents, and explore their impact on policy adjustments. The data form part of my ongoing (since 2005) research into the socio-legal dimensions of reproductive care.\n\n## **Avoidance of the law by the 'virtual' patient communities**\n\nOnline forums are known to have transformed much of the movements and relationships that animate communities of would-be parents. As an example, Waldby (2008) argues that multilingual websites and Internet communication have made international partnerships and the possibilities for patients to travel abroad for treatment much easier. This, together with the legal possibilities offered by the European Union, have facilitated what is sometimes referred to as 'cross-border reproductive care' (CBRC): a form of medical tourism that enables patients to take advantage of differences in reproductive consumer cultures reflected in variations in the cost and availability of treatments, success rates, and waiting lists. The removal of donor anonymity in the UK has resulted in an increasing number of would-be parents going abroad to seek treatment, a movement facilitated by the availability of online support that Waldby describes (Shenfield _et al._ , 2010). Indeed, the Internet plays a significant role here, that I trace through this chapter: through online discussion forums, it facilitates ongoing discourse among the patients, effectively creating a 'virtual patient community', a social aggregation based on common interest and intellectual and affective (rather than geographic) proximity. The virtual patient community that emerged around (and in antagonism to) the changes to the HFE Act did not promote the breaking of law, but instead its avoidance. Community members develop and share strategies to do this (i.e., how best to make use of CBRC), and mobilise around their contentions about the legal shift while remaining anonymous.\n\nIn the UK the online discussion forums hosted by fertility support groups (e.g. Acebabes, fertilityfriends, INUK, More to Life and IVF.Net) or fertility clinics played a salient role (perhaps unintentionally) in helping potential patients to overcome their anxiety about receiving treatment away from home. Following the removal of donor anonymity, these forums were snowed under with posts about treatment opportunities in other jurisdictions. Those who wanted to receive treatment using anonymous gametes and those who were denied access to treatment at home for various reasons (donor shortage, being regarded as unfit parents) exchanged information about the jurisdictions where the treatments they needed would be available (e.g. standards of the clinics, the success rates, costs, quality of the aftercare, whether practitioners or nurses could speak English, and regulatory frameworks). Those who could not afford to travel abroad exchanged information about 'backstreet' arrangements (such as purchasing of sperm online). In addition, these forums became a space in which patients could express their opposition to the new law without making their identities known (e.g. signing online petitions, discussing the HFEA's decision-making processes, sending feedback through forum administrators, talking to journalists and researchers). In what follows, I present data collected during a virtual ethnography on one of these discussion forums: Repromed (www.repromed.co.uk), which was hosted by the website of the Bristol Centre for Reproductive Medicine (BCRM). In doing so, I render visible the discreet ways in which would-be parents avoided, resisted, and ultimately expressed their opposition, to the new law.\n\n## **'Would-be' parents and the discussion forums of Repromed**\n\nRepromed introduced itself as a website for anyone in the UK interested in reproductive medicine. Like many other patient support groups, Repromed was open to anyone, at any stage of their treatment, offering an ideal platform to access to information about treatment techniques, clinics abroad, and other patients' experiences at no or little cost (some patient groups may require a small subscription fee), quickly and anonymously from home. On the home page, there was a link to the discussion area; there, the forum titles were listed. Each topic involved a question, or a brief introduction about the inquiry; for example: 'Is anyone currently trying or pregnant with a donor conceived baby?' Forum administrators and moderators, who had the highest level of control over forum content, could turn forum features on and off, ban and remove users, edit and delete posts, and create users and groups. This gave them some control over the postings, although it might be rather difficult to monitor and manage such a large and continuously evolving content. By the time that I visited the website in 2006, users had posted 27,829 messages, on 704 topics, in 15 forums. The age range of the other 28 users was from 28 to 43. The average age was 35 and there were two modes (the most frequently occurring age in the distribution) at 35 and 40. Although more than 5,000 postings were scanned, the final sample included 900 postings sent between 13 July 2005 and 26 April 2007.\n\nPosts on Repromed included inquiries as to where one could receive treatment abroad by using gametes from anonymous donors, stories about fertility travel, complaints about the long waiting lists and the donor shortage in the UK, and criticism of the new regulations. The data presented here draw mainly on the 'Donor Gamete' Forum on Repromed, which had the fullest coverage of donor anonymity. Although this study does not claim to be a definitive account of UK would-be parents who use online forums, _Repromed_ was clearly one of the most active sites (based on the number of messages posted, and the depth of the threads) during this period and represents an important source of data on voices that were not fully heard in the public debates over ending donor anonymity. The extracts from the forum posts that I present herein are not edited, except for the references to clinics or names of websites where sperm can be purchased.\n\n## **Policy processes and patient reactions**\n\nBefore this forum was set up, fertility travel was being addressed in other discussion forums on the website. One of the patients, Elaine, who was receiving treatment at the BCRM (the clinic that hosts Repromed), had decided to receive treatment in Spain; this was mainly due to the removal of donor anonymity in the UK and consequent lengthening of waiting lists. She shared her experiences with other users of _Repromed_ , eventually meeting in person and befriending six other women users who also ended up receiving treatment in Spain. According to Elaine, the fact that they were discussing their experiences on Repromed discussion forums was not well-received by the host clinic and the administrator set up a new 'Donor Gamete' forum so that people receiving donor gamete treatment could interact in a setting separate from other patients of the clinic. It was established in November 2005 with the following disclaimer:\n\n> After much discussion we have set up the Donor gamete forum. We are a little concerned about the oversees fertility treatment options as these are outside UK ethical and safety regulations, but we fully agree that UK patients receiving donor gamete treatment would benefit from mutual online support.\n> \n> Regards, Admin\n\nMost forum participants were on waiting lists for gamete donation and clearly associated the donor shortage with the change in the law to remove anonymity. Examples of posts expressing this view were as follows (all are excerpted from a thread about the options available for those who attempted to receive treatment in licensed clinics in the UK but for various reasons could not access the treatment they needed):\n\n> Apparently out of 500 applicants to give sperm in the last year only 2 actually do so, predominantly due the anonymity issue and the low sperm counts of the younger applicants. Very frightening stats don't you think?\n> \n> (Mandy)\n> \n> i do think that the change in the law has got every thing to do with it especialy if you are donating eggs or sperm and you are haveing difficulty consiving yourself .i have also been told that sperm donations have gone down because there use to be alot of students that use to donate but have stoped because they dont want a knock on the door in 18 years do you know if it the same if you go out side theuk and have treatment and donate at the same time cause im not sure about that one from the first time i went for icsi to now things seame to have got more compucated and expensive lol.\n> \n> (Brenda)\n\nPrior to the removal of donor anonymity, a 2004 Department of Health survey of UK fertility clinics indicated widespread agreement that lifting anonymity would decrease donations (Bionews, 2004). During the spring of 2004, the HFEA conducted its own survey of clinics to develop a detailed understanding of the current demand for, and provision of, treatment using donated gametes or embryos, as well as to identify relevant trends (Human Fertilisation and Embryology Authority, 2004). The removal of donor anonymity was the most frequently cited factor anticipated negatively to affect sperm and egg donation by the clinics which were already suffering from a donor shortage. In other words, the donor crisis was hardly a surprise to the regulatory bodies.\n\nIn the following post, one user, Tara, drew attention to the fact that the HFEA ignored the evidence provided by the clinics:\n\n> They (HFEA) should have warned clinics to overstock a year ago in case this should happen. From what I could gather from the clinics I rang they had warned the HFEA that this would happen!! The HFEA knew there could be a dip in donors as this happened in Australia and New Zealand and Sweden, I would like to know if they ever recovered! ! All I can say is well done (not) HFEA.\n> \n> (Tara)\n\nIf the donor shortage was predictable, why then did HFEA take precautions? In 2007, the Department of Health and the National Gamete Donation Trust (NGDT) launched the 'Give a Toss' campaign to raise awareness of donation and recruit new donors (critics of the campaign have accused the NGDT of alienating donor parents, would-be parents and donor-conceived children by perpetuating a 'tacky' image) (McVeigh, 2007). Other campaigns were launched throughout the country to promote sperm donation and the British Fertility Society established a working party to examine sperm donation in the UK. Its report was launched at a HFEA meeting in February 2008. The report made a number of recommendations about how the existing pool of donors could be used to more effect: 'the report gives a helpful impression of the current state of sperm donation services in the UK, showing an undeniable mismatch between supply and demand'. The HFEA's response in July 2008 was that the Authority's remit did not extend to the recruitment of donors; this was a matter for the sector and should be addressed as a matter of urgency (Jardine, 2008a).\n\nDuring the donor anonymity public debate, the HFEA made it clear that it supported lifting the anonymity. The then Chair of the HFEA, Suzi Leather noted that '[W]e have been asking for this for a long time. I can understand why people want to know who their genetic parents are' (BBC News, 2004). In the following post Tara expressed her anger about the view that the child's right to know should be championed at any cost, rather than weighed against the competing interests of would-be parents:\n\n> Anyway on another note we should thanks Dame Suzi Leather and her syndicate for causing this mess!! I wouldnt be surprised if a couple took them to the court of human rights \u2013 essentially they are stopping couples having a baby!! To me thats enfringing on our human rights. Of course this is my personal opinion, but I havent yet found anyone who is happy with what the HFEA have done.\n> \n> (Tara)\n\nOne forum user, Sylvia, claimed that in order to resolve the sperm shortage the law should be changed back; she believed that anonymity should be reinstated, and produced a petition to that effect:\n\n> if you think sperm donors should be legally allowed to remain anonymous please sign my petition. i am petitioning the prime minister of the uk to change the law so sperm donors can go back to being anonymous so the sperm shortage can be resolved. if you agree with this please sign at the following link: http:\/\/petitions.pm.gov.uk\/sperm-donor-anon\/.\n> \n> (Sylvia)\n\nLike Tara, she framed the legislative changes as explicitly curtailing her and others' reproductive freedom.\n\n> i think anonymity should be brought back after watching a documentary on it. being someone who possibly may need a sperm donor's help in the future id rather not be waiting for years because of the shortage if you agree could you please sign my petition. i understand the reasons why it shouldn't but why should we prevent a loving couple from having the child they wish to have?\n> \n> come on sign it if you agree :D\n> \n> (Sylvia)\n\nIn her response to Sylvia's post, Dorothy says:\n\n> I have signed your petition as I have used sperm donor in the past and though we were not successful I was shocked for other people at the huge decline in donors when the law was changed.\n> \n> (Dorothy)\n\nSigning petitions is one of the ways that patients can lodge formal complains. However, it might be problematic for those who have confidentiality concerns. For example, one of the users, Hannah had twins through surrogacy arrangements in USA and did not want this to be known. According to UK law, the surrogate is considered a child's legal mother. If the origins of the twins were known, Hannah would have to formally adopt them in order to become legally recognised as their mother. Hence, she was very careful about making formal complaints that might reveal the origins of her children. Another user, Cheryl opted for adoption after failed attempts to achieve pregnancy. She also had confidentiality concerns about making herself known as she did not know how that would affect her chances of adopting a child.\n\nIn order to circumvent the challenges imposed by the law, many would-be parents claimed that they or others might or would go abroad in order to obtain treatment. As Tara put it:\n\n> Now it has happened, where and what do people needing donor sperm do??? If clinics dont start recruiting their own donors then we will have to go to the ends of the Earth to try and get one, thats our only choice, its so unfair.\n> \n> (Tara)\n\nSome commentators consider CBRC as a 'reproductive exile' where patients who are denied access to treatment at home are led to find a legal regime under which the treatment they need will be available (Matorras, 2005). For example, in the Republic of Ireland (where abortion is not permissible unless there is a risk to the pregnant woman's life), between January 1980 and December 2010, at least 147,912 women travelled abroad to terminate their pregnancies. Often due to the complicated travel arrangements these abortions are performed later than is preferred, creating emotional upset for women at an already stressful time. Further, only those who can afford to travel can escape the law's strictures. Again, in Ireland, both illegal abortions, and the use of illegal abortion pills have become a necessity for women who cannot travel abroad to end their pregnancies (McGuinness, 2011). Similarly, in Repromed discussion boards, some of the users ask for others' opinions about 'backstreet' arrangements:\n\n> I have found this company on the web and wanted to know your views: [the website's URL] Having quickly read it through it seems as tho they send you the sperm and you self inseminate (?!) for under _\u00a3_ 400. I know people have been going abroad, but i can't find a place that does D IUI.\n> \n> (Suzanne)\n\nIn one reply, Tara says:\n\n> Also there is another site [the website's URL] as anyone thought or used this one??? I have heard they are fairly good. I have to say it makes me wonder whether all the sperm donors that went through clinics and wanted to remain anonymous are now going to these places in order to help people needing it??!!\n> \n> (Tara)\n\nElaine advices this user to seek treatment in Spain using a cheap airline:\n\n> Try [a clinic in Spain] \u2013 anonymous donor sperm is 260 Euros per go and I flew from Bristol to Barcelona for _\u00a3_ 50 each way!\n> \n> (Elaine)\n\nIn another post, after complaining about the fees that each couple has to pay to undergo treatment, Tara addresses the risks of 'backstreet' arrangements. She notes that the new law leads people to consider less legitimate ways of overcoming their childlessness and to take health risks. For example, one can use the 'turkey baster' method and achieve pregnancy using fresh sperm, but this sperm cannot be screened for certain diseases. Licensed clinics are required to freeze and store the sperm for six months, the incubation period of HIV, before insemination to ensure that it is virus free.\n\n> basically they are ruining couples (like us) chances of having a child using donor or want for a better word 'harder' \u2013 they are actually going to force this underground. Couples will go to other agencies (not so legitimate) and abroad to get donors is that fair??? The sad thing is that the women like us may not have as much in the way of protection against certain things like HIV etc I really think they have made the whole donor issue a lot worse.\n> \n> (Tara)\n\nPatients seem to be aware of risks involved in purchasing sperm online, hence they consult each other about the safety of providers. However, further legislative changes have curtailed even their ability to (legally) assess and respond to such risks themselves. Following the implementation of EUTCD within the UK, Internet sperm providers came to be subject to the same rules as HFEA-licensed clinics. Thus, the procurement, testing, processing or distribution of any embryo or gametes intended for 'human use' must now be licensed by the HFEA or be subject to an agreement with a licensed service. Therefore, even if would-be parents would like to take a risk with fresh sperm, they have to think twice. The current Chair of the HFEA, Professor Lisa Jardine, has 'strongly' advised 'any person who became aware that a person or organisation may be procuring, testing, processing or distributing gametes without a licence to contact the police' (Jardine, 2008b).\n\nOn the other hand, if patients are considering obtaining sperm, eggs or embryos from within the EU, a licensed UK clinic can organise for a transfer to be made from that country. However, the sperm, eggs or embryos transferred must meet UK requirements. Would-be parents can also exercise their rights to travel within the EU to receive gametes or embryos from a clinic even if that clinic does not comply with UK standards.\n\nThus, would-be parents feel that the changes made to the HFE Act enjoins them to seek treatment abroad. One forum user, Suzanne, was a patient at a clinic which found a donor for her; however, he did not match either her physical characteristics or those of her husband. Suzanne and her partner thus chose a donor profile from Spain, to be imported by their licensed clinic, but the clinic informed them that UK law prohibited the purchase of sperm from an anonymous donor. If would-be parents cannot purchase anonymous sperm online, or have the sperm imported by a licensed clinic, they construct themselves as having no option but to go abroad for treatment. Other reasons behind reproductive travel are addressed in the following post by Elaine:\n\n> In the end, I think more and more people will go to Europe for treatment, as it costs the same, is completely anonymous, has no waiting lists and, best of all, not bl**dy HFEA!!! That's not to say they aren't regulated, but by sensible people with our interests at heart!!!\n> \n> (Elaine)\n\nThe following post by Tom summarises many of the issues raised by other users:\n\n> There are many infertility newsgroups out there, so it becomes quite easy to compare the approaches taken by different clinics. Some UK clinics seem to be several steps off the pace of both other UK centres and overseas centres. Treatments used almost routinely in other centres are dismissed with the excuse that it is not current clinic policy, or the treatment isn't proven. The result is that the patient feels that they are not being treated as an individual, and they are left with many unresolved questions about their particular situation, and they are paying a premium price for this unsatisfactory state.\n> \n> In many cases patients find that they are managing their own treatment, pushing the medics to try something outside the standard for that clinic. This is probably the most wearing part of treatment for infertility, dealing with a system that you are not part of. It is probably also the reason why many more women in the UK will turn to overseas centres.\n> \n> Overseas treatment will also be favoured because any form of medical treatment in the UK means that at some stage the patient will come up against the NHS, a hugely expensive medical insurance scheme which has now turned into a useless quivering jelly of administrators, processes and targets; fertility treatment is way down the list of priorities.\n> \n> (Tom)\n\nThe accounts I have presented indicate that forums users had significant difficulties in accessing the treatment they sought; this was deemed to be due to the donor shortage, which was itself accounted for as a consequence of legislative change. Following the removal of donor anonymity, would-be parents who could not access treatment due to the donor shortage, and\/or preferred anonymous donors constructed themselves as having no choice but to seek treatment abroad or purchase gametes (sperm) through the Internet.\n\nAlthough discussion forums on online support groups are \u2013 in theory \u2013 rather democratic places where people can express their opinions freely, they are still subject to control by forum moderators. The power of the moderators was exercised when, in May 2006, Repromed's 'Egg and Sperm Donation' forum was suspended. This action followed the HFEA's warning against reproductive care abroad. In April 2006, the then Chair, Suzi Leather, issued a public statement warning British citizens against the poorly regulated treatment in overseas clinics:\n\n> We know that a relatively small number of people choose to travel abroad to undergo fertility treatment and that sometimes the treatment is packaged as a \"holiday\" where the patient can convalesce in the sun. However, we are concerned about people who choose to have their treatment abroad should know about the potential risks. (...) It is very sad when we receive complaints from patients about their treatment abroad and we are not able to help or reassure them. _We would urge patients to think twice and. consider the risks and implications before going abroad for treatment_.\n> \n> (Leather, 2006, emphasis added)\n\nThe Repromed administrator informed me that the 'Egg and Sperm Donation' forum had been closed because members of the HFEA-licensed centre hosting the site were concerned about the increasing number of posts advising participants about offshore facilities for assisted conception using anonymous gametes. The forum was reopened five months later with a disclaimer stating: 'In view of popular demand the donor gamete forum has been reopened, but anyone thinking of going abroad for treatment should first consider the guidance provided by the HFEA, which is quoted below' The HFEA's warning about fertility treatment abroad and Leather's speech were both quoted. Repromed users posted a range of responses to the disclaimer, expressing discontent about the HFEA's approach to fertility travel:\n\n> I'd like to see the HFEA or perhaps other organisation carry out a full investigation into the standards of clinics abroad, to provide some hard evidence that will allow potential patients to make their minds up once fully informed.\n> \n> (Will)\n\n> I do think the HFEA is very heavy-handed in its approach to this issue, and has also made some very poor regulatory decisions (especially about donor anonymity, for example) which may well end up encouraging people to seek treatment elsewhere. I guess they would open to considerable criticism if it could be shown that, in order to avoid the consequences of their approach, lots of people are travelling abroad for treatment, so now they are trying to discourage them with 'scare tactics'. Not a very grown up approach, really.\n> \n> (Melissa)\n\n> I feel that if donor anonymity was guaranteed in the UK you wouldn't be going to abroad anyway. That's down to the government perhaps in places they shouldn't.\n> \n> (Gemma)\n\nSome individuals grounded their critique of the HFEA with reflections on the high quality treatment they had received abroad; in so doing they (re)positioned foreign clinics not as sites of risk and uncertainty, but as morally progressive and highly professional centres of care:\n\n> I really do get fed up with the HFEA scaremongering about going abroad, especially when they've done so much to prevent successful treatment in the UK, for example, removing donor anonymity and successfully wiping out the entire UK sperm bank in one go!!!!!! I for one can only express my utmost admiration at the extremely expert and professional treatment I received at [name of the clinic], which was vastly better than the treatment I received in the UK and resulted in my gorgeous baby at the first attempt. In total, six of us became pregnant after only one or two transfers after going to [name of the clinic], two with twins, which would never have happened if we'd remained in the UK I think that says it all!\n> \n> (Elaine)\n\n> On the note of the HFEA 'warning' all I will say is that I was offered a very high quality individualised treatment at [a clinic] in Spain. I very carefully researched a number of clinics abroad as there were no further options left for me in the UK \u2013 I did not have 5 years to wait for Egg Donation in the UK. I was offered a service and support at xxx that was not on offer at any of the three clinics I was under in the UK. Will, I agree whole heartedly with you \u2013 it would be good for the HFEA to produce some good statistical evidence of what they claim \u2013 I'd like to know just how many people have complained to them about treatment abroad.\n> \n> (Kelly)\n\nThe HFEA's warning about the risks of offshore treatment \u2013 including countries which were popular EU destinations where assisted conception is highly regulated and standards of fertility clinics are demonstrably as high, if not better, than those in the UK \u2013 seemed to have frustrated the would-be parents; they saw themselves as having no choice but to seek treatment abroad as a consequence of the legal shift promoted by the HFEA itself. It was not until the largest fertility support group of the UK brought patients' views together that their voices were heard.\n\n## **Bringing patients' views together**\n\nFollowing the removal of donor anonymity, while hundreds of fertility patients were protesting against the new policy through various online patient support networks, a 2008 fertility travel survey conducted by INUK (Infertility Network UK, 2008) brought these opinions together for the first time. INUK had 339 responses to its survey, which explored patient experiences of cross-border treatment. The results were striking: 76 per cent of respondents stated that they would consider going abroad for treatment (in particular, to receive egg donation). Further, 120 of the 339 (35 per cent) respondents answered the question \"Which country did you go to for your treatment?\"; the most popular response was Spain, which maintains donor anonymity for fertility treatment. At least three times as many patients went there as anywhere else.\n\nAlthough the INUK study received little recognition by public bodies, its findings were confirmed by the European Society of Human Reproduction and Embryology (ESHRE) Task Force, which undertook the most comprehensive study on cross-border reproductive care to-date. The ESHRE survey contained data from 46 clinics across Europe, including 53 patients travelling from the UK. Thirty-four per cent of UK resident patients \u2013 more than any other nationality within Europe \u2013 cited access issues as their reason for travelling abroad (Shenfield _et al._ , 2010).\n\nFindings from a recent qualitative study of UK residents showed that the four most commonly cited reasons for travel were: donor shortages in the UK (27 cases), cost (13 cases), perceived better success rates overseas (12 cases) and previous unsatisfactory care in the UK (7 cases). Overall, 'a desire for timely and affordable treatment with donor gametes' was evident in 71 per cent of cases, making donor conception the most sought after treatment by the CBRC travellers. Most patients sought treatment within European borders, the most popular destinations being Spain and Czech Republic (Culley _et al._ , 2011).\n\nThe growing evidence around reproductive tourism led the HFEA to move away from solely administering warnings, to critical engagements with its policy. In August 2010, the Authority launched a full, public consultation into sperm and egg donation policies. In the consultation document, the HFEA noted that the landscape of donation had changed since the Authority looked at their rules in 2005, and it emphasized that it would like to update these if necessary. It was acknowledged that the demand for donor treatment in the UK was greater than the supply of donors and that the removal of donor anonymity might have exacerbated the shortage. The Authority further noted that the shortage of donors resulted in long waiting times for treatment, with reports of UK waiting times of three to five years for donor eggs. Concerns were also expressed that some fertility centres overseas might offer substandard treatment, and that increasing numbers of people were purchasing unscreened sperm on the Internet (HFEA, 2011). The HFEA asked if gamete supply could be improved by a change in policy, or withdrawal from commitments under the EUTCD (Blyth, 2011).\n\nThe HFEA received a record number of responses to its consultation, with over 2,400 surveys completed by around 1,100 people. The first set of decisions made by the Authority in response was that the maximum number of families that an individual donor can donate to should not be changed. The HFEA also decided to issue guidance stating that sperm and eggs should not be mixed if they come from very close genetic relatives. Further, the allowable compensation amounts for gamete donors have been increased: egg donors can now receive _\u00a3_ 750 per cycle of donation, while sperm donors can receive a _\u00a3_ 35 fee per individual visit to a clinic (Starr, 2011). Finally, the HFEA decided to take a more proactive approach to donor recruitment and established a National Donation Strategy Group. Chaired by Professor Sheila McLean, the Group will bring together a wide range of experts in order to enhance the information that donors receive, develop new approaches to raising awareness of donation, and ascertain how the care of donors can be improved. However, it is questionable whether any of these measures can improve the UK's donor conception policy in terms of its acceptability to users, unless stakeholders' views are central to the HFEA's policy decisions.\n\n## **Conclusion**\n\nThe virtual ethnography study presented in this chapter offers some potentially important insights into how would-be parents negotiate the issues of donor gamete donation interactively and collectively on discussion boards. The topics generated by users of Repromed provide an indication of those matters they perceive as important or problematic with regards to donation. The most important issue that users raise is that, despite the warnings of a likely donor shortage, the government lifted anonymity at the expense of would-be parents' access to treatment, and the HFEA should have taken measures to avoid the (predictable) crisis that resulted from this legislative shift.\n\nThe continued shortage of donor gametes means that UK patients have to wait longer for treatment and have less choice with respect to donor characteristics. In response, we can see (from discussion threads about fertility travel and online sperm providers) that would-be parents have come to experiment with new ways of contesting the law, and \u2013 more importantly \u2013 novel means of avoiding it. The user accounts indicate that donor shortage problems lead couples to find alternative solutions to their childlessness; in particular, would-be parents in the UK have readily available alternatives in their access to treatment in other EU nations. Success rates, the lower cost of treatment, donor availability, and anonymity make clinics abroad increasingly attractive. Yet, the HFEA warn against their use, creating further frustration.\n\nThe posts on the patient forums that I presented here suggest that patients feel their views are ignored by the HFEA in policy-making. But there is little they can do to have their voices heard without compromising their anonymity. Furthermore, the pursuit of a child is, to many would-be parents, a considerably demanding task: physically, emotionally, and in terms of time. Little space remains for lobbying when pregnancy is still to be achieved. Finally, being a would-be parent is a transitional phase: once pregnant, individuals' priorities might change, and campaigning for new policy to address the needs of others is less salient once one's own needs have been met.\n\nIt is important at this point to clarify that I am not claiming that would-be parents are completely divorced from more influential policy actors and processes. Indeed, patients use online environments to make their voices audible to policy-makers: they participate in online surveys, connect with journalists and researchers who recruit interviewees, and sign electronic petitions. However, in effect, UK would-be parents are caught in a double bind: those who are directly affected by the legal shift may be reluctant to press their claims because confidentiality matters to them. Given this reticence, their voices are not taken into consideration by public bodies, like the HFEA.\n\nSince the passing of the 1990 Act, some academic and political commentators have questioned whether the regulatory model adapted for the HFEA was acceptable to the general public, or to the major stakeholders, those who are directly affected by the law (Callus, 2007). Subsequently, the HFEA has faced several legal challenges and its policy decisions were criticised by the House of Commons Science and Technology Committee: 'democracy is not served by unelected quangos taking decisions on behalf of Parliament authorisation' (House of Commons Science and Technology Committee Report, 2002). In one of these legal challenges in 2004, the Court of Appeal confirmed that the HFEA had the legal authority to regulate these areas: \"Parliament has placed that choice in the hands of the HFEA\". But how does the HFEA make these choices? During the public debate over ending donor anonymity it was clear that the proposed legal shift would create a donor crisis, and as a result, fewer people would have access to treatment. But the would-be parents' interests were not a paramount concern to the Authority. The then Chair of the HFEA made it clear that she viewed the principal of the welfare of the child as fundamental, despite the fact that the HFE Act only requires that this should be 'taken into account' (Dawson, 2004: 3).\n\nAs Dawson argues, the HFEA's inspector role is vital to ensure the clinics are fully licensed and monitored (Dawson, 2004). On the other hand, the recent consultation might be an opportunity to separate or improve the Authority's approach to policy-making. Arguably, the UK assisted conception regulations may benefit from a more democratic decision-making process where the 'conflicting' interests of the stakeholders are fully and equally taken into account. This might include more evidence-based policymaking, and taking full advantage of the Internet and information technology to enhance stakeholder participation especially to reach out those who cannot use transparent and formal modes of communication due to their confidentiality concerns, like the would-be parents in this study.\n\nFinally, it should be noted that the passage of the Public Bodies Act 2011 made it possible to transfer the functions of quangos (including the HFEA and Human Tissue Authority) elsewhere, without opening up the original Act that created them. The HFEA was one of several quangos set to be abolished, in order to meet the UK Government's aim of cutting National Health Services (NHS) costs, although this decision has since been reconsidered. On 28 June 2012, the government launched a consultation on the future of the HFEA to consider whether the regulators' responsibilities should move to the Care Quality Commission and the Health Research Authority (HFEA, 2012). Consequently, the likelihood of current HFEA deliberations into mechanisms to increase donation being successfully translated into practice is uncertain. Regardless, it is clear that future efforts to govern the interactions between law, biomedical knowledge and reproductive practice keep in mind the limitations of authority and the creativity of subjects, and endeavour to seek solutions that satisfy the hopes and aims of both regulators and regulatees in democratically-defensible ways.\n\n## **Acknowledgements**\n\nThis work was supported by a Mildred Blaxter Post Doctoral Fellowship from the Foundation for the Sociology of Health and Illness. I am grateful to Emilie Cloatre and Martyn Pickersgill for their constructive comments.\n\n The Human Fertilisation and Embryology Authority (Disclosure of Donor Information) Regulations 2004 SI 2004\/1511.\n\n European Convention for the Protection of Human Rights and Fundamental Freedoms 1950, ETS 5. The ECHR was incorporated into domestic law in the UK via the Human Rights Act 1998.\n\n Directive 2004\/23\/EC of the European Parliament and of the Council of 31 March 2004 on setting standards of quality and safety for the donation, procurement, testing, processing, preservation, storage and distribution of human tissues and cells [2004] L 102\/48. There are also two supplementary Technical Directives: 2006\/17\/EC; 2006\/86\/EC.\n\n At the time of this study, the clinic behind the Repromed website was the University of Bristol's, Centre for Reproductive Medicine. This centre has now merged with the Southmead Fertility Service.\n\n The Human Fertilisation and Embryology Act 2008 section 33 (or the equivalent HFE Act 1990 section 27 for conceptions before 6 April 2009) provides that the woman who gives birth to a child is the legal mother.\n\n Irish Family Planning Association website.\n\n ibid.\n\n Under the directly effective freedom to provide and _receive_ services under Article 56 Treaty on the Functioning of the European Union: _R v. Human Fertilization and Embryology Authority, ex parte Blood_ [1997] 2 All ER 687 (CA). See T. K. Hervey, 'Buy Baby: The European Union and Regulation of Human Reproduction' (1998) _Oxford Journal Legal Studies_ 18: 207.\n\n _R (on the application of J. Quintavalle on behalf of CORE) v. Human Fertilisation and Embryology Authority_ , [2004] Q.B. 168.\n\n Public Bodies Act 2011 (c. 24) Section 5.\n\n## **References**\n\nBlyth, E. (2011) 'Gamete donation review: not the HFEA's finest hour', _Bionews_ , 605 w\u00adw\u00adw\u00ad.\u00adb\u00adi\u00ado\u00adn\u00ade\u00adw\u00ads\u00ad.\u00ado\u00adr\u00adg\u00ad.\u00adu\u00adk\u00ad\/\u00adp\u00ada\u00adg\u00ade\u00ad_\u00ad9\u00ad4\u00ad0\u00ad1\u00ad2.\u00ada\u00ads\u00adp (accessed 23 September 2013).\n\nCallus, T. (2007) 'Patient perception of the human fertilisation and embryology authority', _Medical Law Review_ , 15: 62\u201385.\n\nCulley, L., Hudson, N., Rapport, F., Blyth, E., Norton, W. and Pacey, A. (2011) 'Crossing borders for fertility treatment: motivations, destinations and outcomes of UK fertility travellers', _Human Reproduction_ , 26: 2373\u20132381.\n\nDawson, A. (2004) 'The human fertilisation and embryology authority: evidence based policy formation in a contested context', _Health Care Analysis_ , 12: 1\u20136.\n\nDonor anonymity to be removed in UK (2004) _Bionews_ , 242, w\u00adw\u00adw\u00ad.\u00adb\u00adi\u00ado\u00adn\u00ade\u00adws\u00ad.\u00ado\u00adr\u00adg\u00ad.\u00adu\u00adk\u00ad\/\u00adp\u00ada\u00adg\u00ade\u00ad_l\u00ad1\u00ad8\u00ad3\u00ad3\u00ad.\u00ada\u00ads\u00adp (accessed 9 September 2013).\n\nHuman Fertilisation and Embryology Authority (2004) 'Sperm, egg and embryo donation (SEED) policy review: findings of the clinic survey', w\u00adw\u00adw\u00ad.\u00adh\u00adf\u00ade\u00ada\u00ad.\u00adg\u00ado\u00adv\u00ad.\u00adu\u00adk\u00ad\/\u00add\u00ado\u00adc\u00ads\u00ad\/\u00adC\u00adl\u00adi\u00adn\u00adi\u00adc\u00ads\u00ad_\u00ads\u00adu\u00adr\u00adv\u00ade\u00ady\u00ad_\u00adS\u00ade\u00ade\u00add\u00ad_\u00adr\u00ade\u00adv\u00adi\u00ade\u00adw\u00ad.\u00adp\u00add\u00adf (accessed 18 September 2013).\n\nHuman Fertilisation and Embryology Authority (2011) 'A review of the HFEA's sperm and egg donation policies', w\u00adw\u00adw\u00ad.\u00adhfea\u00ad.\u00adg\u00ado\u00adv\u00ad.\u00adu\u00adk\u00ad\/\u00add\u00ado\u00adc\u00ads\u00ad\/\u00ad2\u00ad0\u00ad1\u00ad1\u00ad-\u00ad0\u00ad1\u00ad1\u00ad3\u00ad_\u00adD\u00ado\u00adn\u00ada\u00adt\u00adi\u00ado\u00adn\u00ad_\u00adr\u00ade\u00adv\u00adi\u00ade\u00adw\u00ad_\u00adb\u00ada\u00adc\u00adk\u00adg\u00adr\u00ado\u00adu\u00adn\u00add\u00ad.\u00adp\u00add\u00adf (accessed 26 September 2013).\n\nHuman Fertilisation and Embryology Authority (2012) 'Consultation on proposals to transfer functions from the human fertilisation and embryology authority and the human tissue authority', w\u00adw\u00adw\u00ad.\u00add\u00adh\u00ad.\u00adg\u00ado\u00adv\u00ad.\u00adu\u00adk\u00ad\/\u00adh\u00ade\u00ada\u00adl\u00adt\u00adh\u00ad\/\u00ad2\u00ad0\u00ad1\u00ad2\u00ad\/\u00ad0\u00ad6\u00ad\/\u00adc\u00ado\u00adn\u00ads\u00adu\u00adl\u00adt\u00ada\u00adt\u00adi\u00ado\u00adn\u00ad-\u00adr\u00ade\u00adg\u00adu\u00adl\u00ada\u00adt\u00ado\u00adr\u00ads (accessed 27 September 2013).\n\nHouse of Commons Science and Technology Committee Fourth Report (2002) _Developments in Human Genetics and Embryology_ , 18 July 2002 HC 791. Paragraph 18.\n\nJardine, T. (2008a) 'Internet sperm providers are illegal', _Guardian News and Media Limited_ , 23 September 2008, w\u00adw\u00adw\u00ad.\u00adg\u00adu\u00ada\u00adr\u00add\u00adi\u00ada\u00adn\u00ad.\u00adc\u00ado\u00ad.\u00adu\u00adk\u00ad\/\u00adu\u00adk\u00ad\/\u00ad2\u00ad0\u00ad0\u00ad8\u00ad\/\u00ads\u00ade\u00adp\u00ad\/\u00ad2\u00ad3\u00ad\/\u00adu\u00adk\u00adc\u00adr\u00adi\u00adm\u00ade\u00ad.\u00adl\u00ada\u00adw\u00adl (accessed 23 September 2013).\n\nJardine, L. (2008b) ' _BFS report on sperm donation in the UK: HFEA response_ ', _Human Fertilisation and Embryology Authority Paper_ , (HFEA (02\/07\/08) 452. Para 2.2).\n\nLeather, S. (2004) 'Sperm donors to lose anonymity', _BBC News_ , 21 January 2004, h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00adn\u00ade\u00adw\u00ads\u00ad.\u00adb\u00adb\u00adc\u00ad.\u00adc\u00ado\u00ad.\u00adu\u00adk\u00ad\/\u00adl\u00ad\/\u00adh\u00adi\u00ad\/\u00adh\u00ade\u00ada\u00adl\u00adt\u00adh\u00ad\/\u00ad3\u00ad4\u00ad1\u00ad4\u00ad1\u00ad4\u00ad1\u00ad.\u00ads\u00adt\u00adm (accessed 25 September 2013).\n\nLeather, S. (2006) 'Press release: thinking of going abroad? Think twice about going abroad for fertility treatment', w\u00adw\u00adw\u00ad.\u00adh\u00adf\u00ade\u00ada\u00ad.\u00adg\u00ado\u00adv\u00ad.\u00adu\u00adk\u00ad\/\u00ad6\u00ad3\u00ad2\u00ad.\u00adh\u00adt\u00adm\u00adl (accessed 23 May 2012).\n\nMatorras, R (2005) 'Reproductive exile versus reproductive tourism', _Human Reproduction_ , 20: 3571.\n\nMcGuinness, S. (2011) 'A, B, and C leads to D (for delegation!): A, B and C v. Ireland 25579\/05 [2010] ECHR2032', _Medical Law Review_ , 19: 476\u2013491.\n\nMcVeigh, K. (2007) 'Online ads for sperm donors called offensive', _Guardian News and Media Limited_ , 26 March, w\u00adw\u00adw\u00ad.\u00adg\u00adu\u00ada\u00adr\u00add\u00adi\u00ada\u00adn\u00ad.\u00adc\u00ado\u00ad.\u00adu\u00adk\u00ad\/\u00adm\u00ade\u00add\u00adi\u00ada\u00ad\/\u00ad2\u00ad0\u00ad0\u00ad7\u00ad\/\u00adm\u00ada\u00adr\u00ad\/\u00ad2\u00ad6\u00ad\/\u00adn\u00ade\u00adw\u00adm\u00ade\u00add\u00adi\u00ada\u00ad.\u00ada\u00add\u00adv\u00ade\u00adr\u00adt\u00adi\u00ads\u00adi\u00adn\u00adg (accessed 26 September 2013).\n\nShenfield, Rde Mouzon, J., Pennings, G., Ferraretti, A.P., Andersen, A.N., de Wert, G., Goossens, V., ESHRE taskforce on cross border reproductive care (2010) 'Cross border reproductive care in six European countries' _Human Reproduction_ , 25: 1361\u20131368.\n\nStarr, S. (2011) 'HFEA takes first set of decisions following donation review', _BioNews_ , 616, w\u00adw\u00adw\u00ad.\u00adb\u00adi\u00ado\u00adn\u00ade\u00adw\u00ads\u00ad.\u00ado\u00adr\u00adg\u00ad.\u00adu\u00adk\u00ad\/\u00adp\u00ada\u00adg\u00ade\u00ad_\u00ad1\u00ad0\u00ad2\u00ad1\u00ad9\u00ad9\u00ad.\u00ada\u00ads\u00adp (accessed 27 September 2013).\n\nThe Infertility Network UK (2008) _Fertility Tourism Survey Results_ , Bexhill-on-Sea: Infertility Network.\n\nTurkmendag, I., Dingwall, R, and Murphy, T. (2008) 'The removal of donor anonymity in the United Kingdom: the silencing of claims by would-be parents.' _International Journal of Law, Policy and the Family_ , 22: 283\u2013310.\n\nTurkmendag, I. (2012) 'The donor-conceived childs \"right to personal identity\": the public debate on donor anonymity in the United Kingdom', _Journal of Law and Society_ , 39: 58\u201375.\n\nTurkmendag, I. (2013) 'When sperm cannot travel: experiences of UK would-be parents seeking treatment abroad', in Hear, M. L., Farrell, A., Hervey, T. K and Murphy, T. (eds) _(2013) European Law and New Health Technologies_ , Oxford: Oxford University Press.\n\nWaldby, C. (2008) 'Oocyte markets: women's reproductive work in embryonic stem cell research', _New Genetics and Society_ , 27: 19\u201331.\nSection 2\n\n* * *\n\n# Objects and epistemologies in criminal law\n\n* * *\nChapter 4\n\n* * *\n\n# **Unchaining research**\n\nProcesses of dis\/empowerment and the social study of criminal law and investigation\n\n_Barbara Prainsack_\n\n* * *\n\n## **Introduction**\n\nThrough their analyses of legal practices and processes of knowledge production within them, science and technology studies (STS) scholars have made considerable contributions to critical scholarship on law, science, and society in recent decades. One of the most significant achievements of STS scholarship has been its interrogation of the ways in which scientific and technological knowledge and artefacts are embedded in legal systems. Nevertheless, there are other aspects in the domain of law \u2013 and in particular also in criminal justice \u2013 which have not yet seen the same level of critical scrutiny. In these areas, the scope and direction of STS research specifically has been restrained by the tacit acceptance \u2013 the black-boxing \u2013 of some of the categories that other professional experts operate with. In the context of an area such as STS, whose emergence and whose identity is closely linked to the challenging of assumptions and casting light on 'hidden' matters, this is surprising. In this chapter, after discussing in more detail two such under-explored issues within STS and socio-legal scholarship, I will reflect on the conditions and processes within and by which many scholars operate that account for this phenomenon. I will conclude by highlighting two strategies to address these issues.\n\n### **_Unexamined assumptions_**\n\nAbout ten years ago, in an article discussing the first wave of socio-legal studies of wrongful convictions, San Francisco-based legal scholar Richard Leo spoke of the salience of 'unexamined assumptions' within this literature:\n\n> The unexamined assumption in virtually all miscarriages writing and scholarship is that [the reasons examined for miscarriage of justice] are actual causes, and once they are identified, we will know how and why the problem of wrongful conviction occurs. [...] But eyewitness misidentification, false confession, and police and prosecutorial misconduct are not actual root causes. By identifying the causes, we beg the obvious, deeper causal question: What are the causes of eyewitness misidentification? What are the causes of police-induced false confessions? What are the causes of police and prosecutorial conduct? [... Overall, scholars] need to move beyond the legal categories and concepts handed to them by journalists and lawyers and start to reconceptualize the study of miscarriage of justice, most fundamentally, as about the study of human behaviour and human error in social and organizational contexts.\n> \n> (Leo 2005: 213; see also Machado and Prainsack 2012; Heinemann _et al._ 2012)\n\nLeo's critique was primarily aimed at legal scholarship, of course, and both research in STS and socio-legal studies has certainly gone a long way in doing exactly what Leo calls for; namely, challenging the categories and concepts handed to them by actors in the field they study. Seminal STS work on law more generally, and criminal justice in particular, has problematised hitherto unchallenged categories, practices, or technologies such as the belief in the self-corrective ability of science; the rule of law over science; the notion of expertise; the truth claims inherent in scientific, technical, and other kinds of evidence; and the relationship between materiality and conceptuality. Both STS and socio-legal scholarship have also helped to complicate the neat and linear way in which success stories about science and technology in criminal investigation and criminal justice are told.\n\nYet there is still work left to do. We have not yet engaged sufficiently with some unexamined assumptions within criminal law and investigations, obstructing a deeper understanding of norms and practices in the field. In the following section I will discuss two broad issues in the field of criminal justice within which unexamined assumptions are salient and under-explored discourses and practices carry significance, dis\/empowering particular actors, institutions, and technoscientific practice: first, the nature of conviction and of convicts; and second, the role of forensic DNA testing as a technology of individualisation in the sense that they establish a robust link between a person and another construct (here: a crime). These examples have been salient within my own research on forensic DNA technologies and practices, and I hope that they will be able to help to illuminate some of the dynamics underlying the (occasional) ignoring of important social and epistemic assumptions and issues in other areas of STS scholarship.\n\n## **Power, conviction and convicts 8**\n\nIn recent years, the Innocence Project, a US-based litigation and public policy organisation (founded in 1992) dedicated to overturning wrongful convictions, has received a lot of attention in socio-legal scholarship and beyond (see, for example, Risinger 2007; Krimsky and Simoncelli 2011; Krieger 2011; Machado and Prainsack 2012). Although the Innocence Project was founded in the US, and most of its activities are still focused within that region, similar projects and initiatives in other parts of the world have emerged. Besides the obvious and tangible effects that the work of the Innocence Project has on the lives of those who were exonerated due to the _pro bono_ work of its members, it has also had a very pronounced impact on public awareness of the problems and possibilities for errors inherent in the social, legal and technological practices that are employed in criminal investigation and in court. It is no coincidence that calls for more scrutiny of forensic science and practice \u2013 wherever possible according to scientific standards \u2013 are increasingly made by professional institutions and within reports commissioned by governments (e.g. National Academy of Sciences 2009).\n\nThe supposition underpinning such calls, however \u2013 namely that making forensics more 'scientific' will help solve the problem of wrongful convictions \u2013 ignores insights of several decades of STS and socio-legal scholarship, which have challenged this rather simplistic conceptualisation of the role of science in problem solving (see also Edmond 2011). In this respect, such understanding, that 'better' forensic science will solve the problem of wrongful convictions, contradicts the thrust of much STS and socio-legal scholarship.\n\nIn another respect, however, the spirit of calls for a better research culture in forensics _resonates_ with STS scholarship: it implies the admission of a particular kind of mistake on the side of actors in the legal system; i.e., that they have not taken a more critical stance towards expertise and evidence in the past. This admission reinforces STS' critical stance 'through the back door'. And this is indeed how the lesson of the Innocence Project seems to have been received within STS so far: things go wrong when the authority of established experts (in any relevant context: as fingerprint experts, molecular biologists, etc.) is accepted without much scrutiny and reflection (see also Bal 2005).\n\nAccording to an analysis of the causes of the wrongful convictions of the first 239 people that the Innocence Project has helped to exonerate (West 2010), however, it was not the use of high-tech forensic applications, but something very different that was identified as the single most important basis on which the conviction had rested: false eyewitness evidence. This finding has not had much traction in STS research. Perhaps this is to be expected, as STS is by definition the social study of science and technology, and eyewitness testimony does not obviously fall within this remit. But the silence of STS scholarship on this issue could also stem, at least partly, from the fact that the story of the Innocence Project so far \u2013 a narrative of human error (wrongful convictions) that are overturned with the help of a technology (DNA testing) \u2013 has the culprit and the hero in the wrong roles. STS scholars have spent decades countering technological instrumentalism, the view that science and technology are trustworthy in principle and 'tarnished only by the humans using it' (Edmond 2011: 133). Accordingly, the fundamental lesson from the Innocence Project \u2013 i.e., that technology helps undo the mistakes that people make \u2013 conflicts with some of the core tenets of STS. By remaining silent on these issues, the technocratic systems of power that give life to current mechanisms of conviction, are implicitly validated.\n\nThe core narrative of the Innocence Project is also dependent on a clear distinction between 'perpetrator' and 'victim', concepts that have also remained largely unchallenged within STS. It is true that those scholars whose work has helped to debunk assumptions of infallibility of forensic practices and technologies are usually careful not to conflate the label 'convict' with the assumption of guilt, but somehow this has had no wider implications so far. This is even more surprising in the context of a discipline that has so long been engaged in making science and technology more participatory and more 'democratic' through the inclusion of the experience and expertise of people other than social elites and professional experts. Although many convicts have relevant and diverse expertise and experiences in the area of crime scene technologies, this group has not been included in such efforts.\n\nThe experience and expertise of convicts in this respect can be twofold: first, convicts are, in their majority, a marginalised group who have little, if any, influence over how science and technology are used on and against them, while this use has, in many instances, fundamentally life changing effects. Thus many convicts are able to provide perspectives on science and technology from the viewpoint of some of the least powerful actors in a given field. Second, some convicts have relevant 'alternative' professional knowledge pertaining to technology use: this applies to so-called 'career criminals' for example, whose technical and technological knowledge in the realm of forensics applies to avoiding, concealing, or destroying crime scene traces (Prainsack and Kitzberger 2009; Machado and Prainsack 2012).\n\nTo clarify: I am not suggesting that convicts necessarily _should_ be included in our quest for making science and technology more participatory. Bather, I argue that the fact that they have, in general, not yet been included in such endeavours, while other marginalised groups have (e.g. people with disabilities, older adults, ethnic minorities), is instructive. What exactly does this disempowerment tell us? It could be argued that the relative absence of perspectives of convicts from discussions about science and technologies used in criminal investigations and in courts is due to the fact that prisoners are virtually inaccessible as research participants: the obstacles put up by research ethics guidelines, one could say, are so difficult to overcome that they discourage researchers from even trying. This may be a compelling argument insofar as it pertains to prisoners; yet, the group of convicts is much wider, including many people who have never been in prison, and those who were released. Although there may be difficulties in reaching members of these groups as well, these obstacles are regularly no more difficult to overcome than those in respect to working with members of other marginalised groups.\n\nA much more plausible explanation for why convicts have not been listened to in STS scholarship is because there is a tacit acceptance of a wider societal assumption that they should be at the margins of society. This is also what the guilty verdict implies: being guilty means that the guilty person has infringed the social order in some significant way, so that the convicted person is placed outside society, if only temporarily (at the very least, during the commission of the breach; at most, in the literal sense, i.e. being 'locked away' for life) (see also Taslitz 2011). Because they are at the margins \u2013 or even outside \u2013 of society, they have not been included in previous efforts of fostering dialogue and exchange at the interface of science and society. And because the particular distribution of social status and power between actors from criminal justice systems and the supposedly law abiding citizens that they represent on the one hand, and convicts on the other, is tacitly accepted within wide ranges of STS scholarship, other truth claims inherent in criminal justice systems are also taken on board.\n\nExamples of the latter are different valuations of expertise and the power of particular kinds of experts. This, in turn, has two consequences. First, the expertise of convicts pertaining to the science and technology used in legal processes is to be subordinated to, or to be ignored at the cost of, the expertise of other experts. Second, such expertise on the side of convicts is not recognised as professional expertise, even if certain actors do refer to it as such \u2013 typically, 'career criminals', who see their criminal activities as their jobs (Machado and Prainsack 2012).\n\nIn sum, consideration of the nature of conviction and convicts in terms of power and technoscience underscores the importance of refocusing the critical scrutiny of STS and also socio-legal studies. First, an emphasis on what are 'obviously' scientific and technological matters obscures the great degree to which other forms of epistemic and material practices play important roles in the technocratic machinery powering the systems within which conviction occurs (and the debates associated with this). Second, an uncritical acceptance of dominant societal narratives concerning the nature of 'the convict' disempowers convicts as a group and fails to recognise the diverse kinds of expertise they often hold (with implications for both democratic processes and criminological discourse).\n\nIn what follows, I turn now to address an instance where STS scholarship has largely adopted an unexamined assumption underpinning the field of forensic science (specifically, forensic genetics) into its own domain.\n\n## **(Re)producing DNA testing as a technology of individualisation**\n\nA recent article by the geneticist Ant\u00f3nio Amorim makes two propositions: first, that 'the results of forensic DNA testing are limited in both theoretical and practical terms', and second, that an understanding of the reasons for these limitations is crucial even for people outside the field of forensic genetics (Amorim 2012: 260). Most scholars in both STS and socio-legal studies will not be surprised by these suggestions; they have long attended to the limitations of forensic DNA testing. Yet when reading Amorim's argument in detail, some deeper implications of these limitations move to the foreground.\n\nTo begin with, in contrast to fingerprints, for example, which are recorded in the form of visual data (and their annotations), traditional DNA technologies are recorded as a small set of discrete, numeric data (see also Cole 2001; Lynch _et al._ 2008). When we look at a pre-defined set of specific places on the genome, then we find that different people have different numbers of so-called 'short tandem repeats' (STRs, i.e. repetitions of chains of nucleotides) in these places. The larger the number of such places (loci) that are analysed, the smaller the likelihood that two different people _coincidentally_ have the same number of repetitions. This is the basic underlying principle of 'traditional' forensic DNA analysis. It is a principle of stratification, of assigning people to different groups according to their genetic markers. This situation would, in fact, be no different if we did not count the number of repeats in a particular locus but if we noted down the _exact sequence_ of the nucleotides A, T, G, and C themselves as we find them in that locus. The particular order of nucleotides in a given locus describes a geno _type_ that is shared by other individuals (see also Amorim 2012: 261). If we look at the sequence of nucleotides in all possible loci on a person's genome, there is \u2013 again with the exception of monozygotic twins and multiples \u2013 such a small chance that the exact order of all nucleotides in all loci between two individuals matches _by chance_ that it is considered practically (and thus forensically) impossible.\n\nBut stating that something is a practically and thus forensically impossibility does not mean, of course, that it is impossible _in principle_. As Amorim summarises:\n\n> [There are] no unique individuals: everyone belongs to a class shared by many others \u2013 at least potentially, depending on the frequencies of [the nucleotides] A, T, G, and C. Thus, individuals cannot be distinguished from each other at this level of analysis.\n> \n> (Amorim 2012: 261)\n\nForensic DNA analysis never establishes a _certain_ link between a marker and a particular individual, but such a link is always probabilistic. This deficit is, for practical purposes, overcome by the use of statistics. This is necessary because in practice, forensic DNA analysis is mainly used exactly for what it technically cannot do: identifying individuals (via the establishment of a link between a marker from the DNA found at a crime scene and an individual).\n\nThis is not a trivial point; there are few other instances where the connection between a 'marker' and an individual is made by means of statistics and has such severe consequences. In the context of forensics, such a probabilistic link has potentially very grave consequences if accepted in court as 'proof of a link between a person and a crime scene: it may put the narrative of a case in a different light (Johnson and Williams 2004), and in extreme cases, even death sentences can be the result. Even when the DNA profile is treated 'merely' as evidence corroborating other evidence, the consequences can still be very serious. Yet the discussion of the problems inherent in the use of forensic DNA technologies in STS and scholarship on their ethical, legal, and social implications (ELSI) has so far focused primarily on other issues: besides those pertaining to other categories of problems altogether \u2013 such as contamination or flawed analyses of the DNA sample \u2013 discussions about the robustness of DNA analysis have revolved around how to correctly calculate random match probabilities, so that they do not come out as too low or too high (see M'Charek 2005; Aronson 2007; Aronson 2010; Prainsack 2010). The fact that 'true' individualisation is not possible at all on the basis of DNA technologies has been acknowledged but not widely problematised, even within STS scholarship (see M'Charek 2005). This situation compromises the preparedness and effectiveness with which STS arguments can be used to publicly counter claims such as those inherent in the trial of Cleveland Barrett.\n\nIn 2010, Ohio resident Cleveland Barrett was charged with predatory criminal sexual assault of a nine-year-old girl. The evidence presented in the trial against him included both victim testimony and the result of a DNA test. Jurors were told in the closing argument by the Assistant State Attorneys that DNA evidence had been found on the victim that 'matched' Barrett's profile. According to a trial transcript, an Assistant State Attorney posed the rhetorical question to the jury, and provided an answer: 'Who is the major profile in the DNA that's found? The defendant' (quoted from Mills 2012).\n\nWas this a case where the random match probability \u2013 i.e. the chance that the profile obtained from the crime scene matched a randomly drawn person from the general population \u2013 was so small that it was practically impossible for the DNA profile not to have come from Mr Barrett? Was it for this reason that the DNA profiles were simply described as 'matches'? No, and no. The DNA sample found at the crime scene \u2013 on the victim's lips, to be precise -had not resulted in a proper DNA profile at all. When probed, the Illinois State Police crime lab analyst stated during the trial that the DNA test on the sample had only yielded results in one locus; the number of STRs in the other nine genetic loci that are normally tested for could not be determined (this can be the result of degraded DNA, or if insufficient amounts of DNA are available for testing). A one-locus profile, because of the very high likelihood that it could have come from another person who 'accidentally' has the same genotype in that particular locus, is not a suitable 'proof for anything; at best, it could indicate that the defendant cannot be excluded as a source of the trace based on the DNA evidence alone.\n\nThis also explains why the State Prosecutors presented the DNA 'evidence' in such an unusual manner: what the Assistant State Prosecutors had done was to look at how frequently the particular profile from the DNA found on the victim was found in different population groups. In other words, what was the proportion of specific ethnic groups that had this particular genetic variant at this particular locus? These statistical data were then translated into statements about how likely it was that the profile had come from one particular suspect, in this case, Mr Barrett. In this instance, the particular genetic profile found in that one locus that was analysed matched, on average, one in four African-American males, one in eight Hispanic males, and one in nine Caucasian males (Mills 2012). Cleveland Barrett fit the label African American, and thus the DNA profile was sold to the jury as 'matching' Mr Barrett's.\n\nMr Barrett was eventually acquitted, but by then he had already been detained for over a year. How could this happen? How could it be that the minimal DNA 'profile' obtained from the perpetrator's DNA found on the victim was presented to the jury as a 'match' to Mr Barrett's DNA? I would argue that the reason for this was exactly the situation described above, namely that DNA technologies are treated as an individualising method. The fact that DNA technologies cannot provide a certain link to an individual is then overshadowed by the practical and forensic 'impossibility' of adventitious matches if the number of the analysed markers is high enough. Because the scenario of adventitious matches, in many cases, does not matter practically, DNA technologies are seen as 'safe' if the process is done right. This brings us back to exactly the kind of technological instrumentalism that STS has typically challenged, namely that if the application of a technology leads to bad results, the reason for this lies in the application, not in the technology. The effect of this is that everything that carries the label 'DNA test' is seen a reliable method of truth production.\n\nThe question remains open as to why STS scholarship, while having challenged so many other hitherto unexamined assumptions in the field of law and criminal justice, has not prominently highlighted that forensic DNA technologies are tools for stratification (i.e. tools to assign different profiles to different groups), rather than for individualisation. One possible explanation is that STS has, at least to a considerable extent, focused upon 'the usual suspects'. In the area of law and law enforcement, research has been primarily focused on the controversies that lawyers have 'handed us down', to use Richard Leo's terminology once more (see also Edmond 2011).\n\nA key case in point is represented by the so-called 'DNA wars', a heated discussion among experts and practitioners in the aftermath of the first uses of DNA evidence in the 1980s about the scientific 'robustness' of the technology. In this debate, many aspects pertaining to the scientific quality of forensic DNA profiling, and arguments in favour or against of forensic admissibility, were critically examined by STS scholarship (for an overview, see e.g. Aronson 2007; Lynch _et al._ 2008). These examinations, however, remained to a large extent within the existing boundaries of either legal or existing STS discourse. Other discourses, such as that of molecular genetics, for example, while playing a very prominent role in arguments put forward by the participants of the 'DNA wars' themselves, has played a minor role within STS scholarship (for an exception, see Lynch _et al._ 2008; Williams and Johnson 2008; Williams 2010). Also other discourses, including those pertaining to political theory, have not shaped the research questions and methods of STS scholars in this field to the same extent as other discourses have.\n\n## **Unchaining research: how can we get out of 'prison'?**\n\nHow can we solve these problems? How can we 'unchain' STS scholarship in the context of law so that it enables us to see more of the 'root causes' (Leo 2005: 213) for why things are as they are? An approach to a solution will need to contain at least two elements. First, we need to challenge even more what seems unproblematic and obvious. We need to query not only the assumptions and presumptions of the actors in the field that we study, but also our own (e.g. Valverde 2009; Farmer 2010). This means that we need to critically examine (for example) not only how, and according to what standards, truths are produced within the criminal justice system, but we also need to interrogate the practices, norms, and values that render the system legitimate.\n\nSuch an endeavour could be inspired by a commitment to recognise the hegemonic nature of every social order, which, as Chantal Mouffe and other critical discourse theorists (e.g. Mouffe 2005; Howarth 2000) have argued, is always predicated on the exclusion of other possibilities. Applied to the criminal justice system, such an analysis would entail asking how the system could look like if it did not look as it does (see also Lynch 2001; Pollner 1991). Comparative research is one possible way to address such questions, but our thought experiments do not need to be limited by the range of existing scenarios. If fines and prison sentences did not exist \u2013 perhaps because the system operated on the assumption that there was no free will, and that every crime or offence was in some way biologically and\/or environmentally determined \u2013 would such a system operate with a different notion of 'guilt'? What kind of truth would be negotiated in a 'criminal' trial of this sort, and what would be the standards according to which they would be negotiated? Who would make authoritative decisions in such a system? In such a fictitious system, it is likely that science would play a different role, because the truth that this system strived for would serve a different purpose.\n\nI am not advocating that STS and socio-legal researchers should all become armchair philosophers. But keeping in mind that such alternative realities could exist, and that they could provide a fruitful point of reference for our work, can be important (see also Lynch 2011). This process of rendering the familiar exotic is, of course, a task that anthropologists especially have long been concerned with, and STS and socio-legal researchers carrying out ethnographic work are regularly trained to do as well. However, the commitment to challenging what is uncontroversial and obvious to us is something that we need to maintain also outside our immediate field work; in particular, when conceptualising our research and framing the questions we hope it might answer.\n\nA second approach to further enhancing analyses of STS scholarship in the domain of criminal justice consists in rendering explicit the distribution of power and political agency. This is an issue socio-legal scholars have traditionally been highly sensitive to, and much STS research has placed the articulation of agency in the foreground. It will be fruitful to not stop at rendering visible the 'hidden' agency of people and things, but also pose the genealogical question of how particular types of collective action came into being, and how the agency of some entities is more salient than others (see Toom 2012). This requires a particular focus on the role of polities in systems of collective action.\n\nInstances of _situated dis\/empowerment_ (Prainsack and Toom 2010: 1120; see also Haraway 1991: 195; Gad and Lauritsen 2009) may be instructive to bear in mind and locate, in order 'to highlight the importance of \"location, positioning, and situating'\" of practices and contexts, and be perceptive to the ways in which empowering and disempowering processes \u2013 i.e. those which influence the scope of agency of specific actors \u2013 are always intertwined and frequently mutually constitutive. Attention to instances of situated dis\/empowerment enjoins researchers to take seriously the agency of non-human actors while at the same time keeping in view the distribution of political power, where human actors play a different (albeit not necessarily more important) role than non-human ones. Greater attention to rendering explicit the distribution of power and political agency in the fields we study, in conjunction with a continued focus on scrutinising our own tacit assumptions, will go a long way helping us to unpack the remaining unexamined assumptions within STS and socio-legal scholarship.\n\n## **Conclusion**\n\nStudies of technoscience, law, and society have benefited from research and conceptual frameworks emerging from STS, but unexamined assumptions still remain within scholarship in this area, and issues that are salient for the actors occupying the social field of criminal justice continue to be under-explored. In this chapter, I have discussed in detail two broad areas of dis\/empowerment relating to knowledge and technology that I have encountered in the subject area of my own work on the political and social aspects of forensic DNA testing: first, the flows of power intrinsic to the production of the nature of conviction and of convicts; and second, the role of forensic DNA testing as a technology of individualisation capable of linking a specific person to a specific crime. After showing some of the largely unexamined assumptions and omissions within STS scholarship, and discussing some of the consequences of these, I suggested two strategies for enhancing scholarly capacity to open even more black boxes in socio-legal domains. First, continuing to challenge what seems unproblematic and obvious not only to the actors, but to researchers; and, second, explicating the distribution of political power and agency in law and legal processes. Engaging closely with technoscientific practices that structure aspects of law may be an important means of undertaking both of these tasks. In some respects, STS and socio-legal scholars have already attended to them; nonetheless, new and specific emphasis will be a fruitful endeavour.\n\n## **Acknowledgements**\n\nI am grateful to Ant\u00f3nio Amorim, Troy Duster, Marie-Andr\u00e9e Jacob, Corinna Kruse and Victor Toom for their comments on an earlier version of this manuscript, and to Emilie Cloatre and Martyn Pickersgill for having been immensely helpful editors. The usual disclaimer applies. My thanks go to Troy Duster also for bringing the Cleveland Barrett case to my attention.\n\n See, for example, Jasanoff 1995; Lynch _et al._ 2008; Obasogie and Duster 2011.\n\n E.g. Jasanoff 1995; Mercer 2002.\n\n E.g. Smith and Wynne 1989; Jasanoff 1995; Bloor 2000; Cole 2001; Edmond 2004; Lynch and Cole 2005; see also Collins and Evans 2002; Collins and Evans 2007.\n\n For a problematisation of scientific and technical evidence, see, for example, Smith and Wynne 1989; Jasanoff 1995; Angell 1996; Lynch _et al._ 2008; Williams and Johnson 2008; Williams 2010; and Pottage 2011. For a discussion of other kinds of evidence, see also Bloor 2000; Cole 2001; Golan 2004, Lynch _et al._ 2008; M'Charek 2008.\n\n See e.g. Golan 2004; Gillespie 2006; Pottage 2011; Faulkner _et al._ 2012.\n\n See, for example, Cole 2001; Aronson 2007; Jasanoff 2010; Kruse 2012; Prainsack and Toom 2012.\n\n My own research in this area has so far been twofold: it has focused on exploring marginalised perspectives on the role, utility, and broader societal meaning of bioscientific applications in criminal investigation and forensics (e.g. Prainsack and Kitzberger 2009; Machado and Prainsack 2012), and on the shifts in the distribution of power and agency that are intertwined with regulatory measures and approaches (e.g. Hindmarsh and Prainsack 2010; Prainsack and Toom 2010, 2012).\n\n I use the term 'convicts', and not 'offenders', because the latter indicates that they have actually committed an offence or a crime, which the former term, strictly speaking, does not. Moreover, speaking about 'convicts' makes sense with regard to the argument about their experience with forensic technologies; a certain level of contact with criminal investigation and the criminal justice system is required for this to materialise (i.e. somebody who has committed a crime but never been the subject of an investigation is not somebody whom forensic technologies have been used upon, as argued here).\n\n Nucleotides make up human DNA; long chains of DNA are 'rolled' into chromosomes.\n\n The chance that two people do have the same number of repetitions in the same places \u2013 the so-called 'random match probability' \u2013 increases when the number of repeats in a certain loci are relatively typical in the respective population, or when individuals are biologically related. For example, full siblings share 50 per cent of their DNA on average. This means that on average, half of the genetic information (number of repeats) in the loci that forensic tests analyse will be the same. This phenomenon is utilised in a technique called 'familial searching'; the method takes profiles that match in many but not all loci as an indicator that the originator of the profile that is searched for is a biological relative of the one whose profile matched in more loci (than would be expected for unrelated individuals). For more detail, see Murphy (2010).\n\n It should be noted here that 'traditional', STR-based forensic DNA testing does not look at all loci of a person's genome but only to about 10\u201315, depending on the particular genetic loci typed, and the kind of software used, in a given jurisdiction.\n\n Also for this reason, it has been argued that any forensic technique or technology that delivers probabilistic results should not be grounds for a conviction in the absence of convincing other evidence (Mnookin _et al._ 2011: 747).\n\n The delineation of racial categories \u2013 African-American, Hispanic, Caucasian, etc. \u2013 remained an unexamined assumption not only in the trial, but also in the media coverage of the case. There is a growing body of STS scholarship on this topic; see, for example, Duster 2004, 2006a, 2006b, 2008; M'Charek 2008a, 2008b, M'Charek _et al._ 2012; Reardon 2005, and it is also a hotly debated topic in forensic genetics.\n\n This situation is changing in light of growing transnational exchange of DNA profiles within Pr\u00fcm regime (Prainsack and Toom 2010, 2012). The larger the number of profiles that a particular DNA profile is compared with, the higher the chances that matches are not 'real' (i.e. stemming from the same person) but adventitious.\n\n## **References**\n\nAmorim, A. (2012) 'Opening the DNA black box: demythologizing forensic genetics', _New Genetics & Society_, 31(3): 259\u2013270.\n\nAngell, M. (1996) _Science on Trial: The Clash of Medical Evidence and the Law in the Breast Implant Case_ , New York: Norton.\n\nAronson, J.D. (2007) _Genetic Witness: Science, Law, and Controversy in the Making of DNA Profiling_ , New Brunswick, NJ: Rutgers University Press.\n\nAronson, J.D. (2010) 'On trial! Governing forensic DNA technologies in the USA', in R. Hindmarsh and B. Prainsack (eds), _Genetic Suspects: Global Governance ofForensic DNA Profiling and Databasing_, Cambridge, UK: Cambridge University Press, pp. 240\u2013261.\n\nBal, R. (2005) 'How to kill with a ballpoint. Credibility in Dutch forensic science', _Science, Technology, and Human Values_ , 30(1): 52\u201375.\n\nBloor, M. (2000) 'The South Wales miners federation, miners' lung and the instrumental use of expertise, 1900\u20131950', _Social Studies of Science_ , 30(1): 125\u2013140.\n\nCole, S.A. (2001) _Suspect Identities: A History of Fingerprinting and Criminal Identification_ , Cambridge, MA: Harvard University Press.\n\nCollins, H.M. and Evans, R. (2002) 'The third wave of science studies: studies of expertise and experience', _Social Studies of Science_ , 32(2): 235\u2013296.\n\nCollins, H.M. and Evans, R. (2007) _Rethinking Expertise_ , Chicago, IL: University of Chicago Press.\n\nDuster, T. (2004) 'Selective arrests, an ever-expanding DNA forensic database, and the specter of an early-twenty-first-century equivalent of phrenology', in D. Lazer (ed.), _The Technology of Justice: DNA and the Criminal Justice System_ , Cambridge, MA: MIT Press, pp. 315\u2013334.\n\nDuster, T. (2006a) 'Explaining differential trust of DNA forensic technology: grounded assessment or inexplicable paranoia?', _Journal of Law, Medicine & Ethics_, 34(2): 293\u2013300.\n\nDuster, T. (2006b) 'The molecular reinscription of race: unanticipated issues in biotechnology and forensic science', _Patterns of Prejudice_ , 40(4\u20135): 427\u2013441.\n\nDuster, T. (2008) 'DNA dragnets and race: larger social context, history and future', _Gene Watch_ , 21(3\u20134): 3\u20135.\n\nEdmond, G. (2004) _Expertise in Regulation and Law_ , Aldershot: Ashgate Publishing.\n\nEdmond, G. (2011) 'The building blocks of forensic science and law: recent work on DNA profiling (and photo comparison)', _Social Studies of Science_ , 41(1): 127\u2013152.\n\nFarmer, L. (2010) 'Time and space in criminal law', _New Criminal Law Review_ , 12(2): 333\u2013356.\n\nFaulkner, A., Lange, B. and Lawless, C. (2012) Special issue of Journal of Law and Society on ' _Material worlds: intersections of law, science, technology and society_ ', 3(1).\n\nFlear, ML. and Vakulenko, A. (2010) 'A human rights perspective on citizen participation in the EU's governance of new technologies', _Human Rights Law Review_ , 10(4): 661\u2013688.\n\nGad, C. and Lauritsen, P. (2009) 'Situated surveillance: an ethnographic study of fisheries inspection in Denmark', _Surveillance & Society_, 7(1): 49\u201357.\n\nGillespie, T. (2006) 'Engineering a principle: \"end-to-end\" in the design of the internet', _Social Studies of Science_ , 36(3): 427\u2013457.\n\nGolan, T. (2004) 'The emergence of the silent witness: the legal and medical reception of X-rays in the USA', _Social Studies of Science_ , 34(4): 469\u2013499.\n\nHaraway, D.J. (1991) 'Situated knowledges: the science question in feminism and the privilege of partial perspective', in D.J. Haraway (ed.), _Simians, Cyborgs and Women: The Reinvention of Nature_ , London: Free Association Books, pp. 183\u2013202.\n\nHeinemann, T., Lemke, T. and Prainsack, B. (2012) 'Rsky profiles: societal dimensions of forensic uses of DNA profiling technologies', _New Genetics and Society_ , 31(3): 249\u2013258.\n\nHindmarsh, R. and Prainsack, B. (2010) (eds) _Genetic Suspects: Global Governance of Forensic DNA Profiling and Databasing_ , Cambridge: Cambridge University Press.\n\nHowarth, D. (2000) _Discourse_ , Buckingham: Open University Press.\n\nJasanoff, S. (1995) _Science at the Bar: Law, Science, and Technology in America_ , Cambridge, MA: Harvard University Press.\n\nJasanoff, S. (2010) 'Foreword', in R. Hindmarsh and B. Prainsack (eds), _Genetic Suspects: Global Governance of Forensic DNA Profiling and Databasing_ , Cambridge: Cambridge University Press, pp. xix-xxiv.\n\nJohnson, P. and Williams, R. (2004) 'Post-conviction testing: the UK's first \"exoneration\" case? _Science and Justice_ , 44(2): 77\u201382.\n\nKrieger, S.A. (2011) 'Why our justice system convicts innocent people and the challenges faced by innocence projects trying to exonerate them', _New Criminal Law Review_ , 14(3): 333\u2013402.\n\nKrimsky, S. and Simoncelli, T. (2011) _Genetic Justice: DNA Data Banks, Criminal Investigations, and Civil Liberties_ , New York: Columbia University Press.\n\nKruse, C. (2012) 'Legal storytelling in pre-trial investigations: arguing for a wider perspective on forensic evidence', _New Genetics and Society_ , 31(3): 299\u2013309.\n\nLeo, RA. (2005) 'Rethinking the study of miscarriages of justice: developing a criminology of wrongful conviction', _Journal of Contemporary Criminal Justice_ , 21(3): 201\u2013223.\n\nLynch, M. (2001) 'Ethnomethodology and the logic of practice', in T.R. Schatzki, K.K. Cetina and E. Von Savigny (eds), _The Practice Turn in Contemporary Theory_ , London: Routledge, pp. 131\u2013148.\n\nLynch, M. (2011) 'Harold Garfinkel (29 October 1917\u201321 April 2011): a remembrance and reminder', _Social Studies of Science_ , 41(6): 927\u2013942.\n\nLynch, M. and Cole, S. (2005) 'Science and technology studies on trial: dilemmas of expertise', _Social Studies of Science_ , 35(2): 269\u2013311.\n\nLynch, M., Cole, S.A., McNally, R and Jordan, K. (2008) _Truth Machine: The Contentious History of DNA Fingerprinting_ , Chicago, IL: University of Chicago Press.\n\nMachado, H. and Prainsack, B. (2012) _Tracing Technologies: Prisoners' Views in the Era of CSI_ , Aldershot: Ashgate Publishing.\n\nM'Charek, A. (2005) _The Human Genome Diversity Project: An Ethnography of Scientific Practice_ , Cambridge: Cambridge University Press.\n\nM'Charek, A. (2008a) 'Contrasts and comparisons: three practices of forensic investigation', _Comparative Sociology_ , 7(3): 387\u2013412.\n\nM'Charek, A. (2008b) 'Silent witness, articulate collective: DNA evidence and the inference of visible traits', _Bioethics_ , 22(9): 519\u2013528.\n\nM'Charek, A., Toom, V. and Prainsack, B. (2012) 'Bracketing off populations does not advance ethical reflection on EVCs: a reply to Kayser and Schneider', _Forensic Science International: Genetics_ , 6(1): e16\u2013e17.\n\nMercer, D. (2002) 'Scientific method discourses in the construction of \"EMF science\": interests, resources and rhetoric in submissions to a public inquiry', _Social Studies of Science_ , 32(2): 205\u2013233.\n\nMills, S. (2012) 'Weak DNA evidence could undermine justice, experts say', by _Chicago Tribune_ reporter. Online. Available: \u00adh\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ada\u00adr\u00adt\u00adi\u00adc\u00adl\u00ade\u00ads\u00ad.\u00adc\u00adh\u00adi\u00adc\u00ada\u00adg\u00ado\u00adt\u00adr\u00adi\u00adb\u00adu\u00adn\u00ade\u00ad.c\u00ado\u00adm\u00ad\/\u00ad2\u00ad0\u00ad1\u00ad2\u00ad-\u00ad0\u00ad7\u00ad-\u00ad0\u00ad5\u00ad\/\u00adn\u00ade\u00adw\u00ads\u00ad\/\u00adc\u00adt\u00ad-\u00adm\u00ade\u00adt\u00ad-\u00add\u00adn\u00ada\u00ad-\u00adq\u00adu\u00ade\u00ads\u00adt\u00adi\u00ado\u00adn\u00ads\u00ad-\u00ad2\u00ad0\u00ad1\u00ad2\u00ad0\u00ad7\u00ad0\u00ad5\u00ad_\u00ad1\u00ad_f\u00ado\u00adr\u00ade\u00adn\u00ads\u00adi\u00adc\u00ad-\u00add\u00adn\u00ada\u00ad-\u00ada\u00adn\u00ada\u00adl\u00ady\u00ads\u00adi\u00ads\u00ad-\u00add\u00adn\u00ada\u00ad-\u00adp\u00adr\u00ado\u00adf\u00adi\u00adl\u00ade\u00ad-\u00add\u00adn\u00ada\u00ad-\u00ads\u00adc\u00adi\u00ade\u00adn\u00adt\u00adi\u00ads\u00adt\u00ads\u00ad (accessed 26 August2012).\n\nMnookin, J.L., Cole, S.A., Dror, I.E., Fisher, B.A., Houck, M.M., Inman, K., Kaye, D.H., Koehler, J.J., Langenburg, G., Rsinger, D.M., Rudin, N., Siegel, J. and Stoney, D.A. (2011) 'The need for a research culture in the forensic sciences', _UCLA Law Review_ , 58: 725\u2013779.\n\nMouffe, C. (2005) _On the Political_ , Abingdon: Routledge.\n\nMurphy, E. (2010) 'Relative doubt: familial searches of DNA databases', _Michigan Law Review_ , 109(3): 291\u2013348.\n\nNational Academy of Sciences (2009) ' _Strengthening forensic science in the United States: a path forward_. Online. National Research Council', Online. Available: \u00adw\u00adw\u00adw\u00ad.n\u00adc\u00adj\u00adr\u00ads\u00ad.\u00adg\u00ado\u00adv\u00ad\/\u00adp\u00add\u00adf\u00adf\u00adi\u00adl\u00ade\u00ads\u00adl\u00ad\/\u00adn\u00adi\u00adj\u00ad\/\u00adg\u00adr\u00ada\u00adn\u00adt\u00ads\u00ad\/\u00ad2\u00ad2\u00ad8\u00ad0\u00ad9\u00ad1\u00ad.\u00adp\u00add\u00adf\u00ad (accessed 26 August 2012).\n\nObasogie, O.K. and Duster, T. (2011) 'All that glitters isn't gold', _Hastings Center Report_ , 41(5): 15\u201317.\n\nPollner, M. (1991) 'Left of ethnomethodology: the rise and decline of radical reflexivity, _American Sociological Review_ , 56(3): 370\u2013380.\n\nPottage, A. (2011) 'Law machines: scale models, forensic materiality and the making of modern patent law', _Social Studies of Science_ , 41(5): 621\u2013643.\n\nPrainsack, B. (2010) 'Key issues in DNA profiling and databasing: Implications for governance', in R Hindmarsh and B. Prainsack (eds) _Genetic Suspects: Global Governance of Forensic DNA Profiling and Databasing_ , Cambridge: Cambridge University Press, pp. 15\u201339.\n\nPrainsack, B. and Kitzberger, M. (2009) 'DNA behind bars: other ways of knowing forensic DNA technologies', _Social Studies of Science_ , 39(1): 51\u201379.\n\nPrainsack, B. and Toom, V. (2010) 'The Pr\u00fcm regime: situated dis\/empowerment in transnational DNA profile exchange', _British Journal of Criminology_ , 50(6): 1117\u20131135.\n\nPrainsack, B. and Toom, V. (2012) 'Performing the union: the Pr\u00fcm Decision and the European dream', _Studies in the History and Philosophy of Science_ , (in press).\n\nReardon, J. (2005) _Race to the Finish: Identity and Governance in an Age of Genomics_ , Princeton, NJ: Princeton University Press.\n\nRisinger, D.M. (2007) 'Innocents convicted: an empirically justified factual wrongful conviction rate', _Journal of Criminal Law and Criminology_ , 97(3): 761\u2013806.\n\nSmith, R. and Wynne, B. (1989) _Expert Evidence: Interpreting Science in the Law_ , London: Routledge.\n\nTaslitz, A. (2011) 'The rule of criminal law: why courts and legislatures ignore Richard Delgado's rotten social background', _Alabama Civil Rights & Civil Liberties Review_, 2: 79\u2013130.\n\nToom, V. (2012) 'Bodies of science and law: forensic DNA profiling, biological bodies, and biopower', _Journal of Law and Society_ , 39(1): 150\u2013166.\n\nValverde, M. (2009) 'Jurisdiction and scale: legal \"technicalities\" as resources for theory', _Social and Legal Studies_ , 18(2): 139\u2013157.\n\nWest, E.M. (2010) ' _Court findings of ineffective assistance of counsel claims in postconviction appeals among the first 255 DNA exoneration cases_ '. Online. Available: \u00adw\u00adw\u00adw\u00ad.i\u00adn\u00adn\u00ado\u00adc\u00ade\u00adn\u00adc\u00ade\u00adp\u00adr\u00ado\u00adj\u00ade\u00adc\u00adt\u00ad.o\u00adr\u00adg\u00ad\/\u00add\u00ado\u00adc\u00ads\u00ad\/\u00adI\u00adn\u00adn\u00ado\u00adc\u00ade\u00adn\u00adc\u00ade\u00ad_\u00adP\u00adr\u00ado\u00adj\u00ade\u00adc\u00adt\u00ad_\u00adI\u00adA\u00adC\u00ad_\u00adR\u00ade\u00adp\u00ado\u00adr\u00adt\u00ad.\u00adp\u00add\u00adf\u00ad (accessed 20 August 2012).\n\nWilliams, R (2010) 'DNA databases and the forensic imaginary', in R Hindmarsh and B. Prainsack (eds), _Genetic Suspects: Global Governance of Forensic DNA Profiling and Databasing_ , Cambridge: Cambridge University Press, pp. 131\u2013152.\n\nWilliams, R and Johnson, P. (2008) _Genetic Policing: The Use of DNA in Criminal Investigations_ , Cullompton: Willan Publishing.\nChapter 5\n\n* * *\n\n# **Making the colposcope 'forensic'**\n\nThe medico-legal management of a controversial visualisation device\n\n_Gethin Rees_\n\n* * *\n\nThe presence of injury upon the body of a rape assault survivor has been demonstrated as a strong predictor of successful criminal prosecution (Du Mont and White 2007). Forensic practitioners, either doctors (known as Forensic Medical Examiners (FMEs)) or nurses (Forensic Nurse Examiners (FNEs)), spend the majority of the forensic intervention that follows the reporting of a rape assault observing the body and recording anything that they determine constitutes abnormal phenomena (Rees 2011). This observation and recording process is separated into genital and non-genital examinations, and while injuries are often not difficult to observe macro-scopically upon non-genital areas, the same cannot be said for the ano-genital region. Consequently, since the 1990s, a gynaecological technology known as the colposcope has been recommended for use within forensic medical examinations (Slaughter and Brown 1992). The colposcope, initially designed to investigate the cervix and tissues of the vagina and the vulva following the return of an irregular pap smear, is a type of microscope that can magnify the genital (and anal) areas up to 30 times (Rogers 2004). It comes complete with a light for illumination and a camera for recording the examination. According to those promoting the use of the colposcope within forensic medical examinations, magnification and illumination is likely to increase the probability of observing signs of genital injury; meanwhile, the camera enables improved contemporaneous recording of the examination, producing better evidence for criminal justice as well as medico-legal research (Brennan 2006). However, not all forensic practitioners are convinced that this device is useful or appropriate, and some challenge its incorporation on both victim-centric and scientific grounds.\n\nSuch concerns are indicative of the difficulties that arise when new technological interventions are introduced into the legal milieu. While the practitioners of both science and the law promote reliance upon rationality and objectivity in their knowledge production (Smith and Wynne 1989) and are keen to employ new devices in order to ascertain that knowledge, as other chapters in this volume have testified, implementing new technologies is often complex and controversial. While (medico-) legal practitioners may endeavour to utilise technologies that are the state of the art as part of a thorough investigation, the use of the technologies are also required to meet the ethical, legal and moral standards required by law. The colposcope is such an example of a device where it was considered to offer (what I will go on to label) evidential benefit at the expense of victim care and as a result an ongoing controversy in forensic medicine ensued regarding whether the colposcope should become established best practice.\n\nControversies have been a fruitful area of study in science and technology studies (STS) as they demonstrate science in the making (Mercer 1996). Harry Collins (1975, 1981) employed the metaphor of the ship in the bottle to explain the ways that controversies can help the analyst of scientific knowledge and practice. Once the ship (a scientific claim or practice) is in the bottle (accepted as 'knowledge' or 'best practice'), then for all intents and purposes it appears to have always been there, or 'ready assembled' (Collins 1981: 45). In other words, it becomes seen as self-explanatory that a particular practice is the most efficacious method. However, close observation of the controversy \u2013 either historically, when the analyst knows the outcome of the debate (see for example Barnes and Shapin 1979), or ethnographically when the controversy is on-going (Collins 1981) \u2013 draws attention to the complex negotiations and social influences required to generate consensus around the knowledge claim or practice (for instance, not publishing papers in high profile journals, or not inviting certain scientists to important conferences; Barnes _et al._ 1996, Kuhn 1996).\n\nIn many ways the colposcope controversy mirrors a larger dialectic, often considered to be at the heart of forensic medicine: whether the forensic medical practitioner is an objective evidence-gatherer or the provider of therapeutic care to a victim. This 'dual role' inherent in the original title of the forensic practitioner \u2013 i.e., the 'Police Doctor' (Savage _et al._ 1997) \u2013 necessitates that practitioners choose to emphasise either the evidential or the therapeutic aspect of their work when performing a forensic intervention. As I will show in the first half of this chapter, both those who promote and those who oppose the routine implementation of the technology are drawing upon a particular rhetoric surrounding the forensic medical role (i.e. evidential or therapeutic) to make their case. However, as with nearly all the scientific controversies previously studied, there has been no decisive 'crucial experiment' or announcement marking an end to the debate; rather, research interests have moved on, leaving forensic practitioners to make local decisions regarding the best ways to employ the technology. Accordingly, the device is routinely employed in forensic work, but in circumscribed ways: practitioners aim to achieve the most benefit from the colposcope while simultaneously attempting to limit the harm caused by the device to the survivor.\n\nIn the second half of the chapter I will draw upon interview data from two studies of forensic medical practitioners (one with FMEs, the other with FNEs) to argue that when making their decisions about the best ways to use the colposcope, practitioners are not acting in a 'dual role' but rather in a specially 'forensic' mode (Mulla 2011, Rees forthcoming). Such a modality of practice does not involve the fluctuation between evidence gathering and therapy, but involves a unique style of praxis that is simultaneously alert to and incorporative of both aspects of the work traditionally characteristic of FM\/NEs. Further, I will argue that the introduction of new technologies into the legal milieu is not without its problems; in this case the requirement for both therapeutic and evidential practices necessitated innovation in the ways that the colposcope was employed, most notably the ways that the images produced by the device were stored and utilised. Such local negotiations and workarounds are common for the successful application of scientific and medical technologies in the legal milieu.\n\nThe data presented in this chapter derive from two studies: one investigating the role and work of FMEs in Scotland, conducted between 2006 and 2008, and the second investigating FNEs in England in 2010. Interviews were performed between 2006\u201310 with 13 FMEs from four constabularies in Scotland and five FNEs from five hospitals in England (the total population of FNEs employed at the time). Alongside the interview data, both studies also involved analysis of the textbooks, journal articles and practice guidelines that FM\/NEs use in training, research and work. Drawing upon these data for the remainder of this chapter, I will set out the colposcopy controversy before focusing on the way that it has played out and been resolved in local contexts.\n\n## **The colposcopy controversy**\n\nThe controversy centred around two studies by Laura Slaughter and colleagues in the 1990s (Slaughter and Brown 1992, Slaughter _et al._ 1997), which advocated the use of the colposcope due to the evidential benefit provided by its improved illumination and magnification. However, many practitioners were unconvinced by Slaughter _et al._ 's recommendations. I will outline the details of the controversy, beginning with those who promoted the device for its _evidential benefit_.\n\n### **_Evidential benefit_**\n\nGynaecologists were the first to identify the potential benefits of the colposcope for forensic medical investigations (Teixeira 1980). Due to their experience working with the device in order to perform clinical investigations of the cervix, gynaecologists were aware of the improved magnification and postulated that this would enable forensic practitioners to examine genital regions with similarly greater precision. This early recommendation is often elided in practitioner histories of the colposcope (e.g., Sommers _et al._ 2005), and Teixeira's (1981) colposcopic study of 500 victims of sexual offences is normally cited as the first medico-legal study to identify the efficacy of the technology. Teixeira reported that his method of observing the hymen via the colposcope resulted in greater clarity than macroscopic observation alone, arguing the colposcope: 'to be irreplaceable in the examination of the victims of sexual offences and recommends its adoption by all medicolegal services' (Teixeira 1981: 214).\n\nAlthough Teixeira's conclusions have since been labelled 'dubious' and 'confusing' (Lincoln 2001: 212), his findings nevertheless demonstrated that the colposcope could provide improved visibility for genital examinations. This was deemed to have implications for criminal justice, such as the identification of injuries unseen by macroscopic observation and the ability to identify observed phenomena as naturally (or otherwise) occurring.\n\nTeixeira's findings appear to have been ignored and forgotten about during the 1980s, and the question of the incorporation of the colposcope was not broached again until the publication of Slaughter and Brown's (1992) study. Concerned with the reported low observation rate of genital injuries (only in 10 per cent to 30 per cent of cases), the authors compared the colposcope with a nuclear stain, 'Toludine Blue', customarily placed on complainants' genitalia in order to highlight signs of injury (albeit leaving the complainant with a stain on her genitalia that came to be known as the 'the mark of the rape victim' (Lincoln 2001: 215)). Slaughter and Brown colposcopically examined 131 women who had reported non-consensual penile-vaginal intercourse taking place within the past 48 hours, observing genital injury in 87 per cent of cases. They concluded that the colposcope was not only superior in observing signs of injury, but was also, by attaching a camera, of significant value to criminal investigations and the development of a medico-legal evidence base for genital injury:\n\n> The clarity afforded by colposcopic magnification will also allow us to further characterize genital changes associated with rape. The findings in this study support the opinion that rape victims often sustain mounting injuries. The ability to detect and document these injuries through colposcopic photography provides valuable medical and legal information.\n> \n> (Slaughter and Brown 1992: 86)\n\nThis quotation exemplifies much of the argument for the colposcope (although it of course developed as the controversy progressed). Slaughter and Brown emphasised the different types of evidence that the device could provide: not only would it result in magnified and clearer images of the genitalia that would enable experts to make more precise observations, but it also enabled recording of those images (as photographs, or later moving visual images on film and digital recordings), which could be used by other actors in the criminal justice process.\n\nMoreover, those same images could also be collated and used as scientific evidence in the development of forensic medical knowledge about the effects of forced sexual intercourse on female genitalia, which in turn would enable more probative expert evidence. Of course, precautions would have to be taken in order to maintain the anonymity and confidentiality of such images, and supporters of the colposcope drew attention to the logistical concerns that needed to be overcome in order to enable routine use of the device (e.g., Rogers 1996). However, these difficulties were dwarfed by the 'benefits that will accrue if there are fewer or shorter trials because of the improved quality of the evidence' (Rogers 1996: 121). For medico-legal practitioners such as Slaughter, Brown and Rogers, the improved illumination, magnification and recording of the ano-genital region would certainly improve the quality of the evidence that forensic medical experts could provide in rape cases.\n\n### **_Victims and the absence of injuries_**\n\nA year after the publication of Slaughter and Brown's article in the _American Journal of Obstetrics and Gynaecology_ , a letter was published in the same journal stating that colposcopy as a routine part of the forensic medical intervention was 'a regressive step in the management of rape victims whose histories may not be believed unless demonstrable colposcopic injuries are seen' (Patel _et al._ 1993: 1334). Patel and colleagues were voicing a concern, growing amongst the forensic medical community at the time, that it was common for victims of rape to report without any signs of injury.\n\nThis position was made explicit a year later with the publication 'It's normal to be normal' (Adams _et al._ 1994). Based on a review of child and adolescent sexual assault cases where the victims had been colposcopically examined shortly after reporting, the authors found that very few of the cases had observable signs of ano-genital injury. While Adams _et al._ 's study was paediatric, the findings \u2013 and, more importantly, the slogan \u2013 were quickly mobilised by forensic practitioners primarily concerned with adult examinations (Brennan 2006).\n\nPatel and colleagues were clearly sympathetic (if not pre-emptive) to the idea of 'it's normal to be normal' and feared that the routine implementation of the colposcope would only serve to reinforce popular misconceptions that all victims present with signs of injury, and therefore fail to benefit those who do not have colposcopically observable ano-genital injuries. In fact they went further, voicing concerns that the device could become the arbiter of truth in criminal cases, with decisions over the validity of allegations being decided upon the presence of injuries alone. Lincoln (2001) made a similar argument in relation to resource management. She was concerned with the prohibitive costs of the device and argued that if the colposcope did achieve the 'gold standard' many victims would be disadvantaged as some medico-legal teams would be unable to afford it.\n\nPatel and colleagues developed their critique of colposcopy further, arguing that its routine application would constitute a second assault upon the survivor by requiring them to undergo an invasive genital examination, potentially against their wishes. They argued that the potential for the survivor to develop (what was then known as) 'Rape Trauma Syndrome' as a result of the anxiety caused by a colposcopic procedure should stay forensic practitioners' hands from routinely implementing the technology. Patel _et al._ thus chose to criticise the routine employment of the device on victim-centric\/therapeutic grounds: while the argument was not framed around the provision of therapy, they postulated that routine use of the device could result in further harm, first to survivors' cases (with the potential for signs of colposcopically recorded genital injury to become the arbiter of rape cases), and second to the survivors themselves (due to the possibility of 'an increase in psychological trauma' (Patel _et al._ 1993: 1334)). In opposition to Slaughter and colleagues' promotion of the device based upon the evidential benefit, Patel _et al._ advanced a counterargument grounded in the therapeutic (or medical) aspects of medico-legal work.\n\n### **_Injury interpretation_**\n\nAlongside therapeutic concerns, still more professionals chose to confront Slaughter and Brown in evidentiary terms. In particular, Bowyer and Dalton (1997) were concerned by the large quantity of injuries found in Slaughter and Brown's (1992) study and questioned whether they were all the result of non-consensual intercourse. The authors cited Norvell _et al._ 's (1984) colposcopic study of women who had engaged in consensual sexual intercourse six hours before examination and displayed signs of genital 'microtrauma'. Bowyer and Dalton postulated that Slaughter and Brown had included injuries that were not necessarily the result of non-consensual intercourse including 'tiny lacerations, abrasions, bruises and swelling' (Bowyer and Dalton 1997: 619), and challenged the extent to which one could necessarily distinguish between 'microtrauma' caused by consensual intercourse and injuries that were the result of non-consensual intercourse. In effect, Bowyer and Dalton challenged Slaughter and Brown's claim that the device would provide more probative evidence of non-consensual intercourse.\n\nSlaughter _et al._ (1997) attempted to address the question of the relationship between injuries and consent in their follow-up study, 'Patterns of genital injury in female sexual assault victims'. The objective was to ascertain whether it would be possible to identify, colposcopically, a particular injury or collection of injuries specific to non-consensual sexual intercourse. They reviewed the colposcopic recordings of 311 victims and compared them with findings in 75 women who were examined 24 hours after (what was judged to have been) consensual sexual intercourse. The study concluded that genital injury was only identifiable in a small number (11 per cent) of the consensual intercourse sample, while injury was more commonplace amongst the victim sample (89 per cent).\n\nThese results led Slaughter and colleagues to assert: 'Although coital injury seems to be associated with insertion of the penis, its prevalence is significantly associated with a history of non-consensual intercourse' (Slaughter _et al._ 1997: 615). They were nevertheless cautious in drawing conclusions about consent from their findings, declaring that 'Further investigation is needed to determine whether there is a finding or group of findings that can distinguish nonconsensual and consensual activity' (Slaughter et _al_. 1997: 615). However, they repeated their recommendation that the colposcope should be routinely used as part of 'best practice' for generating forensic evidence in sexual assault cases.\n\n### **_The new quest for the rape injury_**\n\nThe last paper to expressly address the controversy was Brennan's (2006) review of the debate. Brennan reported having used the colposcope since 1999, and strongly advocated for its routine use in order to develop an evidence base of 'normal' genitalia from which the 'abnormal' (i.e., assaulted) could then be identified. Such evidence was necessary to collect, since:\n\n> At this point in time, research in adult sexual assault is probably at the stage where child sexual abuse was two decades ago: individual clinicians working in isolation (without accountability or peer review), who were compelled to interpret what they found... with variable standards in forensic qualifications and clinical experience, and without the benefits of normative standards for the appearance of the hymen, peri-hymenal tissues, or perianal tissues.\n> \n> (Brennan 2006: 199)\n\nPerforming colposcopic examinations with the accompaniment of visual recording would, she argued, develop a collection of images that could be used to produce an evidence base; this would help the development of medico-legal knowledge around sexual assault and could also be used in the training of new practitioners. Enabling colleagues to watch contemporaneous recordings of each other's work would provide a basis for the judgement of an individual's practices as well as the quality of their interpretations.\n\nBrennan was aware of the problems with the colposcope, and accepted that the standardised performance of the colposcopic exam might not be in the best interests of the survivor; she emphasised that healthcare requirements should be identified and addressed first. Moreover, she empathised with the survivors, noting that they might not wish to have invasive images of their genitalia produced (whether for medical, scientific or legal reasons), and strongly encouraged medico-legal teams to take great care with any images produced:\n\n> Since photographs of the genital area, taken for valid medical purposes, can be similar to pornographic images, the question of ownership and control of the photographs is of concern to victims, but more so to those responsible for taking them. Once consent is granted and images come into existence, these can end up in hospital records, a police station, a journal article, or a court room, with no clear boundaries as to who might get to view them.\n> \n> (Brennan 2006: 196)\n\nNevertheless, like Rogers before her, Brennan believed the device could provide substantial evidential benefit and that the concerns she outlined could be resolved: it was acceptable to conduct a colposcopic examination as long as survivors gave their consent.\n\nIt would be a misrepresentation to characterise the debate around the role of the colposcope in forensic examinations as _solely_ about evidential benefit versus medical clinical utility; however, in making their arguments either for or against the device, contributors to journal debates about its use relied upon one or other aspect of the dual role of forensic medicine. For instance, Slaughter and colleagues continually emphasised its evidential aspect, whereas Patel _et al._ focused on the dangers to victims, and even those who demonstrated awareness of the other argument (e.g. Brennan's and Rogers' worries about the dissemination of colposcopic images) eventually subsumed these concerns to the benefit of the greater good, i.e. the expected probative evidence.\n\nAs with so many scientific controversies, this debate appears to have come to an end as the community has moved on to other research questions, isolating those who continue to debate as their opinions are not discussed in the literature. The question of whether or not to use the colposcope is no longer addressed in the forensic medical literature; today, it is common to see large-scale studies and reviews of colposcopic examinations following consensual and\/or non-consensual intercourse in an attempt to discover the still elusive 'rape injury' (e.g. Anderson _et al._ 2009, Astrup _et al._ 2012, Hilden _et al._ 2005). However, concerns still remain, and as there has not been a decisive statement on whether or not to use the device routinely, local adoption has been an _ad hoc_ affair, with forensic practitioners finding their own ways to negotiate the difficulties caused by the technology. In the remainder of this chapter, I will outline some of the workarounds practitioners have employed in order to manage the evidential benefits while at the same time limiting harm.\n\n## **Managing the colposcope**\n\nThe colposcope is routinely used in much of England and Scotland, and, in attempting to localise the device, practitioners are at pains to reap its benefits while limiting further harm to the survivor. Drawing on my interview data, I demonstrate that there are two ways in which practitioners can make the colposcope a successful 'forensic'\/therapeutic-evidential tool (Mulla 2011, Rees forthcoming): first, choosing when to record, and second, limiting access to the images.\n\n### **_Recording decisions_**\n\nBoth FMEs and FNEs were keenly aware of the potential for the colposcope to enable the forensic medical practitioner to perform the genital examination more clearly:\n\n> The facility with the colposcope that we have got there... I think it's a fantastic machine and the clarity and the pictures you get are fantastic you know, and you will really see stuff that you probably wouldn't have seen before.\n> \n> (FME C)\n\n> I always examine with the colposcope because the examination with the colposcope picks up, depending on how many studies you are looking at, about 10 to 30 per cent of the injuries that are not visible to the naked eye, so it is far better to have that. If I see an injury then I would record it, if not I would just be using the colposcope to just look for injuries as a light source and magnification.\n> \n> (FME A)\n\nLike those arguing for the evidential benefit before them, FMEs remarked upon the way the technology enabled them to observe the genital area more clearly, enabling them to see injuries that they would not have been able to observe macroscopically and\/or confirm the existence of abnormal phenomena:\n\n> Genital injuries what I would do is... if it was external I would first put the colposcope on because you can do that without recording, and it could well be that they've shaved their pubic hair, it could be a shaving rash, which I would look more clearly at because the colposcope would magnify, and again I would say to them 'well I'm just going to take a look at this' and I would say 'well actually when you look at that it is an in-growing hair or whatever'. \n> \n> (FNE B)\n\nWhile both FME A and FNE B emphasise the benefit of the device in enabling them to view and clarify, they also draw attention to the fact that they are doing so while the colposcope's camera is switched off; i.e. they are using the device for illumination and magnification, not to record. Maintaining discretion over whether an examination should be recorded enables practitioners to limit the potential for further harm to come to the survivor; as highlighted by Rogers (1996) and Brennan (2006), the security of the colposcopic images once recorded is paramount and can place resource pressures on a medico-legal team to ensure that they have the storage capacity to maintain the security and confidentiality of the recordings. In the words of one FME:\n\n> May I say though the trend is... to use the colposcope more routinely, more regularly. We may not take photographs with it, but at least we use it for as I say, value of illumination and magnification. So it is being used more and more frequently... but no recording carried out, because that gives us the headache of having to store the video and the rest of it.\n> \n> (FME B)\n\nConsequently, practitioners make a decision on a case-by-case basis regarding whether to switch on the recording aspect of the device: some wait until an injury is observed that requires recording (FME A above for instance), while others have their own classificatory framework to identify which victims it would be appropriate to record. For instance:\n\n> Use the colposcope when you feel you are going to get more information; so for example, if you are looking at somebody who has had six children and had sex last week, the chance of finding any injury are limited, so you wouldn't bother with the colposcope, but if you are looking at somebody who's 18, never had sex before, then you want to record the details in minute, minute detail.\n> \n> (FME B)\n\nWhichever way practitioners decide if and when to switch on recording, an explicit reason for maintaining this discretion is to limit unnecessary further harm to the survivor (such as that which can result from images not being maintained securely).\n\nIn the same vein, before FM\/NEs can carry out the examination, they must first obtain consent to recording it. While no guidance has been provided on whether the device should be used routinely, the Faculty of Forensic and Legal Medicine has produced advice on the collection, storage and use of any images produced via colposcopic recording. The practitioner is required to inform the survivor that the images are likely to be viewed by other medical practitioners, and, if in England and Wales (although not Scotland for reasons discussed below), that there is the exceptional chance that the images will also be shown in the courtroom. On receiving this information, the survivor is deemed to have been informed, and can choose whether or not to allow recording.\n\nPractices such as maintaining discretion over when to record and asking for the survivor's consent before recording are not significantly victim-oriented. For instance, they do not address Patel _et al._ 's fears that routine colposcopic examination could result in the greater possibility of psychological harm to the survivor, something that troubled one of the FMEs in the study:\n\n> Yes we usually use a colposcope to check for injuries, but genital injuries are usually rare... but it's a usual technique to use the colposcope... so when you go straight according to the guidelines you have to use the colposcope, so that's a second rape now.\n> \n> (FME D)\n\nYet, at the same time, consenting for recording does not have an overtly evidential basis; for instance, obtaining consent potentially reduces the number of colposcopic images, which might in turn diminish the quantity of evidence produced during the examination. As a result, the measures introduced to make the colposcope useful do not fit with either of the two traditional discourses about forensic medicine; they constitute a compromise, attempting to employ the colposcope for its benefits (illumination and magnification) but at the same time ensuring that they only have as many images as they can securely manage in order to maintain a duty of care to the survivor. Such compromises between the evidential and the therapeutic are essential in forensic work, as further demonstrated by another strategy used to make the colposcope beneficial but not harmful: controlling who has access to the images once they have been produced.\n\n### **_Access to images_**\n\nBrennan (2006) drew attention to the potential for the colposcope to assist with training and peer review. For the training of FNEs in particular, the colposcope has been used as a means to record the trainee's examination and review their practice in conversation with their trainer. During the early phases of training, the more experienced practitioner is physically present while the trainee is performing the forensic examination, guiding and advising the process. As part of this early phase, the trainer pays attention to the way the trainee uses the colposcope as well as the details pertaining to the other parts of the examination. This includes:\n\n> not only how you physically do the cases, how you record it as well, if the light's not right, if your hands are in the way or if the focus is not right then it's basically a worthless recording, you might as well have not done it. So it's the technique with the colposcope.\n> \n> (FME H)\n\nExtensive initial training ensures that the experienced practitioner will not be required at later examinations performed by the trainee; they can instead rely on the colposcopic recording (or 'DVD') as the focus of training conversations and feedback.\n\nApart from cases where the survivor did not provide consent, all examinations are recorded during training in order to provide feedback for development. Of course, once a practitioner is deemed competent it is down to their own discretion to determine when it is appropriate to record. Nevertheless, the cases chosen to be recorded by the practitioners are important for their own development, with medico-legal teams holding regular peer review meetings to discuss interesting cases, comment on each other's skills and share ideas of best practice. As one FME reflected: 'We have a monthly meeting for peer review... and if anyone's got an interesting DVD they bring that to the meeting... so there is an on-going [peer review]' (FME H).\n\nWhile the technology is of notable benefit for the training and development of medico-legal practitioners, it does come at the cost of the sharing of images which the survivor may not want distributed to the practitioner's colleagues. As discussed above, the Faculty of Forensic and Legal Medicine have provided clear guidance that the practitioner should explain the various ways that the documents may be used, including within the context of the courtroom; however, FMEs in Scotland have found another way to limit the number of people who will see the recordings.\n\nScottish FMEs are adamantly opposed to the practice of showing colposcopic recordings in the courtroom as has been done in other jurisdictions (Faculty of Forensic and Legal Medicine 2010, _State of West Virginia v. William. M._ 2010) and have an agreement with the Scottish prosecuting body, the Crown Office and Procurator Fiscal Service, not to be asked to show the images:\n\n> I find some women feel that this [colposcopic recording] is going to be brandished in court, _NO WAY_ , we only allow another medical expert, the defence medical expert to see, they will come to us and I will not stay in the room it will be a police officer who will stay with them, they will see this thing, see my report and then go.\n> \n> (FME A, respondent's emphasis)\n\n> If we're using colposcopy... then there is a CD now which is recording the examination, now that CD is not available to anybody but the doctor, doctors who've taken it... So the Crown Office [and Procurator Fiscal Service], the courts, nobody, and we have an assurance from the Crown Office that they will not ask for it to be produced in evidence, as has been done in England, it has been used as evidence in England. In Scotland we believe, there is a woman in the witness box and the next minute you are showing 'Glorious Technicolors' of her private parts, that's not on. Now you may say 'Well, what about the defence?' Now if the defence have their own experts they are allowed to see the video, together with the doctors who have removed it, taken it... And that's how it is done and we've had no problems in Scotland, the fact that we cannot produce it in evidence... But that [colposcopic recording] we guard with our lives almost, in no way is that going to be divulged, given to anybody.\n> \n> (FME B)\n\nDue to the potential for harm to the survivor via the observation of the video within a public forum like the court, Scottish FMEs have negotiated with the Crown Office and Procurator Fiscal Service to ensure that such images are not publicly displayed. In order to do so they have developed a strategy to maintain one of the evidential benefits of the colposcope, the allowing of others in the criminal justice process to view a contemporaneous representation of the examination. However, they have permitted access only to those who will read the images in a medico-legal manner; other forensic practitioners to whom the survivor has already provided consent. The strategy is yet another example of a forensic compromise: attempting to retain the benefits of the technology (e.g. maintaining the perception of fairness and justice by enabling the defence to have access to the recordings), while at the same time ensuring the dignity of the survivor. It is with some final thoughts about the relationship between the evidential and therapeutic aspects of forensic medical work and what that means for the incorporation of new technologies into legal spaces that I will conclude.\n\n## **Conclusion**\n\nIt is no longer the case that forensic medical practitioners solely focus on either evidential or therapeutic concerns; although there are certain times in the forensic medical examination when one of these may be emphasised over the other, studies have demonstrated that practitioners are always cognisant of both during the overall conduct of their work (Mulla 2011, Rees forthcoming). Consequently, it is not possible for technologies such as the colposcope to address only one aspect of the evidential-therapeutic spectrum. As the colposcope controversy and the ensuing local decisions about practice have identified, it is not enough for a new technology to provide only an evidential benefit if it is also harmful to victims.\n\nWhile it was assumed by early proponents of the technology that the colposcope would revolutionise rape trials, enabling practitioners to identify more injuries and thereby provide more certain evidence, in reality the routine use of the device has only caused greater confusion over what the presence of genital injuries actually represents. Routinisation of colposcopic examinations despite continuing uncertainties regarding what its use represents testifies to the recently consolidated consensus around the broad import of the device. Yet, the colposcope was only considered useful once practitioners had found ways to ensure that they could successfully manage and securely store the large numbers of images it produced, i.e. by only recording when they believed it to have the greatest potential evidential benefit, ensuring the survivor's consent, and negotiating for the images only to be admissible in the courtroom in exceptional circumstances (and in Scotland, attaining agreement not to show them at all). It was only after such manoeuvring was complete that the colposcope achieved the status of a successful 'forensic' device.\n\nThe colposcope is indicative of the difficulties of incorporating technological devices into the legal milieu. While the institutions of science and law are interrelated and reliant upon each other in order to maintain credibility (Smith and Wynne 1989), the nexus at which they meet throws up complex sets of difficulties that require workarounds, as other contributions to this volume indicate. While many assumed that the colposcope would be of significant benefit to legal fact-finding as it was, when faced with the realities of victims' experience and criminal justice it was deemed necessary to adapt the ways through which the device was used. It is not enough to transfer scientific technologies wholesale into the legal arena; rather to be successful, devices must be made forensic and slowly integrated, taking account of the ways such artefacts can be differently interpreted, utilised and understood within legal contexts.\n\n The assumption that evidence gathering and care of the complainant are mutually exclusive categories has not only been prevalent in the forensic medical intervention, but the police and prosecuting and defending barristers have also, for many years, believed that the only way to get at the truth of whether an allegation happened was by appearing sceptical or disbelieving to the complainant. i.e. emphasizing evidential needs over care (Horvath and Brown 2009, Kelly _et al._ 2005, Rees 2010, Temkin 2000, 2005). Fortunately such practices are nowadays less common, especially in the police (McMillan and Thomas 2009), however, even the recent Stern Review (Stern 2010) can be characterised as stating that complainants can either receive fair treatment (known as procedural justice) or have a strong case (criminal justice), i.e. the care versus evidence dichotomy produced anew.\n\n Throughout this chapter I will use the terms 'survivor' and 'victim' interchangeably to connote the subject upon which the colposcope is used. While I am aware of the various difficulties with the choice of preferred noun for the survivor of sexual assault, and in particular the use of the word 'victim', I believe that these represent a far less disbelieving attitude towards the subject than the term 'complainant', normally used to identify someone who has reported a sexual assault but which leaves dubiety over the veracity of their allegation.\n\n The team were assessing the veracity of sexual violence cases based upon whether the hymen was intact, a highly circumspect method of interpretation.\n\n Teixeira's study is not cited in Slaughter and Brown (1992) and in fact they explicitly state that '[n]o published studies have evaluated the usefulness of the colposcope in the routine assessment of rape victims' (Slaughter and Brown 1992: 83). Of course such an evaluation was exactly what Teixeira was trying to achieve.\n\n STS scholars have drawn attention to the importance of images in scientific and medical work, emphasising the ways in which representations enable work, particularly when the object of investigations is not visible to the naked eye, but also recognising that such images are not the product of 'natural realism', but rather are constructed along established disciplinary conventions (Frow 2012, Lynch 1985). The production of images and the ways that they come to be interpreted can be very important in legal contexts given the significant power representations have in terms of their 'visual persuasiveness' (Burri 2012: 52). The belief that images do portray naturalistic realism gives them the status of objective facts, enabling others, for instance jurors, to see for themselves. Given that images are constructions however, they are also framed with particular meanings, drawing attention to certain aspects of phenomena at the expense of others. As a result images can provide very persuasive evidence (Burri 2012, Jasanoff 1998).\n\n Although as Brennan notes, the extension of the slogan to the adult cases had taken place 'without a significant research base' (Brennan 2006: 195).\n\n This has unfortunately come to pass; see White and Du Mont (2009).\n\n The study has been criticised for its methodology; cases were allocated to either the 'sexual assault' or 'consensual sexual intercourse' cohort based upon whether a case had been corroborated (or not). Consequently, if there existed substantiating evidence for sexual assault, for instance evidence of genital injury, the case would be added to the victim cohort. Conversely, if an account was not corroborated, then the complainant (who had originally been colposcopically examined as a victim) would be re-categorised within the consensual sexual intercourse group. As a result, some of the cases located within the consensual category might have been inappropriately classified. The implication of this is that a large proportion of the consensual sex category could represent women who had, in fact, been sexually assaulted, but whose cases resulted in insufficient corroboratory evidence to substantiate the allegation.\n\n As Lloyd _et al._ 's (2005) study recording the genital measurements of 50 women reported, female bodies are more diverse than most medical practitioners assume.\n\n Such assumptions about the type of cases where injuries are more likely to appear are not without their difficulties (Rees 2010).\n\n## **References**\n\nAdams, J.A., Harper, K., Knudson, S. and Revilla, J. (1994) 'Examination findings in legally confirmed child sexual abuse: it's normal to be normal', _Paediatrics_ , 94: 310\u2013317.\n\nAnderson, S.L., Parker, B.J. and Bourguignon, C.M. (2009) 'Predictors of genital injury after nonconsensual intercourse', _Advanced Emergency Nursing Journal_ , 31: 236\u2013247.\n\nAstrup, B.S., Ravn, P., Lauritsen, J. and Thomsen, J.L. (2012) 'Nature, frequency and duration of genital lesions after consensual sexual intercourse \u2013 implications for legal proceedings', _Forensic Science International_ , 219: 50\u201356.\n\nBarnes, B. and Shapin, S. (1979) _Natural Order_ , Thousand Oaks, CA: Sage Publications.\n\nBarnes, B., Bloor, D. and Henry, J. (1996) _Scientific Knowledge: A Sociological Analysis_ , London: Athlone.\n\nBowyer, L. and Dalton, M.E. (1997) 'Female victims of rape and their genital injuries', _British Journal of Obstetrics and Gynaecology_ , 104: 617\u2013620.\n\nBrennan, P.A.W. (2006) 'The medical and ethical aspects of photography in the sexual assault examination: why does it offend', _Journal of Clinical Forensic Medicine_ , 13: 194\u2013202.\n\nBurri, R.V. (2012) 'Visual rationalities: towards a sociology of images', _Current Sociology_ , 60: 45\u201360.\n\nCollins, H.M. (1975) 'The seven sexes: a study in the sociology of a phenomenon or the replication of experiments in physics', _Sociology_ , 9: 205\u2013224.\n\nCollins, H.M. (1981) 'Son of seven sexes: the social destruction of a physical phenomenon', _Social Studies of Science_ , 11: 33\u201362.\n\nDu Mont, J. and White, D. (2007) _The Uses and Impacts of Medico-Legal Evidence in Sexual Assault Cases: A Global View_ , Geneva: World Health Organisation.\n\nFaculty of Forensic and Legal Medicine (2010) _Guidance for Best Practice for the Management of Intimate Images that May Become Evidence in Court_. Available at: h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00adf\u00adf\u00adl\u00adm\u00ad.\u00ada\u00adc\u00ad.\u00adu\u00adk\u00ad\/\u00adl\u00adi\u00adb\u00adr\u00ada\u00adr\u00ady\u00adb\u00ady\u00ad\/\u00adt\u00ady\u00adp\u00ade\u00ad\/\n\nFrow, E.K. (2012) 'Drawing a line: setting guidelines for digital image processing in scientific journal articles', _Social Studies of Science_ , 42: 369\u2013392.\n\nHilden, M., Schei, B. and Sidenus, K. (2005) 'Genitoanal injury in adult female victims of sexual assault', _Forensic Science International_ , 154: 200\u2013205.\n\nHorvath, M. and Brown, J. (2009) _Rape: Challenging Contemporary Thinking_ , Cullompton: Willan Publishing.\n\nJasanoff, S. (1998) 'The eye of everyman: witnessing DNA in the Simpson trial', _Social Studies of Science_ , 28: 713\u2013740.\n\nKelly, L., Lovett, J. and Regan, L. (2005) _A Gap or a Chasm? Attrition in Reported Rape Cases, London_ : Home Office Research, Development and Statistics Directorate.\n\nKuhn, T.S. (1996) _The Structure of Scientific Revolutions_ , Chicago: University of Chicago Press.\n\nLincoln, C. (2001) 'Genital injury: is it significant? A review of the literature', _Medicine, Science and the Law_ , 41: 206\u2013216.\n\nLloyd, J., Crouch, N.S., Minto, J.L., Liao, L.M. and Creighton, S.M. (2005) 'Female genital appearance: \"normality unfolds\"', _British Journal of Obstetrics and Gynaecologists_ , 112: 643\u2013646.\n\nLynch, M. (1985) 'Discipline and the material form of images: an analysis of scientific visibility', _Social Studies of Science_ , 15: 37\u201366.\n\nMercer, D. (1996) _Understanding Scientific\/Technical Controversy_. Available at www.uow.edu.au\/arts\/sts\/research\/STPPapers\/Occpaper-1.html\n\nMcMillan, L. and Thomas, M. (2009) 'Police interviews of rape victims: tensions and contradictions in contemporary rape cases', in M. Horvath and J. Brown (eds), _Rape: Challenging Contemporary Thinking_ , Cullompton: Willan Publishing.\n\nMulla, S. (2011) 'Facing victims: forensics, visual technologies, and sexual assault examination', _Medical Anthropology_ , 30: 271\u2013294.\n\nNorvell, M.K., Benrubi, G.I. and Thompson, G.I. (1984) 'Investigation of microtrauma after sexual intercourse', _Journal of Reproductive Medicine_ , 29: 269\u2013271.\n\nOstrander, S.A. (1995) 'Surely you're not in this just to be helpful', in R. Hertz and J. B. Imber (eds), _Studying Elites Using Qualitative Methods_ , Thousand Oaks: Sage.\n\nPatel, H.C., Courtney, G.M. and Forster, G.E. (1993) 'Colposcopy and Rape', _American Journal of Obstetrics and Gynecologists_ , 168: 1334\u20131335.\n\nRees, G. (2010) '\"It is not for me to say whether consent was given or not\": Forensic medical examiners' construction of \"neutral reports\" in rape cases', _Social and Legal Studies_ , 19: 371\u2013386.\n\nRees, G. (2011) '\"Morphology is a witness that doesn't lie\": diagnosis by similarity relation and analogical inference in clinical forensic medicine', _Social Science and Medicine_ , 73: 866\u2013872.\n\nRees, G. (forthcoming) 'Contentious roommates? Spatial constructions of the therapeutic-evidential spectrum in medico-legal work', in I. Harper, T. Kelly and A. Khanna (eds), _The Clinic and the Court: Medicine, Law and Anthropology_ , Cambridge: Cambridge University Press.\n\nRogers, D. (1996) 'Physical aspects of alleged sexual assaults', _Medicine, Science and the Law_ , 36: 117\u2013122.\n\nRogers, D. (2004) 'The genital examination', in M. Dalton (ed.), _Forensic Gynaecology: Towards Better Care for the Female Victim of Sexual Assault_ , London: Royal College of Obstetricians and Gynaecologists Press.\n\nSavage, S.P., Moon, G., Kelly, K. and Bradshaw, Y. (1997) 'Divided loyalties? \u2013 the police surgeon and criminal justice', _Policing and Society_ , 7: 79\u201398.\n\nSlaughter, L. and Brown, C.R.V. (1992) 'Colposcopy to establish physical findings in rape victims', _American Journal of Obstetrics and Gynecology_ , 166: 83\u201386.\n\nSlaughter, L., Brown, C.R.V., Crowley, S. and Peck, R. (1997) 'Patterns of genital injury in female sexual assault victims', _American Journal of Obstetrics and Gynecology_ , 176: 609\u2013616.\n\nSmith, R. and Wynne, B. (1989) _Expert Evidence: Interpreting Science in the Law_ , London: Routledge.\n\nSommers, M.S., Fisher, B.S. and Karjane, H.M. (2005) 'Using colposcopy in the rape exam: health care, forensic, and criminal justice issues', _Journal of Forensic Nursing_ , 1: 28\u201334.\n\nState of West Virginia v. William M. (2010) Available at www.state.wv.us\/wvsca\/docs\/Spring10\/35130.htm\n\nStern, V. (2010) _The Stern Review_ , London: Home Office.\n\nTeixeira, W.R. (1980) 'Hymenal colposcopic examination in sexual offences', _American Journal of Forensic Medicine and Pathology_ , 2: 209\u2013215.\n\nTemkin, J. (2000) 'Prosecuting and defending rape: perspectives from the bar', _Journal of Law and Society_ , 27: 219\u2013248.\n\nTemkin, J. (2005) _Rape and the Legal Process_ , Oxford: Oxford University Press.\n\nWhite, D. and Du Mont, J. (2009) 'Visualizing sexual assault: an exploration of the use of optical technologies in the medico-legal context', _Social Science and Medicine_ , 68: 1\u20138.\nChapter 6\n\n* * *\n\n# **Telling tales**\n\nSome episodes from the multiple lives of the polygraph machine\n\n_Andrew Balmer_\n\n* * *\n\n## **Introduction**\n\nThis chapter is about the polygraph device, more popularly known as the 'lie detector'. The device has changed over time, though it has retained a central logic: specifically, physiological measures (e.g. blood pressure) are used as proxies for emotional responses created by the act of lying. However, this simple description might already go too far in prescribing a particular account of the device, since it will be the argument of this chapter that much of what the polygraph does can only be understood in relation to the contexts of its application. My argument focuses on the use of the polygraph in United States criminal trials and governance. Though it is largely assumed that the device is inadmissible in the US, the actual picture is more complex. Moreover, the adoption of the polygraph outside of criminal trials is similarly intricate, with the technology playing diverse roles in government departments, police investigations, dispute resolutions, post-probation programmes, surveillance, private investigation, family conflicts, media campaigns, in films and television, and a range of other contexts.\n\nThere are now several histories of the lie detector, which provide a fascinating account of how it developed, spread and found its way into systems of governance. These descriptions often tell a kind of 'origins' story. As the scholarship has developed, the accounts have worked further and further back into the device's history, to its antecedents and to the epistemological spaces into which it was born. They have also developed lines of connection between the polygraph and the recent emergence of brain-based devices. These analyses have argued that there are significant similarities in the assumptions underlying the way that these ostensibly different technologies are seen to work, or in the ways in which these technologies are rhetorically positioned as solutions to problems of governance, justice and social order.\n\nFor example, Alder affirms the significance of the polygraph in solving the problem of (dis)trust in American workplaces, criminal procedure and governance (Alder, 2002; Alder, 2007). One important feature of this argument regards police use of 'third degree' practices of physical abuse to obtain confessions. The politicisation of violence in policing practices allowed developers of the polygraph to rhetorically position use of the device as enlightened, humane and progressive (Littlefield, 2009; Littlefield, 2011; Bunn, 2012).\n\nSuch tales often recount the troubled emergence of the polygraph, during which it struggled to gain scientific and legal legitimacy. They highlight that the machine has undergone a series of alterations to its material components and that the protocols for its application during an examination have been revised on a number of occasions in the quest for credibility. Despite this story of technical change, such efforts to legitimise the polygraph, and demonstrate its reliability, appear to have been largely in vain. The poly-graph's status is as contested as it ever was. This proves to be the point of friction that gets the socio-historical fires burning: how is it that a device that has been doggedly attacked over a hundred years has found application in so many varied and vitally significant contexts?\n\nBunn (2012: 7\u201329), for instance, describes the significance of the notion of the 'born criminal' for the development of criminology and how this concept proved a barrier to the emergence of lie detection since it treated lies as a symptom of an underlying pathology, rather than a common social behaviour. Charting a similar territory, Weber (2008: 68\u201370) cogently illustrates how scientific work on fear was vital to establishing a link between lying, the body and the measurement devices available at the time. Ronald Thomas (1999: 23) traces the role of detective fiction in the emergence of a 'new literacy' of the body, and argues that the developers of polygraphy 'aimed at achieving in the field of law enforcement the same feat that detective writers sought to produce in the literary imagination: reading the truth directly inscribed in the criminal body.' Drawing on Thomas' insights, Littlefield (2011: 68) highlights the way in which a history of 'mind reading' in fiction was similarly important in creating a body that could be read and notes the continued importance of this narrative in contemporary brain-based deception detection.\n\nIn this regard, the scholarly literature has perhaps largely assumed that the story of the polygraph somehow needs to be made sense of. Though these social histories revel in the many twists and turns of the polygraph's life story, they nonetheless often point towards the contradiction between the device's status and its application as being in need of resolution. Such an impetus to explain the polygraph's success embeds a certain notion that the world can indeed be explained coherently; that seemingly incongruent phenomena can be resolved if enough context is excavated from the historical archives.\n\nThough these stories of lie detection are already vibrant and rigorous, I aim to further nuance this work by taking inspiration from scholarship in science and technology studies (STS) that has evidenced the ways in which technologies and social order are co-produced (Latour, 1987; Jasanoff, 2004a). Co-production is 'shorthand for the proposition that the ways in which we know and represent the world (both nature and society) are inseparable from the ways in which we choose to live in it. Knowledge and its material embodiments are at once products of social work and constitutive of forms of social life' (Jasanoff, 2004a: 2). I also draw on work that has celebrated the complexities and multiplicities of technological artefacts and knowledge-making practices (see the following collections for examples: Law and Mol, 2002; Ang, 2011). In this regard, I adopt the notion that multiple realities co-exist and that each individual context comes with its own set of practices for making sense of technologies, bodies, texts, and so on. Mol (2002: 6) argues that, ' _ontology_ is not given in the order of things, but that, instead, _ontologies_ are brought into being, sustained, or allowed to wither away in common, day-to-day, sociomaterial practices' (emphasis in original'. Following Mol and others, Pickersgill (2013) has drawn attention to the ways in which technoscientific innovation can profoundly reshape social and legal ontologies that in turn become implicated in innovative processes themselves and in complex ways that escape linear narratives. In what follows, I draw inspiration from these positions to tell a few tales of the polygraph from its lives in and out of the courtroom. Using the extant literature, I argue that the numerous practices in which the polygraph has found application produce a complex story at the intersections of law and science that cannot be easily resolved into a coherent and sensible trajectory.\n\n## **Assessing the legal status of the polygraph in the USA**\n\nIn thinking about the legal history of the polygraph, it is useful to reflect on the development of physiological studies relating to (criminal) behaviour. In particular, let us consider the history of blood pressure. From the 1600s to the late 1800s this phenomenon was constituted through a great many developments and paradigmatic changes in understanding and apparatus, much of which we owe to Stephen Hales, a biologist working in the eighteenth century. In 1733 he took measurements directly from the artery of a horse by use of a 9ft glass tube (Booth, 1977) and thus made the first measurements of the pressure of blood. One hundred and fifty years later, in 1896, the first non-invasive blood pressure cuff was invented by Scipione Riva-Rocci (Noyes, 1936). Around this time, Angelo Mosso first made connections between blood pressure and fear. Mosso observed that a patient with a cranial fissure showed increased brain pulsations when he was fearful. For example, when the patient accidentally missed saying a prayer at 12pm Mosso's measuring device showed an increase in pulsations. This helped form the basis of studies of the correlation of the blood pressure with emotional changes (Raichle, 1999). Mosso was a student and colleague of the infamous Cesare Lombrosso, who theorised that criminal behaviour was a form of atavism and that criminals could be identified by their anatomical features (Bunn, 2012). Lambrosso was the first to apply this knowledge in assisting police interrogation of criminal suspects (Trovillo, 1939). Thus, by the beginning of the twentieth century, the links between emotions, criminality and physiology had been born, but had only taken small steps.\n\nHugo Munsterberg's work at Harvard in 1915, conducted with his student William Moulton Marston, brought Mosso's early observations into dialogue with understandings of deception and truth. Marston published his work on the systolic blood pressure symptoms of deception in 1917 claiming that: 'The behavior of the blood pressure [... ] constitutes a practically infallible test of the consciousness of an attitude of deception' (Marston, 1917: 162). Littlefield (2011: 52\u201356) shows how Marston developed what he called the 'significant lying curve', a graphical representation of bodily responses, which he connected to his concept of the 'deceptive consciousness'. Most important in this work was Marston's creation of experimental protocols, still used today, that are designed to have subjects perform known lies at known times (having subjects, for example, commit a 'mock crime' by 'stealing' an object from a known room and then lie about it). Weber argues that the performance of mock crimes and the staging of the examination were crucial to invoking fear in the subject and that this emotional response was vital to the measurement of lying as conceived by Marston and colleagues (Weber, 2008: 74\u201375 and 149). Thus were established the key ontological connections that made lie detection in its early years. Importantly, these ontological connections between emotion, the body, and lying invoked emerging ideas from physiology, but also psychology and psychiatry.\n\nSo it was that, in the 1920s, what Alder (2007) calls 'the American obsession' with lie detectors began. The brutal 'third degree' interrogation practices of police of the time, coupled with widespread corruption of the justice system, proved a fertile ground for the emergence of this nascent forensic science (Alder, 2007). The work on deception continued into the late 1920s and it was then that Marston called his first version of the lie detector 'the systolic blood pressure deception test' (Grubin and Madsen, 2005). Shortly after this, it was Marston himself that administered the test in what came to be seen as a landmark decision in _Frye v. United States_ (1923) 293 F. 1013 D.C. Circuit Court.\n\n_Frye_ set the standard of admissibility for expert evidence, and through it began the long battle to establish the polygraph as admissible evidence for criminal trials. In 1922, William Moulton Marston \u2013 one of the developers of the lie detector \u2013 examined James Alphonse Frye in a now infamous case of combined robbery and murder. Marston determined that the test results exonerated Frye of wrongdoing. However, his testimony was deemed inadmissible by the lower trial court and the judgment was upheld by the D.C. Court of Appeals. The opinion of the appeal court created the now well-known and broadly applied _Frye_ standard, which held that, for expert testimony to be admissible, the scientific principle from which the testimony was deduced must be 'sufficiently established to have gained general acceptance in the particular field in which it belongs' ( _Frye_ , 1014). The Court determined that the systolic blood pressure deception test, later to become 'the polygraph', had 'not yet gained such standing and scientific recognition among physiological and psychological authorities' ( _Frye_ , 1014).\n\nIn this regard, the device's scientific immaturity, based not only on the still emerging science of physiology but also of psychology and psychiatry, proved an important technical contribution to the determination of its inadmissibility. _Frye's_ focus on general acceptance and reliability exemplified this:\n\n> Just when a scientific principle or discovery crosses the line between the experimental and demonstrable stages is difficult to define. Somewhere in this twilight zone the evidential force of the principle must be recognized [... ] We think the systolic blood pressure deception test has not yet gained such standing and scientific recognition among physiological and psychological authorities as would justify the courts in admitting expert testimony deduced from the discovery, development, and experiments thus far made.\n> \n> ( _Frye v. United States_ (1923) 54 App. D.C. 46; 293 F. 1013, 1014)\n\nThe Court's decision in _Frye_ was therefore directly, though not fully, determined by the status of the device and of the scientific fields with which the polygraph was affiliated. In physiology, the practices of measurement were progressing towards standardisation but had really only just been developed ten years prior. Similarly, in 1923, the psychological notion of the emotional body and the psychiatric understanding of the unconscious were still in development and these ideas were central to the polygraph's claim to be objective in the determination of truth and falsehood (Weber, 2008). The lie detector was thus similarly in an early stage of development, with new physiological measures being added, new practices for the examination being devised, and with little or no standardisation in the devices, its usage, or the training and backgrounds of operators. The ambiguous nature of the lie detector, then, ensured that it would not be used in court at this time; more profoundly, however, this same ambiguity played a key role in the production of stable evidentiary procedure that would resonate through the courts for decades to come.\n\n## **The shifting terrain of admissibility**\n\nThe _Frye_ ruling developed across the states into a _per se_ exclusion, meaning that the polygraph was routinely refused admissibility without much hearing of its reliability, validity or technological development. This began to change in the 1970s and 1980s, when a few cases emerged at the Federal level in which the polygraph was granted limited admissibility (Giannelli, 1997). Such cases were largely heard in relation to specific issues, such as perjury or cases in which the defendant's taking of a polygraph test was relevant to some matter outside of the central issue of the trial. As such, it was specificities of the crime under examination that allowed some courts to grant admission of the evidence from polygraph tests. Many of these developments were overturned and by the late 1980s the response to the polygraph remained largely exclusionary at the Federal level.\n\nAt the state level, however, a different picture had emerged. Whilst some courts retained the _per se_ ban on admissibility, others took a different path in their development of a 'prior-stipulation' mechanism for admitting polygraph results and testimony. The 'prior-stipulation' rule required that all parties to the trial agreed to the admission of the findings of the examination prior to it being conducted. This meant that defendants might find themselves hearing evidence to which they themselves had stipulated but which claimed they were guilty. This unusual situation continues to this day at the state level, with twenty-nine states retaining the _per se_ ban, fifteen states adopting prior stipulation and only one state, New Mexico, allowing polygraph evidence without prior stipulation (Shniderman, 2011).\n\nThis emergence of the prior-stipulation rule was facilitated by _Frye's_ focus on reliability and general acceptance. Cases seeking admissibility of the polygraph were able to argue that technical developments in the period since 1923 had led to increased reliability, whilst narrowing down the relevant scientific community to a select group of experts on polygraph science (Balmer, 2011(. Advocates of the polygraph, it seemed, had finally found a route into criminal trials.\n\nWhilst the prior-stipulation rule had been developing, there was also mounting dissatisfaction with _Frye_ (Jasanoff, 1995). For example, some courts were having trouble applying the rule as regards determining the appropriate field in which to evaluate the evidence and struggled to determine exactly what would constitute general acceptance (Giannelli, 1980). Since the Rules Enabling Act of 1934, the Supreme Court had been able to promulgate rules of practice and procedure but nonetheless took a rather light-handed approach to developing rules of evidence. In 1975, the Federal Rules of Evidence (FRE) were approved by Congress and President Ford; these had been hashed-out and re-developed over an achingly long period between 1934 and 1975. A number of committees were established and disbanded, reports written and re-written, and a series of political controversies troubled the formation of rules for evidentiary decision-making (Scallen, 2001).\n\nDespite such trouble, it was not until 1993 \u2013 seventy years after _Frye_ \u2013 that the Supreme Court finally decided to hear a case directly regarding the admissibility of expert evidence, in _Daubert v. Merrell Dow Pharmaceuticals_ (1993) 509 U.S. 579. In that decision, the Supreme Court held that general acceptance 'is not a necessary precondition to the admissibility of scientific evidence under the Federal Rules of Evidence' ( _Daubert_ , 597). As such, the _Daubert_ decision moved away from _Frye_ (though it retained a 'general acceptance' test as one of five criteria to be considered) and firmly adopted the FRE as the principal mechanism through which admissibility decisions were to be made at the Federal level (see also Solomon and Hackett, 1996).\n\n_Daubert_ was modified in two further Supreme Court cases, _General Electric Co._ v. _Joiner_ (1997) 522 U.S. 136 and _Kumho Tire Co. v. Carmichael_ (1999) 526 U.S. 137, which now constitute the ' _Daubert Trilogy_ ' used to determine admissibility at the Federal level under the FRE. Each state in the US may choose to adopt or adapt the Trilogy, and many had already adopted or adapted the FRE to their own local context before _Daubert_. However, the Trilogy has been far from uniformly implemented at the state level. Indeed, _Frye_ remains in force in its home Court in the District of Columbia, and in a number of states, including some of the larger ones like California, New York, Pennsylvania, Illinois and Florida. Other states have adopted the Trilogy in part or in whole but the uneven spread of _Daubert_ has been unpredictable (Kritzer and Beckstrom, 2007).\n\nFrom this summary of decisions, we begin to see that the United States operates something of a hodgepodge of practices for determining admissibility of the polygraph at the state level. The emergence of _Daubert_ in a period of relative success for the polygraph (as regards prior stipulation) has added to the options available to any court seeking to determine the admissibility of the technique. Moreover, as with the emergence of prior stipulation, it is the specificity of the case that often influences the way in which the polygraph is handled by the trial court.\n\nTake, for example, the case of _State of New Mexico v. Shane Glen Harrison_ (2000) 129 N.M. 328; 2000 NMSC 22; 7 P.3d 478. In New Mexico, polygraph evidence is currently admissible without prior stipulation \u2013 a situation that has developed over a number of cases in the past twenty years. Given a set of five rules to which the evidence must adhere, the polygraph charts and examiners can be admitted to the trial. One of these rules relates to the expertise of the examiner. In _Harrison_ , the Court had to determine what constituted a sufficient level of expertise for a polygraph examiner. The defendant claimed that the polygraph expert was not qualified to determine the effect that the defendant's condition of eye irritation and associated pain would have on the results of the test. The Court decided that the expertise of the polygraph examiner was sufficient to determine the results of the test, despite it not extending to the influence the eye irritation may have on those results. Rather, the Court instructed, such information had to be taken into account in weighing the significance of the evidence from the polygraph in relation to the rest of the evidence in the trial. In this regard, the state rules, based on elements of the FRE, interacted with the particular trajectory of cases heard on the polygraph in New Mexico to that date, notions of expertise, the specificities of the defendant's body and medical history, and the other evidence in the case.\n\nAdmissibility practices are thus constantly shifting and are open to re-interpretation. The corollary of this is that the meaning and import of polygraph evidence and of the polygraph's reliability and validity is similarly shifting and multiplying. We can see how the ontological status of the polygraph is not only given by the practices of which it forms a part but similarly influences the direction and consolidation of those practices. The ambiguous material and practical technoscientific status of the polygraph contributed to the creation of practices for the determination of admissibility of expert evidence more generally, which in turn pushed the device's development outside of the criminal trial and into the legal periphery. In the following section, I tell another tale of the polygraph that recounts some of how its ontological status changed as it multiplied out into\/formed a new world as a surveillance technique.\n\n## **The world of surveillance: the polygraph and social order**\n\nFollowing the exclusions of the lie detector from the criminal courts, innovation with the device nevertheless continued. Although Marston had done much to get the lie detector started, and would later contribute significantly to its popularisation, it was John Larson, a physiologist and medical student, and Leonarde Keeler, a 'man-about-town and entrepreneur' (Alder, 2002: 8) who oversaw its development and dissemination into American industry, government and the justice system. The pair worked under August Vollmer, the Berkeley Police Chief who was responsible for the professionalization of the police force in the early part of the century. Vollmer's pledge was to do away with the third degree and found policing anew, this time on the basis of objectivity and a dispassionate pursuit of justice. Larson and Keeler's skill with the techniques of physiology were central to that programme since the emergence of criminology meant that crime was increasingly studied through biological lenses (Rafter, 2008; Bunn, 2012). Whilst each of the men contributed to the development of the polygraph, it was Keeler who patented an 'apparatus for recording arterial blood pressure' in 1931 (Keeler, 1931) and who, reacting to the exclusion of the polygraph, steered the device into policing, government and industry.\n\nIndeed, it was Keeler who did the most to popularise the polygraph, drawing on rhetorics of science, magic, law and order in his marketing of the device (Bunn, 1997). For example, in 1929 he suggested subjecting department store employees to regular polygraph exams, believing that this would substantially reduce the losses suffered from pilfering (Alder, 2007). Within twelve months Keeler had begun administering the polygraph in banks, where he found 10\u201325 per cent of employees confessed to theft of one sort or another. The rate of the confessions was startling, and the outing of employees as untrustworthy connected with the emergence of a post-Taylorist demand for efficiency and workforce management (Alder, 2002). Despite a managerial will to fire these individuals, Keeler proposed that they instead be retained and re-tested regularly, a strategy that would certainly make him more money and position the polygraph not simply as a detector of deceit but also as a psychological deterrent to criminal activity. It was through this practice of regular testing and the notion of prevention, spurred by the strategies for producing social order in industry, that the polygraph became a surveillance technique. Contrary to Foucault's arguments regarding the panopticon's visual force (Foucault, 1977), the polygraph's key potency in the emerging practices of corporate surveillance was its temporal effects. As Weber (2008) argues, the polygraph promised to instil a feeling of being observed by the threat of the upcoming examination. This depended on a characterisation of the criminal mind and body: that one's crimes were no longer eroded by the ever flowing waters of time but were inscribed into the emotional body, later to be read.\n\nThis re-emphasis on the polygraph for employee screening had an impact on Keeler's development of training and examination. Previously, the number of examinations he could conduct was limited and every new trainee he produced became a competitor for consultations. The sale of the polygraph device did not net him nearly as much profit as would the consultations lost to his own progeny of investigators (Alder, 2002). The trouble was that in order to get a patent on the device, Keeler had drawn too strong a line between himself as 'inventor' of the device and its increasing depiction in the media (Alder, 2007; Bunn, 2012; Littlefield, 2010). As such, Keeler shied away from mass production of the device, worried that it would tarnish his image if it was taken up by a range of unqualified and untrained individuals and ultimately cost him in sales since he was the only one with the patent. However, his patent ran out after the war and \u2013 no longer able to monopolise the developing market \u2013 Keeler switched from a stratagem of direct consultation and selling of his own expertise, to one of selling the device and training-up more operators. He began running two-week training courses on how to use the polygraph, which rapidly increased the number of examiners in the USA (Alder, 2007). By the 1950s, various parts of American industry were routinely employing the polygraph to secure the reliability of its white collar workforce, with over 2,000,000 examinations taking place in 1950 alone (Alder, 2002).\n\nThe 1940s and 1950s thus saw a massive expansion in the use of polygraphy, particularly during the McCarthyite period. What Keeler had developed in the context of industrial efficiency now took up residence in the heart of the government's security policies as the device became used to police against the threat of hidden communist sympathisers and agents. The first government agency to engage the practice became the Atomic Energy Commission in 1947 (Cook and Robinson, 1947; Shaw and Sapirie, 1947) but in the decades following it has been picked-up in the Department of Energy, the Department of Defense, the National Security Agency, and a range of other large departments in the USA with responsibilities for (inter)national security. The primary capability of the polygraph in this context thus became its use as a deterrent to espionage and as an investigative technique to expose it. Fear around the stability of American social order demanded new technologies for securing against its breakdown; the polygraph seemed to answer social and political concerns, helping to facilitate the expansion of practices of surveillance and interrogation. In this respect, the polygraph helped to shape the notion of American security from the 1940s onwards.\n\nA combination of factors thus influenced the shift in Keeler's strategies: the disappointing outcome that had been _Frye_ and the need to find a new home for the polygraph outside of the legal system; the struggle to attain a patent; the need to profit from the device without the patent; and the various political practices that constituted the culture of suspicion and mistrust during the 1940s and 1950s. Keeler played a prominent role in creating a new ontology for the polygraph by shifting his sales strategy. The emerging McCarthyist climate helped in stabilising this new world for the machine by adopting the polygraph as a surveillance device in government practices. The polygraph similarly proved influential in sustaining these practices. In the following section, I outline another context in which the polygraph came to act as a surveillance device (the Department of Energy) and show how the application of the device within these practices caused protest, leading to a damning report on its reliability and to its simultaneous entanglement with the 'war on terror'.\n\n## **The world of terror: challenges to the polygraph at the DOE**\n\nThe DOE was created in 1977 to coordinate the various energy enterprises and research centres across the USA in order to respond to the energy crisis crippling the US economy (Rudman _et al._ , 1999). The scientific work conducted in some of the DOE's most important organisations is classified and can be traced back to the Manhattan Project. However, counterintelligence (CI) reports over its first few decades consistently claimed that the DOE was failing to maintain high-level security. This was a problem since the 'nuclear secret' had become a fetish of the Cold War and the DOE had hundreds of millions of pages of classified material to manage (Masco, 2002).\n\nA Presidential Commission was created to investigate the matter in 1998, which declared the Department to be a 'Byzantine, and bewildering bureaucracy' (Rudman _et al._ , 1999) that would have to be fundamentally reformed to fortify itself against the threat of security breaches. The structure of the DOE was thus overhauled by Bill Clinton with Presidential Decision Directive 61. That directive created the DOE Counter Intelligence Office, which became operational in 1998 (Anon, 2000). It also included provisions for polygraph testing and so in the same year 800 DOE Federal and contractor employees were scheduled to be screened (Cumming, 2007) by the CI for a number of 'espionage-related reasons' (Richardson, 1999).\n\nThe screening found very little acceptance by the energy scientists working in the major DOE laboratories (Anon, 2000). Four public hearings were held and over 100 written comments were received during 1999 on the subject of polygraph screening (Richardson 1999). Even so, in the autumn of the same year Congress approved legislation expanding the number of employees subject to polygraph screening to 13,000, which was met with further criticism. Due to continued protest about the programme and the influence of Senator Jeff Bingaman \u2013 at the time, a prominent democrat on the Senate Energy and Natural Resources Committee (Bingaman, 1999) \u2013 an 18 month report from the National Research Council (NRC) was commissioned by the DOE with funding of $860,000 (McKee, 2000) to investigate the scientific status of the polygraph.\n\nThe report's findings painted a worrisome picture for DOE security and for the polygraph's legitimacy. It argued that the proportion of spies and terrorists that would be subject to DOE polygraph screening would, presumably, be extremely low, since they would account for only a small percentage of the much larger legitimate scientific workforce. The polygraph, it was argued, does not have adequate sensitivity to demarcate such a small number of threatening subjects. The hypothetical population of ten spies in 10,000 employees working at a reasonably generous accuracy rate of 80 per cent detection of deceivers would, according to the report, produce 1,606 failed tests, of which eight would be spies. The rest would be false positives that misidentified trustworthy scientists as spies (NRC, 2003). The report thus questioned the validity and reliability of the polygraph and criticised many of the scientific articles supporting the device for, amongst other factors, not having adequate experimental controls. Overall, the report is a rather damning assessment of the potential of the polygraph, as two of the authors have themselves tried to make clear elsewhere (Fienberg and Stern, 2005).\n\nThe report, however, concedes that the polygraph may still have utility in deterring application by candidates that present a security risk and that it may go some way to reducing security violations and increasing confessions. Thus, the use of the device in security settings was nonetheless (partially) preserved by virtue of its confession-producing ability, which the report implies has little to do with any scientific principle. Instead, the potency of the polygraph is attributed to the continued _perception_ of the device as valid and accurate and its use in leveraging confessions from examinees.\n\nDespite retaining some of the promise of the polygraph, the NRC report recommended a programme of scientific research that might validate, supplement or supplant the device: 'National security is best served by a broad research programme on detecting and deterring security threats, not a narrow focus on polygraph research' (NRC, 2003(. Congress responded to this recommendation by allocating $500,000 for the National Science Foundation and Office of Science and Technology Policy through the 2004 Intelligence Authorization Act. The funding was for research on 'behavioral, psychological, and physiological assessments of individuals' (Public Law 108\u2013 177-December 13th, 2003, Section 375.).\n\nIn 2002 the National Defense Authorization Act directed the DOE to develop a new polygraph programme in light of the NRC report (Cumming, 2007) and the DOE dutifully responded in the Federal Register of April the following year (Anon, 2003). As regards the criticisms of the NRC, the DOE acknowledged the lack of scientific evidence for polygraph screening whilst retaining its programme because it was mandated to do so by the Congress. The DOE argued that the polygraph was not as central to the determination of security as seemed to be the case. It explained that it was just one test used as part of an investigative process and that the number of false positives was tolerable since the technique would nevertheless identify suspect individuals and innocent individuals would not be affected by a false positive. The Secretary of Energy, Spencer Abraham, commented that 'in light of the current national security environment, the ongoing [sic] military operations in Iraq and the war on terrorism it was appropriate to retain the existing system' (Anon, 2006). Thus a number of political factors were implicated in the continued use of the polygraph machine at the DOE despite the protests from prominent scientists. The device was constituted as part of a programme of measures being taken to ensure the security of the United States and so its role as a surveillance device in homeland security was expanded along new lines to become an important tool in the war on terror.\n\n## **Conclusion**\n\nAs Mol and Law contend, investigations into the organisation of the modern world lead us to discover that we are 'living in two or more neighbouring worlds, worlds that overlap and coexist' (2002: 8). In this chapter I have shown how the status of the polygraph is importantly contextualised within the practices of its use in different socio-legal spaces and how it similarly influences their development.\n\nThe establishment of admissibility rules in _Frye_ , their influence on prior stipulation, the shifting positions on prior stipulation across the US state courts, the way in which the FRE was adopted in _Daubert_ , the subsequent creation of the Trilogy and the individual specificity of cases brought before trial courts all evidence a tangled, multiple and complex picture for the polygraph as regards its legal status. Within this heterogeneous network of relations, any attempt to pin-down the exact nature or pattern of polygraph admissibility is made impossible by the tensions that emerge between the various worlds of legal evidence decision-making.\n\nIf this was not enough complexity, the various multiplications of the poly-graph's existence in the legal periphery and government of the United States make any representation of the device ever more unstable. The adoption of the lie detector by the US government during the 1950s and 1960s was concurrent and overlapped with its heterogeneous extensions in the worlds of law. Moreover, the movement of the polygraph into the world of contemporary practices of war is only one of many new developments in the lives of the polygraph. In particular, the choices I have made in this chapter have obscured our view of the polygraph's use with sex offenders. Such use has been developing since the 1960s and has further interacted with admissibility decisions and surveillance practices (Balmer and Sandland, 2012). I might also have chosen to explore the connections made between these episodes and the development of brain-based techniques of deception detection (Littlefield, 2009). The polygraph's lives in the media have resonances and discordances with these other worlds of representation and use. There are more and more tales to be told.\n\nThe tendency in the literature has been to see these various episodes as progressions in the journey of the device, from failure at trial to success in the government or to try to explain how the device 'doesn't work' scientifically (or at least it is highly contested) but does work as a surveillance tool (or at least rhetorically). But there are non-coherences (Law, 2002) in the narratives analysts have articulated about the polygraph, and these must be recognised as such. Accepting non-coherence in our accounts of the device is to begin to produce a polygraphic story of polygraphy, which might lead us to novel perspectives, analytical tools and representational strategies to understand the way in which the polygraph navigates and constitutes relations between science, law and social order.\n\n## **References**\n\nAlder, K. (2002) 'A social history of untruth: lie detection and trust in twentieth-century America', _Representations_ , 80(1): 1\u201333.\n\nAlder, K. (2007) _The Lie Detectors: The History of an American Obsession_. New York, Free Press.\n\nAng, I. (2011) 'Navigating complexity: from cultural critique to cultural intelligence', _Continuum_ , 25(6): 779\u2013794.\n\nAnon (2000) _Improving Counterintelligence Capabilities at the Department of Energy and the Los Alamos, Sandia, and Lawrence Livermore National Laboratories_.\n\nAnon (2003) 'Office of counterintelligence: polygraph examination regulations 10 Cfr Part 709', _Federal Register_ , 68(71): 17886\u201317890.\n\nAnon (2006) 'Counterintelligence evaluation regulations department of energy 10 Cfr Parts 709\u2013710', _Federal Register_ , 71(1989): 57386\u201357397.\n\nBalmer, A.S. (2011) _Where the Truth Lies: The Development and Legal Status of Lie Detection for Use in U.S. Criminal Trials_ , PhD Thesis, University of Nottingham.\n\nBalmer, A.S. and Sandland, R. (2012) 'Making monsters: the polygraph, the plethysmograph and other prctices for the performance of abnormal sexuality', _Journal of Law and Society_ , 39(4): 593\u2013615.\n\nBingaman, J. (1999) 'Speaking in congress on the national defense authorization act for fiscal year 2000 conference report 22 September 1999', _Congressional Record_ : S11189\u2013S11201.\n\nBooth, J. (1977) 'A short history of blood pressure measurement', _Proceedings of the Royal Society of Medicine_ , 70: 793\u2013799.\n\nBunn, G.C. (1997) _The Hazards of the Will to Truth: A History of the Lie Detector_ , York: York University.\n\nBunn, G.C. (2012) _The Truth Machine: A Social History of the Lie Detector, Baltimore_ , MA: Johns Hopkins University Press.\n\nCook, R.W. and Robinson, J.C. (1947) _Continued and Extended Use of the Polygraph, Department of Energy Declassification Review_ Online]. Available: [h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ada\u00adn\u00adt\u00adi\u00adp\u00ado\u00adl\u00ady\u00adg\u00adr\u00ada\u00adp\u00adh\u00ad.\u00ado\u00adr\u00adg\u00ad\/\u00add\u00ado\u00adc\u00adu\u00adm\u00ade\u00adn\u00adt\u00ads\u00ad\/\u00adR\u00adH\u00adT\u00adG\u00ad4\u00ad0\u00ad4\u00ad2\u00ad2\u00ad.\u00ads\u00adh\u00adt\u00adm\u00adl\u00ad (a\u00adc\u00adc\u00adevs\u00ads\u00ade\u00add\u00ad \u00ad2\u00ad4\u00ad \u00adJ\u00adu\u00adl\u00ady\u00ad 2\u00ad0\u00ad1\u00ad0).)\n\nCumming, A. (2007) _Polygraph Use by the Department of Energy: Issues for Congress_. _Library of Congress Washington DC, Congressional Research Service Report for Congress_.\n\nFienberg, S. and Stern, P. (2005) 'In search of the magic lasso: the truth about the polygraph', _Statistical Science_ , 20(3): 249\u2013260.\n\nFoucault, M. (1977) _Discipline and Punish_ , London: Allen Lane.\n\nGiannelli, P.C. (1980) 'The admissibility of novel scientific evidence: Frye v. United States, a half-century later', _Columbia Law Review_ , 80: 1197\u20131250.\n\nGiannelli, P. C. (1997) 'Polygraph evidence: post-Daubert', _Hastings Law Journal_ , 49: 895\u2013924.\n\nGrubin, D. and Madsen, L. (2005) 'Lie detection and the polygraph: a historical review', _Journal of Forensic Psychiatry & Psychology_, 16(2): 357\u2013369.\n\nJasanoff, S. (1995) _Science at the Bar: Law, Science, and Technology in America_. Cambridge, MA: Harvard University Press.\n\nJasanoff, S. (2004) ' _The idiom of co-production_ ', in S. Jasanoff (ed.), _States of Knowledge: The Co-Production of Science and Social Order_ , New York: Routledge.\n\nKeeler, L. (1931) _Apparatus for Recording Arterial Blood Pressure_.\n\nKritzer, H.M. and Beckstrom, D.C. (2007) 'Daubert in the states: diffusion of a new approach to expert evidence in the courts', _Journal of Empirical Legal Studies_ , 4(4): 983\u20131006.\n\nLatour, B. (1987) _Science in Action: How to Follow Scientists and Engineers through Society_ , Boston, MA: Harvard University Press.\n\nLaw, J. (2002) ' _On hidden heterogeneities: complexity, formalism and aircraft design_ ', in J. Law amd A. Mol (eds), _Complexities: Social Studies of Knowledge Practices_ , London: Duke University Press.\n\nLaw, J. and Mol, A. (eds) (2002) _Complexities: Social Studies of Knowledge Practices_ , London: Duke University Press.\n\nLittlefield, M.M. (2009) 'Constructing the organ of deceit the rhetoric of fmri and brain fingerprinting in post-9\/11 America', _Science, Technology & Human Values_, 34(3): 365\u2013392.\n\nLittlefield, M.M. (2011 ) _The Lying Brain: Lie Detection in Science and Science Fiction_ , Ann Arbor, MI: University of Michigan Press.\n\nMarston, W.M. (1917) 'Systolic blood pressure symptoms of deception', _Journal of Experimental Psychology_ , 2(2): 117\u2013163.\n\nMasco, J. (2002) 'Lie detectors: on secrets and hypersecurity in Los Alamos', _Public Culture_ , 14(3): 441\u2013467.\n\nMckee, J. (2000) '$860,000 Department of Energy study to evaluate polygraphs', _Albuquerque Journal_ , 5 December.\n\nMol, A. (2002) _The Body Multiple: Ontology in Medical Practice_. London: Duke University Press.\n\nNoyes, B. (1936) 'The history of the thermometer and sphygmomanometer', _Bulletin of the Medical Library Association_ , 24(3): 155\u2013165.\n\nNRC (2003) _The Polygraph and Lie Detection_. Washington, DC: National Academies Press.\n\nPickersgill, M.D. (2013) ' _Sociotechnical innovation in mental health: articulating complexity_ ', in Flear, M. L., Farrell, A., Hervey, T. K. and Murphy, T. (eds) (2013) _European Law and New Health Technologies_ , Oxford: Oxford University Press.\n\nRafter, N. (2008) _The Criminal Brain: Understanding Biological Theories of Crime_ , New York: NYU Press.\n\nRaichle, M.E. (1999) 'Modern phrenology: maps of human cortical function', _Annals of the New York Academy of Sciences_ , 882: 107\u2013118.\n\nRichardson, B. (1999) _Memorandum for Heads of Departmental Offices 'Counter Intelligence Polygraph Implementation Plan' the Secretary of Energy Washington DC 20585_ [Online]. Available: www.fas.org\/sgp\/othergov\/polygraph\/ polyimp.html (accessed 24 July 2010).\n\nRudman, W.B., Caracristi, A.Z., Drell, S.D. and Friedman, S. (1999) _Science at Its Best, Security at Its Worst: A Report on Security Problems at the U.S. Department of Energy_. _President's Foreign Intelligence Advisory Board_ , Washington: United States Government.\n\nScallen, E.A. (2001) 'Analyzing the politics of (evidence) rulemaking', _Hastings Law Journal_ , 53: 843\u2013884.\n\nShaw, D.F. and Sapirie, S.R (1947) _Continued and Extended Use of the Polygraph, Department of Energy Declassification Review_ Online]. Available: [h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ada\u00adn\u00adt\u00adi\u00adp\u00ado\u00adl\u00ady\u00adg\u00adr\u00ada\u00adp\u00adh\u00ad.\u00ado\u00adr\u00adg\u00ad\/\u00add\u00ado\u00adc\u00adu\u00adm\u00ade\u00adn\u00adt\u00ads\u00ad\/\u00adR\u00adH\u00adT\u00adG\u00ad4\u00ad0\u00ad4\u00ad2\u00ad2\u00ad.\u00ads\u00adh\u00adt\u00adm\u00adl\u00ad (a\u00adc\u00adc\u00adevs\u00ads\u00ade\u00add\u00ad \u00ad2\u00ad4\u00ad \u00adJ\u00adu\u00adl\u00ady\u00ad 2\u00ad0\u00ad1\u00ad0).)\n\nShniderman, A. (2011) 'You cant handle the truth: lies, damn lies, and the exclusion of polygraph evidence'', _Albany Law Journal of Science and Technology_ , 22(2): 433\u2013473.\n\nSolomon, S.M. and Hackett, E.J. (1996) 'Setting boundaries between science and law: lessons from Daubert v. Merrell Dow Plarmaceuticals, Inc.', _Science, Technology, & Human Values_, 21(2): 131\u2013156.\n\nThomas, R.R (1999) _Detective Fiction and the Rise of Forensic Science_. New York: Cambridge University Press.\n\nTrovillo, P.V. (1939) 'A history of lie detection (concluded)', _Journal of Criminal Law and Criminology_ , 30(1): 104\u2013119.\n\nWeber, S. (2008) _The Hidden Truth: A Sociological History of Lie Detection_ , PhD Thesis, London School of Economics.\nSection 3\n\n* * *\n\n# Regulation, ethics and values\n\n* * *\nChapter 7\n\n* * *\n\n# **Through the thicket and across the divide**\n\nSuccessfully navigating the regulatory landscape in life sciences research\n\n_Graeme Laurie and Shawn H.E. Harmon_\n\n* * *\n\n## **Introduction**\n\nThe metaphor of the 'regulatory landscape' may be over-used, but its utility ensures its continuation. The metaphor is particularly apt for lawyers (working around the life sciences) because much of the responsibility for navigating the landscape is theirs. Within this landscape, law is often characterised \u2013 or caricatured \u2013 as both creator of problems and bringer of solutions. We argue that both accounts are true, but only in half-measure. While it is often said that law struggles to keep up with rapid advances in the life sciences, we focus on a particular aspect of this struggle, viz, it is not due to an absence of law but rather to the significant (and ever-expanding) space that law has come to occupy. The result is that law is often seen as a 'thicket': a complex, fragmented, and cumulative collection of instruments, institutions, and mechanisms that requires ever greater knowledge, time and capital to navigate, thereby imposing disproportionate costs on actors and requiring inordinate amounts of effort to move through.\n\nWe posit that a root cause of many of the problems associated with the regulatory thicket is the piecemeal and reactive ways in which law has emerged; it has grown rapidly under the guidance of those largely uninformed about the workings and trajectories of modern life sciences and has tended to operate in a compartmentalised\/atomised fashion \u2013 focusing on particular _objects_ of regulatory concern, and building frameworks around those objects, rather than taking the time and adopting a broader perspective to view technological trajectories holistically (cf., Metzler and Webster, 2011). Two particular regulatory objects have caught regulators' attention: human tissue, and personal data. In scientific terms, these might be seen as connected and indistinguishable in relation to their potential and value as sources of new knowledge, but in legal terms they have become subject to vast and disconnected structures of control involving local, national and international authorities and instruments, and a plethora of judicial decisions on how their collection, storage and use does or does not impinge on the legal rights of those from whom they are derived. In short, regulatory systems have grown up around tissue and data with little, if any, consideration of how these regulatory objects relate to each other or, indeed, to the interests that the systems are designed to protect.\n\nThis chapter argues that the unanticipated and worrying consequence of these processes is a net failure to deliver regulatory systems that work to protect citizens' rights and to promote scientific development (where such is warranted). The undesirable result is that, despite the complex of regulatory rules and standards, the risks regulated for persist, sometimes living in the many gaps between regulatory instruments, and new ones are generated. Focusing on the life sciences, and more specifically the (controversial) storage and use of newborn bloodspot collections, we explore the problem of unintelligent regulatory design. We advocate a re-visioning of design: one that is driven by appropriate objects of attention, that is proportionate, reflexive, value-based and interest-achieving, and that does not attempt to impose anticipatory control of science trajectories but rather provides an adaptive _regulatory continuum_.\n\n## **'Thickets' and 'divides' in life science regulation**\n\nAn examination of life sciences regulation discloses a general drift towards a 'decentred' approach (Black, 2001) whereby control\/influence is exercised by a diffuse collection of actors which includes but goes beyond the state. In short, despite injunctions to focus regulation, particularly state efforts, on specific problems or objectives (Sparrow, 2000) and to ensure that efforts are 'responsive' (Baldwin _et al._ , 2012), the life sciences are characterised by multiple authority-wielding actors bringing to bear different and often conflicting interests and values from perspectives that are frequently narrow, all of which get captured in elements of the landscape. Moreover, actors call on others to intervene (or contribute to the landscape) despite not fully appreciating the propriety of what might be delivered. The law is frequently called upon to so intervene, and it has frequently done so; regimes governing the life sciences have burgeoned in the last half century. The legal landscape relevant to human tissue is a good illustration of this growth.\n\nThe _Human Tissue Act 1961_ was 2.5 pages long and consisted of four sections. Its successor, the _Human Tissue Act 2004_ (HRA, 2004), has 61 sections (and seven Schedules) and is 25 times longer than the original Act. It established a regulatory system, including the establishment of a statutory governing authority, the regulation of consent, storage and use, provisions on property, and the articulation of criminal offences. Moreover, for those conducting research on human material, the HTA 2004 is but one of many legislative landmarks that sit at the centre of a mesmerising landscape of legislative provisions, professional guidelines, and ethical approval mechanisms. Those dealing with human reproductive material must navigate an entirely different legal landscape as articulated by the _Human Fertilisationand Embryology Act 2008_ (HFEA, 2008). This erects separate institutional and licensing arrangements despite the fact that, in practice, the distinction between reproductive\/non-reproductive materials is not always clear.\n\nFurther, research in the life sciences must begin and end with willing participants whose personal data will inform to greater and lesser extents the establishment of basic research resources or, indeed, downstream research and products. The processing of personal data is governed by a European regime which in the UK is implemented by the _Data Protection Act 1998_ (DPA, 1998). All forms of handling and storing personal data are caught by its provisions. While anonymised data are not caught by the regime, the processes of securing 'adequately anonymised' data are elusive, and all handling of data up to and including anonymisation are covered by the DPA 1998.\n\nVarious additional science-dependent considerations complicate the picture further. Research that is conducted with a view to delivering products or benefits to the market must satisfy various regimes as governed by the Medicines and Healthcare products Regulatory Agency (MHRA) and the European Medicines Agency (EMA). These can include conformity with the European Clinical Trials Directive (designed to establish safety and efficacy in medicinal products) and\/or the medical devices regimes (designed to test safety and performance). At various junctures, researchers will encounter and have to satisfy research ethics committees \u2013 of which there can be many, arising within each of the above regimes \u2013 as well as local, more _ad hoc_ arrangements that are driven by particular concerns (e.g., the protection of research participants' privacy). Finally, all of these regimes and the conduct that they are designed to regulate must be in conformity with human rights. In Europe, this regime stems from the Council of Europe's Convention on Human Rights (1950), as incorporated into UK law by the Human Rights Act 1998.\n\nThe increasingly expansive nature of law's participation in (if not dominion over) life sciences regulation has not been tempered by a well-informed and proportionate approach. Accordingly, the problem of the regulatory thicket is well-recognised in the research community, which has frequently lamented the regulatory burden under which it labours. Giving evidence to the House of Lords Science and Technology Committee on Genomic Medicine in 2009, Andrew Morris noted: 2006)\n\n> The Department of Health guidance suggests that this domain is affected by 43 relevant pieces of legislation. There were 12 sets of relevant standards and eight professional codes of conduct. What this has bred is a culture of caution, confusion, uncertainty and inconsistency.\n> \n> (HL, para 6.15)\n\nThe Academy of Medical Sciences (AMS) has twice reported on the issue, first in the context of using health information in medical research (AMS, 2006) and then in a wider call for more proportionate governance regimes (AMS, 2011). A common theme in these analyses is that confusion about law and the constraints and flexibilities that it allows is a recurring problem; a further cross-cutting theme is the call for a shift in regulatory culture as part of the solution. This has resonance with the recommendations of the Data Sharing Review (2008), which reported that:\n\n> We have found that in the vast majority of cases, the law itself does not provide a barrier to the sharing of personal data. However, the complexity of the law, amplified by a plethora of guidance, leaves those who may wish to share data in a fog of confusion.\n> \n> (Thomas and Walport, 2008: Foreword)\n\nIt was addressing the information governance environment only; in cases where downstream market considerations come into play, the complexities multiply.\n\nTo demonstrate how these thickets can evolve, and to exemplify the deeply unhelpful divides that are erected\/perpetuated, we consider one particular undertaking: the ambition to conduct research using newborn bloodspot collections, or so-called 'Guthrie cards'. Newborn screening programmes began in many western countries in the 1960s. They involve a heel-prick sample of blood from the infant that is retained on manual cards. Blood samples are initially taken in the health interests of the infant to diagnose and treat conditions that can be detected and eradicated or effectively managed at this early stage of life (e.g. phenylketonuria (PKU) or hypothyroidism). It was common to retain these cards, however, and their value has grown over the decades both with advances in genetic analysis and with the prospect of effective linkage to medical and other records (Couzin-Frankel, 2009).\n\nPerhaps unsurprisingly, the nature and balance of the interests implicated have changed, as have the social attitudes and the legal regimes involved. Whereas the idea of asking for consent was anathema in the 1960s, consent has ascended to a principal role in patient management and research regulation (Laurie, 2002). Equally, the advent of tissue and data protection regimes has changed the calculus (Lewis _et al._ , 2011; Hu, 2012); they raise important questions about the legal status of these collections, which contain both tangible (blood spots as tissue) and intangible (personal data) elements, and which have been recognised as de facto state-held DNA databases (Nuffield Council on Bioethics, 2007).\n\nA failure to engage families and publics around the existence and possible uses of such resources has resulted in numerous high-profile destructions of entire collections. For example, in _Beleno v. Texas Dept. of State Health Services_ (2009), the failure to obtain parental consent for the retention and use of blood spots was raised as an infringement of Constitutional rights (Drabiak-Syed, 2011). As part of the out-of-court settlement, the Department of Health ordered that four million samples be destroyed. In Canada, _LD and ED (Guardian Ad Litem) v. British Columbia Women's Hospital and Health Centre and British Columbia Children's Hospital_ (2010), was initiated for breach of privacy. Most recently, the Irish Department of Health and Health Service Executive announced plans to destroy its collection unless appropriate consent for retention was obtained ( _Irish Times, 7_ March 2013). This decision was revised shortly after, and the destruction of the collection was put on hold until further expert advice was gathered ( _Irish Times_ , 26 March 2013). The argument underlying all of these machinations is that data protection law does not permit retention and use without this specific consent.\n\nThere have, of course, been attempts to address the challenges that new uses and changing regulatory regimes create. But these initiatives all seem to face a common problem alluded to above: existing legal regimes have developed in a sector-specific manner. Thus, data protection and the common law of confidentiality govern data sharing and linkage, while the HTA 2004 and common law developments on property in the body dominate tissue research (Harmon and Laurie, 2010). These legally-created and embedded artificial distinctions are not only unhelpful but potentially obstructive to the core interests they purport to protect.\n\nLooking beyond law, human geographer Bronwyn Parry has argued that we need to collapse distinctions in order to explore questions about who controls 'bio-informational' transactions in the global economy (Parry, 2004). While Parry's concern is to understand the downstream intellectual property and benefit-sharing regulations that apply once transformations have taken place, we are more concerned with the logically prior and upstream question of regulating the 'raw' materials and data _before_ any commercial value is realised. The perspectives are inherently connected, however, because ultimately we must all consider the importance of a coherent regulatory continuum from basic research to delivery of benefit. Having said that, the upstream considerations are particularly important for two reasons. First, complete 'decontextualisation' of the materials and data has not yet taken place in that connections to persons remain, necessitating regulatory regimes that avert to the protection of their rights and interests in the handling of said materials. Second, serious limits or outright failures in regulatory regimes in the early stages of research and technological development will thwart efforts to realise any downstream benefits.\n\nA well-designed regulatory continuum can help facilitate the translation of knowledge to product but bridging the tissue\/data divide is critical. At present, legal frameworks and those working within them interact insufficiently and fail to see the bigger picture despite literatures on data sharing and tissue-use being replete with common governance challenges (e.g., propriety and expectations around consent, overly complex legal provisions and disproportionate regulatory burdens, concerns around public and research participant trust, and conflicting ambitions for public and participant engagement).\n\n## **The widening 'divide' in data and tissue**\n\nA crucial step in delivering appropriate and proportionate governance is the vertical and horizontal trimming of the regulatory thicket. The current data\/tissue regulatory divide is deeply unhelpful; not only can it lead to further cross-sector regulatory burden when the same artefact is caught by different regimes, but it can also lead to regulatory responses that are almost wholly divorced from the technological and scientific realities of dealing with tissues and data. From a biomedical perspective, tissue and data are simply different points on a continuum; tissue is less valuable for its tangible qualities than for the precious data it yields. Equally, data is important not necessarily in and of itself, but because of the aggregate value that derives turning raw data into _information._ Law fails completely to appreciate this relative value and its implications.\n\nConsider the example of the European data protection regime, currently under review. The focus for regulatory attention is the processing of 'personal data'. This is defined as 'data which relate to a living individual who can be identified \u2013 (a) from those data, or (b) from those data and other information which is in the possession of, or is likely to come into the possession of, the data controller' (DPA, 1998, sl(1)(e)). The details of the regulatory requirements are immaterial for present purposes; rather, three observations can be made.\n\nFirst, the regime is premised on a semblance of continuing connection between data being processed and the person to whom they relate. This is so irrespective of the nature or degree of any risks (such as to privacy) to the said person; moreover, it is blind to the motive or interest of the data controller. In essence, it treats the snooper in the same way as the researcher who is only interested in the aggregate value of information about a cohort of persons rather than any one individual (although see the limited provisions of the 'research exemption' discussed below).\n\nSecond, the definitions imply that if individuals are not identifiable from the data then the regime does not apply (i.e., anonymised data are not caught). This, however, raises two questions: What level of anonymisation is sufficient, given that such processes are more art than science? What is the position if identifiable data _are_ required for the scientific robustness of research? This is an on-going issue and the subject of considerable uncertainty and variation in practice. It has most recently been addressed in the UK by a Code of Practice from the Information Commissioner's Office (ICO, 2012). The Code provides recommendations to minimise risks but leaves a margin of appreciation with respect to what counts as _effective_ anonymisation. It does not have legal force and its reach is limited to practices within the UK. In one key respect, however, it provides assistance in conducting risk assessments with respect to anonymisation practices: it recommends considering privacy threats relative to the 'motivated intruder', that is, what would be the likelihood of success if someone sought to use anonymised data to re-identify individuals when they were a reasonably competent actor with access to public resources and who would make enquiries of third parties who could provide additional information? The test does not assume specialised knowledge or skills (e.g., hackers), nor access to specialist equipment, nor resort to criminality. Thus, the threshold is set midway between the 'relatively inexpert' member of the public and the 'skilled specialist with prior knowledge'. The test operates by way of a benchmark against which data controllers can test the robustness of their anonymisation practices, noting importantly that motives, threats and risks can change over time and therefore policies and practices must be kept under regular review \u2013 adding to the administrative and regulatory burden of policing the use of anonymised data.\n\nThird, as to research uses of identifiable data, the data protection regime provides a limited research exemption that can apply to data obtained for one purpose when later used for a research purpose so long as two crucial criteria are met:\n\n> (i)... the data must not be processed to support measures or decisions with respect to particular individuals, and (ii) the data must not be processed in such a way that substantial damage or substantial distress is, or is likely to be, caused to any data subject.\n> \n> (DPA, 1998, s. 33)\n\nEffectively anonymised data clearly meet these requirements. As to identifiable data, the justification must be made. If applicable, the consequence is that data can be retained indefinitely (normally data must be destroyed after original purposes for processing are met), and data subjects need not be granted access to their data (otherwise a norm in the regulations) _so long as_ research results are not published in an identifiable form. This is an attempt to relieve the regulatory burden of managing access requests, but this has been undermined by the fact that the Code suggests that granting such access is good practice (2012: 46). Moreover, the reduction in burden might be slight because while consent to research uses is not mandated, data subjects must still have adequate notice of the fact that data are being used for research. And if consent is not obtained and research data cannot be published in an effectively anonymised form, then subject access must be granted and the researcher is exposed to an action for breach of data protection (unless it could be shown that there was no alternative but to publish the results in an identifiable form).\n\nThe result is that research involving personal data continues to be constrained by the 'consent or anonymise' paradigm that has been so heavily criticised by the research community (AMS, 2006). This might be compounded further if current proposals to develop a European Data Protection Regulation are adopted in their current form. A January 2013 report from a European Parliament Committee advocated that research involving health data should only be conducted with data subject consent, which should be 'freely-given, specific, informed and explicit' (EP, 2013: 198). If this is not possible then research would only be permitted if it served 'an exceptionally high public interest' and data must then be 'anonymised or at least pseudonymised using the highest technical standards' (EP, 2013: 198). This trajectory for regulation is disconcerting because it suggests a widening, not a closing, of the gap between law and scientific practice contrary to accumulated experience. The slow pace and political charge of European legislative change are notorious, and the time to worry for data protection and health-related research has not yet come, but this phenomenon speaks to the deeper issues that concern us in this chapter.\n\nA further example from data protection \u2013 this time from the European Court of Human Rights (ECtHR) \u2013 indicates that a proportionate and common-sense approach is not prevailing, making the imperative for change all the stronger. _S and Marper v. United Kingdom_ (2008) concerned the lawfulness of the blanket policy in England and Wales to retain indefinitely and without consent DNA profiles and samples taken from persons arrested for (but never convicted of) criminal offences, however minor. It was held that such an indiscriminate policy was a breach of human rights that could not be justified. In the wake of the decision, the law was changed to embody a more proportionate approach involving time limits for retention (three years usually), but with the possibility of longer retention depending on whether there was a prior record of serious offending and\/or the seriousness of the instant offence (Protection of Freedoms Act 2012, Chapter 1). The reforms will result in around 1 million samples being destroyed, slated to be completed by May 2013 (Hansard, 2012).\n\nThe significance of _Marper_ lies in two key findings of the ECtHR. First, mere retention of DNA \u2013 even without use \u2013 is an interference with the human right to respect for private life because of the possible implications that future uses can have for individuals _(Marper_ , paras 67\u201377). This can be justified in the public interest so long as any interference meets a social need and is necessary and proportionate. Second, the samples, DNA profiles and any information derived from them constitute 'personal data' within the data protection regime ( _Marper_ , para 67), a fact previously accepted by the UK government. These findings can be contrasted with 2007 Opinion of the Article 29 Working Party which exists as an independent advisory body on matters of European data protection:\n\n> Human tissue samples (like a blood sample) are themselves sources out of which biometric data are extracted, but they are not biometric data themselves (as for instance a pattern for fingerprints is biometric data, but the finger itself is not). Therefore the extraction of information from the samples is collection of personal data, to which the rules of the Directive apply. The collection, storage and use of tissue samples themselves may be subject to separate sets of rules.\n> \n> (Article 29 Working Party, 2007: 9)\n\nHerein lies the rub. One European body suggests that DNA samples are 'personal data' in their own right while another advises contrarily and claims that separate rules might apply. Consequently, there is legal uncertainty about the status of tissue samples with respect to whether they are caught by the data protection regimes while, at the same time, there are additional legal regimes being applied to human tissue alone. The implications are potentially profound. On either perspective we have growth of the regulatory thicket. Either a cautious approach will be adopted to the regulation of health research involving data and tissues \u2013 in which case both regimes will be considered to apply \u2013 or the divide between data and tissues will be considered self-evident, in which case separate regimes will evolve that necessarily have overlapping effect.\n\nWe do not argue that data protection ought or ought not to apply to human tissues. Rather, we make a plea to develop and implement regulatory regimes that adopt a functional approach both with respect to the protection of core human rights and interests as well as the likely cumulative effect of their operation on the scientific endeavour. From a scientific perspective the data\/tissue dichotomy is a false one. The distinction can be collapsed with little consequence. Can we achieve this view in law?\n\n## **Collapsing the 'divide' between data and tissue**\n\nAlthough the regulatory objects in these diverse regimes are data and tissue, the core concern of the law \u2013 quite rightly \u2013 is the perceived link to the individual and the protection of his or her rights and interests. Seen in this way, regulatory regimes can be found to have much in common. For example, each regime (tissue and data) is far less onerous if anonymisation is achievable (i.e., if the link to the individual is broken). When the link remains, there is a tendency to place the individual at the centre of regulatory propriety (i.e., individual consent is needed to legitimise dealings with the tissue and data). We have argued elsewhere that consent is neither necessary nor sufficient to protect individual concerns in the health-research context (Laurie and Postan, 2012).\n\nMoreover, the tendency in law to fix the 'consent moment' by the use of the now ubiquitous consent form is dangerous and limiting for many reasons, not least that it can only represent a snapshot of expectations and responsibilities. This ignores the fact that expectations and responsibilities can and will change over time. To the extent that prior consent says anything about an individual's future interests, it can have an ossifying effect on their protection, foregoing the opportunity to develop dialogue and partnership with research participants. Where consent has a role to play, then, it should be seen as a continuing process necessitating genuine interaction throughout the life of the project (Laurie and Postan, 2012).\n\nAs we argued above, though, consent is not always possible or appropriate. Accordingly, it is here that law probably fails most to take a proportionate view of what is at stake. If the function of law is to protect core rights and interests, then what _arc_ the legitimate residual interests that individuals have in their tissue and data once these find their way into a research setting? Privacy is undoubtedly a key concern, yet many concerns around privacy can be addressed through adequate security and access arrangements. Such measures can reduce the risks involved, as well as serve to screen for future inappropriate motives in access and use.\n\nNotwithstanding these technical responses\/solutions, some have argued that the importance given to individual autonomy should deliver a degree of continuing control over data and samples to originators of that tissue or data (Beyleveld and Histed, 2000). For example, one might find it objectionable if one's tissue or data were used to further research objectives that one finds offensive or immoral. While the existence of affront might be real, it is not obvious what the legitimate _legal_ interest is that would be compromised. It is not privacy, which is about intrusion into a state of separateness from others (Laurie, 2002), nor is it bodily integrity since samples are no longer part of the individual and the data never were. But the control metaphor looms large because we tend to talk about both samples and data as if they were extensions of ourselves. The trend toward autonomy reinforces this, most particularly in the fetishisation of consent; moreover, there are some judicial mutterings about property in the body which reinforce these intuitional perspectives (Harmon and Laurie, 2010).\n\nWhere does this leave law? The consent-based approach to research regulation has strength in its focus on the importance of the individual and in its strong suggestion of an element of control. There is, in fact, an elegant simplicity in such an approach. But much is lost as a result. First, the power that consent provides is largely the power to say 'No' (i.e., to refuse). Second, it is not clear that consent can ever operate alone or in a primary role to protect research participants' interests, and the role of consumer protection laws is a helpful analogy here. Third, the simplicity of the solution leads us to ignore\/overlook much that is at stake especially in terms of the overall interests in play.\n\nWe contend that the appropriate regulatory ambition for law is to _correctly_ identify and facilitate the core interests and values at stake, not just private interests (of research participants) but also public interests (such as promoting scientifically sound and ethically robust research). This demands a clearer and more sustained commitment to the wider range of underlying values that promote this broad spectrum of public and private interests (Harmon, 2010; Harmon, 2011; Laurie _et al._ , 2012; Nuffield Council, 2012). Here, an important but often un-vindicated value is that of solidarity, which encourages us to adopt a more social or communitarian stance and to consider what duties might flow from the individual in support of valuable undertakings aimed at the collective (Harmon, 2006; Nuffield Council, 2011).\n\nAccordingly, we advocate a functional approach to regulation in this area that is constructed around both the nature and scope of interests at stake, and which is calibrated by a consideration of the relative risks to said interests. For example, if the core continuing interest that individuals have in research using tissue or data derived from them is privacy, then an appropriate regulatory response would be to address security and access. If research can be demonstrated to have a reasonable prospect of promoting public interests it should be supported. Yet, the trend towards the non-negotiability of consent militates against this. Ironically, it does not guarantee that core interests will be better protected and it runs the considerable risk that public interests will be thwarted. This is not to suggest that consent never has a role to play. Rather, we propose that consent must be deployed as one in a set of regulatory options relating to the range of interests and objectives at stake. Paying due attention to the importance that many people place on consent is an important element in fostering trust in research. Equally, however, trust in that enterprise is dependent on the research delivering on its promises, and this can and does take commitment and investment over time.\n\nGiven this reality, governance regimes must be adaptive and reflexive; that is, capable of picking up on signals about shifting parameters, including participant expectations and scientific possibilities, and responding to these quickly and effectively (Laurie, 2011; Harmon _et al._ , 2013). A reflexive system of governance is based on constant interaction between stakeholders, including participants and publics, and, importantly, it is not entirely dependent on law. At best, law can provide structural architectures within which reflexive governance can be delivered, but it is a crude tool to provide what is needed on the ground and on a case-by-case basis (Laurie, 2013). It further suggests the need for multi-disciplinary design in regulatory regimes, at both the conceptual and the operational stages. A first step in this is collapsing any false distinctions such as that between tissue and data in health-related research.\n\n## **Crossing the 'divide': newborn screening programmes redux**\n\nAs indicated above, Guthrie card collections are a prime example of the challenges that face contemporary biomedical research. The hurdles to (research) success are compounded by the legal regimes that have grown up in an _ad hoc_ manner in the decades since their establishment. While some have turned their attention to the need for appropriate guidance (Botkin _et al._ , 2013; Chrysler _et al._ , 2011), there has been far less consideration of the lessons to be learnt for law. The long-term nature of the collection poses considerable dilemmas about whether and how it is appropriate or possible to impose contemporary governance requirements \u2013 such as consent \u2013 on collections that were established at a very different time (and in this regard we note the literature on museums and the politics of display: Lohman and Goodnow, 2006).\n\nThe value of the resource has also changed over time and will continue to do so, as will the nature of any interests people have in samples and data, facts which also complicate their appropriate regulation. Arguably, with respect to the tissue originators, direct health interests diminish and are replaced by core privacy interests. On our analysis, if the research value of the collection can be demonstrated and if adequate security and access provisions are put in place, this interest is met. In such a case, must consent still be sought (assuming it is at all practicable to do so)? There is no reason to suggest that consent from parents of present and future newborns should not be sought for inclusion in the resource since this is the dominant contemporary ethical paradigm. But whether retrospective consent should be sought or even contemplated is another matter. The logistics aside, we posit that on-going governance of such resources need not and should not place consent as its central focus.\n\nAnd what is the legal status of such collections in light of the argument above? The decision in _Marper_ suggests that data protection applies, but the tangible nature of the collection will also be caught by tissue legislation where this exists. Accordingly, in England and Wales both the DPA 1998 and the HTA 2004 must be observed. In Scotland, the collection of 4 million samples is treated as part of the medical record (i.e., for the informational value that it represents). Arguably, there is no need for an additional layer of legal regulation simply because tangible cards\/spots are involved (and we acknowledge the literature on materiality as exemplified by Faulkner _et al._ , 2012, and the special issue it introduces). To require consent represents an example of disproportionate governance relative to the risks and interests at stake. Far more importantly for the future and for the effective protection and promotion of both private and public interests, is the need for dynamic governance arrangements that both engage publics about the existence and value of such resources and deliver transparent and adaptive mechanisms for their management (Douglas _et al._ , 2012). Arguments for the primacy of consent are not self-evident, either practically (given the scale and age of the resources) or as a matter of principle (it is unlikely to deliver a balance of all interests).\n\nAs a robust alternative to consent, there is the role of independent oversight bodies (Rothwell _et al._ , 2011), perhaps with lay representation as appropriate (albeit this raises important questions about what is meant by both 'lay' and appropriateness of representation). Denmark and other European countries have adopted this model, charging a suitably-constituted body with the stewardship of the resource (N\u00f8rgaard-Pedersen and Hougaard, 2007). We support such an approach and suggest further that central to its operation must be a commitment to the twin principles of openness and reflexivity.\n\n## **Conclusion**\n\nAfter much deliberation, the UK government announced that it will not abolish the Human Tissue Authority or the Human Fertilisation and Embryo Authority as had been proposed. Rather, it will conduct bespoke reviews of the functions of these regulatory entities with a view to reducing regulatory burden (Department of Health, 2013, para 69). This is only the UK context and we must take into account European and international regulatory pressures; yet, we strongly suggest that any such reviews in the near or longer-term future must consider both the vertical and horizontal branches of the regulatory thicket. We advocate an approach that is cognizant of the full range of underlying values and interests at stake in modern health research and which embraces and reflects key features that deliver both responsiveness and proportionality in regulation. Legal architectures that embody a command and control approach are unlikely to reflect these dimensions. This is a plea to recognise the limits of law and a call to deliver intelligent regulatory design at the nexus of science and law that reflects scientific realities and changing understandings of what counts as social goods.\n\n For example, the Health Research Authority was created in 2011 with a specific remit to: 'create a unified approval process and to promote proportionate standards for compliance and inspection within a consistent national system of research governance' (HRA website). We have also developed a regime in the Wellcome Trust Scottish Health Informatics Programme that adopts a principles-based approach to the sharing and linkage of health-related data and that delivers a proportionate approval scheme that matches applications to link and shares data to appropriate governance pathways based on a risk-based approach (Laurie and Sethi, 2013).\n\n We have advocated this elsewhere in the context of data linkage governance (Laurie and Sethi, 2013).\n\n## **References**\n\nAcademy of Medical Sciences(2006) _Personal Data for Public Good: Using Health Information in Medical Research_ , London: AMC.\n\nAcademy of Medical Sciences(2011) _A New Pathway for the Regulation and Governance of Health Research_ , London: AMS.\n\nArticle 29 Data Protection Working Party, Opinion 15\/2011 on the Definition of Consent (2011). _Available at_ : http:\/\/ec.europa.eu\/justice\/policies\/privacy\/docs\/wpdocs\/2011\/wp187_en.pdf (accessed 12 April 2013).\n\nBaldwin, R., Cave, M. and Lodge, M. (2012) _Understanding Regulation_ , 2nd edn., Oxford: Oxford University Press.\n\nBeyleveld, D. and Histed, E., (2000) 'Betrayal of confidence in the Court of Appeal', _Medical Law International_ , 4: 277\u2013311.\n\nBlack, J. (2001) 'Decentring regulation: the role of regulation and self-regulation in a \"Post Regulatory\" world', _Current Legal Problems_ , 54: 103\u2013146.\n\nBotkin, J., Goldenberg, A., Rothwell, E., Anderson, R. and Lewis, M. (2013) 'Retention and research use of residual newborn screening bloodspots', _Pediatrics_ , 131: 120\u2013127.\n\nChrysler, D., McGee, H., Bach, J., Goldman, E. and Jacobson, P. (2011) 'The Michigan BioTrust for Health: using dried bloodspots for research to benefit the community while respecting the individual', _Journal of Law and Medical Ethics_ , 39 Supp 1: 98\u2013101.\n\nCouzin-Frankel, J. (2009) 'Newborn blood collections: science gold mine, ethical minefield', _Science_ , 324: 166\u20138.\n\nDepartment of Health (2013) _Government response to the consultation on proposals to transfer functions from the Human Fertilisation and Embryology Authority and the Human Tissue Authority_ , 25 January2013.\n\nDouglas, C., van El, C., Faulkner, A. and Cornel, M. (2012) 'Governing biological material at the intersection of care and research: the use of dried blood spots for biobanking', _Croatian Medical Journal_ , 53: 390\u20137.\n\nDrabiak-Syed, K. (2011) 'Legal regulation of banking newborn blood spots for research: how Bearder and Beleno resolved the question of consent', _Houston Journal of Health Law & Policy_, 11: 1\u201346.\n\nEuropean Parliament (Committee on Civil Liberties, Justice and Home Affairs: Rapporteur Jan Philipp Albrecht) (2012) ' _Draft report on the proposal for a regulation of the European Parliament and of the Council on the protection of individual with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation)_ ', 17 December2012.\n\nFaulkner, A., Lange, B. and Lawless, C. (2012) 'Introduction: material worlds: intersections of law, science, technology and society', _Journal of Law & Society_, 39: 1\u201319.\n\nHarmon, S. (2006) 'Solidarity: a (new) ethic for global health policy', _Health Care Analysis_ , 14: 215\u201336.\n\nHarmon, S. (2010) 'Regulation of stem cell and regenerative science: stakeholder ppinions, plurality and actor space in the Argentine social\/science setting', _Law, Innovation & Technology_, 2: 95\u2013114.\n\nHarmon, S. (2011) 'Ambition and ambivalence: encouraging a science culture in Argentina through engagement and regulatory reform', _Studies in Ethics, Law & Technology_, 5: 1\u201326.\n\nHarmon, S. and Laurie, G. (2010) 'Yearworth v. North Bristol NHS Trust: property, principles, precedents and paradigms', _Cambridge Law Journal_ , 69: 476\u201393.\n\nHarmon, S., Laurie, G. and Haddow, G. (2013) 'Governing risk, engaging publics and engendering trust: new horizons for law and social science?', _Science & Public Policy_, 40: 25\u201333.\n\nHu, S. (2012) 'Minnesota Supreme Court hears whether the Genetic Privacy Act protects newborn blood spot samples obtained under the state's newborn screening statutes \u2013 _Bearder v. State of Minnesota', American Journal of Law & Medicine_, 38:225\u20137.\n\nIrish Times, 'Calls for \"priceless resource\" of heel-prick test results to be saved', 7 March2013. _Available at_ : www.irishtimes.com\/newspaper\/health\/2013\/0205\/1224329653693.html.\n\nIrish Times, 'Reilly does U-turn, order HSE not to destroy newborn blood samples', 26 March2013. Available at: www.irishtimes.com\/news\/health\/reilly-does-u-turn-ordering-hse-not-to-destroy-newborn-blood-samples-1.1338666.\n\nInformation Commissioner's Office(2011) _Data Sharing Code of Practice_ , Cheshire: ICO.\n\nLaurie, G. (2002) _Genetic Privacy_ , Cambridge: Cambridge University Press.\n\nLaurie, G. (2002) 'Better to hesitate at the threshold of compulsion: PKU testing and the concept of family autonomy in Eire', _Journal of Medical Ethics_ , 28: 136\u20137, doi:10.1136\/jme.28.3.136.\n\nLaurie, G. (2011) 'Reflexive governance in biobanking: on the value of policy led approaches and the need to recognise the limits of law', _Human Genetics_ , 130: 347\u2013356.\n\nLaurie, G. (2013) 'Governing the spaces in-between: law and legitimacy in new health technologies', in Flear, M. L., Farrell, A., Hervey, T. K. and Murphy, T. (eds) (2013) _European Law and New Health Technologies_ , Oxford: Oxford University Press.\n\nLaurie, G., Harmon, S. and Arzuaga, F. (2012) 'Foresighting futures: law, new technologies, and the challenges of regulating for uncertainty', _Law, Innovation & Technology_, 4: 1\u201333.\n\nLaurie, G. and Postan, E. (2012) 'Rhetoric or reality: what is the legal status of the consent form in health-related research?', _Medical Law Review_ , 1\u201344, doi: 10.1093\/medlaw\/fws031.\n\nLaurie, G. and Sethi, N. (2013) 'Towards principles-based approaches to governance of health-related research using personal data', _European Journal of Risk Regulation_ , 1: 43\u201357.\n\nLewis, M., Goldenberg, A., Anderson, R., Rothwell, E. and Botkin, J. (2011) 'Blood samples: state laws regarding the retention and use of residual newborn screening blood samples', _Pediatrics_ , 127: 703\u201312.\n\nLohman, J. and Goodnow, K. (eds) (2006) _Human Remains and Museum Practice_ , London: UNESCO.\n\nMetzler, I. and Webster, A. (2011) 'Bio-objects and their boundaries: governing matters at the intersection of society, politics, and science', _Croatian Medical Journal_ , 52: 648\u201350.\n\nN\u00f8rgaard-Pedersen, B. and Hougaard, D. (2007) 'Storage policies and use of the Danish Newborn Screening Biobank', _Journal of Metabolic Disease_ , 30: 530\u20136.\n\nNuffield Council on Bioethics(2007) _Forensic Uses of Bioinformation: Ethical Issues_ , London: NCB.\n\nNuffield Council on Bioethics(2011) _Solidarity: Reflections on an Emerging Concept in Bioethics_ , London: NCB.\n\nNuffield Council on Bioethics(2012) _Emerging Biotechnologies: Technology, Choice and the Public Good_ , London: NCB.\n\nParry, B. (2004) _Trading the Genome: Investigating the Commodification of Bio-information_ , New York: Columbia University Press.\n\nRothwell, E., Anderson, R., Burbank, M., Goldenberg A., Lewis, M., Stark, L., Wong, B. and Botkin, J. (2011) 'Concerns of newborn blood screening advisory committee members regarding storage and use of residual newborn screening blood spots', _American Journal of Public Health_ , 101: 2111\u201316.\n\n_S and Marper v. United Kingdom_ (2009) 48 EHRR 50.\n\nSalter, B. and Faulkner, A. (2011) 'State strategies of governance in biomedical innovation: aligning conceptual approaches for understanding \"Rising Powers\" in the global context', _Global Health_ , 7: 1\u201314.\n\nSparrow, M. (2000) _The Regulatory Craft_ , Washington, DC: Brookings.\n\nTaylor of Holbeach (Lord), _Statement: Protection of Freedoms Act 2012_ , Hansard, 13 December2012: Column WS161.\n\nThomas, R. and Walport, M. (2008) _Data Sharing Review Report_ , at www.connectingforhealth.nhs.uk\/systemsandservices\/infogov\/links\/datasharingreview.pdf\nChapter 8\n\n* * *\n\n# **Misconduct hunting**\n\nResearch integrity via law, science and technology\n\n_Marie-Andr\u00e9e Jacob_\n\n* * *\n\nToday many elite scientists and experts from different parts of the world are part of virtual global networks, sharing more with one another than with many colleagues from their own countries (Haas 1992; Boulton _et al._ 2012; Riles 2011; Nielsen 2012). Actors within elite science receive similar education, read and write in the same journals, attend the same conferences, obey the same courtesy rules (Lam 2010: 15); in other words, they share a culture (Strathern 2000). A notable element of this shared culture is well-disseminated guidelines about 'good research practice'. There is a lot to be said about what 'good research practice' has come to mean in the last few decades, and why it has been so minutely articulated in some contexts and left purposefully vague in others. In fact these 'good research practice' norms are currently being renegotiated, and hence should not be taken for granted (Biagioli 2003, 2012; Jacob 2013). For instance, there have been calls for expanding the definitions of research misconduct comprised of Fabrication, Falsification and Plagiarism (FFP) to include self-plagiarism, ghost authorship, non-disclosure of conflicts of interest, amongst others (Farthing and Wells 2008; Goodstein 2010). Montgomery and Oliver (2009) have traced three 'institutional logics' mobilised to deal with deviance in science in the United States. They note that prior to 1975, the discourse was about norms and counter-norms of 'normal practice of science'. It then moved, between 1975 and 1990, to a focus on the prevention of scientific misconduct; and from the 1990s to the present, to promoting research integrity.\n\nIn the last couple of years in the UK, various consultation papers and reports have been lamenting that the regulation of research is too burdensome, bureaucratic, and expensive. There have been calls from the government to cut the bureaucracy, with prominent actors and institutions advocating a 'radical simplification' (DoH 2010: 9) of the legal frameworks regulating research. The Department of Health 'Report on the arm's-length bodies review' (2010) \u2013 which followed the publication of the White Paper 'Equity and excellence: Liberating the NHS' (2010) \u2013 has set out proposals to make the health sector more efficient and less bureaucratic by increasing accountability and transparency and reducing the number and cost of quangos. Amongst other things, the Report mandated the Academy of Medical Sciences to review current regulatory frameworks and plan a new pathway for regulating research. The Academy's subsequent 2011 report suggested increasing the speed of decisionmaking, reducing complexity, and eliminating unnecessary bureaucracy and costs (see also Laurie and Harmon, this volume). In this context \u2013 where bureaucracy almost became a dirty word, and where traditional modes of state regulation by law are seen as inefficient and expensive \u2013 self-regulation through the use of science and technologies takes on special relevance (Lessig 1999; Black 2001). But when it comes to regimes of governance, do legal tools and motifs ever recede in the background?\n\nIn contrast to the efforts of those interested in preventing misconduct, this chapter will instead examine how the idea of 'research integrity' itself has been translated into a regulatory agenda. What interests me is how 'legal' and 'technological' tools that govern contemporary science act to differentiate research 'integrity' from 'misconduct,' and hunt for misconduct. I shall further argue that these research integrity tools cannot, in fact, be neatly classified as either legal or technological. By drawing parallels between legal practices and peer deliberations by scientists, I also show how the specificity of the 'legal' and the 'scientific' is bound to get blurred. This has consequences for our own theoretical (or disciplinary) perspectives regarding the study of research integrity.\n\nSTS scholar Mario Biagioli suggests that norms within science, such as norms on scientific authorship, act as 'a para-legal discourse predicated on the absence (perhaps the impossibility) of a Law' (Biagioli 2007: 147). A first objective of this chapter is to modulate this claim, by showing _how_ methods of regulation and moderation are not only used as a substitute for a law in this area but also how, in the area of research integrity, these methods themselves mobilise typically legal resources and modes of enunciation. Drawing upon ethnographic fieldwork in the Committee on Publication Ethics (COPE) over 2010\u20132014, I suggest that the work of participants in present-day research integrity agendas resonates with an old kinship shared between legal and scientific practices.\n\nThe regulatory agenda of scientific research integrity is characterised by striking similarities between the practices of regulatees and that of regulators. This phenomenon is distinct, and arguably far more intense, than that of 'regulatory capture' (Braithwaite and Ayres 1992) as it implies instead the sharing of the very same activities, tools, documentation practices and apprentices, between both parties to the regulatory encounter. Hence a second objective of this chapter is to explore the implications of one such example of simultaneous sharing of practices between regulators and regulatees, coming from the domain of the regulation of scientific conduct.\n\nMy third objective is to compare past and current regulatory strategies to hunt scientific misconduct, and reconsider what sets them apart. In particular, the chapter seeks to provoke a re-thinking of the difference between so-called 'old' legal tools, and so-called 'new' scientific-technological regulation (Brownsword 2005; Lessig 1999) that have been applied to the governance of scientific work. Using the example of moderation of nineteenth-century medical publishing, I deflate the novelty (and hype) of regulation by technologies, and then reflect back on the evolving practices of COPE.\n\nLaw and society scholars Silbey and Ewick (2003) have declared that the new 'authenticating' site of science, where scientific truth is lodged, is no longer the lab, but the text, the publication itself. The milieu of scientific publication (and of publication ethics) has also become the site where some of the most interesting regulatory and self-regulatory initiatives within science take place. Before turning to this regulatory archipelago (Rose and Miller 2008: 218) that moderates the conduct of researchers, and hunts instances of misconduct, let me briefly contextualize my chapter within the movements that inspire this collection.\n\n## **Engaging socio-legal studies and STS**\n\nSTS and socio-legal scholars are increasingly exchanging their tools, methods and concepts, in order to study their respective objects of inquiry (Valverde 2005; Lezaun 2006; Cloatre and Dingwall 2013). The scrutiny of STS scholarship has opened some of the black boxes of law and regulation, and cast fresh light on socio-legal domains more generally (e.g. Prainsack, Turkmendag, Shapiro, this volume). Socio-legal scholars, on their part, have effectively unpacked various ways of 'making order' in science. 'Legal-pluralist' analyses that decentre 'enforcement by the state' (Braithwaite 1993; Black 2001) have made socio-legal approaches eminently relevant to study normative orders beyond state law, and tackle regulatory webs that involve all sorts of actors, including scientific ones.\n\nPerhaps because of their trained eye and sensitivity to tailored legalities, lawyers and socio-legal scholars are well situated to study multiple forms of 'disputing' (Abel 1973) within science. For example, their acquaintance with texts that 'tell us what to do' (Constable 2008), may turn out to be useful when it comes to unpacking standards (Bush 2012) and other forms of 'regulatory objectivity' (Cambrosio _et al._ 2006) that proliferate within, and increasingly typify, biomedicine. Their familiarity with matters of statecraft and legislative modes of enunciation may help to tackle the power and authority of the government in managing dissent, consensus-making and reordering in science.\n\nPerhaps most importantly, what the joint streams of STS and socio-legal works have shown us is that if one 'attempts to capture the dynamics of knowledge processes', rather than labelling knowledges statically or categorically as either scientific, or legal, or social, 'one sees new things' (Valverde 2005: 421). The cross-fertilization between the two fields gives hope for less of pre-conceived separation of units and more open-ended inquiries into what law and science can mean.\n\nIn the following pages, I use my object of inquiry to foreground a found intimacy between scientific and legal reasoning. Misconduct hunting, as we will see, is an evocative interplay between law, science and society.\n\n## **'Witnessing' and deliberating science**\n\nI start by sharing some initial thoughts about innovative methods of deliberation used by the Committee on Publication Ethics (COPE), an organisation in which I have been conducting ethnographic observations on a quarterly basis since 2010. I will show that the deliberations of these participants can be analogised with forms of experimental work performed in laboratories (cf. Bogner 2012), and that through their ways of 'witnessing' science, these participants juggle scientific but also long-standing legal traditions in their deliberations.\n\nA number of organisations are involved in preventing misconduct, many of them groups of editors, including the International Committee of Medical Journal Editors (ICMJE), the World Association of Medical Editors (WAME), and the European Association of Science Editors (EASE). One of the most notable is certainly COPE, an international organisation which started in April 1997 as an informal group of medical editors convened by Michael Farthing (then editor of _Gut_ , and now Vice Chancellor of the University of Sussex), Richard Smith (then editor of the _British Medical Journal_ ), and Richard Horton (of _The Lancet_ ). At that first meeting, about 15 medical editors gathered together at the British Medical Association (BMA) House in London: 'We discussed cases, and I think that we found it interesting and all learnt something. It was a very informal atmosphere, and we laughed and had fun', recalls Smith (Horton _et al._ 2012: 5). They kept meeting regularly to 'tell each other stories' (Horton _et al._ 2012: 5) but soon feared being perceived as a 'kangaroo court' (Horton _et al._ 2012: 5) and being sued for libel and slander. Following the advice of eminent professor of medical law Ian Kennedy, they began to anonymise all the cases under discussion, and made clear that they 'weren't making decisions [...] simply offering advice' (Horton _et al._ 2012: 5).\n\nInitially a local group, today COPE is a large international network (though still physically operating from the UK). It is well known and consulted by the scientific and publishing community across the English-speaking world. Its Code of Conduct, and flowcharts 'designed to help editors follow COPE's Code of Conduct and implement its advice when faced with cases of suspected misconduct', have been translated into many languages including Spanish, Portuguese, and Persian. A truly cosmopolitan organisation, COPE has Council members from Europe, Australia, Brazil, China, and Iran, and an ombudsman from Australia, who is flown over to London quarterly for Council meetings.\n\nCOPE is funded by annual subscriptions from a very wide range of publishers and individual journals, and works on the basis of membership (as of 2012, it had more than 7,000 members). Originally, journals would become members individually, but now publishers like Blackwell and Elsevier have registered all their journals, and therefore some journal editors might now be officially members of this organisation without knowing about it. Indeed, I myself became an Associate Member (available for those not working as editor but who have an interest in publication ethics) of COPE, for _\u00a3_ 50 per year, in order to gain better access to its materials and benefit from reduced rates to attend its seminars.\n\nIn 2012, COPE received an award from the Council of Science Editors (CSE) for its work on improving scientific communication and promoting high editorial standards. The award was displayed to all participants at a COPE meeting. The CSE award tells something about how COPE sees itself and is seen by others in the field: as a crucial actor in the promotion of good practice in science. The prize is also indicative of the increasing professional-isation of the field of misconduct studies, as Biagioli has discussed. (Biagioli 2012: 467). Further to this, Richard Horton has recently said that COPE 'has become a national treasure' (Horton _et al._ 2012: 7). This might also indicate the implicit and under-examined Britishness of COPE (in spite of its international membership).\n\nCOPE's distinctive activity is its Forum (where my on-going ethnographic fieldwork is taking place). COPE members call it a 'self-help' group for journal editors. The Forum is a place where suspicions of questionable conduct are discussed openly, and where participants share their experiences and offer suggestions of responses and advice as to what to do next: 'It's an open forum for the sharing of views, we don't decide or judge', as one member explained to me. The Forum meets quarterly \u2013 on the day of the administrative meetings of the Council \u2013 to discuss cases submitted to them by journal editors. Members of the Forum include Council members of COPE (the Council includes medical writers, journal editors, publishing consultants, academics, one lawyer), but anyone who has editorial functions for a journal member of COPE, or who works for a publisher member of COPE, can also come and contribute to the exchange in the Forum.\n\nAny member can submit a case to COPE in advance of the Forum meeting, after which the case is anonymised, summarised in writing and circulated to all members prior to the Forum. At the Forum itself, the person who submitted the case presents orally, in person or by phone or videoconference, to the participants, and then participants openly discuss the case, share their own experience with similar cases, raise questions, share views, and debate -but, as the Chair pointed out to me, 'without rendering a judgment'. The open-endedness is further highlighted by the format with which COPE publishes the cases: 'Anonymised Text of the Case submitted'; 'Advice'; 'Follow-Up'; sometimes an 'Advice on the Follow-up'; and 'Resolution'.\n\n_Figure 1.1_ Committee on Publication Ethics (COPE) Forum\n\nSource: Jeannie Wurz\n\n'COPE is an experiment', wrote Michael Farthing in his first annual report (1998). Drawing on the seminal works of Shapin and Schaffer (1985), and Shapiro (2002), I now show how COPE can be understood as a regulatory space, and a laboratory.\n\nIn _Leviathan and the Air-Pump_ , Shapin and Schaffer show how, in the seventeenth century, Robert Boyle and his colleagues from the Royal Society formulated ways of conducting and reporting experiments that propelled 'not only new regimes of fact-finding but also new regimes of governance' (Jasanoff 2007: 22). The 'witnessing' of science practised through peer-review, for instance, helped to democratize science, and gradually imposed itself hand in hand with other democratic means of deliberation that would relieve citizens from the absolutisms of the monarchy. Shapiro (2002) points out that within the Royal Society, during this transformation of the conduct of experiments in the laboratory, long-standing legal conventions of legal witnessing and fact-finding were integrated in scientific enterprise and reasoning. Experiments and specific observations were debated within the language of law. Boyle himself often used legal terminology, for example arguing that 'matters of fact had to be brought to trial' (Shapiro 2002: 243). Shapin and Schaffer describe how he defined the laboratory:\n\n> In Boyle's programme there was to be a special space in which experimental natural philosophy was done, in which experiments were performed and witnessed. This was the nascent _laboratory_.\n> \n> (Shapin and Schaffer 1985: 334, their emphasis)\n\nShapin and Schaffer also draw a connection between experimental knowledge, and legal knowledge produced in criminal court, in that both are based on 'collective witnessing'. In law and science 'the reliability of testimony depended upon its multiplicity' (Shapin and Schaffer 1985: 334).\n\nThey explain further:\n\n> The thrust of the legal analogy should not be missed. It was not merely that one was multiplying authority by multiplying witnesses (although this was part of the tactic): it was that _right action_ could be taken, as seen to be taken, on the basis of these collective testimonies.\n> \n> (Shapin and Schaffer 1985: 56, their emphasis)\n\nThe Royal Society debated 'matters of fact' not only with the language of truth, but also of moral certainty. Members were interested in normative questions about thresholds of credibility themselves. Latour (1993), discussing Boyle's use of legal metaphors (Shapin 1984), describes the scene almost ethnographically: 'credible, trustworthy, well-to-do witnesses gathered at the scene of the action can attest to the existence of a fact, the matter of fact, even if they do not know its true nature' (Latour 1993: 18).\n\nI suggest that the COPE Forum similarly uses a legally-influenced mode of 'witnessing', as well as deliberation, as tools of regulation. The Forum appears to constitute a kind of informal dispute resolution arena, but also, a laboratory, an experimental space, a 'place where this multiplicity of interests [is] addressed, acquitted, and drawn together' (Shapin and Schaffer 1985: 340).\n\nCirca 2010\u20132014, members of the COPE Forum came to examine 'matters of fact' and matters of conduct in specific contexts of disputes. They try to do both: determine what is, what has happened, but also how to characterise (or categorise) what is now to be done. They ask what is and what has happened by following the plot closely, asking follow-up questions about chronology, dates of exchanges, specific locations and jurisdictions. For example, members are interested in knowing where the research takes place, where the researchers come from (this alludes to the controversy over whether different scientific standards of conduct might apply to different countries; see Jacob 2013). COPE members also use analogies with previous cases to feed into and shape their discussion. In the course of case discussions, the first question the Chair addresses the Forum with is, recurrently: 'has anyone experienced something similar before?'\n\nThe Royal Society was commonly confronted with conflicting testimony: gentlemen could conflict in the Royal Society 'as they often did in the course of lawsuits' (Shapiro 2002: 56). Members and spokesmen often compared the work of the Royal Society to the work of the court. However, in practice they carefully 'avoided making collective judgments about matters of fact' (Shapiro 2002: 56). Specific scientists may have asked the Royal Society to rigorously 'investigate, examine, compare and also to declare judgment' (Shapiro 2000: 126) in cases of disputes and controversies. However, the Society preferred to suspend judgment, instead making statements about the instruments used by the disputed scientists, their propensity or not as to 'willingly falsify an observation' (Shapiro 2002: 257). This suggests that the Society wished to restrict their mandate to examining the facts of methods and the facts of conduct of scientific experiments, but _not_ scientific merit _per se_.\n\nLikewise, COPE members are explicit about their wish not to take the role of an adjudicator: as one member observed to me, 'We don't judge, it's very open-ended'. In the deliberations, issues of scientific merit are not discussed. What is key is the facts of conduct and of potential misconduct of scientists, not the worth of the science. Donald Kennedy (1997) points to that distinction as a feature of the tension between legal and scientific approaches to scientific conduct in the context of adjudication of scientific fraud allegations. In the context of such inquiries, he notes, the subject matter itself never gets disputed. It is the sequence of events, the order of manuscript drafts (or of email exchanges), and the minute details of records that come under scrutiny. Kennedy alleges that these are the kind of things that scientists absorbed with the technical content (of their science) might miss (or misinterpret); however they would hardly be missed by lawyers (Kennedy 1997: 232; see also Latour 2002).\n\nI interpret the COPE Forum as a contemporary practical case that makes Boyle's thoughts and experiment look suddenly much less part of a distant past. COPE emphasises that it is precisely because there are many views expressed, and many testimonies, and experiences shared, that it can offer the right advice (they are careful to state that this is not a _judgment_ ). Moreover, in the COPE Forum, it seems that there is an epistemic affinity between research and the mode of moderating it. Here it is deliberation, discussion of a case on the basis of the sharing of multiple accounts by multiple witnesses: an old legal technique, in renewed form? In other words, the scientific, experimental method proposed by Boyle and which, according to Shapin and Schaffer, can be analogised to natural philosophy and to criminal law, is re-performed in the Forum, to control and moderate deviations in science.\n\nUntil recently, the COPE Forum had been meeting face to face around a large oval table, in a boardroom in central London, inviting editors to present their cases in person, and also taking cases via phone from editors working all over the world. In March 2013 COPE held its very first virtual Forum meeting via webinar. The head of COPE, Virginia Barbour from _PLoS Medicine_ , chaired the Forum from an office in Cambridge, with the technical assistance of COPE staff. For two hours, 49 participants from different parts of the world, presumably sitting in front of a computer, headsets on, read cases and guidelines on screen as pointed to by the Chair, and listened to editors presenting their cases. Some took a more active role, by 'raising their hand' with one-click button in order to technically be 'unmuted' by the moderator and to share their thoughts to the group via microphone. I was conducting my ethnographic observation sitting at a desk in my university office in Keele. The meeting seemed to have gone well and similarly to previous meetings in many ways, in the sense that opinions and disagreements were aired politely and firmly, in the usual manner. Parts of the flow of the Forum's collective discussion, including spontaneous reactions, the occasional joke and ensuing collective laughter, the raised eyebrows and head-shaking that also characterise the Forum table, were for sure clogged by the media of the webinar. From COPE's perspective, this meeting was a success, and made the Forum more accessible to a broader range of international members. COPE expressed that its intention to hold both virtual and face-to-face Forums from now on.\n\nPeople make new social media in their own image, and not the other way around; yet, I cannot help but wonder: what will such virtual meetings mean for the future of this type of legally-influenced, 'Royal Society style' face-to-face deliberation about scientific research conduct? How will the 'conduct of conduct of conduct' (cf. Rose 2000) be reshaped in the process?\n\n## **Regulating (with) science and technologies**\n\nWhat happens when science and technology are _simultaneously_ (rather than alternatively) tools and targets? Can the governance of scientists' conduct gain legitimacy, become more workable and efficient, precisely because it is grounded on science and technology (Miller and Rose 2008: 218)? Or: can control, regulation and moderation of research conduct be effective because, and perhaps only because, it is effectuated by a method of moderation that is grounded on, and mimics research itself?\n\nTechnologies of research integrity verification, such as electronic tracking, textual similarity algorithms and statistical analysis of submissions, are now well integrated in the machinery of scientific publishing. Based on the premise that technology can detect, prevent, repair and undo the mistakes (see Prainsack, this volume) but also the misconduct and improprieties people make, these technologies and the body of quantitative studies that has developed on and with them (e.g. Fanelli 2010) receive a lot of attention in research integrity circles.\n\n'Technologies of integrity verification' exemplify what lawyers Brownsword and Yeung, borrowing from STS, call regulatory 'technological fixes' (2008), in that they make use of knowledge and technology to regulate and thus cure the problem of research integrity. Technoscience can now alternatively be a _tool_ as much as a _target_ of regulation (Brownsword and Yeung 2008; see also Flear and Pickersgill 2013; Cloatre and Dingwall 2013). Framed as tools to detect fraud in research, technologies of integrity verification entail what Biagioli calls a 'depersonalisation and delocalisation of the evidence of misconduct' (Biagioli 2012: 468\u2013469): instead of focusing on the grievance of a particular author who has been plagiarised or otherwise wronged by misconduct, these technologies prevent misconduct at large by disciplining all scientists (see also Lewis and Atkinson 2011). In other words, these regulatory technologies aim at upholding an ethos, rather than preventing or compensating for harm. This is consistent with other contexts of professional self-regulation, as it emphasises the requirement that professional values (of integrity and ethics) be _shown_ (Jacob 2012: 7).\n\nOne example of a technology of verification is statistics. It is currently a standard procedure for the most important medical journals to have a statistician on their editorial board, and to have any papers with numerical figures reviewed by statisticians. A statistician also sits on the board of the UK Research Integrity Office. Statisticians are able to detect data fabrication and falsification, on the basis of the concept of digit preference (Evans 2008). Digit preference theory asserts that particular numbers are preferred to be recorded or chosen, rather than having a uniform distribution. It is thus difficult to invent, or alter data without leaving 'fingerprints' (Evans 2008: 164). Statistics can also be useful to verify data integrity in the case of multi-centres treatment trials: here, the statistician can compare the data sets and to detect anomalous distribution of data.\n\nTechnologies of integrity verification also seem to have 'computerized the detective's role' (Biagioli 2012: 468). Perhaps more dramatically, in a collection edited by prominent figures in the UK milieu of medical research integrity, statistician Evans describes his own discipline as an 'omniscient adjudicator' so convincing that 'no corroboration is needed' (Evans 2008: 161). Statistics, as a discipline, is thus being self-promulgated by its practitioners as a judge. This judge behaves secretly, disinclined to reveal its precise methods of detection. The rationale for the vagueness found in writings about statistical methods to detect fraud in medical research, claims Evans, is that it is important to prevent fraudsters becoming more sophisticated and thus able to circumvent detection methods; when there is a code, there is a code breaker (Machado and Prainsack 2012: 57\u201372).\n\nTextual similarities software is another mode of integrity verification. Focusing on the large-scale quantitative detection of textual similarity, these tools search for textually similar articles, whether published by different authors, or by the same authors. Hence these electronic tools are used to detect plagiarism, and also 'self-plagiarism' (i.e., 'redundant' or duplicate publications). Because there are far more instances of duplicate publications than of plagiarised articles, the engineers and proponents of textual similarity algorithms significantly increase their markets and audiences by mixing the 'apples of plagiarism with the oranges of self-plagiarism' (Biagioli 2012: 467) without pausing and asking what these forms of deviance have in common and what sets them apart.\n\nThe idea of self-plagiarism is noteworthy in two key senses. First, the term itself indicates again this notion that misconduct self-regulation is not about protecting others from harm, but about a scientific ethos and, here more particularly perhaps, about the relationship of a scientist to herself. The category 'self-plagiarism' suggests that a scientist's respect for professional standards is a matter for regulatory remit. Second, with self-plagiarism, we might wonder whether the regulator, at least in part, might be technologically constructing its own regulatee and inflating its own regulatory target, hence creating the need for more technological fixes (cf. Biagioli 2012).\n\n## **Guarding against the quacks**\n\nIn spite of the alluring novelty of integrity verification entrepreneurship, nineteenth-century medical publishing practices suggest that the use of science and technology to regulate knowledge production is no recent innovation. In this sense it is helpful to examine medical journals' past responses to the phenomenon of quackery. Throughout history, the quack has often been referred to as cunning, and as a confidence trickster, fraudster, and faker (Bartrip 1995: 191; Porter 2003: 313). It was every practitioner's responsibility to guard the gullible public against it (Porter 2003: 312; see also Warner 2005), just as today scientists are responsible for making sure that people are not deceived by poor science (Biagioli 2012: 465).\n\nThere were dilemmas among medical reformers of the nineteenth century as to whether to suppress or eradicate quackery, or try to educate the quacks (Porter 2003: 309). The editor of _The Lancet_ , Thomas Wakley, accused the medical establishment itself of being the source of quackary, and physician Thomas Percival had taken a tough approach towards quackery in his influential 1803 _Medical Ethics_. A petition to parliament was presented to outlaw unqualified practice, to protect legally qualified practitioners and suppress quacks (Porter 2003: 309). The British Medical Association's work against quack medicine included campaigning for legislation forcing manufacturers to reveal ingredients of secret remedies, the restriction of sales, prosecution of those in breach, as well as public relations and persuasion of mainstream newspapers not to publish advertisements for 'imposters'. Practitioners, politicians, corporations and medical editors allied together to propel legislative changes, but also felt there was a need to move beyond the law.\n\nCampaigns against quackery related to the 'politicisation of medicine' (Porter 1989: 222). The late nineteenth and early twentieth century was a critical period for the professionalisation of medicine, and medical innovators and the 'old guard' of the traditional elite were fighting for control over the provision of medical services and cures (Bynum and Porter 1986). This was also a period of growth for medical journals, which, for financial reasons, included plenty of advertisements of quack medicines. Bartrip (1995) reports that in 1904 Dawson Williams, then head of the BMA, commissioned an analysis of proprietary medicines by reputable pharmacist Edward Harrison, in order to identify and reveal the ingredients of these medicines. Harrison's results, along with data on the price of these medicines were published by Williams as a series in the _British Medical Journal_ in 1904\u20131908, thus uncovering that a range of medicines with little ingredients of value were sold at high prices. Bartrip notes how great an achievement this had been for the _British Medical Journal_.\n\nOne constant between twenty-first-century technologies of research integrity verification and nineteenth-century pharmaceutical science's detection of counterfeit medicines is that science and technology themselves are, and were, being used in self-regulation settings, to regulate and monitor the behaviour of scientists. The alleged reasons for doing so are, in both settings, to protect the public, and (hence) maintain public trust. More salient here is the epistemic affinity between the surveillance of research and the research itself, between the methods of controlling and the conduct to be controlled (cf. Rose and Miller 2008). Through such internal audit, science also succeeds in keeping positive law at bay.\n\nAccording to recent fieldwork by Lewis and Atkinson, laboratories have already become sites where 'technologies of surveillance, measurement and monitoring' (Lewis and Atkinson 2011: 382) make visible the methods (as opposed to the object) of scientific work. Could we imagine operational means \u2013 and the ideological and legal set of mind \u2013 that would allow us to prevent fraud in research from happening altogether? In the name of security, integrity and so forth, regulation envisaged by 'technologies of integrity verification' could eventually get beyond the paperwork, audits, and checks on the reported material of the scientists, and immerse itself into the interstices of the scientific work itself. For instance, technologies could regulate science _ex ante_ (Kerr 2013) by preventing human interventions, in certain components of the research. Or alternatively, as is already happening in the field of bioinformatics (for purposes of protection of intellectual property), there could be engines that trace everything the scientist does in the lab, instead of relying on the human scientist to record _post-facto_ what she does aside from her work, on a log sheet. This form of built-in regulation could even look like what Larry Lessig (1999) refers to as regulation through 'code,' meaning that the scientists might eventually not know that their lab work is being regulated in such ways. What would this all mean for regulation? But more intriguingly, what would this mean for science and for scientists as critical human agents (cf. Rouvroy 2013)? What would science look like, stripped of the _possibilities_ of human misconduct and fraud?\n\n## **Conclusion**\n\nAs we saw above, deliberations amongst the citizens of 'the republic of science' (Jasanoff 1997: 93) as well as old legal techniques of collective decision-making, take on surprising, rejuvenated forms in COPE in London, circa 2010\u20132012. Taken together, the profile of the seventeenth-century Royal Society's and the COPE Forum's respective members, rules of procedure, sponsors\/benefactors\/funders, physical setting and geographical location (London), elicit a unique, almost inescapable parallel between the two organisations.\n\nIn this context it seemed crucial to highlight the presence of good old legal tools and legal knowledge, found rejuvenated, in innovative forms, in the area of research governance and management of scientific misconduct. 'Witnessing' and deliberation do seem to survive and to show their sustained, renewed relevance. Legal knowledge and modes of enquiry can be found in innovative sites (Jacob 2011), for example in the COPE Forum, and hence are still very much pertinent to the governance of research conduct. This means that in spite of high-tech strategies deployed to govern scientific integrity, scientists and misconduct hunters perhaps do not quite live yet in the 'post-regulatory world' (Black 2001: 103).\n\nThe nature of COPE's work for research integrity is itself changing, in that it increasingly works with technology, and is supplemented by a panoply of high-tech integrity verification tools. These tools in turn resonate with older technological responses to deviance in medicine. Many dimensions of the multivalent work of misconduct hunting have yet to be unpacked. I have only highlighted some aspects here, namely, the sharing of practices between the regulator and its regulatees, the kinship between 'scientific' and 'legal' means of regulation, and the commonalities between 'traditional' versus 'novel' engagement with deviance.\n\nThe discourses of 'research integrity' keep gaining momentum, and getting more and more professionalised. Studying research integrity will thus require the foregrounding of knowledge itself, without letting the mapping of that knowledge's provenance ('is this legal? or scientific? or is this 'just' social?') pre-establish our grid of analysis (cf. Latour 1993), and also without letting single perspectives, and perspectivism (Lavi 2011: 815), condense our understanding of the multiple incarnations research integrity takes.\n\n## **Acknowledgements**\n\nThis chapter draws on research conducted under the AHRC Fellowship AH\/J008338\/1, as well as on ethnographic fieldwork conducted in the Committee on Publication Ethics, with support from the Centre for Law Ethics and Society at Keele University. Approval has been received from Keele Ethics Review Panel in October 2010. The author would like to thank the editors Emilie Cloatre and Martyn Pickersgill, as well as Martin Wasik, Barbara Prainsack, Hyo Yoon Kang, Vaios Karavas, Andrew Webster and Ginny Barbour for their helpful comments. Earlier versions have benefited from comments by participants at numerous events including at Universit\u00e4t Luzern, University of Nottingham Rhetorics of Moderation ESRC Seminar, Birkbeck College of Law, and the workshop 'Regulating Health Technologies' (supported by the Wellcome Trust and held at the Royal Society of Medicine, London, 12\u201313 May 2011).\n\n 'Good research practice' norms aim to practically implement values pertaining to research integrity, such as honesty, transparency and fairness in research. 'Good research practice' refers to norms regarding, for instance, acknowledgement of authorship, data storage and data sharing, disclosure of conflicts of interests, or the use of images in scientific publications, to name a few (Jacob 2013).\n\n These norms, proposed by sociologist Robert Merton (1979), were: communalism, universalism, disinterestedness, originality and scepticism, and the counter norms: solitariness, particularism, interestedness, and dogmatism.\n\n _Report of the UK Research Integrity Futures Working Group_ , chaired by Janet Finch, September 2010; _Liberating the NHS: Report on the Arm's Length Bodies Review_ , July 2010. _The AMS Report, A New Pathway for the Regulation and Governance of Health Research_ , January 2011.\n\n Committee on Publication Ethics Flowcharts: h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00adp\u00adu\u00adb\u00adl\u00adi\u00adc\u00ada\u00adt\u00adi\u00ado\u00adn\u00ade\u00adt\u00adh\u00adi\u00adc\u00ads\u00ad.o\u00adr\u00adg\u00ad\/r\u00ade\u00ads\u00ado\u00adu\u00adr\u00adc\u00ade\u00ads\u00ad\/\u00adf\u00adl\u00ado\u00adw\u00adc\u00adh\u00ada\u00adr\u00adt\u00ads\u00ad\n\n See Committee on Publication Ethics h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00adp\u00adu\u00adb\u00adl\u00adi\u00adc\u00ada\u00adt\u00adi\u00ado\u00adn\u00ade\u00adt\u00adh\u00adi\u00adc\u00ads\u00ad.o\u00adr\u00adg\u00ad\/\u00ada\u00adb\u00ado\u00adu\u00adt\u00ad\/h\u00adi\u00ads\u00adt\u00ado\u00adr\u00ady\u00ad (accessed 10 January 2013).\n\n For instance, the General Medical Council's codes of practice and guidance evoke the importance of maintaining a good standard of practice and are in order to justify the trust placed upon registered doctors.\n\n I thank Hyo Yoon Kang for drawing my attention to this fact.\n\n## **References**\n\nAbel, R. (1973) 'A comparative theory of dispute institutions ', _Law and Society Review_ , 8(2): 217\u2013347.\n\nBartrip, P. (1995) 'Secret remedies, medical ethics, and the finances of the British Medical Journal ', in R. Baker (ed.), _The Codification of Medical Morality: Historical and Philosophical Studies of the Formalization of Western Medical Morality in Eighteenth and Nineteenth Centuries_ , Dordrecht: Kluwer.\n\nBiagioli, M. (2003) 'Rights or rewards? changing frameworks of scientific authorship', in M. Biagioli and P. Gallison, _Scientific Authorship: Credit and Intellectual Property in Science_ , New York: Routledge, pp. 253\u2013280.\n\nBiagioli, M. (2007) 'Documents of documents: scientists names and scientific claims ', in A. Riles (ed.), _Documents: Artifacts of Modern Knowledge_ , Ann Arbor: Michigan University Press.\n\nBiagioli, M. (2012) 'Recycling texts or stealing time? plagiarism, authorship, and credit in science ', _International Journal of Cultural Property_ , 19(3): 453\u2013476.\n\nBlack, J. (2001) 'Decentring regulation: understanding the role of regulation and self-regulation in a \"post-regulatory\" world ', _Current Legal Problems_ , 54(1): 103\u2013147.\n\nBogner, A. (2012) 'The paradox of participation experiments ', _Science, Technology & Human Values_, 37(5): 506\u2013527.\n\nBoulton, G.et al. (2012) _Science as an Open Enterprise_ , London: The Royal Society.\n\nBraithwaite, J. (1993) 'Transnational regulation of the pharmaceutical industry', _Annals of Academy of Political Sciences_ , 525: 12\u201330.\n\nBraithwaite, J. and Ayres, I. (1992) _Responsive Regulation: Transcending the Deregulation Debate_ , Oxford: Oxford University Press.\n\nBrownsword, R (2005) 'Code, control, and choice: why east is east and west is west ', _Legal Studies_ , 25(1): 1\u201321.\n\nBrownsword, R and Yeung, K. (2008) 'Regulating technologies: tools, targets and thematics ', in R Brownsword and K. Yeung (eds), _Regulating Technologies: Legal Futures, Regulatory Frames and Technological Fixes_ , Oxford: Hart, pp. 3\u201322.\n\nBush, L. (2012) _Standards: Recipes for Reality_ , Cambridge, MA: MIT Press.\n\nBynum, W. and Porter, R. (1986) _Medical Fringe and Medical Orthodoxy, 1750\u20131850_ , London: Routledge.\n\nCambrosio, A., Keating, P., Schlich, T. and Weisz, G. (2006) 'Regulatory Objectivity and the Generation and Management of Evidence in Medicine ', _Social Science and Medicine_ , 63(1): 189\u2013199.\n\nCloatre E, and Dingwall, R. (2013) '\"Embedded regulation:\" the migration of objects, scripts, and governance ', _Regulation and Governance_ , 7(3): 35\u201338.\n\nConstable, M. (2008) 'On the (legal) study methods of our time: Vico Redux ', _Chicago-Kent Law Review_ , 83(3): 1303\u20131332.\n\nEvans, S. (2008) 'Can statistical analysis reveal research misconduct? ', in F. Wells and M. Farthing (eds), _Fraud and Misconduct in Biomedical Research_ , London: The Royal Society of Medicine Press.\n\nFanelli, D. (2010) 'How many scientists fabricate and falsify research? A meta-analysis of survey data', _PLoS ONE_ 4 (5): e5738. doi:10.1371\/journal.pone.0005738.\n\nFarthing, M. and Well, F. (eds) (2008) _Fraud and Misconduct in Biomedical Research_ , 4th edn., London: Royal Society of Medicine.\n\nFlear, M.L. and Pickersgill, M.D. (2013) 'Regulatory or regulating publics? The European Unions regulation of emerging health technologies and citizen participation ', _Medical Law Review_ , 21(1): 39\u201370.\n\nGoodstein, D. (2010) _On Fact and Fraud, Cautionary Tales from the Front Lines of Science_ , Princeton, NL: Princeton University Press.\n\nHaas, P. (1992) 'Introduction: Epistemic communities and international policy coordination ', _International Organization_ , 46(1): 1\u201335.\n\nHorton, R, Farthing M. and Smith R (2012) 'In the beginning... a Q&A with COPEs founders ', _Ethical Editing_ , 4(1): 5\u20138.\n\nJacob, M.-A. (2011) 'Knowledge games, truthfulness, and organ transplants regulation ', _BioSocieties_ , 6(2): 243\u2013261.\n\nJacob, M.-A. (2012) _Matching Organs with Donors: Legality and Kinship in Transplants_ , Philadelphia, PA: University of Pennsylvania Press.\n\nJacob, M.-A. (2013) _Scientific Research Integrity: Background Paper_ , London: Nuffield Council on Bioethics.\n\nJasanoff, S. (1997) _Science at the Bar: Law, Science, and Technology in America_ , Cambridge, MA: Harvard University Press.\n\nJasanoff, S. (2007) _Designs on Nature: Science and Democracy in Europe and the United States_ , Princeton, NJ: Princeton University Press.\n\nKennedy, D. (1997) _Academic Duty_ , Cambridge, MA: Harvard University Press.\n\nKerr, I. (2013) 'Prediction, pre-emption, presumption: the path of law after the computational turn ', in M. Hildebrandt and K. de Vries (eds), _Privacy, Due Process and the Computational Turn: The Philosophy of Law Meets the Philosophy of Technology_ , London: Routledge, pp. 91\u2013120.\n\nLam, A. (2010) 'From \"ivory tower traditionalists\" to \"entrepreneurial scientists\"? academic scientists in fuzzy university-industry boundaries ', _Social Studies of Science_ , 40(2): 307\u2013340.\n\nLatour, B. (1993) _We Have Never Been Modern_ , trans. C. Porter, Cambridge, MA: Harvard University Press.\n\nLatour, B. (2002) _La fabrique du droit. une ethnographie du Conseil d'\u00c9tat_ , Paris: La d\u00e9couverte.\n\nLavi, S. (2011) 'Turning the tables on \"law and...\": a jurisprudential inquiry in contemporary legal theory ', _Cornell Law Review_ , 96(4): 811\u2013838.\n\nLessig, L. (1999) _Code and other Laws of Cyberspace_ , New York: Basic Books.\n\nLewis, J. and Atkinson, P. (2011) 'The surveillance of cellular scientists practice ', _BioSocieties_ , 6(3): 381\u2013400.\n\nLezaun, J. (2006) 'Creating a new object of government: making genetically modified organisms traceable ', _Social Studies of Science_ , 36(4): 499\u2013531.\n\nMachado, H. and Prainsack, B. (2012) _Tracing Technologies: Prisoners' Views in the Era of CSI_ , Farnham: Ashgate.\n\nMerton, R. (1979) _The Sociology of Science: Theoretical and Empirical Investigations_ , Chicago, IL: University of Chicago Press.\n\nMiller, P. and Rose, N. (2008) _Governing the Present: Administering Economic, Social and Personal Life_ , Cambridge: Polity.\n\nMontgomery, K. and Oliver, A. L. (2009) 'Shifts in guidelines for ethical scientific conduct how public and private organizations create and change norms of research integrity,' _Social Studies of Science_ , 39(1): 137\u2013155.\n\nNielsen, M. (2012) _Reinventing Discovery: The New Era of Networked Science_ , Princeton, NJ: Princeton University Press.\n\nPorter, R. (1989) _Health for Sale: Quackery in England, 1660\u20131850_ , Manchester: Manchester University Press.\n\nPorter, R. (2003) _Quacks: Fakers and Charlatans in Medicine_ , London: Tempus.\n\nRiles, A. (2011) _Collateral Knowledge: Legal Reasoning in the Global Financial Markets_ , Chicago, IL: University of Chicago Press.\n\nRose, N. (2000) 'Government and Control, ' _British Journal of Criminology_ , 40(2): 321\u2013339.\n\nRouvroy, A. (2013) 'The end(s) of critique: data behaviourism versus due process ', in M. Hildebrandt and K. de Vries (eds), _Privacy, Due Process and the Computational Turn: The Philosophy of Law Meets the Philosophy of Technology_ , London: Routledge, pp. 143\u2013168.\n\nShapin, S. (1984) 'Pump and circumstance: Robert Boyles literary technology ', _Social Studies of Science_ , 14(1): 481\u2013520.\n\nShapin, S. and Schaffer, S. (1985) _Leviathan and the Air-Pump: Hobbes, Boyle, and the Experimental Life_ , Princeton, NJ: Princeton University Press.\n\nShapiro, B.J. (2000) _A Culture of Fact: England 1550\u20131720_ , Ithaca, NY: Cornell University Press.\n\nShapiro, B.J. (2002) 'Testimony in seventeenth-century english natural philosophy: legal origins and early development ', _Studies in History and Philosophy of Science_ , 33(2): 243\u2013263.\n\nSilbey, S. and Ewick, P. (2003) 'The architecture of authority: the place of law in the space of science ', in A. Sarat, L. Douglas and M. M. Umphrey (eds), _The Place of Law_ , Ann Arbor: Michigan University Press, pp. 75\u2013108.\n\nStrathern, M. (2000) (ed.), _Audit Cultures_ , London: Routledge.\n\nUK Department of Health (2010) _Liberating the NHS: Report on the Arm's Length Bodies Review_. Available at: w\u00adw\u00adw\u00ad.g\u00ado\u00adv\u00ad.u\u00adk\u00ad\/\u00adg\u00ado\u00adv\u00ade\u00adr\u00adn\u00adm\u00ade\u00adn\u00adt\u00ad\/\u00adu\u00adp\u00adl\u00ado\u00ada\u00add\u00ads\u00ad\/\u00ads\u00ady\u00ads\u00adt\u00ade\u00adm\u00ad\/\u00adu\u00adp\u00adl\u00ado\u00ada\u00add\u00ads\u00ad\/\u00ada\u00adt\u00adt\u00ada\u00adc\u00adh\u00adm\u00ade\u00adn\u00adt\u00ad_\u00add\u00ada\u00adt\u00ada\u00ad\/\u00adf\u00adi\u00adl\u00ade\u00ad\/\u00ad1\u00ad5\u00ad2\u00ad0\u00ad1\u00ad6\u00ad\/\u00add\u00adh\u00ad_\u00ad1\u00ad1\u00ad8\u00ad0\u00ad5\u00ad3\u00ad.\u00adp\u00add\u00adf\u00ad.p\u00add\u00adf\u00ad.\n\nValverde, M. (2005) 'Authorizing the production of urban moral order: appellate courts and their knowledge games ', _Law & Society Review_, 39(2): 419\u2013456.\n\nWarner, M. (2005) _Publics and Counterpublics_ , Cambridge, MA: MIT Press.\nChapter 9\n\n* * *\n\n# **Financial derivatives and the challenge of performation**\n\nWhere contingency meets contestability\n\n_Donatella Alessandrini_\n\n* * *\n\n## **Introduction**\n\nThe attention financial derivatives have attracted with the 2007 crash has noticeably subsided as a result of the focus on austerity and deficit reduction which has come to dominate public discourse since 2010. However, concerns over the financial speculation they have originated and its effects on the real economy \u2013 that is, the domain within which goods and services are produced \u2013 have not disappeared. No longer in the limelight, derivatives nevertheless continue to populate the world of regulators with technical standards and rules aimed at preventing their speculatory 'excesses'. One important manifestation of such excess concerns the value they express: this became apparent in 2008 when the amount of outstanding derivatives worldwide exceeded USD 1.114 quadrillion, 22 times the GDP of the entire world (Matai, 2009), raising critical questions about their relationship with the sphere of the real economy.\n\nAt the risk of oversimplifying a very complex debate, two approaches have emerged in relation to the role of financial derivatives in the economy, and the possibility of curbing their excesses. One has characterised them as dangerous innovations, with nothing to contribute to the economy but the distortion (if not destruction) of 'real' values that has tremendous consequences for people's lives. This argument resonates with the position in the social sciences which has expressed concerns over the growth of financial markets generally and financial derivatives in particular (Strange, 1997; Lipuma and Lee, 2004). While important work has been carried out on financial capitalism and its violent effects (see Marazzi, 2010), my intention is to demonstrate how, rather than being parasitical and unproductive, financial derivatives act as a powerful technology able to affect current value-making processes. They do certain things to the value they are supposed to measure and it is important to take this process into account so as to appreciate both the way in which the real and the financial spheres of the economy are deeply entangled and the kind of intervention at their intersection which are possible and desirable.\n\nThe second approach sees derivatives playing a beneficial role in the economy, allowing firms and governments to hedge (i.e. protect) themselves against the risk of price volatility, but also carrying certain risks, particularly over-the counter (OTC) derivatives which are not traded on official exchanges. This is certainly the position underlying the regulatory approach, which recognises their role in providing continuity to global production while also attempting to limit the risks they present. Although regulation remains an important tool in reducing the level and type of speculation, my interest lies in demonstrating how (in a manner not too dissimilar from the previous approach) the regulatory response remains predicated on a conceptual separation between the real and the financial sphere of the economy which does not attend to the complex value-making processes to which financial derivatives contribute. Moreover, focusing on the productive power of derivatives allows us to see how regulation always exceeds the realm of formal rules and processes that aim to affect human behaviour (Cloatre and Dingwall, 2012: 7), in this case that of market participants. Regulation encompasses the wide 'range of socio-technical practices that shape institutions and subjectivity either deliberately or otherwise' (Pickersgill, 2013: 325), and, from this angle, financial derivatives emerge as a regulatory device able to shape economic values and activities.\n\nIn this chapter, I explore the technology of financial derivatives and analyse the modalities through which they participate in value-making processes; that is, the processes through which economic value is conceptualised and produced in today's (mainly Western) financialised economies, with a three-fold aim; first, to illustrate the contribution that the STS-inflected Social Studies of Finance (SSF) have made to our understanding of the complex workings of economic markets in general and financial innovations in particular. Drawing on STS insights into the impossibility of separating material objects and social practices, SSF have shown how social values are embedded in the tools and techniques that make up the supposedly neutral realm of finance. Second, to reflect on the limitations emerging from this approach when the question of how to act in the midst of such complexity comes to the fore: it is my contention that, while opening up many black boxes of financial practices, including those of financial derivatives, STS-inflected SSF have up until now refrained from thinking about how to 'format' markets otherwise; and finally, to point to the potential for collaboration with constructivist approaches in critical and feminist political economy that share a commitment to complexity, contingency and contestability, thereby bringing to the fore the question of how values (in)form action.\n\nThe chapter is organised as follows: the first section problematises current approaches to financial derivatives by showing how this technology has called into question the possibility of neatly distinguishing between a real and a financial sphere of the economy, and highlighted the difficulty inherent in disentangling real and financial values. In this respect, it contributes to the rich STS tradition of troubling long established boundaries both empirically and conceptually. The second section takes the financial uncertainty revealed by derivatives as its starting point for a consideration of the contribution that STS-inflected SSF have made to our understanding of how economic models, technologies and innovations do not simply mirror or represent an underlying economic reality but actively make or perform it. The third section brings the insights of critical and feminist political economy on the contingency and contestability of values to bear on the SSF, particularly in their performativity of economics manifestation, in thinking about desirable forms of intervention. I conclude by referring to an instance of such (always imperfect) intervention in the context of currency futures.\n\n## **The real and the financial life of derivatives**\n\nMuch has been written about financial derivatives from a variety of perspectives and long before the crisis erupted in 2007 (Maurer, 2002; Knorr Cetina and Preda, 2005; Bryan and Rafferty, 2006). I look at them from a particular angle, one which focuses on their relationship to the value they express. Indeed, the standard definition is that of contracts whose value is derived from an underlying asset (i.e., a commodity or a financial asset such as shares, interest rates and exchange rates, or an index). However, this definition does not capture the different nature, characteristics and functions of the various types of derivatives in circulation (such as futures, swaps, options, collaterised debt obligations, and so on). Derivatives are different things; for instance, whereas _futures_ are contracts according to which two parties commit to buy or sell a certain asset at a future date and at a specified price, _options_ confer the right to buy or sell a certain asset at a certain date and at a specified price. The variety of derivatives in circulation makes any comprehensive definition unsatisfying.\n\nWith this in mind, one way of understanding derivatives is to look at their function: what they _are_ might be best characterised through what they _do_. Historically, their realm has been that of commodities \u2013 in particular, agricultural commodities \u2013 and they have been important in ensuring continuity in the production cycle by allowing farmers, for instance, to hedge against price fluctuations. Alongside this hedging function, however, derivatives also carry a speculative role; this is because there will always be discrepancies at any time on the market between the demand and offer of hedging opportunities. In other words, these markets need speculators to take on the risks and this pertains to commodity as well as the more recent financial derivatives. Indeed, as Lipuma and Lee (2004: 20\u201321) have put it, although derivatives have existed for a long time, it is as a result of the collapse of the Bretton Woods monetary management system and the move towards floating exchange rates in the 1970s, of the lifting of capital controls, and of the proliferation of contractual outsourcing, that old risks such as counterparty and interest rates risks have increased and new ones such as currency and 'socio-political' risks have emerged.\n\nI want to pause on Lipuma and Lee's work, _Financial Derivatives and the Globalisation of Risks_ , because it is representative of the perspective in the social sciences that recognises the role of derivatives in providing the means to address price volatility in an increasingly uncertain and interconnected world. This perspective acknowledges that since exchange and interest rates have been made to fluctuate in the market, derivatives have become a crucial tool to manage the risks of global investing. However, it also sees derivatives as the chief instruments of a 'speculative' capital that remains distinct and distinguishable from industrial and commercial capital. In their words, 'technologically driven derivatives detach the value, cost and price of money \u2013 manifest in exchange and interest rates \u2013 from the fundamentals of the economy, particularly the state of production, the social welfare of the producers, and the political needs of citizens' (Lipuma and Lee, 2004: 2). However, this account of derivatives as instruments detaching 'the value, cost and price of money' from the so-called fundamentals of the economy requires further scrutiny as it relies on two problematic assumptions about derivatives that have been called into question by the reality of their trading.\n\nThe first assumption is that it is possible to effectively distinguish between hedging on the one hand and speculation on the other, so that by curtailing the latter (or at least its 'unnecessary' manifestations) derivatives can be kept in line. This is the reason why regulators distinguish between the speculation that is necessary for granting liquidity to these markets, and arbitrage, which is the practice of buying and selling to profit from changes in market prices. Thus, the argument often made is that it is necessary to curb the latter form of speculation so as to preserve the healthy purpose derivatives serve in hedging against risks. The problem with this position is that in actual markets the behavioural differences between the two types of market participants, hedgers and speculators, as well as between necessary speculation and arbitrage, are difficult to detect. Studies have not only found a range of reasons for (or types of) both hedging and speculation, but also that, at any point in time, participants are engaged in both activities (Bryan and Rafferty, 2006, 2002). It is therefore extremely difficult to distinguish between these different categories.\n\nThe second assumption concerns the standard definition of a derivative as a contract whose value is based on that of the underlying asset. The problem is that, however one measures it, the value of a derivative does not derive from the value of the underlying asset, but from the trend of that value in the market. This is important since the trend in market prices can be influenced by market players, which means that the mechanisms that determine the value of the derivative are not linear. This is not only a matter of individual players who influence the trend of the value of the 'underlier'; rather, it is their agreement, the derivative itself, that can have this effect. The point is that we are dealing with contracts in which to anticipate or speculate on the value of an asset is also to intervene in, influence or make that very value. The challenge this poses to regulation is enormous: introducing clarity and limiting the amount of speculation is certainly important. However, regulation cannot avoid, save the banning of derivatives altogether, the circular value-making process between the underlying asset and the derivative whereby the latter is in actual fact participating in the production of the value it is supposed to be merely based on.\n\nThe fact that the derivative makes the value of, as well as deriving it from, the underlying asset creates a complex circularity which raises important questions exceeding the regulatory realm. Investors refer to 'fundamental values' in their estimation of whether the underlying asset is to be considered worth investing in. For businesses, information includes revenues, assets, liabilities and growth. For governments it includes interest rates, budget deficit, the balance of trade account, inflation rates, the state of the banking and financial sector and consumer confidence, among other things. But if value is acknowledged as being produced at the intersection between the two spheres of the economy, what are we to make of the 'fundamental' values to which financial entities such as derivatives are supposed to be referring?\n\nExpectations have long been recognised to play a role in shaping the value of financial entities (see Marx, 1981: 597). However, what is becoming increasingly evident is that today's 'asset markets have a reflexive nature in that prices are generated by traders' expectations, but these expectations are formed on the basis of anticipations of _others'_ expectations' (Arthur _et al._ , 1996: 21). This is a completely different scenario from the one depicted by orthodox financial theory where stock prices are not supposed to greatly diverge from the fundamental value of the firm, that is the value 'equal to the discounted stream of future expected dividends': today's market prices depend to a great extent on agents' anticipation of others' expectations (Orlean, 2005: 3).\n\nThe self-referentiality thesis therefore points to the first degree of abstraction from 'fundamental values' operated within financial markets. This is why some scholars have argued that market participants should revert to an analysis based on fundamentals, i.e. one that grounds expectations in the economic situation of the firm issuing the assets (Graham and Dodd, 1940: 21\u201321 in Mackenzie, 2006: 76\u201377). However, the difficulty this position confronts is that not only are prices today so distorted, but firms' balance sheets and state accounting are so compromised that assessing the health of a company \u2013 let alone that of an economy (as in the case of exchange rates) \u2013 is extremely challenging. Moreover, financial derivatives add, through the circularity just described, another layer of difficulty in accounting for the 'real' or 'fundamental' value of an asset. This is yet more problematic when we consider that they are financial entities relating to other financial entities which have also undergone the first degree of abstraction at the level of expectations.\n\nTo sum up, rather than simply deriving their value from that of the underlying asset, derivatives participate in its actual construction. This points to the extreme difficulty of disentangling 'real' and 'financial' values, and therefore reveals a deep level of uncertainty in the determination of 'fundamental' values in today's financialised economies. It also poses a further question: does acknowledging such uncertainty imply we should give up altogether on 'fundamental values' as the anchor between the real and the financial spheres and take the market as the sole arbiter and maker of values? I do not think so: it is my contention that acknowledging contingency in value-making processes means neither ignoring the constraints within which such processes take place nor doing away with judgments about their desirability. In what follows, I draw on the theoretical resources offered by the performativity of economics thesis to show how thinking about contingency (within constraints) in value-making processes characterised by financial uncertainty is a necessary (albeit not sufficient) condition for thinking about desirable forms of intervention in such processes.\n\n## **Performing the economy**\n\nSTS-inflected research that came to be called the Social Studies of Finance (SSF) emerged in the 1990s, ostensibly motivated by a desire to move away from the emphasis of classical economic sociology on human action (as embedded in institutions, norms, relationships and structures) to explain economic activities. Scholars working within SSF turned their attention to the ways in which such action takes place through hybrids made of humans and technical and material devices; subsequently, this body of work has particularly focused on economics, arguing that economic models and theories actively shape and format \u2013 rather than merely describe \u2013 the economy (cf. Callon, 1998: 2). The _performativity thesis_ that has emerged since has traced the multifarious ways in which economic theories and models become a material force embodied in market practices, arrangements and infrastructures. Contrary to self-fulfilling prophecy developed by Merton in 1948 (1968: 477), performativity does not imply falsehood in the (human) beliefs that come to be performed. Performativity is concerned 'with the incorporation of economics in the infrastructures of markets' so that it exceeds beliefs and mindsets, however important these are. As Mackenzie puts it\n\n> an aspect of economics that is incorporated only into beliefs \"in the heads of economic actors\" may have a precarious status. A form of incorporation that is in some sense deeper is incorporation into algorithms, procedures, routines, and material devices. An economic model that is incorporated into these can have effects even if those who use them are skeptical of the model's virtues, unaware of its details, or even of its very existence.\n> \n> (Mackenzie, 2006: 19)\n\nTwo aspects of the performativity argument are worth reflecting on to illustrate the contribution that STS-inflected SSF have made to our understanding of the co-constitutive relationship between economics and the economy, theory and practice, and I would also say between the financial and the real spheres of the economy. The first is that performativity exceeds the sphere of human beliefs as the mere 'psychological' realm. It highlights the role of material devices with which the theory or model interacts. At the same time, and this is the second important aspect, it brings to light the constraints within which the model operates. Thus, to say that the Black-Scholes-Merton formula was performative, as Mackenzie does in a classic SSF story, 'is not to make the crude claim that any arbitrary formula for option prices, if proposed by sufficiently authoritative people, could have 'made itself true' by being adopted. Most such formulas could not do so, at least other than temporarily' (Mackenzie, 2006: 20). Even if initially adopted widely, the formula would soon have been abandoned had it led agents to consistently lose money or conduct arbitrage so as to gain from systematic price discrepancies.\n\nThe issue becomes one of understanding why certain formulas have durable consequences and this requires expanding as far as possible the radar of research so as to delve into the reality in which formula, models, and devices have become immersed. The concept of _agencement_ is meant to shed light on how markets take different forms in different places: ' _Agencements_ denote sociotechnical arrangements when they are considered from the point [of] view of their capacity to act and to give meaning to action' (Callon and Caliskan, 2005: 24\u201325 quoted in Mackenzie, 2009: 20\u201321). _Agencement_ therefore places the emphasis on the tools that enable performativity. Mackenzie for instance argues that the performativity of financial theory has become incorporated in financial markets in three specific ways. From a technical perspective, the growth of financial markets would have been impossible without the development of infrastructures, and without financial theory built in software so as to enable risk calculation and semi-automatic fast-trading. From a linguistic perspective, financial theory provided words and concepts with which to manage and reduce the complexity inherent in market transactions. For instance 'implied volatility' is an engineered concept that reduced the complexity inherent in, and made possible by, the pricing of options. Finally, from a regulatory or legitimising perspective, financial markets would not have developed without the active intervention of economists and regulators who 'purified' previously contentious practices and instruments from their association with gambling by making them efficient and rational realms of transactions (Latour, 1993; de Goede, 2005). Thus, by opening up to scrutiny the complex web of socio-material practices constituting markets, work on the performativity of economics points to contingency within constraints, material and technical as well as human.\n\nYet, the performativity thesis has not gone uncontested: if engaging socio-material practices always necessarily entails a partial account, however serious the attempt is to widen up the reality to account for, concerns have been expressed in relation to the kind of entanglements that have ended up being privileged. Thus Mirowski and Nik-Khah have lamented the fact that in their attempt to do away with structures, forces and intentionality, performativity scholars and STS-inflected SSF more generally have ended up concentrating on machine metaphors, theories and models in a way that has consistently diverted attention from other important processes. In _Do Economists Make Markets_?, a collection edited by Mackenzie, Muniesa and Siu (2007), Mirowski and Nik-Khah take issue with the account of the Federal Communication Commission (FCC) spectrum auctions which Callon, Muniesa and Guala provide as a clear instance of successful performativity (successful since economists' theoretical accounts of auctions shaped the format of the auctions adopted). Their argument is that concentrating on economists' theories has led these authors to ignore both the role of the government and that of telecommunication companies 'in orchestrating the outcome' of the auctions (Mirowski and Nik-Khah, 2007: 216\u2013217).\n\nAs they recognise, this 'partial' account might have depended on the fact that performativity scholars relied on economists' own accounts of the event (ibid., 202) rather than on a careful description of the phenomena. This would have required them to follow all 'actants', and not only the economists and their account of the 'successful' auction. Indeed, placing the emphasis on _agencement_ would have required opening up to scrutiny the 'sociotechnical arrangement', i.e. the FCC actions in its various articulations. Furthermore, I would argue, nothing prevents one from talking about powerful actors as long as that power is carefully traced in action (cf. Latour, 2005: 177). Similarly, there is nothing in the performativity thesis that prevents us from seeing that certain theories\/models cannot be made to work. Mackenzie has certainly acknowledged this point in relation to the Black-Scholes-Merton model when he has argued that not all formulas for pricing options could have worked.\n\nThis connects to the more general point emerging from their critique of the performativity school, and that is the fact that recognising contingency within constraints, however important, is not sufficient to start thinking about, and devising, ways to organise markets differently. What is crucial is to enter the debate about their desirability, and this means confronting economists on the reality they attempt to construct through their theories\/models. Interesting in this respect is the recent abandonment by neo-classical economists of their comparison of the price system to a natural mechanism, and the revelation of their ambition is to ' _fabricate_ markets, and not simply treat them as States of Nature' (Mirowski and Nik-Khah, 2007: 215\u2013216). In essence, they have admitted that markets are achieved and normative rather than neutral and natural, making more apparent the fact that the struggle has always been one over the reality which different actors attempt to construct. This is a point which Callon acknowledges in the concluding chapter to the edited collection where he proposes to explore different market formations as 'struggles of performation [...] between competing programs which make the disassembling and reassembling process possible, necessitating investments that measure up to those by which actual markets were formatted' (Callon, 2007: 349; see also Mackenzie, 2007: 80).\n\nYet, I would argue, work on the performativity of economics has up until now fallen short of making substantial 'investments' that measure up to those which have formatted financial markets so far, and which have furthered a vision based on the privatisation of risks and uncertainty. This is the challenge the performativity of economics confronts if it is not to become redundant: not to provide a blueprint for action, but to think of partial, limited and always precarious attempts to re-assemble markets differently. This is a point that is emerging more generally within STS: as the 'inherently contextual nature of not only the application of scientific knowledge but also its means of production' (Pickersgill, 2012: 599) is re-emphasised, questions about ethics and technology announce a possible 'turn to politics and the normative' (Bijker _et al._ , 2012: xxvi; Johnson and Wetmore, 2009; Collins and Evans, 2002; Jasanoff, 2003; Wynne, 2003).\n\n## **Struggling for different performations**\n\nHow does this work on performativity relate specifically to the discussion of financial derivatives? STS-inflected SSF have been mainly concerned with the second order economy, that is, with circulating 'goods' such as contracts (security, currencies, bonds, derivatives), rather than with the production side of the economy. Much is to be gained by concentrating on both sides, however. Indeed, the case of financial derivatives shows how the two are deeply entangled: their trading clearly impacts on production as decisions on investment in the so-called real economy are made based on valuation processes taking place in financial markets. Acknowledging that value is produced at the intersection between the real and the financial sphere of the economy presents a dual task: it requires us to take into account the contingency (within constraints) of such a process as well as to pose the question of its desirability.\n\nThis crucial task is not new: it has preoccupied feminist and critical political economists for quite some time. Indeed, if neo-classical economists have only recently acknowledged that markets are fabricated, feminist and critical political economists have since the 1970s put forward an understanding of political economy as the nexus between economy and society, showing how these two realms have been kept separate by a problematic belief in an economic order governed by natural laws that exist 'out there' and that society had to discover and implement (for instance, the law of demand and offer, the law of equilibrium, and certain interpretations of the Marxist law of value) (see Dalla Costa and James, 1972; Federici, 1980; Mirowski, 1991). This reconceptualisation of political economy led them to reject both objective and subjective theories of value, and to put forward an understanding of it as a radically contingent process. Through their work on reproductive labour in particular, feminist political economists showed how value is actively made and measured rather than being objectively determined. This further enjoined them to explore alternative value-making processes. For instance, they put forward alternative institutional arrangements so as to instil different processes of valorisation. An example of this was the wages for housework initiative which required the state to recognise domestic labour as 'productive'. The campaign was not without its problems as it risked re-enforcing the valorisation of domestic labour as women's labour as well as extending the same measurement process they were critiquing. As a provocative demand, however, this form of intervention was considered and deployed in light of its potential to shift exactly those value-making processes (see Alessandrini, 2012).\n\nAcknowledging the contestability of value remains today a difficult task: it is because of this contingency that many economists, for example, prefer to focus on money and prices rather than value per se. However, as Bryan and Rafferty (2006: 36) have noted, explanations in terms of money and prices end up being unsatisfactory: hence the need for valuation of something more 'determinant' and 'underlying' (such as fundamental value) always re-occurs, as it has in the current crisis when the exorbitant value of derivatives' markets has been called into question. The body of critical and feminist political economic work I have briefly referred to can provide crucial resources for interrogating and intervening in today's value-making processes. This is because it illustrates how seeing value as always 'contingent, hermeneutic, negotiable and non natural' (Mirowski, 1990: 7) and its measurement as socially constructed does not mean rejecting _any_ measure of value. Rather, it leads us to recognise that while there is no one 'correct way for a society to measure a commodity [...] the way its measurement is _instituted_ has important consequences' (Mirowski, 1991: 568; emphasis added).\n\nIndeed, the growth of financial derivatives has not happened overnight. In addition to the technical, linguistic and legitimatory mechanisms Mackenzie has identified in relation to the performativity of financial theories, the role institutional arrangements have played cannot be overestimated. It was with the demise of the Bretton Woods monetary management system that old risks previously managed by the state were revealed and became privatised while new ones started to proliferate; and firms and governments' increasing need to protect themselves against price volatility was met by the proliferation of financial instruments such as derivatives. And the significance of such a process is not limited to the commercial sphere of firms and governments. As states have gradually withdrawn from the realm of social provisioning, financial markets have increasingly acquired the role of social insurers (Fumagalli and Mezzadra, 2009: 210). This means that an individual's present and future has become ever more reliant on the value of assets invested in financial markets, pension funds being just one example. As well as illustrating the role of institutional arrangements in generating change, this process brings to light the productive power of a technology such as derivatives, showing how although financial innovations, economy and society 'may be shaped through regulatory processes, technology can also act as a form of regulation in its own right since it forces certain kinds of human behaviour while removing the conditions of possibility for others' (Pickersgill, 2013: 329).\n\nNow, we have seen how in the process of allowing firms, governments and individuals to hedge themselves against an ever growing number of risks, derivatives have the potential of affecting the value of the underlying asset, raising the more general issue about the impossibility of retrieving any 'pure, 'real' value untainted by the financial sphere. We can therefore appreciate how crucial it is \u2013 if a world in which financial markets are the arbiters of value is to be avoided, not because they are unreal but because the reality they produce is undesirable \u2013 to start thinking of ways to deal with such uncertainty differently, and this includes thinking of alternative institutional arrangements which might be able to affect such value-making processes. I turn briefly to one such example to illustrate the possibility of dealing with financial uncertainty and the contestability of value revealed by financial derivatives differently. This should not be taken as a blueprint for macroeconomic action but as an (always imperfect) attempt to _institute_ a different way of determining and measuring value in today's financialised economies. In 2010, I conducted research on the Regional Monetary Agreement (RMA), a proposal for dealing with exchange rates tabled by the Republic of Ecuador at the UN. Although the proposal has since been abandoned, I refer briefly to it since it represents an attempt to take financial uncertainty seriously by recognising, while participating in, the co-production of the financial and real spheres of the economy (Ecuador, 2008). Exchange rates are quite significant from a perspective that considers the contestability of value as they are supposed to represent the worth of one currency (and therefore the state of one economy) in terms of another. But how does one think about this relation of equivalence? How does one measure it?\n\nNeo-classical theory talks of Purchasing Power Parity and fundamental values but in financial markets actual exchange rates are determined on the basis of the supply and demand of the currencies and it is here that expectations play an important role. Although these expectations are supposed to be based on so-called real variables such as price index, interest rates, macro-economic stability and inflation rates, and leaving aside the question of how real an account of the state of the economy these variables can ever give, the point is that in today's financial markets expectations about the future impact of such variables, including anticipations of others' expectations, are always already affecting today's value (Harvey, 2005: 5\u20136). This level of uncertainty shows exactly the limits of regulation: measures can and should be introduced to limit speculative attacks against currencies. However, as long as derivatives provide such a crucial hedging function there will be a need for them and as long as there is such a need, there will be speculation. Furthermore, even if the impact of derivatives on prices were reduced, the value exchange rates are supposed to express (the average prices of goods and services) remains contestable. Indeed, how is the 'real' value of goods and services, whose relation of equivalence exchange rates are supposed to express, to be measured in the first place?\n\nThese are the two aspects the RMA aimed to address, first, by proposing to deal with exchange rates in a cooperative manner. This was through mutually agreed bands based on the sharing of information about macroeconomic policies among states (Ecuador, 2008: 2\u20133), and it aimed to remove a critical source of uncertainty from the realm of private dealings. Thus, it would have reduced the need for firms and governments to hedge their operations against the risks of price volatility in the first place. Second, by acknowledging, while trying to shape, expectations, and agreeing on optimal exchange rates within bands, policymakers intended to immerse themselves in the construction of exchange rates, thereby removing this process from the exclusive realm of financial markets. The proposal envisaged interaction between governments and market agents in the following way: governments would give an indication of exchange rate through bands; the rates would however be left free to fluctuate within these bands until governments' projections would have been called into question by market agents. Then a re-assessment would have taken place and a decision by governments been made again. The proposal therefore envisaged a different kind of arrangement between states and markets, recognising the participation of both in the determination of exchange rates' value.\n\nWhat is interesting is that, although the proposal talked about anchoring expectations, the actual aim was not that of making exchange rates reflect equilibrium prices (in particular the average price of goods and services). When I asked in Quito (23 March 2010) what 'real' variables exchange rates' expectations should be anchored to, Pedro P\u00e1ez, the President of the Ecuadorean Technical Commission, pointed out that 'equilibrium prices' have not existed for a long time, particularly if one takes into account the 'historic matrix of dependence' that has its roots in the colonial period: hence the difficulty with setting any (exchange rate) parity at the beginning of the institutional arrangement. However, this did not imply abandoning the attempt to connect exchange rates with the economy (in this case average prices); on the contrary, it meant participating more actively in its making. This is why, as the proposal made clear and P\u00e1ez was adamant to emphasise, the financial architecture could not be thought of in isolation from reforms that need to occur in areas ranging from food, agriculture, housing, health and education, to energy, technology, environmental and bio-services.\n\nIt is not my intention to idealise such an arrangement nor to suggest it as a blueprint for future action in the international management of exchange rates. I have briefly referred to it as an instance of 'investment' in the economy that rejects the separation between the real and the financial spheres: its emphasis was rather on the economy conceived of not in isolation from the other realms of life but as the terrain on which the (re)production of life takes places. I have also referred to it as an instance of an arrangement that shows how seeing value as contested does not mean that any value-making process goes but on the contrary requires actors (i.e. policy makers, market participants, etc.) to engage more fully in its construction. From this perspective, this notion of 'fundamental' value need not be abandoned but rather reconceptualised; no longer seen as that which is pure and untainted by financial activity \u2013 and therefore outside financial markets which attempt to capture its substance \u2013 but rather as that which is constantly made. Finally I have conceived of this arrangement as a 'provocation', and I am proposing to think of institutional arrangements more generally as provocations, in three interrelated respects: first, in the tradition of feminist autonomists, so as to think of desirable forms of intervention in the economy capable of engendering alternative valorisation processes. Second, to acknowledge that non-humans as well as humans participate in valorisation processes and that the reality provoked by these interventions might be very different from that envisaged by humans. Third, to bear in mind that whatever the 'goals' we aim to achieve through interventions, the act of formatting markets will always comport violent exclusions.\n\n## **Conclusion**\n\nThis chapter has argued against the view of financial derivatives as parasitical and unproductive financial innovations. It has claimed they can be best understood as a technology, and one which acts as a 'form of regulation in its own right', facilitating certain practices and obstructing others (Pickersgill, 2013: 329). Thus, while embedding the institutional drive for the privatisation (and proliferation) of risks and uncertainty, derivatives have also affected the process through which value is produced at the intersection between the real and the financial spheres of the economy, with important consequences for the investment decisions of firms, governments and individuals. Tracing the complexity of such value-making processes has simultaneously brought to light the porosity of the boundaries between the real and the financial spheres of the economy, the economy and society, formal and informal regulation, and the intended and unintended effects of the latter, thereby extending the important contribution STS have made in troubling traditional boundaries to the domain of financial practices (cf. Cloatre, 2013). Of paramount importance in this respect has been the contribution of STS-inflected SSF: scholars working within this field have opened up to scrutiny the many financial innovations which until recently have been considered as either neutral devices or mere receptacles of values. Notwithstanding this important work, I have argued that SSF have up until now refrained from making substantial 'investments' in reformatting the economy. Such investments include, but are not limited to, struggles over different kinds of regulation, institutional and otherwise. Regulation is here conceived of not as a blueprint for action but as a provocation, that is the partial, limited and always precarious attempt to re-assemble markets (and reality) differently, with an appreciation that the reality thus provoked might be different from the one we envisage in action.\n\n1. The research was supported by the British Academy (grant number SG090138).\n\n## **References**\n\nAlessandrini, D. (2012) 'Immaterial labour and alternative valorisation processes in Italian feminist debates: (re)exploring the \"commons\" of re-production', _Feminists@Law_ , 1(2): 1\u201328.\n\nArthur, B.W., Holland, J.H., LeBaron, B., Palmer, R. and Tayler, P. (1996) 'Asset pricing under endogenous expectations in an artificial stock market, Social Sciences Working Paper'. Available: h\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ads\u00ads\u00adr\u00adn\u00ad.\u00adc\u00ado\u00adm\u00ad\/\u00ada\u00adb\u00ads\u00adt\u00adr\u00ada\u00adc\u00adt\u00ad=\u00ad2\u00ad2\u00ad5\u00ad2 (accessed 5 July 2013).\n\nBijker,W., Hughes, T.P. and Pinch, T. (2012) _The Social Construction of Technological Systems: New Directions in the Sociology and History of Technology_ , Cambridge, MA: MIT Press.\n\nBryan, D. and Rafferty, M. (2006) _Capitalism with Derivatives: A Political Economy of Financial Derivatives, Capital and Class_ , London: Palgrave Macmillan.\n\nCallon, M. (1998) _The Laws of the Markets_ , Oxford: Blackwell.\n\nCallon, M (2007) 'What does it mean to say that economics is performative?, in D.A. Mackenzie, F. Muniesa, and L. Siu (eds), _Do Economists Make Markets: On the Performativity of Economics_ , Princeton, NJ: Princeton University Press.\n\nCollins,H.M. and Evans, R. (2002) 'The third wave of science studies: studies of expertise and experience', _Social Studies of Science_ , 32(2): 235\u2013296.\n\nDalla Costa, M. and James, S. (1972) _The Power of Women and the Subversion of the Community_ , Bristol: Falling Wall.\n\nCloatre, E. (2013) _Pills for the Poorest: An Exploration of TRIPs and Access to Medication in Sub-Saharan Africa_ , London: Palgrave Macmillan\n\nCloatre, E. and Dingwall, R. (2012) 'The migration of objects, scripts and governance', _Regulation and Governance_ , first published online: 10 October2012.\n\nde Goede, M. (2005) _Virtue, Fortune, and Faith: A Genealogy of Finance_ , Minneapolis: University of Minnesota Press.\n\nEcuador, Permanent Mission to the United Nations (2008) 'The Ecuadorian proposal for a crisis response agenda from the south', _New York: UN_.\n\nFederici, S. (1980) 'Wages against housework', in E. Malos (ed.), _The Politics of Housework_ London: Allison and Busby.\n\nFumagalli, A. and Mezzadra, S. (2009) _Crisi Dell'Economia Globale: Mercati Finanziari, Lotte Sociali e Nuovi Scenari Politici_ , Verona: OmbreCorte-UniNomad.\n\nGraham,B. and Dodd, D. (1940) _Security Analysis:Principles and Technique_. New York and London: McGraw-Hill Book Company, Inc.\n\nHarvey, J.T. (2005) ' _Post Keynesian versus neoclassical explanations of exchange rate movements: a short look at the long run_ ',Working Paper, Texas: Department of Economics, 1996.\n\nJasanoff, S. (2003) 'Breaking the waves in science studies: comment on H.M. Collins and Robert Evans, the third wave of science studies', _Social Studies of Science_ , 33(3): 380\u2013400.\n\nJohnson, G.G. and Wetmore, J.M. (2009) _Technology and Society: Building our Sociotechnical Future_ , Cambridge, MA: MIT Press.\n\nKnorr Cetina, K. and Preda, A. (eds) (2005) _The Sociology of Financial Markets_ , Oxford: Oxford University Press.\n\nLatour, B. (1993) _We Have Never Been Modern_ , Brighton: Harvester Wheatsheaf.\n\nLatour, B. (2005) _The Politics of Nature: How to Bring the Sciences into Democracy_ , Cambridge: Harvard University Press.\n\nLipuma, E. and Lee, B. (2004) _Financial Derivatives and the Globalisation of Risks, Durham_ , NC and London: Duke University Press.\n\nMackenzie, D. (2006) _An Engine, Not a Camera: How Financial Models Shape Markets_ , Cambridge, MA: MIT Press.\n\nMacKenzie, D. (2009) _Material Markets: How Economic Agents are Constructed_ , Oxford: Oxford University Press.\n\nMackenzie, D., Muniesa, F. and Siu, L. (2007) _Do Economists Make Markets: On the Performativity of Economics, Woodstock_ , Oxfordshire: Princeton University Press.\n\nMarazzi, C. (2010) _The Violence of Financial Capitalism_ , London: MIT Press.\n\nMarx, K. (1981) _Capital_ , Vol. III, London: Penguin Classics.\n\nMatai, D.K. (2009) ' _The size of the derivative bubble = $190k per person on the planet_ ', Global Research. Available: www.globalresearch.ca\/the-size-of-the-derivatives-bubble-190k-per-person-on-planet\/12753 (accessed 5 July 2013).\n\nMaurer, B. (2002) 'Repressed futures: financial derivatives theological unconscious', _Economy and Society_ , 31: 15\u201336.\n\nMerton, R.K. (1968) _Social Theory and Social Structure_ , New York: Free Pres.\n\nMirowski, P. (1990) 'Learning the meaning of a dollar: conservation principles and the social theory of value in economic theory', _Social Research_ , 57(3): 689\u2013717.\n\nMirowski, P. (1991) 'Postmodernism and the social theory of value', _Journal of Post-Keynesian Economics_ , 13(4): 565\u2013582.\n\nMirowski, P. and Nik-Khah, E. (2007) 'Markets made flesh: performativity and a problem in science studies, augmented with the consideration of the FCC auctions', in D. Mackenzie, F. Muniesa, and L. Siu (eds), _Do Economists make Markets: On the Performativity of Economics_ , Woodstock, Oxfordshire: Princeton University Press.\n\nOrlean, A. (2005) 'The self-referential hypothesis in finance', in J.P. Touffut (ed.), _The Stability of Finance in Europe_ , Paris: Albin Michel.\n\nPickersgill, M.D. (2012) 'The co-production of science, ethics and emotions', _Science Technology Human Values_ , 37(6): 579\u2013603.\n\nPickersgill, M.D. (2013) 'Sociotechnical innovation in mental health: articulating complexity' in Flear, M. L., Farrell, A., Hervey, T. K. and Murphy, T. (eds) _(2013) European Law and New Health Technologies_ , Oxford: Oxford University Press.\n\nStrange, S. (1997) _Casino Capitalism_ , Manchester: Manchester University Press.\n\nWynne, B. (2003) 'Seasick on the third wave? subverting the hegemony of propositionalism', _Social Studies of Science_ , 33(3): 401\u2013417.\nSection 4\n\n* * *\n\n# Law, technoscience and the stabilization of knowledge\n\n* * *\nChapter 10\n\n* * *\n\n# **Epistemic jurisdictions**\n\nScience and courts in regulatory (de)centralisation\n\n_David E.Winickoff_\n\n* * *\n\nRegulatory science can be defined as the scientific foundation of public policy decisions, whether regulatory, legislative, or judicial (FDA 2013). As a practice, it includes, among other things, technical reasoning such as risk analysis and modeling; the operation of expertise in policy; and procedures for validating knowledge in courts and other governing institutions. The politics and operation of these forms of technical reason within public process has been the subject of STS scholarship for some time (e.g. Nelkin 1984). This work has helped uncover, in nuanced ways, the politics of who makes authoritative knowledge, using whose methods, for whom, and subject to what political consequences (e.g. Jasanoff 1987; Gieryn 1999; Sarewitz 2004). As STS scholars have shown, regulatory science does not represent knowledge free of value. Often sitting in a context of uncertainty and contestation, it combines scientific reasoning with social and political judgment (Jasanoff 1990: 229).\n\nScholars turning their attention to regulatory science at the global and international levels have largely focused on its operation, broadly speaking, within single institutions. A central finding has been that these institutions can act as hegemonic centers of calculation and assessment, specialized knowledge, and embodied expertise. This work has analyzed, among other things, the use of science for environmental impact assessments at the World Bank (Goldman 2005); the use of self-governance and peer review at the IPCC (Edwards and Schneider 2001); and power dynamics in the Millennium Ecosystem Assessment (Miller 2007). Such work has also examined how expert communities function and acquire power at the global scale (Oreskes and Conway 2010; Haas 1992).\n\nMuch less has been said about how science operates through governance systems that traverse levels, scales, and sectors to address collective problems (Ostrom 2009; Biermann and Pattberg 2008). It is well understood that regulatory approaches to global issues, from climate change to food safety, are increasingly moving away from globalizing systems and instead operating through pluralistic, overlapping, and fragmented regimes (Boyd 2010). These forms of \"polycentric\" or \"complex\" governance are hailed by some scholars and policy-makers (e.g. Ostrom 2009; Rabe 2004). But they can result in jurisdictional overlap and competition. Regulatory science is often a focal point of conflict (e.g. Peel 2010).\n\nThe jurisdictional politics of ethanol regulation is a good illustration. In the absence of strong Federal leadership, U.S. cities, states, and regions are enacting efficiency incentives and purchasing mandates to mitigate their carbon footprints. California recently enacted a \"low-carbon fuel standard\" which was cited as a model by the European Union in its passage of the 2009 Renewable Energy Directive. However, in a lawsuit currently being heard in an appellate court in California, private actors have sued California on the grounds that national standards pre-empt California's technical standards on federalist grounds. Meanwhile, at the global level, Brazil and the U.S. are beginning to challenge the EU Directive at the World Trade Organization (\"WTO\") on the grounds that its sustainability standards constitute an unfair technical barrier to trade. In this process, technical indicators of sustainability themselves will likely go on trial.\n\nHow is regulatory science operating in the jurisdictional politics of complex regulatory regimes, and what are the implications for democratic legitimacy? A number of cases of cross-level legal friction and multi-level formation help me explore these questions. First, I attempt to theorize the role of regulatory science within multi-level governance through the STS framework of coproduction, drawing attention to what I call \"epistemic jurisdiction.\" Second, I analyze legal encounters with regulatory science in three case studies, including food safety adjudication at the WTO, the Clean Air Act at the U.S. Supreme Court, and California Global Warming Solutions Act. Through these cases, I argue that the creation of epistemic jurisdictions and the authorization of particular forms of regulatory science are performing critical political work in shaping multi-level regimes.\n\n## **Coproduction, multi-level governance, and the law**\n\nImportant strands of work on regulatory science focus on the decision-making processes within expert communities, the contingency of knowledge, and political influences on knowledge production (e.g. Edwards and Schneider 2001; Oreskes 2010). In this chapter, I wish to examine not only how regulatory science is shaped in contexts of multi-level conflict, but also how it is _productive_ of new jurisdictional arrangements.\n\nWork in the STS framework of coproduction (Jasanoff 2004) is useful here because it understands science and technology on the one hand, politics and law on the other, as co-constitutional. STS scholars have explored the co-emergence of political and epistemic order in a wide array of contexts, including climate change (Miller 2005), the protection of endangered species (Thompson 2004), patenting in the life sciences (Parthasarathy 2007), and racial categorization in drug development (Kahn 2012). As a theory, coproduction draws on Michel Foucault's insight that knowledge is not just a factor in governance but a constitutive element (e.g. Foucault 1971 [1966]).\n\nThe coproductionist framework encourages us to view the development of regulatory science and the formation of jurisdictional architecture in the same breath. This architecture is in a period of rapid change and so deserves analysis and scrutiny. In the U.S., it is an era of devolution. Power is actively shifting, broadly speaking, from the center back toward the States and towards even lower levels in arenas like environmental policy. In the European Union, the process of integrating economic and regulatory functions has birthed a politics of \"subsidiarity\" (Nicolaidis and Howse 2001: 1). There is, in general, a move to complex regulatory architectures that depart from the clearly demarcated levels of regulatory jurisdiction found in pure federalism (Hooghe and Marks 2001). For this reason, many legal scholars have given new power sharing agreements across levels the name of \"dynamic federalism\" (Ruhl and Salzman 2010: 66) or prefer to speak of \"complex governance.\" Multiple levels of government are interacting in regulatory processes (Engel 2006), and these interactions require more attention within the field of STS.\n\nCoproduction points to how processes both of political centralization and decentralization may depend upon, more than has been acknowledged, the production of new forms of science and expert authority. In particular, evolving multi-level governance raises questions about the allocation and reallocation of \"epistemic jurisdiction\": _the power to produce, interpret or warrant technical knowledge for a given political community, topical arena, or geographical territory_.\n\nThe production of new epistemic jurisdiction can support political centralization. For instance, the mandate to liberalize trade has called into being new global standard setting bodies that have a taken a share of regulatory authority from sovereign states. Through appeals to a universalist expertise and regulatory science, these bodies can operate as depoliticized vectors of globalization (Moore _et al._ 2011). But epistemic jurisdiction can also be invoked to oppose political centralization. For instance, within the arena of climate change, sub-national actors have resisted international regimes, and the policies of their own states, through appeals to \"local knowledge\" (Long-Martello and Jasanoff 2004). In other words, the appeal to local expertise can underwrite claims of localized political autonomy.\n\nUnderstanding the ways in which centralization and decentralization depend on regulatory science and the configuration of epistemic authority requires more systematic investigation at different sites of practice. Legislative history, rule making, and adjudication are promising sites for they stage the interaction of legal and epistemic forms of authority, and are productive of new regulatory architectures. Legal process within the regulatory arenas of food safety and climate change will be the focus here. These cases show the different ways in which new jurisdictions\u2014both epistemic and political\u2014are built through legal entanglements with regulatory science.\n\n## **International food safety regime**\n\nThe Sanitary and Phytosanitary (SPS) Agreement, one of the suite of agreements adopted under the WTO in 1994, is the dominant source of food safety governance at the global level. The agreement grew out of widespread concerns that food safety regulations were being used to protect domestic production against foreign competition. The aim of the SPS Agreement is actually more ambitious than that of the core Global Agreement on Tariffs and Trade (GATT), for it moves beyond the principle of non-discrimination to one of regulatory harmonization (Charnovitz 2000). Harmonization is not homogenization, and the agreement recognizes the need to preserve a degree of regulatory autonomy for member states (SPS Prologue). Even so, harmonization does imply the operation of a common rationality that can serve to promote convergence and cooperation.\n\nThe negotiation history, text, and legal contestations of the SPS Agreement suggest how regulatory science has been a constitutive element of the food safety regime. Harmonization as embodied in the agreement was only achieved via a technocratic ideology, a scientistic epistemology, and the delegation of epistemic jurisdiction to a global agency. Further, as adjudication under the SPS shows, the system of power sharing within the global regime has co-emerged with interpretations of what counts as valid science.\n\n### **_The ascent of risk assessment and a global regulatory jurisdiction_**\n\nWhile regulatory harmonization was a goal of the Uruguay round of trade negotiations, it was not easy to achieve (Winickoff and Bushey 2010). Food, after all, is a highly regulated area sitting squarely within the traditional police powers of the sovereign state. Safety standards of all kinds, let alone ones in the arena of food, are value-laden and indicative of ethical choices (Busch 2012; Timmermans and Epstein 2010). To develop global standards, then, is to enact a form of global political and cultural order (Nils Brunsson and Bengt Jabobsson 2000).\n\nThe negotiating history of the SPS Agreement shows how science became the primary ideological resource for producing agreement (Winickoff and Bushey 2010). As the lone superpower in the mid-1990s, the U.S. pushed a universalist notion of \"scientific justification\" as the primary tool to combat bans on beef hormones in European states and other non-tariff barriers to the food trade. Negotiators from Canada, Australia, New Zealand, and the U.S. identified formal risk assessment as the earmark of scientific justification, even though it was not yet well established in food safety regulation worldwide. Concurrently, the European Commission\u2014Europe's trade representative at the WTO\u2014was engaged in its own project of harmonizing \"social regulation\" across EU member states (Joerges 1997), and was ultimately willing to embrace risk analysis as a harmonizing force within the SPS Agreement. Given the alignment of interests across the U.S. and the EC to harmonize food safety regulation, scientific evidence was a plausible enough idea around which to forge agreement.\n\nThe focus on science is evident in the text of the agreement. The SPS's core disciplines require that WTO members ensure that any food safety measure \"is based on scientific principles and is not maintained without sufficient scientific evidence\" (SPS 2.2). Final regulatory judgments of states are given safe harbor, so long as they are \"rationally based\" on technical risk assessments. Scientific validity, as established through a particular type of regulatory science, is thus necessary to establish legal validity.\n\nHarmonizing regulatory logic does not address the substance of standards themselves, so SPS negotiators looked for existing international food standards as a possible resource for substantive harmonization. They found the Codex Alimentarius Commission, a little-known bureau of the FAO and WHO that had been producing voluntary food safety standards on residues, pesticide use, etc., since the 1960s. Accordingly, within the SPS Agreement, the Codex was designated one of three \"relevant international organizations\" around whose standards the signatories would attempt to harmonize (SPS 3.4). The guidelines and recommendations of Codex, if adopted by nations, would be considered to satisfy the obligations under the agreement (SPS 3.2). Ironically, it was only the prospect of designation as a \"scientific\" agency of the WTO that the Codex began to formalize its own risk assessment protocols (Winickoff and Bushey 2010). Far from drawing upon a well-established set of epistemic authorities, the WTO regime was a midwife to them.\n\nThe negotiation and text of the SPS Agreement reveal how within multilevel governance systems, authorizing regulatory science and new epistemic jurisdictions can underwrite projects of centralization. Within this regime, derogation from known international standards requires special scientific justification. Epistemic jurisdiction has evolved from an exclusively national one to a multi-level one in which state agencies are held accountable to certain methodological universals.\n\n### **_Adjudicating GMOs at the World Trade Organization_**\n\nThe fact that the SPS Agreement did not settle questions of power sharing is evidenced by the litigation that has steadily gone on since its execution. It is unsurprising that these legal disputes have tested the requirements of \"scientific justification,\" given its pivotal role in the regime. Cases involving hormones in beef, imported salmon, apples, and genetically modified crops show how the patterns of regulatory power sharing within the global regime are co-evolving with interpretations of validity in regulatory science.\n\nIn particular, the GMO dispute under the SPS Agreement illustrates this process. The context of the dispute is well known. In the 1980s, the U.S. developed a \"products approach\" for regulating agricultural biotechnology. This approach assumes that transgenesis entails no additional risk to consumers or agriculture and requires no additional scientific evidence of safety so long as the new product is deemed \"substantially equivalent\" to its conventional counterparts. All new GMO crops in the 1980s and 1990s met this standard. In the EC and European states, regulators have taken a more precautionary \"process approach.\" This approach assumes that transgenesis may entail novel and unique risks to human health or the environment and puts an extra scientific burden on producers to demonstrate safety (Jasanoff 1995).\n\nIn 2003, the WTO Dispute Settlement Body had to address the growing trade conflict on GMOs, and adjudicate these distinct approaches to evidentiary burden and scientific justification. Europe had begun importing Round-up Ready soy and other GMOs in the mid-1990s. In the face of rising consumer concerns in the late 1990s, European states had delayed the approval process for new GMO applications, ostensibly waiting until the EU could pass new legislation on traceability and labeling (Winickoff _et al._ 2005). The United States, Canada, and Argentina initiated dispute settlement procedures at the WTO against the EC for delaying approvals of GM crops.\n\nThis dispute became the so-called _Biotech Products_ case, and it involved the interpretation provisions of the SPS Agreement concerning \"scientific justification\" and \"risk assessment.\" In _Biotech Products_ , the U.S. challenged the scientific basis of various European actions preventing the importation of GM crops and food products. In particular, the U.S. alleged that reversals of GM regulatory policy within the EU illustrate the EU's departure from a fixed body of sound science and constituted \"unreasonable\" or \"undue delay\" under Annex C(1)(a) of the agreement. In response, the EC focused its argument on the safe harbor provision of SPS Article 5.7\u2014which permits members to impose provisional or precautionary measures \"in cases where relevant scientific evidence is insufficient\"\u2014arguing that GM presented novel scientific and social questions and therefore, at the time, required a more developed scientific record of safety.\n\nIn the process of the adjudication, the WTO took up a detailed review of the risk assessment processes of European states, putting it in the position of certification or rejection of competing models of technical reason (Winickoff _et al._ 2005). The Dispute Settlement Body (DSB) found against individual EU member states and the European Commission for procedural irregularities in their risk assessment processes (Bonneuil and Levidow 2012). The DSB held that a number of European countries had conducted risk assessments, and the conclusions of these assessments did not support a \"precautionary\" moratorium on crop approvals. Rather, their procedure entailed an \"undue delay\" and therefore a violation of the agreement. Further, the DSB disagreed that \"relevant scientific evidence was insufficient\" for a valid risk assessment, an interpretation that seems to narrow the availability of defenses based on the precautionary principle.\n\nThe _Biotech Products_ case achieved an important result in the food safety regime, effectively strengthening the SPS Agreement as a sword against nationally-enacted food safety regulations. The international level of governance was strengthened within the multi-level regime at the same time as risk assessment was reinforced as the master regulatory science.\n\nPerhaps less predictably, the ruling produced new regulatory order at lower levels. A follow-on result was that _Biotech Products_ helped the European Commission consolidate its epistemic jurisdiction over food safety _across_ Europe. The EC had recommenced GM approvals before the resolution of _Biotech Products_ , and had adopted the sharpened risk-based approach emerging out of Codex and WTO case law. As the opinion found non-compliance of individual member states, there has been a shift in epistemic jurisdiction from the national to the European level. Since _Biotech Products_ , there has been an ongoing struggle between the European Food Safety Authority (EFSA), which nominally holds the authority to conduct the science-based risk assessment, and national authorities. Although the ability of member states to make a political judgment concerning the acceptability of a given set of risks (as outlined by EFSA) has not been questioned, the idea that the EC would take from member states the power to assess, i.e., conduct the technical analysis itself, has led commentators and activists to label this as an unwarranted grab at state sovereignty.\n\nPolitical legitimacy has been a key concern of EFSA's power to assess. This assignment of epistemic authority over food safety is perceived to be part of a larger pattern whereby Europe attempts \"to integrate new geographic spaces and populations not by overt coercion, but by instituting a host of harmonized regulations, codes, and standards\" (e.g. Dunn 2005). For some, centralization of these \"powers to know\" are a back channel for constituting Europe as a unified political community without the proper political debate and without democratic accountability (Wickson and Wynne 2012).\n\n## **Climate federalism in the United States**\n\nThe food safety case evinces a redistribution of power operating through particular scientific logics and the redistribution of epistemic authority. It has been a top-down process, although lower levels are pushing back in Europe through the \"safeguard clause\" of the European GMO directive (Dir. 2001\/18\/EC). In the arena of climate change, regime formation is not occurring from the top down, but from the bottom up, creating innovative approaches to carbon mitigation, but also introducing conflicts between lower and higher regulatory authorities. Here, too, science and epistemic jurisdiction are important terrain for negotiating multi-level power sharing.\n\nClimate change is currently a crucial site in the construction of a more complex global order (Boyd 2010). Networks of cooperation are moving vertically and horizontally across cities, states, regions, and nations worldwide. In the U.S.\u2014where national policy has moved at snail's pace\u2014regions, states, counties, and cities are taking matters into their own hands, producing what legal scholars have called \"climate federalism\" (Farber 2008). Advocates of this _de facto_ decentralized approach argue that it can foster a laboratory of democratic experimentation and render governance more accountable (Rabe 2004; Chimerinski _et al._ 2008). Critics see a multi-level approach as potentially causing unnecessary patchworks of regulation (Farber 2008).\n\nRegardless of the relative merits of climate federalism in the U.S., it has led to jurisdictional friction. The mismatch in scale of biogeophysical systems and management systems can cause regulatory overlaps, gaps, and controversies (Cash _et al._ 2006). Science and technical reason are focal points both for cooperation and conflict. Jurisdictional conflicts, often playing out in the courts, are sites in which particular forms of knowledge, assessment, and expertise are made authoritative. In this domain as in food safety, the construction of epistemic jurisdiction\u2014especially as it operates over evidentiary issues, the choice of analytic frameworks, and methods of accounting\u2014is a driver of multi-level reformation.\n\n### **_Carbon regulation at the U.S. Supreme Court_**\n\nThe early 2000s were a time of political logjam and intense acrimony in U.S. climate policy. The arrival of George W. Bush as the 43rd U.S. President meant the beginning of an era of overt obstructionism towards reducing GHGs in the U.S. The Environmental Protection Agency (EPA), the lead environmental agency on climate, was complicit. Faced with an intransigent White House and EPA, 19 non-government organizations petitioned the agency to begin regulating GHGs as an \"air pollutant\" under the Clean Air Act (CAA). After holding a number of public hearings on the issue, the EPA decided to do nothing. A large number of State and local governments joined the original petitioners to file suit in the Federal Courts, alleging that the agency's decision not to regulate violated the law. The State of Massachusetts became the lead plaintiff in the case, and the U.S. Supreme Court eventually took the case.\n\nNarrowly speaking, there were two legal issues in the resulting case, _Massachusetts v. EPA_. First, did Massachusetts have legal standing to sue the EPA? To establish standing in U.S. Federal Courts, a plaintiff must be able to show, among other things, that it \"suffered a concrete and particularized injury that is either actual or imminent.\" Plaintiffs must also show that the injury is fairly traceable to the defendant, and that the injury is capable of redress. Second, was the EPA's judgment not to regulate carbon as a \"pollutant\" \"arbitrary and capricious,\" the standard of judicial review under the Clean Air Act?\n\nComplex questions of federalism and epistemic jurisdiction lurk underneath these seemingly narrow legal issues. In its complaints, the State of Massachusetts appealed to a global science and global experts to establish local harms, making both a procedural and substantive challenge to regulatory authorities at the Federal level. The EPA, the expert agency to whom the U.S. Congress delegated power, denied that existing scientific knowledge justified legal standing, and claimed that the Court itself lacked sufficient legal and scientific authority to overturn its definition of \"pollutant.\" The questions of epistemic authority and federalism were inseparable.\n\nIn a 5\u20134 decision, the U.S. Supreme Court decided that Massachusetts had demonstrated its injury and was granted standing. The Court also held that the EPA's conclusion that carbon cannot be an air pollutant was an \"arbitrary and capricious\" judgment, and therefore subject to rejection by the Court under the terms of the Clean Air Act. Further, the Court ruled that the EPA could \"avoid taking further action only if it determines that greenhouse gases do not contribute to climate change or if it provides some reasonable explanation as to why it cannot or will not exercise its discretion to determine whether they do\" (533). In other words, the EPA would need to justify its inaction: the Court shifted the scientific burden of proof to the EPA.\n\nWriting for the majority, Justice John Paul Stevens led not with law but facts about climate change as understood by \"respected scientists\" (505). He narrated a history of growing agreement about the scope and cause of global warming, indicating the ways in which accreting knowledge had grounded government actions in the U.S. and the Kyoto Protocol. Stevens drew together a large network of actors to attest to the harm Massachusetts has suffered from global warming and the EPA's inaction\u2014from the Mauna Lao observatory and its atmospheric carbon measurements dating back to 1959, to ancient ice cores, to the United Nations climate institutions, to rising coastal sea levels (507\u2013509).\n\nUnder the Clean Air Act, epistemic jurisdiction clearly lies with the EPA, for EPA regulators have the power to assess risk and produce regulations. However, Stevens explicitly cited the Intergovernmental Panel on Climate Change (IPCC) five times, describing it as \"a multinational scientific body organized under the auspices of the United Nations [...] drawing on expert opinions from across the globe\" (508). The effect of his scientific discussion is to simultaneously validate an account of nature and privilege an epistemic authority that transcends state or nation.\n\nWriting for the minority, Justice John Roberts rejected the state's arguments on both logical and empirical grounds. He said that deriving a causal inference between the EPA's failure to regulate and Massachusetts' harm is \"pure conjecture\": the complexity of feedbacks in the system make drawing this causal inference to _local_ harm impossible. Furthermore, Roberts attacked the empirical basis of the injury finding, noting that the measured amount of Massachusetts sea level change is within the computer model's conceded margin of error (542).\n\nIn the face of such a strong dissent, Stevens probably felt the need to shore up his epistemic case with a moral argument about federalism\u2014they work hand in glove. As part of the federalist bargain in the U.S., he said, individual States like Massachusetts have effectively waived certain \"sovereign prerogatives\", including ones relevant here, e.g., the power \"to force reductions in greenhouse gas emissions, to negotiate emissions treaties with developing countries, and (in some circumstances) to exercise the police power to reduce motor-vehicle emissions\"; these powers \"are now lodged in the Federal Government\" (519). Because of these significant forfeitures, States as \"quasi sovereigns,\" empowered to protect their land and citizenry, must have special consideration as they call the Federal Government to action in pre-empted regulatory spheres.\n\nIn effect, Stevens struck an interesting new balance of Federal power: the exclusive power to assess science and regulate has not been dislodged from the Federal Government, but States\u2014as a class of plaintiffs challenging regulatory judgments\u2014should be shown special consideration because of their traditional sovereign roles. In aid of his argument, Stevens reached far back to a 1907 opinion by the famous Supreme Court Justice, Oliver Wendell Holmes, in the case of _Georgia v. Tennessee Copper Co._ (206 U.S. 230). In that case, Georgia sought to protect its citizens from air pollution originating outside its borders. Quoted at length in the opinion, Justice Holmes stated that,\n\n> [t]his is a suit by a State for an injury to it in its capacity of quasi-sovereign. In that capacity the State has an interest independent of and behind the titles of its citizens, in all the earth and air within its domain. It has the last word as to whether its mountains shall be stripped of their forests and its inhabitants shall breathe pure air.\n> \n> (518\u2013519)\n\nStevens not only quoted Holmes, but also echoed his ominous tone. \"Rising seas,\" he says, \"have already begun to swallow Massachusetts' land\" (499). Stevens seemed to be thoroughly persuaded by the IPCC and scientific depositions that Massachusetts, perhaps like Georgia in 1907, was under the kind of environmental assault that required calling back the old doctrine of \"quasi-sovereignty.\" This standing granted Massachusetts a right to challenge the EPA's epistemic judgments.\n\n### **_San Bernardino County under California's Global Warming Solutions Act_**\n\nJustice Roberts' dissent in _Massachusetts v. EPA_ , especially the idea that the very concept of global warming is inconsistent with recognizing the standing of localities, cuts against the grain of recent policy. Indeed, there is now quite significant regulatory activity on climate change at the state and local level in the U.S. As of 2011, seven regional climate initiatives were underway, 23 U.S. states had established GHG emissions targets or goals, and over a thousand mayors had committed their cities to try and meet the U.S. Kyoto Protocol targets (Pew 2011).\n\nCalifornia stands out among these efforts: it has legislated ambitious emissions targets and developed novel regulatory mechanisms. Signed into law in 2006 by Governor Arnold Schwarzenegger, the Global Warming Solutions Act is the capstone of the California's climate policy. Best-known as AB 32, the law is explicit that although it is acting as one state, its passage is vital to both local and global interests:\n\n> [a]ction taken by California to reduce emissions of GHGs will have far-reaching effects by encouraging other states, the Federal Government, and other countries to act [...] [AB32] will continue [California's] tradition of environmental leadership by placing [it] at the forefront of national and international efforts to reduce emissions of GHGs.\n> \n> (2(c)\u2013(e))\n\nAB 32 creates a statewide GHG emissions registry and mandates a cap of GHG emissions at 1990 levels by 2020. The law does not specify the regulatory mechanisms for achieving these goals, but it directs the California Air Resources Board (CARB) to spearhead the regulatory process in collaboration with other agencies.\n\nOpponents to California's aggressive approach to climate change have attacked the law on federalist grounds. As referenced above, the Low-Carbon Fuel Standard (LCFS) developed by the California Air Resources Board regulates the kinds of fuels that can be imported and used in the state, deploying particular accountings of \"carbon intensity.\" The state's LCFS was challenged in the Federal Courts by a coalition of actors in the transportation industry who claimed that the Federal approach to fuel standards should pre-empt California's stricter carbon intensity approach. Because it regulated exhaust emissions before the Federal Government, California has an exemption under the Clean Air Act for setting emissions standards of fuels higher than those of the Federal Government. California defended its fuel standard on these grounds.\n\nCalifornia's scope for derogating from Federal regulatory standards turned on contested meanings of carbon intensity: does regulating \"carbon intensity\" of fuels amount to a regulation of the fuel itself, thereby triggering California's exemption, or is it merely a regulation on production methods and therefore not exempt? The Federal District Court ultimately affirmed that carbon intensity constituted a fuel characteristic, allowing the California LCFS to stand. The Court simultaneously answered a technical and legal question, constituting fuel and federalism in the same judgment.\n\nJust as California's carbon regulations have come into conflict with a higher level of government, so have they brought conflict with lower ones. Under the California Environmental Quality Act (CEQA), counties or other public agencies undertaking a project with the potential to harm the environment must prepare an environmental impact report (EIR) that analyzes foreseeable environmental effects of the project and measures available to mitigate those effects. Located in Southern California, San Bernardino County covers a dry terrain of over 20,000 square miles, twice the size of Massachusetts but with one-third of the population (roughly 2 million). On March 13, 2007, the Board of Supervisors of San Bernardino County approved its 25-year Growth Plan and EIR, which lays out the policies underlying future land use and transportation. It projected a 64 percent increase in population from 2000 to 2020 (Osofsky 2009).\n\nWith California's climate law freshly in the books, State-level regulators submitted formal comments criticizing the county for failure to include an assessment of how this growth would impact GHG emissions. In its formal reply, San Bernardino County claimed that these impacts were \"currently speculative\" (607), and that it lacked methodological guidance from State authorities for carrying out such a calculation. In April 2007, unimpressed by the country's response, the California Attorney General initiated _The People of California v. San Bernardino County_ , a lawsuit alleging regulatory violations under CEQA for insufficient disclosure and analysis of the project's impacts on climate change.\n\nThe dispute amounted to a test of the relative regulatory power of the State and the county within the emerging California climate regime. Can a state environmental law on the control of GHG emissions be used to compel local government to change the way it approaches urban growth? What should be the appropriate balance of regulatory power across States and their counties?\n\nIn August 2007, both parties to the lawsuit agreed to settle the case out of court. As part of the settlement, San Bernardino County agreed to develop an inventory of all known sources of GHGs currently existing in its territory, an estimate of the current amount of gases emitted per year, a projection of emissions based on current growth plans, and a new policy target for mitigation. The county agreed to all of this, though it got the State to concede that no data sources currently exist to build this accounting system.\n\nPolitical settlement was achieved through a new system of GHG accounting and was predicated on a form of shared epistemic jurisdiction. On the one hand, the State holds the power to determine the kinds of gases that need to be counted and mandate the methods of assessment. California, in other words, sets the frame for technical analysis. The county, on the other hand, has the authority to collect the data and perform the analysis, as well as to develop a mitigation strategy based on that analysis.\n\nThis was simultaneously a political and epistemic settlement of climate change's jurisdictional politics. What are the rights and obligations of localities in the face of globally caused phenomena, and what forms of legal and epistemic standing should they have? This was the question in _Massachusetts v. EPA_ and it reappears in San Bernardino County. In each case, we see how the constitution of multi-level carbon regimes is disputed on the terrain of regulatory science and shaped by the assignment of epistemic jurisdiction. But just like political jurisdiction in an age of complex and dynamic governance, epistemic jurisdiction is divided and shared.\n\n## **Conclusion**\n\nThis chapter represents a response to what Sovacool and Brown (2009) have called \"entanglements of scale\" in the dual keys of STS and legal studies. Although overlapping and \"polycentric\" forms of governance offer certain advantages, they also introduce forms of cross-level conflict that are giving rise to new multi-level formations. Science is rarely absent as regulatory power is negotiated across levels, sometimes functioning as a foundation for agreement. More often it is a sticking point.\n\nFor legal and STS scholars alike, the cases in this chapter help underscore the politics of epistemic jurisdiction\u2014the power to produce, interpret, or warrant technical knowledge for a given political community, topical arena, or geographical territory\u2014and how it is involved in processes of political (de)centralization. The food safety case illustrates how the operation of technical reason and the polycentric architecture of governance are co-evolving. _Massachusetts v. EPA_ illustrates how the construction of a new vision of split-sovereignty grows from the epistemic authority claimed by the Court and assigned to the IPCC. In the San Bernardino case, California went on the offensive against one of its constituent counties, mandating a new regime of carbon accounting, while relinquishing the basic power to interpret that data.\n\nThe production and negotiation of epistemic jurisdiction, of course, raises critical questions for democracy. In ways that are not always obvious, processes of forming and deploying epistemic jurisdiction enact\u2014both tacitly and explicitly\u2014principles of subsidiarity, standing, pre-emption, and quasi-sovereignty. The technical veil of epistemic jurisdiction is worrisome given the tangible stakes for political representation and accountability. However, the development of shared epistemic jurisdiction within new federalisms suggests an interesting solution to an old problem in STS: the concern about how diverse and distributed worldviews can be accommodated in larger political and economic integrations. If the Federal principle is applied to epistemic jurisdiction, one begins to have a political language for the sharing of the power to know across scales.\n\n Interview with members of the SPS Secretariat, Geneva Switzerland, 2006\u20132007.\n\n The others are enumerated as the International Office of Epizootics and the Secretariat of the International Plant Protection Convention. (SPS, Annex A(3)).\n\n Formally called, _European Communities\u2014Measures Affecting the Approval and Marketing of Biotech Products_ (Dispute DS291), initiated Aug. 8, 2003 (WT\/DS291\/23).\n\n See ICLEI \u2013 the global cities network. \u00adw\u00adw\u00adw\u00ad.\u00adi\u00adc\u00adl\u00ade\u00adi\u00ad.\u00ado\u00adr\u00adg\u00ad\/\u00ad\n\n Codified at 42 U.S.C. \u00a77401 et seq.\n\n _Massachusetts v. Environmental Protection Agency_ , 549 U.S. 497 (2007).\n\n _Rocky Mountain Farmers Union v. Goldstene_ 843 F.Supp.2d 1042, (E.D.C.A.) 2011.\n\n California Environmental Quality Act (\"CEQA\"; Pub. Resources Code, \u00a7 21000 et seq.).\n\n## **References**\n\nBiermann, F. and Pattberg, P. (2008) \"Global environmental governance: taking stock, moving forward,\" _Annual Review of Environmental Resources_ , 33: 277\u2013294.\n\nBonneuil, C. and Levidow, L. (2012) \"How does the world trade organisation know? The mobilization and staging of scientific expertise in the GMO trade dispute,\" _Social Studies of Science_ , 42(1): 75\u2013100.\n\nBoyd, W. (2010) \"Climate change, fragmentation, and the challenges of global environmental law: elements of a post-Copenhagen assemblage,\" _University of Pennsylvania Journal of International Law_ , 32: 457\u2013550.\n\nBrunsson, N. and Jacobsson, B. (2000) _A World of Standards_. Oxford: Oxford University Press.\n\nBusch, L. (2012) \"Standards: recipes for reality,\" _Rural Sociology_ , 77: 490\u2013492.\n\nCash, D. W., Adger W., Berkes, F., Garden, P., Lebel, L., Olsson, P., Pritchard, L., and Young, O. (2006) \"Scale and cross-scale dynamics: governance and information in a multilevel world,\" _Ecology and Society_ , 11(2): Article 8 online] [\u00adw\u00adw\u00adw\u00ad.\u00ade\u00adc\u00ado\u00adl\u00ado\u00adg\u00ady\u00ada\u00adn\u00add\u00ads\u00ado\u00adc\u00adi\u00ade\u00adt\u00ady\u00ad.\u00ado\u00adr\u00adg\u00ad\/\u00adv\u00ado\u00adl\u00ad1\u00ad1\u00ad\/\u00adi\u00ads\u00ads\u00ad2\u00ad\/\u00ada\u00adr\u00adt\u00ad8\u00ad\/\u00ad (accessed 31 July 2013).\n\nCharnovitz, S. (2000) \"The supervision of health and biosafety regulation by world trade rules.\" _Tulane Environmental Law Journal_ , 13: 271\u2013302.\n\nChimerinski, E., Daniels, B., Hardy, B., Profeta, T., Schroeder, C. H. and Siegel, N. S. (2008) \"California, climate change, and the constitution,\" _The Environmental Forum_ , 25(4): 50\u201363.\n\nDunn, E. C. (2005) \"Standards and person-making in East Central Europe,\" in A. Ong and S.J. Collier (eds). _Global Assemblages: Technology, Politics, and Ethics as Anthropological Problems_ , Oxford: Blackwell, pp. 173\u2013193.\n\nEdwards, P. N. and Schneider, S. H. (2001) \"Self-governance and peer review in science-for-policy: the case of the IPCC second assessment report,\" in C. Miller and P. N. Edwards (eds), _Changing the Atmosphere: Expert Knowledge and Environmental Governance_ , Cambridge, MA: MIT Press, pp. 219\u2013246.\n\nEngel, K. H. (2006) \"Harnessing the Benefits of Dynamic Federalism in Environmental Law,\" _Emory Law Journal_ , 56: 159\u2013188.\n\nFarber, D. (2008) \"Climate Change, Federalism, and the Constitution,\" _Arizona Law Review_ , 40: 879.\n\nFDA (Food and Drug Administration), \"Advancing regulatory science,\" \u00adw\u00adw\u00adw\u00ad.\u00adf\u00add\u00ada\u00ad.\u00adg\u00ado\u00adv\u00ad\/\u00ads\u00adc\u00adi\u00ade\u00adn\u00adc\u00ade\u00adr\u00ade\u00ads\u00ade\u00ada\u00adr\u00adc\u00adh\u00ad\/\u00ads\u00adp\u00ade\u00adc\u00adi\u00ada\u00adl\u00adt\u00ado\u00adp\u00adi\u00adc\u00ads\u00ad\/\u00adr\u00ade\u00adg\u00adu\u00adl\u00ada\u00adt\u00ado\u00adr\u00ady\u00ads\u00adc\u00adi\u00ade\u00adn\u00adc\u00ade\u00ad\/\u00add\u00ade\u00adf\u00ada\u00adu\u00adl\u00adt\u00ad.\u00adh\u00adt\u00adm\u00ad. (accessed 30 January 2013).\n\nFoucault, M. (1971) [1966]. _The Order of Things: An Archaeology of the Human Sciences_. New York: Pantheon Books.\n\nGoldman, M. (2005) _Imperial Nature: The World Bank and Struggles for Social Justice in the Age of Globalization_ , New Haven, CT: Yale University Press.\n\nGieryn, T. F. (1999) _Cultural Boundaries of Science: Credibility on the Line_. Chicago, IL: University of Chicago Press.\n\nHaas, P. M. (1992) \"Introduction: Epistemic communities and international policy coordination.\" _International Organization_ , 46(1): 1\u201335.\n\nHooghe, L. and Marks, G. (2001) \"Types of multi-level governance.\" European Integration online Papers (EIoP) 5(11), \u00adh\u00adt\u00adt\u00adp\u00ad:\u00ad\/\u00ad\/\u00ade\u00adi\u00ado\u00adp\u00ad.\u00ado\u00adr\u00ad.\u00ada\u00adt\u00ad\/\u00ade\u00adi\u00ado\u00adp\u00ad\/\u00adt\u00ade\u00adx\u00adt\u00ade\u00ad\/\u00ad2\u00ad0\u00ad0\u00ad1\u00ad-\u00ad0\u00ad1\u00ad1\u00ada\u00ad.\u00adh\u00adt\u00adm\u00ad (accessed 29 July 2012).\n\nJasanoff, S. (ed.)(1987) \"Contested boundaries in policy-relevant science,\" _Social Studies of Science_ , 17: 195\u2013230.\n\n\u2014\u2014 (1990) _The Fifth Branch: Science Advisors as Policymakers_. Cambridge, MA: Harvard University Press.\n\n\u2014\u2014 (1995) \"Product, process, or programme: three cultures and the regulation of biotechnology,\" in M. Bauer (ed.). _Resistance to New Technology_ , Cambridge: Cambridge University Press, pp. 311\u2013331.\n\n\u2014\u2014 (2004) _States of Knowledge: The Co-production of Science and Social Order_. London: Routledge.\n\nJoerges, C. (1997) \"Scientific expertise in social regulation and the European Court of Justice: Legal frameworks for denationalized governance structures,\" in C. Joerges, K.-H. Ladeur, and E. Vos (eds), _Integrating Scientific Expertise into Regulatory Decision-making_ , Baden: Nomos Verlagsgesellschaft, pp. 295\u2013324.\n\nKahn, J. (2012) _Race in a Bottle: The Story of BiDil and Racialized Medicine in a Post-Genomic Age_. New York: Columbia University Press.\n\nLong-Martello, M. and Jasanoff, S. (eds) (2004) _Earthly Politics: Local and Global in Environmental Governance_ , Cambridge, MA: MIT Press.\n\nMiller, C. A. (2005) \"New Civic Epistemologies of Quantification: Making Sense of Indicators of Local and Global Sustainability,\" _Science, Technology, & Human Values_, 30(3): 403\u2013432.\n\n\u2014\u2014 (2007) \"Democratization, international knowledge institutions, and global governance,\" _Governance_ , 20: 325\u2013357.\n\nMoore, K., Kleinman, D.L., Hess, D. and Frickel, S. (2011) \"Science and neoliberal globalization: a political sociological approach,\" _Theory and Society_ , 40: 505\u2013532.\n\nNelkin, D. (1984) _Controversy: Politics of Technical Decisions_ , 2nd edn., Beverly Hills, CA: Sage Publications.\n\nNicolaidis, K. and Howse, R. (eds)(2001) _The Federal Vision: Legitimacy and Levels of Governance in the US and the EU_ , New York: Oxford University Press.\n\nOreskes, N. and Conway, E. M. (2010) _Merchants of Doubt: How a Handful of Scientists Obscured the Truth on Issues from Tobacco Smoke to Global Warming_ , New York: Bloomsbury Press.\n\nOsofsky, H. M. (2009) \"Is climate change 'international'?: litigation's diagonal regulatory role,\" _Virginia Journal of International Law_ , 49: 585\u2013650.\n\nOstrom, E. (2009) \" _A polycentric approach for coping with climate change_ ,\" Report prepared for the WDR2010 Core Team, Development and Economics Research Group, World Bank, October 2009, \u00adw\u00adw\u00adw\u00ad-\u00adw\u00add\u00ads\u00ad.\u00adw\u00ado\u00adr\u00adl\u00add\u00adb\u00ada\u00adn\u00adk\u00ad.\u00ado\u00adr\u00adg\u00ad\/\u00ade\u00adx\u00adt\u00ade\u00adr\u00adn\u00ada\u00adl\u00ad\/\u00add\u00ade\u00adf\u00ada\u00adu\u00adl\u00adt\u00ad\/\u00adW\u00adD\u00adS\u00adC\u00ado\u00adn\u00adt\u00ade\u00adn\u00adt\u00adS\u00ade\u00adr\u00adv\u00ade\u00adr\u00ad\/\u00adI\u00adW\u00ad3\u00adP\u00ad\/\u00adI\u00adB\u00ad\/\u00ad2\u00ad0\u00ad0\u00ad9\u00ad\/\u00ad1\u00ad0\u00ad\/\u00ad2\u00ad6\u00ad\/\u00ad0\u00ad0\u00ad0\u00ad1\u00ad5\u00ad8\u00ad3\u00ad4\u00ad9\u00ad_\u00ad2\u00ad0\u00ad0\u00ad9\u00ad1\u00ad0\u00ad2\u00ad6\u00ad1\u00ad4\u00ad2\u00ad6\u00ad2\u00ad4\u00ad\/\u00adR\u00ade\u00adn\u00add\u00ade\u00adr\u00ade\u00add\u00ad\/\u00adP\u00adD\u00adF\u00ad\/\u00adW\u00adP\u00adS\u00ad5\u00ad0\u00ad9\u00ad5\u00ad.\u00adp\u00add\u00adf\u00ad (accessed 29 July 2012).\n\nParthasarathy, S. (2007) _Building Genetics Medicine: Breast Cancer, Technology, and the Comparative Politics of Health Care_ , Cambridge, MA: MIT Press.\n\nPeel, J. (2010) _Science and Risk Regulation in International Law_. Cambridge: Cambridge University Press.\n\nPew Center on Global Climate Change (PCGCC) (2011) _Climate Change 101: State Action_. \u00adw\u00adw\u00adw\u00ad.\u00adc\u00ad2\u00ade\u00ads\u00ad.\u00ado\u00adr\u00adg\u00ad\/\u00ads\u00adc\u00adi\u00ade\u00adn\u00adc\u00ade\u00ad-\u00adi\u00adm\u00adp\u00ada\u00adc\u00adt\u00ads\u00ad\/\u00adc\u00adl\u00adi\u00adm\u00ada\u00adt\u00ade\u00ad-\u00adc\u00adh\u00ada\u00adn\u00adg\u00ade\u00ad-\u00ad1\u00ad0\u00ad1\u00ad\/\u00ads\u00adt\u00ada\u00adt\u00ade\u00ads\u00ad\n\nRabe, B. (2004) _Statehouse and Greenhouse: The Emerging Politics of American Climate Change Policy_ , Washington, DC: Brookings Institution Press.\n\nRuhl, J. B. and Salzman, J. (2010) \"Climate change, dead zones, and massive problems in the administrative state: a guide for whittling away,\" _California Law Review_ , 98: 59\u2013120.\n\nSarewitz, D. (2004) \"How science makes environmental controversies worse,\" _Environmental Science & Policy_, 7: 385\u2013403.\n\nSovacool, B. K and Brown, M. A. (2009) \"Scaling the policy response to climate change,\" _Policy and Society_ , 27: 317\u2013328.\n\nThompson, C. (2004) \"Co-producing CITES and the African elephant,\" in S. Jasanoff (ed.), _States of Knowledge: The Co-production of Science and Social Order_ , London: Routledge, pp. 67\u201386.\n\nTimmermans, S. and Epstein, S. (2010) \"A world of standards but not a standard world: toward a sociology of standards and standardization,\" _Annual Review of Sociology_ , 36: 69\u201389.\n\nWickson, F. and Wynne, B. (2012) \"The anglerfish deception.\" _EMBO Reports_ , 13(2): 100\u2013105.\n\nWinickoff, D. E. and Bushey, D. (2010) \"Science and power in global food regulation: the rise of the codex alimentarius,\" _Science, Technology, and Human Values_ , 35: 356\u2013381.\n\nWinickoff, D. E., Jasanoff, S., Busch, L., Grove-White, R. and Wynne, B. (2005) \"Adjudicating the GM food wars: science, risk, and democracy in world trade law,\" _Yale Journal of International Law_ , 30: 81\u2013123.\nChapter 11\n\n* * *\n\n# **Un-knowing exposure**\n\nToxic emergency housing, strategic inconclusivity and governance in the US Gulf South\n\n_Nicholas Shapiro_\n\n* * *\n\n> There is not one but many silences, and they are an integral part of the strategies that underlie and permeate discourses.\n> \n> (Michel Foucault 1978: 27)\n\nLate in the summer of 2005, two hurricanes\u2014Katrina, then Rita\u2014tore into the central Gulf Coast of the United States. The damage spanned hundreds of miles of coastline, from East Texas to the Southwestern tip of Alabama. Between 1 and 1.5 million people were displaced. Regionally available rental units, hotels, and motels quickly reached capacity. After months spent shifting between the homes of friends, relatives, and shelters, some 300,000 internally displaced persons were eventually housed in approximately 120,000 travel trailers supplied by the Federal Emergency Management Agency (FEMA). These 250-square foot emergency housing units were deployed both to the driveways of homes-in-ruin, and to federally run trailer parks for those that possessed no real estate of their own.\n\nWithin a fortnight of the inundation of New Orleans and the ravaging of the central Gulf Coast, a single manufacturer, Gulf Stream Inc., had secured over $500 million in contracts to build temporary housing units. New plants were opened to accommodate these rush orders. As one Gulf Stream employee recounted to the House Committee on Oversight and Government, \"There would be supervisors and plant managers walking down the line with bullhorns screaming at us, go faster, you're not doing your job, you don't deserve your paycheck.\" Many laborers on the assembly floor experienced compromised health while working to meet these large expedited orders; symptoms ranged from constant sinus infections and shortness of breath to bloody eyes and noses. When employees complained of the stench and \"stickiness\" of the fiberboard and lauan paneling, they were met with little explanation or reaction from their supervisors other than noting that it came from a \"new supplier\" (Majority Staff Analysis 2008: 10\u201311).\n\nAs displaced Gulf Coast residents began to inhabit these emergency housing units and fall ill, complaints began to trickle back to FEMA and trailer manufactures. \"There is an odor in my trailer that will not go away,\" read a message sent in March 2006 by an inhabitant of a Gulf Stream-made FEMA trailer to the company via a comment feature on their website. The complaint continued, \"It burns my eyes and I am getting headaches every day. PLEASE, PLEASE HELP ME!!\"\n\nConcurrently, other trailer residents began to investigate the chemical constitution of their new homes after experiencing increased asthma attacks, bloody noses, insomnia, chronic fatigue, headaches, diarrhea or dermatological irritation. On his first morning in his new FEMA trailer \u2013 installed in front of what little was left of his Mississippi home \u2013 Paul Stewart woke up to find his pet cockatiel inactive on the floor of his cage, his wife bleeding from her nose, and a burning sensation in his eyes, nose and throat. Paul, a businessman and environmental activist, began researching inexpensive means of assessing the quality of the trailer's indoor air. He found a $39 mailorder formaldehyde test kit, which, when run in his home, yielded a reading of 0.22 parts per million (ppm), more than twice the 0.1ppm maximum indoor level recommended by the U.S. Environmental Protection Agency (EPA). In collaboration with another local environmental activist, Becky Gillette, and equipped with funding from the Sierra Club (an American grassroots environmental organization), the two assessed the atmospheric formaldehyde content of 69 FEMA trailers in Mississippi. They found 61 units, or 88 percent, to contain indoor formaldehyde in concentrations higher than the EPA's 0.1ppm threshold.\n\nThis informal citizen-science assessment drew international media attention and incited both a crisis in post-disaster governance and a multi-million dollar lawsuit. In this chapter I unpack the multiplicity of scientific processes that refute, dilute, and disqualify claims of FEMA trailer-related formaldehyde exposure and ensuing negative health outcomes in both the immediate governmental response and later court proceedings. I document how multiple and divergent technical practices weave together into a larger regime of chemical exposure and ensuing illnesses obfuscation, what I later call \"un-knowing.\" The methods of un-knowing analyzed herein span the realms of both federal controversy governance and the application of science in court. While the actors, instruments and techniques of un-knowing shift between these domains, the outcomes remain the same. Such an analysis brings into relief the negative epistemological space that shapes the contours of both technoscientific and legal facts.\n\n## **Techniques of un-knowing**\n\nHistorian of science Michelle Murphy observed a similar uniformity of toxic injury dismissal in her wide-ranging study on the chemical exposure complaints of female office workers beginning in the 1980s, a phenomenon that came to be referred to as \"sick building syndrome\" (Murphy 2006). Murphy contends that the refutation of toxic harm by the dominant means of scientific evaluation was and still is structured by \"regimes of imperceptibility.\" Regimes of imperceptibility make sense of the patterned outcomes of scientific explanatory systems that time and time again found low-dose chemical exposures to be inconclusively toxic, undeserving of redress, and without an onus for remediation. Yet, attending to the existence of regimes of imperceptibility is not the same as explicating _how_ such conclusions were reached. This chapter thus elaborates Murphy's analysis by enumerating and theorizing the diverse means of un-knowing by which regimes of imperceptibility are achieved. Some facets of toxic inconclusivity are the inevitable affordances of different scientific objects or exposure events, others are byproducts of shifting diagnostic criteria, while many more are strategically sculpted.\n\nScholarship on the latter category, the production of scientific indeterminacy, is rapidly developing within science and technology studies (STS) and affiliated disciplines. These studies largely document practices of inconvenient knowledge avoidance or contesting the causal relationship between products and adverse health outcomes in the pharmaceutical (Corrigan 2002, Avorn 2006, Abraham and Davis 2006) and tobacco (Proctor 2006, Michaels 2008) industries. Linsey McGoey asserts that studies on the strategic deployment of ignorance are of increasing urgency in a democratizing and neoliberalizing world as such manipulations of the known are particularly well suited in contexts where public scrutiny threatens the viability of plain-old state and corporate secrecy (McGoey 2007, cf. Galison 2004).\n\nThe empirical contribution of this chapter to the emergent literature on strategic unknowns is an apprehension of the multiple techniques of inducing ignorance surrounding a single issue. My analysis moves through several scales, arenas of governance, and scientific practices. The various modalities of scientific incomprehension, enumerated below, are part and parcel of the active and dynamic process of moving away from perception, admittance, or substantiation of mass chemical exposure, a discursive current I refer to collectively as \"un-knowing.\"\n\nUn-knowing encapsulates the techniques by which ignorance is deployed, the privileges of secrecy are exercised, and the methods by which the facticity of exposure-related illnesses are called into question. Un-knowing includes both intentional strategic ignorance and unwitting discursive ignorance. As I shift my analytical gaze from harried federal denial of potential chemical health threats to the meticulously choreographed testimony of expert witnesses in the ensuing litigation, the filaments of un-knowing charted in this chapter scale down from the brash and systematic exercise of governmental control to more sophisticated means of disavowing negative health outcomes. What is at stake in these techniques of un-knowing is how we cohabitate with low-dose toxicity, how we understand the human health effects of these exposures, and the allocation of responsibility for such impairments.\n\nApprehending the multiple and protean tactics that aggregate into unknowing connections between chemical exposure and health effects demands a diversity of theoretical approaches. I employ a triad of perspectives in the hopes of amassing a forensic toolbox for the scrutiny of un-knowing. I first make use of the STS focus on the avoidance of inconvenient knowledge, then utilize the Foucauldian analytics of the subjugation of knowledge (Foucault 2003) to understand the manipulation of toxicological thresholds in a federal assessment of chemical levels in the FEMA trailers. Transitioning into the courts, I track the ontological choreography (Cussins 1996) used by expert witnesses in the courtroom to evaporate causal links between formaldehyde exposure and disease outcomes, specifically asthma. The chapter closes with a focus on how the pharmaceuticalization (Biehl 2007, Whitmarsh 2008) of asthma facilitates the exclusion of environmental triggers, such as formaldehyde, from equations of disease causation. This unintentional and reductionistic facet of biomedicine's definition of asthma is then strategically utilized by a defense expert witness in court.\n\nThis chapter draws on participant-observation in one of the law offices representing the plaintiffs in the FEMA trailer formaldehyde litigation, deposition transcripts, expert affidavits, internal FEMA emails disclosed during a congressional investigation, interviews in Arkansas, Louisiana, Massachusetts, and Georgia and is situated within a two-year ethnographic study tracking the social, scientific, and economic lives of the FEMA trailers.\n\n## **Forestalling assessment**\n\nIn early spring of 2006, as many families were still moving into FEMA trailers, the Sierra Club circulated a press release about their grassroots indoor formaldehyde testing campaign and media attention began to mount. In response, a FEMA spokesperson announced in May, \"FEMA and industry experts are monitoring the small number of cases where odors of formaldehyde have been reported, and we are confident that there is no ongoing risk.\" Behind the serenity of their official statements, FEMA was divided in planning their response to the issue. Since mid-March internal emails were circulating among FEMA field offices which advised an immediate and proactive response.\n\nOn May 18 a lawsuit, which would eventually swell to over fifty thousand plaintiffs, was filed in the Easter District of Louisiana. FEMA trial attorney Patrick \"Rick\" Edward Preston was assigned to the case in June and was simultaneously propelled to the forefront of FEMA's scientific inquiry into the formaldehyde issue (Minority Staff Report 2008: 10). One day after he was appointed to the litigation, Preston sent an email vetoing a swift evaluation:\n\n> Do not initiate testing until we give the OK. While I agree that we should conduct testing, we should not do so until we are fully prepared to respond to the results. Once you get results and should they indicate some problem, the clock is ticking on our duty to respond to them.\n\nThe institutional liability of scientifically verifying the hazards of chemical exposure outweighed such knowledge's utility in informing citizens or interdicting the corporeal risks of exposure that continue to accrue with time. This form of forestalling the scientific and systematic corroboration of FEMA trailer related illnesses was the first and most straightforward instance of the Agency's un-knowing of exposure.\n\nIn June 2006, FEMA, the Agency for Toxic Substances and Disease Registry (ATSDR), the Environmental Protection Agency (EPA), and the Centers for Disease Control and Prevention (CDC) began conducting biweekly interagency conference calls to address escalating public relations, public health, and litigation concerns. Through these calls it was eventually decided in August that the EPA would run tests in September and October of _unoccupied_ trailers to test what methods of ventilation were most efficacious. Environmental activist Becky Gillette, a central organizer of the Sierra Club's testing efforts, felt triumphant for garnering federal attention rather than weary of investigations shaped in part by FEMA; as she related to me in her Arkansas home, \"I was happy when I heard the EPA was testing in October, I was glad that the professionals were coming in to verify.\"\n\nGillette's faith began to ebb as the New Year came and went and no results were made public; as she recounted, \"they really dragged their feet and so I wrote a FOIA [Freedom of Information Act letter] in February to force them to release the results.\" While the EPA had conducted the tests, they had then delivered the data unanalyzed to FEMA who was then supposed to forward them to ATSDR for analysis. In early December of 2006 ATSDR was still awaiting delivery of the test results from Rick Preston's Office.\n\nSoon thereafter Preston passed on the data with the stipulation that all ATSDR analyses would remain confidential \"No information should be released to any third party without my express permission,\" Preston mandated in an email to ATSDR. Duly, the report remained secret until FEMA issued a press release in May of 2007 \u2013 three months after ATSDR sent its final report to Preston's office and eight months after the tests were conducted. In addition to avoiding chemical assessment, secrecy was also utilized to the extent permissible by contemporary democratic structures.\n\n### **_Leveling concern_**\n\nSetting the correct exposure threshold for FEMA trailers has been contentious from the start. When the Agency for Toxic Substances and Disease Registry (ATSDR) analyzed the potential hazards posed by the indoor air quality of the FEMA trailers in 2007, they disregarded their own intermediate (14\u2013364 days) and long-term (over 1 year) formaldehyde minimum response levels of 0.03ppm and 0.008ppm, respectively. Instead, in drafting the report, the authors crafted their own standard, which they dubbed \"the level of concern\" (ATSDR 2007) and has, according to former ATSDR director Dr. Howard Frumkin, \"little or no operational meaning\" (Minority Staff Analysis 2008: 17). The level of concern was set at 0.3ppm\u2014ten to thirty-seven times the concentration of formaldehyde that the Agency deemed capable of causing adverse health effects. This 0.3ppm threshold dismissed the illnesses that stricter guidelines were intended to prevent, labeling them mere \"nuisance\" symptoms. The Agency's report found that by opening all windows, static vents, and exhaust fan vents, the indoor formaldehyde levels of FEMA trailers would, on average, fall below their \"level of concern\" and, therefore, the trailers should not be expected to produce adverse health effects (ATSDR 2007).\n\nThis tidy conclusion served to, temporarily, quell rising concern about the chemical consequences of inhabiting a FEMA trailer. The study was designed not to assess the chemical levels of FEMA trailers in the state in which people lived but rather to see if it was possible to bring the indoor formaldehyde levels within a \"safe\" range. The conclusion of safety was based upon a best-case scenario that was inappropriate for brutal Gulf Coast summers, when trailer windows need to be closed and air conditioning set at full tilt in order to fend-off withering heat and humidity.\n\nFrank Renda (a pseudonym), then a senior environmental scientist at ATSDR, was alarmed by the report when it landed on his desk for approval:\n\n> [The consultation] was dated February 1st [2007], I saw it on about the 17th of February; within about two hours I called my senior management and said that we had a problem. The consultation as it was written was misleading, possibly misleading, and a potential public health threat.\n\nAs Renda recalled over lunch in Atlanta, he traced the study design's lineage to the FEMA attorney, Rick Preston, who requested ATSDR analysis:\n\n> They had been directed by FEMA first of all not to share it with anyone [and secondly] that they were only to address the shorter term. That was the thing, I didn't have to go into any in-depth review to know that we had missed the boat, we missed the mark, that FEMA had gotten what they had asked for.\n\nRenda's technical misgivings and the patent influence of a lawyer managing FEMA's liability did not move Renda's superiors into amending the consultation. It was sent to FEMA and only resurfaced later in the summer when, in Renda's words, \"things began to break loose.\" In May the report was used in a FEMA press release to justify inaction. Following this, the two junior scientists that penned the consultation were raked over the coals by senior management. ATSDR's senior leadership then turned to Renda and asked him to lead an across-Agency work group to develop recommendations for remediating the report. Renda's committee produced a revised document that bore increased precautions:\n\n> One of the points stated, \"given the hazards posed by [...] the trailers and the formaldehyde exposures, that efforts should be undertaken in the areas of health education and that appropriate measures to interdict exposures should be implemented.\" And after that there was all of a sudden some sudden displeasure with that I had done. It was removed from my oversight and the executive summary was revised to say, \"analyze\" as opposed to \"implement.\" Paralysis by analysis.\n\nRenda was not only removed from his role supervising the revision of the FEMA trailer consultation but was soon notified of his removal from his position at ATSDR and his placement on a \"a performance improvement plan\"\u2014a slow, bureaucratic means of firing.\n\nFurther criticism was leveled against the 0.3ppm \"level of concern\" by Dr. Vincent Garry, a pathologist and reviewer of ATSDR's 1999 toxicological profile on formaldehyde. In a March 2008 letter to Representative Brad Miller, Chairman of the Subcommittee on Investigations and Oversight, Garry pointed out the inapplicability of such a threshold of concern because of the authors' reliance on occupational exposure studies which are based upon eight hours of exposure per day for five days a week, whereas, for many trailer residents, \"this is a 24 hr per day 7 days per week exposure\" (Garry 2008: 1). In addition to disjunctures between the exposure durations of the industrial exposure literature drawn upon and the residential case which it was applied to, Garry noted the physiological differences between the largely healthy adult males that composed the occupational studies and the wideranging biological makeup of trailer inhabitants, \"For example, children under age 2 have a short trachea and breathe faster than adults (30\u201340 breaths per minute), therefore, process more formaldehyde into the body and are probably less efficient in the metabolism of the chemical\" (ibid.). He concludes his letter by diagnosing a lack of peer-reviewing and robust intra-Agency communication as the underlying issue that led to the usage of this technical miscalculation as the central benchmark in ATSDR's report.\n\nThese issues, highlighted by Garry and Renda, are not isolated to this particular health consultation but rather are symptoms of a longstanding lack of oversight and shortcomings in study design. Such underestimations of harm are the result of techniques of toxicological investigation employed by ATSDR since its creation.\n\nFrom its early days in the 1980s through to the present, the Agency has been accumulating complaints regarding the irresolution of their toxic exposure assessments. The U.S. General Accounting Office, in a 1991 review, found ATSDR's results to be \"seriously deficient as public health analyses\" (GAO 1991: 2). The Environmental Health Network and the National Toxins Campaign Fund asserted in a 1992 study that ATSDR's studies were \"inconclusive by design,\" elucidating their argument with multiple accounts of toxic exposures in the Gulf Coast that were deemed by ATSDR to pose no health risk (Russell _et al._ 1992). Corroborating this statement, one current ATSDR scientist, bearing witness anonymously, testified that, \"It seems like the goal is to disprove the communities' concerns rather than actually trying to prove exposures\" (ATSDR Staff Report 2009: 2).\n\nThe technical means of disavowing the potential chemical harm of these emergency housing units falls squarely within a domain of inquiry that the French philosopher Michel Foucault referred to as \"subjugated knowledges\"; knowledges \"that have been disqualified as non-conceptual knowledges, as insufficiently elaborated knowledges [...] that are below the required level of erudition or scientificity\" (Foucault 2003: 7). The 0.3ppm level of concern disqualifies the \"nuisance\" effects of formaldehyde exposure as non-conceptual. It submerges the constant headaches, fatigue, rashes, coughing, and diarrhea experienced by FEMA trailer residents under an ostensibly higher level of scientificity, disavowing their reality as a true or significant health consequence of exposure.\n\nThe analytics of power\/ignorance\u2014largely focusing on knowledge avoidance\u2014and subjugated knowledges are complimentary as evinced by the study design of the ATSDR FEMA trailer health consultation. The EPA-collected data on trailers that were ventilated to an extent beyond practical possibility was an apparatus of hazardous knowledge avoidance, while the parameter of analysis\u2014'the level of concern'\u2014was an instrument of subjugation. The threshold and the policy decisions it informed subjected tens, if not hundreds, of thousands of Gulf Coast residents to sustained formaldehyde exposures, for \"at least one year longer than necessary\" (Minority Staff Report 2008: 1). It further disqualified the existing scientific recommendations for maximum domestic formaldehyde concentrations\u2014including those of ATSDR itself\u2014as overly precautious and attempted to historically mask the hazards manifest in the exposure data it was charged with adjudicating. These technical maneuvers are tempered by the Agency's distinct history of evaporating concerns about the public's cohabitation with toxics, privileging liabilities of fellow federal agencies and industry over population health.\n\n## **Ontologies of expertise**\n\nThis section moves from an analysis of the often-blunt means of liability avoidance, deferral, or disqualification deployed as techniques of governance, to attend to the more indirect ways through which scientific ambiguity and inconclusivity is leveraged vis-\u00e0-vis the ever-shifting ontology of the \"expert witness.\" This is documented through an exemplary test case, technically known as a \"bellwether trial.\" The trail of focus here revolves around Christopher Cooper, who was eight years old when Katrina hit and the flood protection system of New Orleans failed, destroying his family home.\n\nChristopher was diagnosed with asthma when he was three. Since his diagnosis he would use a corticosteroid \"rescue\" inhaler one or two times a month during (mostly sports-induced) asthma attacks. After the hurricane and six-months of displacement in Florida, Christopher, his mother, and his elder sister moved into a FEMA trailer parked in the driveway of their flooded home. Soon thereafter, the frequency of Chris' asthma attacks increased fourfold. This worsening of his asthma is the basis of Chris' litigious claim for redress.\n\nDr. Kenneth Smith is the director of the intensive care unit at East Jefferson General Hospital, just West of New Orleans, and a practicing respiratory disease specialist at Jefferson Pulmonary Associates. Smith was hired by the defense council of the FEMA trailer formaldehyde litigation to refute claims of both general formaldehyde-asthma causation and the specific causation of the exacerbated asthma of Christopher Cooper. In the following excerpt from his deposition, a plaintiff attorney asked Smith whether a series of statements were true or false\u2014a line of questioning often used to establish the baseline opinions of expert witnesses:\n\nQ. | First statement: There's a cause-and-effect relation between formaldehyde and asthma.  \n---|---  \nA. | That's a qualified maybe.  \nQ. | Is there a cause-and-effect relationship between formaldehyde and initiation of asthma?  \nA. | Possibly.  \nQ. | Formaldehyde is an irritant to the respiratory system?  \nA. | Possibly. Again, these are all concentration issues. They are potential irritants at the appropriate concentration. We've seen several patients who lived temporarily in FEMA trailers that have asthma and have other illnesses. But that [allergen and toxin exposure] was a very common occurrence in this city after Katrina. [...]  \nQ. | Formaldehyde exposure can have harmful effects on developing lungs in children?  \nA. | I would say that I'm unaware of any data that would suggest that that is the case.  \nQ. | So is that a no?  \nA. | That's a qualified no because I'm unaware of it.\n\n(Metairie, Louisiana, 10 July 2009)\n\nThis strained back and forth continued on for several minutes. To questions revolving around whether or not the scientific literature demonstrated that low-dose formaldehyde exposure can exacerbate asthma, Smith responded \"maybe,\" \"possibly,\" \"I'm unaware of it.\" He also, and to a lesser extent, answered with, \"false.\"\n\nThe latter was his ruling on the statement, \"There is a correlation between formaldehyde exposure and the risk of allergic sensitization.\" Allergic sensitization is one of three potential pathophysiological links between formaldehyde exposure and asthma attacks. After Smith's firm negative answer, the questioning attorney pressed further:\n\nQ. | A definitive not true or you haven't seen any studies that suggest that?  \n---|---  \nA. | I haven't seen any studies that suggest that, and the studies that I have seen suggest that that's not the case.\n\nIn Smith's list of reliance materials, a technical term for everything that he has consulted in formulating his opinion, he listed only seven articles. Of these articles none were studies that scrutinize allergic sensitization in the presence of formaldehyde, and most had little to no bearing upon the trial's claimed concurrence of FEMA trailer residency and childhood asthma exacerbation. Smith produced a firm account of causal uncertainty through his listless engagement with the scientific literature. His window upon the scientific literature was not just small but also strategically sculpted, as evidenced through his responses to questions from the plaintiff attorney:\n\nQ. | The studies that you reviewed that are in your Reliance File, how did you get those? Did you go out and hit the books or hit the computer or did the lawyers provide them? Did someone else provide them to you?  \n---|---  \nA. | Most of the\u2014the documents that I referenced in my opinion paper, except the ones for Up-To-Date, which is an on-line textbook, mega textbook\u2014  \nQ. | Sure.  \nA. | \u2014were provided by defense counsel.\n\nThus, Smith did not produce his expert opinion alone. The intentional ignorance of Smith's testimony is shaped by the tandem efforts of defense attorneys' strategically sanitized reading list and his own restrained research. By way of this combined effort, Smith was able to aver, without fear of perjury, formaldehyde's non-causation of asthma. It was through this calibrated epistemic avoidance that his testimony achieved an ontological fixity of the non-existence of a relationship between formaldehyde and asthma.\n\nSmith's claims were further buttressed by demarcating his expertise: \"I'm not a study wonk,\" he remarked as a plaintiff attorney questioned him about exactly what study it was that refuted sensitization to formaldehyde. Smith continued, when an attorney pressed him on the specifics of a study, \"That's not my deal. I'm a clinician.\" Although Smith has treated at least 50,000 patients with respiratory ailments since he finished his training in 1978, he reported that not a single one of his asthmatic patients have claimed formaldehyde-induced aggravation. He centered his expertise upon his capabilities and knowledge in practice and not his methods of literature inquiry or memory of textual specificities.\n\nContradictorily, however, Smith later stated that his clinical instinct is not what he relies upon. When asked how he formulates his opinions Smith circled back to the literature, which is in conflict with his clinical common-sense:\n\n> Well, it's based on my personal experience, but it's also based on literature and people that have done this. You know, quite frankly, quite frankly, I\u2014I would have thought intuitively, intuitively that any irritant can\u2014can exacerbate asthma. All right? If you ask me as a clinician, I'd say, yeah, virtually any irritant. So it was real interesting to me as a\u2014as a scientist to\u2014to read literature that's from good places that seem to be good literature where that in fact is not the case [...]\u2014that's somewhat surprising, quite frankly. But I believe it.\n\nSmith did not base his expert opinion of formaldehyde-induced asthma exacerbation on what he saw day-in-and-day-out as a clinician, nor did he claim to base his views on the literature alone. Rather, he coordinated his authority between clinic and library\u2014practicing pulmonologist and scientist. This movement \"ontologically choreographed\" (Cussins 1996) Smith's different claims on expertise; i.e., he forged \"a functional zone of compatibility that maintain[ed] referential power between\" his different medical existences (Cussins 1996: 600). Smith's dogged coordination of his expertise deflected further cross-examination from plaintiff attorneys and his expertise was not called into question by the presiding judge.\n\nSmith's expertise arose from a repeated oscillation between the clinical and the scholarly, yet as he jockeyed back and forth he persistently located the seat of his knowledge in his _other_ medical existence. Behind the onto-logical choreography of Smith's expertise that sufficed to maintain his legitimacy in court, was a deep-seated \"phantomatic\" quality to the ontology of his expertise. In her work on waterborne single-celled microorganisms that were simultaneously the suspected and vindicated culprit of large-scale fish die-offs, Astrid Schrader developed the notion of \"phantomatic ontology.\" While the term was coined to describe scientific objects that \"reshape their configurations in different contexts\" (Schrader 2010: 278), I extend it here to the producers, or at least brokers, of scientific knowledge whose own expertise bears \"the paradoxical existence of a specter as neither being nor non-being\" (ibid.).\n\nSmith coordinated his polyvalent expertise to successfully shield his data avoidance from contempt and vulnerability. Yet, on closer inspection the ontology of his expertise that was so readily choreographed, also revealed the vacancy of his epistemic practices. As Lezaun and Woolgar (2013: 321\u2013323) note, ontology and epistemology are not easily parsed. Both are implicated and entangled within broader practices of un-knowing links between formaldehyde and asthma.\n\n## **Pharmaceutical logics**\n\nIn the move from speaking about general causation to specific causation, Smith repositions himself as a clinician. While he concedes that Christopher Cooper did experience more frequent attacks with greater severity while living in the FEMA travel trailer, he does not \"know specifically what caused it.\" He frays the claimed aetiology of formaldehyde exposure not only with his \"will to ignorance\" (McGoey 2007) but also with his clinical knowledge of asthma provocation. As with his testimony negating the general causality of asthma by formaldehyde, these accounts of specific causality are also produced by drawing upon an emergent pharmacological re-definition of asthma's aetiology. This section transitions from an analysis of the likely intentional ignorance of expert witnessing to pervasive biomedical logics that are without personal intent yet are also braided into processes of un-knowing, specifically the exclusion of environmental exposures from definitions of asthma.\n\nIn May of 2009, Christopher Cooper and his mother, Alana Alexander, traveled to National Jewish Hospital on a trip paid for by the plaintiff steering committee. The respiratory wing of this hospital is widely regarded as a global leader in pulmonary care. There, Dr. Karen A. Pacheco ran a battery of tests on Christopher: pulmonary function tests, a challenge with a bron-choconstrictive agent, skin-prick allergy testing, and CT scans of his sinuses and chest. In her affidavit, she concludes that it is \"medically plausible that [Christopher's] asthma would have been aggravated by occupancy of a water damaged trailer with formaldehyde off-gassing\" (2009: 7) but given the lack of baseline medical records (many of which were lost to the 2005 flood waters) she cannot attest to the precise degree of aggravation. \"Nonetheless,\" she continues:\n\n> The patient's test results document moderately severe bronchial hyperresponsiveness as well as a component of fixed airflow obstruction. This likely reflects a long history of asthma that has been suboptimally treated. Although it is common to wish to minimize medication use, especially in children, in those with asthma this is not necessarily a good policy. Chronic untreated asthma can lead to airways remodeling and scarring, with fixed airflow obstruction that no longer completely reverses with inhaled bronchodilator.\n> \n> (Pacheco 2009: 7)\n\nPacheco reads Christopher's pulmonary landscape as sculpted by a longstanding lack of pharmaceutical intervention. This explanation of Christopher's current airway topography falls in line with an even longer history of the biomedical imagination's use of asthma pharmaceuticals to subtract environmental factors from equations of asthma causation or exacerbation.\n\nIn his wide-ranging history of asthma and allergies in the US, Gregg Mitman notes that following the conclusion of the Second World War, \"engineering the body in ways that would overcome the peculiarities of place and environmental change became a consumer and corporate dream\" (2007: 211). In other words, bronchodilating inhalers were used to replace considerations of environmental exposures when thinking about asthma or allergies. This shift in the way asthma was conceptualized aligned the interests of patients-cum-consumers and the pharmaceutical industry and obfuscated many of the underlying environmental triggers. Whitmarsh also notes that, \"increasingly in the early-twentieth-century United States, pharmaceutical intervention was used to define [asthma]\" (Whitmarsh 2008: 61). Alleviation of bronchoconstriction by inhaled pharmaceuticals became foundational in defining what respiratory illnesses were identifiable as asthma. It is upon this foundation that lack of treatment becomes more of a cause of asthma than the environment. In a circular fashion, the treatment of asthma has become part of its definition, and a lack of treatment becomes the cause of asthmatic damage to the lungs. The original impetus for respiratory distress and pharmaceutical innovation, the environment, is displaced by its response. It is under this rubric of pharmaceuticalization (Biehl 2007), that we must understand the assignation of non-environmental causality.\n\nOnly Pacheco's affidavit appears in court, not her person. Her expert report is brought in as an exhibit and then, as an expert witness, Smith translates her medical terminology into more accessible English. At the beginning of his analysis he pauses to \"applaud\" her reading of Christopher's severe bronchial hyperresponsiveness and fixed airflow obstruction as a function of a lack of pharmaceutical intervention rather than the result of a pathophysiology exacerbated by environmental substances. Indeed, Christopher's (and his mother's) responsibility for the changes to his airways is, Smith posits, evinced by \"[t]he fact that he has no alteration in forced expiratory volume in one second that doesn't go normal after the use of a bronchodilator.\" In other words, how much air he can exhale at a baseline is the same as how much air he can exhale immediately after a chemically triggered bronchoconstriction is countered by a fast-acting asthma inhaler. Christopher's test responses place his bronchial sensitivity snugly within the pharmaceuticalized definition of asthma. He should, therefore, be able to remove environmental encounters with asthmogens from asthma's causal horizon.\n\nSmith implies below that pathophysiologies are not asthma; they are merely risk factors for asthma. Asthma becomes a lack of human control of bodily risks in Smith's testimony, \"Christopher actually has mild asthma. He has severe bronchial hyperresponsiveness which may put him at risk for more severe asthma or may\u2014if he's not controlled, he may\u2014he could have long-term sequelae.\" Christopher's asthma is itself mild, although his lungs are very reactive. The control necessary to prevent \"more severe asthma\" is a control of Christopher and not a control of the environment. As Smith renders the ontology of asthma, asthma is not in the body or influenced by the body's surroundings but rather comes into being by how one medicates or does not sufficiently medicate the body.\n\nIn his expert testimony, Smith takes the helm of diffuse practices of pharmaceuticalization and aligns their logics with his own argument. Yet, the momentum of such unwieldy processes affords only minimal maneuverability. The primary agency of separating environmental asthma triggers from rubrics of causation lies beyond individual or institutional agency, as outlined in previous sections, and resides in scientific discourse itself. These pharmaceutical logics are the largest scale and most de-centered aspect of un-knowing documented in this chapter.\n\n## **Conclusion**\n\nIn this chapter I have enumerated a compendium of techniques of un-knowing as they emerged at the interface of the post-disaster governance, law, and the health sciences. In doing so I hope to contribute a toolbox of approaches to the growing body of work on intentional ignorance that is developing in STS and render visible the ways in which socio-legal venues and practices mediate the un-knowing of evidence.\n\nThe practices of un-knowing illnesses in the ill-fated emergency housing units of study are mercurial. They take multiple forms across scale and time, including: secrecy, assessment postponement, scientific disqualification\/knowledge subjugation, knowledge avoidance, and the ontological obfuscation of environmental triggers.\n\nThe above techniques weave together to form the process I have dubbed \"un-knowing.\" Un-knowing answers the technical question \"how did regimes of imperceptibility come to veil the potential harm of formaldehyde in FEMA trailers?\" By documenting the ways in which knowledge can be parried, submerged under thresholds of significance, concealed, and ignored. While the STS literature on \"strategic ignorance\" semantically imputes intentionality, a focus on processes of un-knowing highlights how willful knowledge insufficiencies and unpremeditated discursive blinders, such as pharmaceuticalization, work hand-in-hand and towards similar ends.\n\n## **Acknowledgments**\n\nI would like to thank the editors of this volume and Nadine Levin for their comments on an earlier draft and Fadhila Mazanderani for commenting on an ancestral version of this chapter.\n\n Displaced persons owning or renting apartments and those with homes that did not have sufficient space on their lot for a trailer were also allocated units on federally run group sites.\n\n Over 2 billion dollars in emergency housing contracts would eventually be awarded to over 60 manufacturers.\n\n This digital complaint was forwarded from Scott Pullin to Dan Shea, both employees of Gulf Stream Inc., on March 21, 2006. This and all further cited internal FEMA emails were disclosed during the discovery process of a congressional investigation into the mishandling of the FEMA trailer formaldehyde issue.\n\n The Stewarts had a friend air the trailer out for two weeks prior to their return to Mississippi. The cockatiel made a full recovery after being removed from the trailer.\n\n While this threshold is a useful, albeit conservative, yardstick for indicating the level at which potential negative health effects can ensue, it is not a regulatory benchmark, as domestic air quality is not regulated.\n\n These lay\/activist tests were later largely corroborated, see Maddalena _et al._ 2009.\n\n Nigel Thrift (1996: 97\u2013120) details a \"social unknowing\" which divides unknowns into five groups. His usage aligns with the noun \"unknowingness\" as he is interested in states of unknowing and not my verbal and processual usage of un-knowing.\n\n My one-sided focus on defensive techniques of weakening connections between chemical exposure and illness, as opposed to those intending to strengthen causal links, puts me at risk of accusations of what Woolgar and Pawluch (1985) refer to as \"ontological gerrymandering,\" or the asymmetrical leveling of critique. The unbalanced problemitizations of this chapter result from a focus on the triumphant epistemic practices, which in the case at hand happens to be those of disavowal and disqualification. Michelle Murphy (2006) meticulously documented the dominance of these regimes of imperceptibility; my goal here in is to understand the processes by which such regimes are constituted.\n\n _Hillard v. United States Government_ , Complaint, Civil Action 06\u20132576, U.S. District Court, Eastern District of Louisiana, filed on May 18, 2006.\n\n Internal E-mail sent from Patrick Preston to Jill Igert, Jordan Fried, and Kevin Souza on June 15, 2006.\n\n Internal E-mail sent from Patrick Preston to Scott Wright on November 30, 2006.\n\n They did so in spite of EPA, National Cancer Institute and the Occupational Safety and Health Administration assertions that formaldehyde exposure as levels three times lower (0.1ppm) can cause not just irritation but coughing, skin rash, and severe allergic reactions.\n\n The press release is available at: \u00adw\u00adw\u00adw\u00ad.\u00adf\u00ade\u00adm\u00ada\u00ad.g\u00ado\u00adv\u00ad\/\u00adn\u00ade\u00adw\u00ads\u00ad\/\u00adn\u00ade\u00adw\u00ads\u00adr\u00ade\u00adl\u00ade\u00ada\u00ads\u00ade\u00ad.f\u00ade\u00adm\u00ada\u00ad?\u00adi\u00add\u00ad=\u00ad3\u00ad6\u00ad0\u00ad1\u00ad0\u00ad, issued on May 4th 2007. Release Number: HQ-07\u2013061.\n\n## **References**\n\nAbraham, J. and Davis, C. (2006) \"Testing times: the emergence of the Practolol disaster and its challenge to British drug regulation in the modern period.\" _Social History of Medicine_ , 19 (1): 127\u2013147.\n\nATSDR (2007) \"ATSDR health consultation: revised formaldehyde sampling at FEMA temporary housing units, October 2007.\"\n\nAvorn, J. (2006) \"Dangerous deception\u2014hiding the evidence of adverse drug effects.\" _New England Journal of Medicine_ , 355 (21): 2169\u20132171.\n\nBiehl, J. G. J. (2007) \"Pharmaceuticalization: AIDS treatment and global health politics.\" _Anthropological Quarterly_ , 80 (4): 1083\u20131126.\n\nCA OEHHA (2001) _Prioritization of Toxic Air Contaminants: Formaldehyde_.\n\nCorrigan, O. P. (2002) \"A risky business: the detection of adverse drug reactions in clinical trials and post-marketing exercises.\" _Social Science & Medicine_, 55 (3): 497\u2013507.\n\nCussins, C. (1996) \"Ontological choreography: agency through objectification in infertility clinics.\" _Social Studies of Science_ , 26 (3): 575\u2013610.\n\nFoucault, M. (1978) _The History of Sexuality Volume I: An Introduction_. New York: Pantheon.\n\nFoucault, M. (1980) _Power\/Knowledge: Selected Interviews and Other Writings_ , 1972\u20131977. New York: Pantheon.\n\nFoucault, M. (2003) _\"Society Must Be Defended\": Lectures at the College de France_ , 1975\u20131976. London: Picador.\n\nGAO (1991) _[Title obscured] B-244536_.\n\nGalison, P. (2004). Removing knowledge. _Critical Inquiry_ , 31 (1).\n\nGarry, V. (2008) \"Formaldehyde and FEMA provided mobile home trailers,\" (letter to Congress by scientific expert).\n\nGraeber, D. (2006) \"Beyond power\/knowledge an exploration of the relation of power, ignorance and stupidity.\" Malinowski Memorial Lecture, London School of Economics.\n\nLezaun, J. and Woolgar, S. (2013) \"The wrong bin bag: a turn to ontology in science and technology studies?\" _Social Studies of Science_ , 43 (3): 321\u2013340.\n\nMaddalena, R., Russell, M., Sullivan, D. P. and Apte M. G. (2009) \"Formaldehyde and other volatile organic chemical emissions in four FEMA temporary housing units.\" _Environmental Science & Technology_, 43 (15): 5626\u20135632.\n\nMajority Staff Analysis (2008) _Trailer Manufacturers and Elevated Formaldehyde Levels. Washington, DC_ : United States House of Representatives, Committee on Oversight and Government Reform.\n\nMathews, A. (2005) \"Power\/knowledge, power\/ignorance: forest fires and the state in Mexico.\" _Human Ecology_ , 33 (6): 795\u2013820.\n\nMcGoey, L. (2007) \"On the will to ignorance in bureaucracy.\" _Economy and Society_ , 36 (2): 212\u2013235.\n\nMichaels, D. (2008) _Doubt Is Their Product_. Oxford: Oxford University Press.\n\nMinority Staff Report (2008) _Toxic Trailers\u2014Toxic Lethargy: How The Centers for Disease Control and Prevention has Failed to Protect Human Health_. Washington, DC: U.S. House of Representatives.\n\nMitman, G. (2007) _Breathing Space: How Allergies Shape Our Lives and Landscapes, 1st edn_. New Haven, CT: Yale University Press.\n\nMol, A. (1999) \"Ontological politics: a word and some questions,\" in John Law and John Hassard (eds), _Actor Network Theory and After_. Oxford: Blackwell, pp. 74\u201389.\n\nMurphy, M. (2006) _Sick Building Syndrome and the Problem of Uncertainty: Environmental Politics, Technoscience, and Women Workers_ , 1st edn. Durham, NC: Duke University Press.\n\nOffermann, F. J. (2009) _Ventilation and Indoor Air Quality in New Home_. California Energy Commission.\n\nPacheco, K. (2010) _Clinical Summary: Cooper_. Denver, CO: National Jewish Hospital.\n\nProctor, R. N. (2006) '\"Everyone knew but no one had proof: tobacco industry use of medical history expertise in US courts, 1990\u20132002.\"' _Tobacco Control_ , 15 (suppl_4): iv117\u2013iv125.\n\nRussell, D., Lewis, S. and Keating, B. (1992) _Inconclusive by Design: Waste, Fraud and Abuse in Federal Environmental Health Research_. The Environmental Health Network and National Toxics Campaign Fund.\n\nSchrader, A. (2010) \"Responding to Pfiesteria piscicida (the fish killer) phantomatic ontologies, indeterminacy, and responsibility in toxic microbiology.\" _Social Studies of Science_ , 40 (2): 275\u2013306.\n\nTattersfield, A. E., Knox, A. J., Britton, J. R. and Hall, I. P. (2002) \"Asthma.\" _The Lancet_ , 360 (9342): 1313\u20131322.\n\nThompson, C. M. and Roland, C. G. (2008) \"Mechanistic considerations for formaldehyde-induced bronchoconstriction involving S-nitrosoglutathione reductase.\" _Journal of Toxicology and Environmental Health, Part A: Current Issues_ , 71 (3): 244.\n\nThrift, N. J. (1996) _Spatial Formations_. Thousand Oaks, CA: Sage.\n\nWhitmarsh, I. (2008) _Biomedical Ambiguity: Race, Asthma, and the Contested Meaning of Genetic Research in the Caribbean_. Ithaca, NY: Cornell University Press.\nChapter 12\n\n* * *\n\n# **A likely story**\n\nHIV and the definition of disability in UK employment equality law, 1996\u20132005\n\n_Emily Grabham_\n\n* * *\n\n## **Introduction**\n\nIn the late 1990s, HIV-positive people in the UK, buoyed by recent developments in anti-retroviral therapies, found themselves having to argue their own potential decline in health in order to gain access to discrimination claims for HIV-related harassment or dismissal. This chapter analyses how legal and clinical ideas about time contributed to these claims. It focuses on how a test of 'likelihood' of future impairment became a significant legal requirement for such individuals, and how lawyers interacted with clinicians to obtain medical reports about claimants' futures. This analysis is thus about 'prognosis', 'uncertainty' and the (newfound) promise of anti-retrovirals. It is also about the role of non-human actors: HIV tests, T-cell counts, viral loads, medical reports, and anti-retrovirals, and their role in confabulating multiple, often contradictory, legal temporalities which had significant effects in the lives of people living with HIV.\n\nThe conceptual question that this chapter raises, therefore, is about law and time. Scholarship in this area continues to grow (Douglas 2011; Keenan 2012; Mawani forthcoming; van Marle 2003). Time is central to Nikolas Luhmann's systems theory (as temporalisation) as much as it animates Derrida's theories of deconstruction and difference (Cornell 1990; Philippopoulos-Mihalopoulos 1999). Within law and society scholarship, social horizons of time influence legal actors' approaches to formal and informal legal regimes (Greenhouse 1996; Engel 1987; Melissaris 2005); multiple and contrasting temporal orders can fail to co-exist as much as they might combine or cohere (Cooper 2013), and legal constructions of temporal thresholds are heavily influenced by technoscientific practices (Beynon-Jones 2012).\n\nContributing to these conversations, my current research investigates how legal actants create and sustain legal temporalities. At a basic level, this line of inquiry has been anticipated within social theory and social history for some time. Our relationship with objects and technologies such as clocks and calendars has created particular segmented understandings of time, which, sociologists and anthropologists remind us, are culturally specific (Zerubavel 1985). My perspective here also draws heavily on work in science and technology studies (STS) which aims to provide accounts of the significance of objects and material artefacts in culture (Henare _et al._ 2007; Pottage and Mundy 2004; Latour 2010). Legal temporalities, from this perspective, are produced through new and changing social relationships, and through mutating interactions between human and non-human actors. In essence, non-human actors, technoscientific practices, objects, and legal technicalities, can be just as important as human actors in creating legal ontologies of time (Beynon-Jones 2012).\n\nThe first section of this chapter situates debates about law and time alongside work in STS which engages with questions of legal processes, networks, paradigms and dilemmas, and with questions of materiality. The following sections, and the conclusion, set out, and reflect upon, research I conducted into how lawyers and activists argued that HIV was a disability for the purposes of discrimination claims in the late 1990s (drawing on semi-structured interviews with activists, lawyers, and policy makers, and documentary and case analysis). These sections explain the context for the claims, and the strategies of activists and advisors in trying to situate their claimants within the relevant legal test. I conclude with some thoughts about how this kind of analysis of time can help us to develop understandings of how temporalities emerge within networked relationships, legal and non-legal.\n\n## **From law and time to material legal temporalities**\n\nBruno Latour argues that temporality is just a way of connecting entities and 'filing them away' (Latour 1993: 75). Crucially, if we change the way this connecting and filing happens, then we get a different temporality. In other words, time is created through exchange and action, not, as modernity would have it, through calendars, flows, and progress (ibid.). As Latour puts it:\n\n> We have never moved either forward or backward. We have always actively sorted out elements belonging to different times. We can still sort. _It is the sorting that makes the times, not the times that make the sorting_.\n> \n> (Latour 1993: 76)\n\nIt is to be expected that the sorting to which Latour refers in this argument is embedded within what has been termed 'material worlds' (Faulkner _et al._ 2012). In other words, the relationships between human and non-human actors create and sustain the temporalities that are seen, in some quarters, as 'natural' or 'social'. Yet as actor-network theory (ANT) and other scholarly approaches have demonstrated, tracing the role of the material in socio-technical or socio-legal relationships is not at all straightforward, especially in the legal field (Pottage 2012). What is a 'thing', after all? Alain Pottage and Martha Mundy have traced the distinctions between persons and things that legal concepts, themselves, have initiated (Pottage and Mundy 2004). The distinction between persons ( _personae_ ) and things ( _res_ ) that was central to Roman law remains influential within codified, as well as common law, traditions, and even shapes current political theory (Pottage, 2004, drawing on the work of Yan Thomas). For this reason, it is also important, when thinking about the role of things in creating legal concepts, to remember the role of law in classifying phenomena as _res_ or _personae_ in the first place.\n\nAddressing the study of law, science, and technology in particular, Alex Faulkner and colleagues use the ANT-inspired concept of 'socio-materiality' to describe the interaction of human and non-human actors within performative legal and scientific networks (Faulkner _et al._ 2012). Drawing on the work of Michel Callon, they describe the process through which theories and material worlds, and a range of actors, are co-ordinated and co-constituted as 'socio-technical agencement', a type of performative material agency (Faulkner _et al._ 2012: 10). For these authors, law is part of the material world, which can support, or be prepared for, particular developments in science and technology. However, within ANT-inspired accounts, law is often not visible, because the analytic intention is to avoid bestowing upon processes, actors, or objects particular categories such as 'social' and 'legal' in the first place (Rooke _et al._ 2012). As Catriona Rooke and colleagues put it:\n\n> The material and its link to regulation are conceived of in a more interdependent way where law is both the result of socio-technical assemblages and becomes part of specific materials, so the 'things' followed during analysis are themselves shaped and defined by the legal and regulatory frameworks that they carry.\n> \n> (Rooke _et al._ 2012: 43)\n\nFollowing 'legal things' is therefore part of an approach to social relations that requires being able to trace when objects enter, leave, or are shaped by heterogeneous networks and encourages the researcher to look past what has already been stabilised as the subject of legal regulation.\n\nHow much can such a view help us to understand the creation of legal temporalities? One example can be found in Latour's own work. In his ethnography of the French administrative court, the Conseil d'\u00c9tat, Latour argues that legal processes within the Conseil produce a sort of homeostasis, a sense of everything being covered 'completely and seamlessly', unlike scientific processes, which leave voids for future researchers to investigate (Latour 2004: 114). Homeostasis evokes law's ability, in this context, to maintain a constant through the adjustment of other features of a system. The temporal horizon of homeostasis does particular work within Latour's analysis. Legal processes and individual conseillers labour under an obligation to ensure legal predictability (or securit\u00e9 juridique). However, paradoxically, in Latour's analysis, legal predictability happens through the fabrication of doubt and distancing \u2013 in other words, through the strategic _avoidance_ of certainty:\n\n> At the Conseil d'\u00c9tat, every effort is made to sustain doubt for as long as possible, but when a decision is reached it is made once and for all.\n> \n> (Latour 2004: 94)\n\nHaving suspended certainty, and indeed actively fabricated doubt, a curious completeness then takes over law: a type of all-encompassing, self-adjusting, temporality (Latour 2004: 113). From the present (post-judgment) vantage point, the legal principle confirmed through proceedings at the Conseil is as it has always been, despite the fact that the entire process has been pursued through means of a graduated suspension of certainty.\n\nAn alternative analysis of legal temporalities can be found in Alain Pottage's work on genetic patrimony and human rights. Pottage focuses on a petition created by two parliamentary members of the Council of Europe in May 2000, which protested against the European Union's Directive on the Legal Protection of Biotechnological Inventions. The protestors argued that this Directive authorised patenting of human gene sequences and that instead, it should be recognised that 'the human genome is the common patrimony of humanity', hence no patents on the genome should be granted (Pottage 2004: 249). According to Pottage, the phrase, 'common patrimony of humanity, which was mobilised within the petition, uses a particular type of 'forever' time, associated with the socio-legal concept of inheritance, to establish the inviolability of the genome. Thus, the use of the concept of genetic patrimony, or inheritance, has specific temporal effects:\n\n> [It] produces a horizon of inheritance that resembles an aevum of mediaeval historiography. That is, it unfolds as a sort of permanent present, a period in which all events occur within the same existential frame or medium; all events have, so to speak the same essence.\n> \n> (Pottage 2004: 264)\n\nWithin accounts of legal temporalities such as these, time is not, or not merely, an intrinsic aspect of the social field, but instead it is created through social relationships, including legal processes, in which the role of nonhuman actors is key. For these, and many other reasons, deploying an analytics of materiality when tracing legal temporalities is a complex process, involving layers of excavation that I can only begin to attempt here. At the very least, however, analysing material worlds in relation to law and schemes of legality can and should involve understanding the temporalities inaugurated and sustained by human and material actors within legal networks (Beynon-Jones 2012). The following sections constitute one attempt to do so, tracing the multi-layered temporalities of HIV found in the field of disability law activism in the late 1990s and early 2000s when, as treatments improved, people's experiences of HIV changed radically, and new legal routes became available to HIV-positive people.\n\n## **1980\u20131996: legal-temporal ontologies of crisis**\n\nAs many will be aware, the new 'crisis' of what was soon to be termed AIDS first emerged as a public health problem in countries of the global North, including the United States and United Kingdom, in the early 1980s. The early years of understanding and isolating the virus, establishing the link between HIV and AIDS, and scoping testing and treatments for HIV, were characterized by a number of intense credibility struggles, as scientific teams advanced different explanations for the causes of this new disease, and clinicians and government agencies advanced a wide range of often conflicting responses (Patton 1990; Epstein 2009). Activists were key to the development of new knowledges, treatments, and new relationships between clinicians and health populations in the context of a political climate all too hospitable to homophobic and racist fears about the new disease (Cohen 1999; Epstein 2009; Patton 1990).\n\nBy the mid-1990s, a causal link between HIV and AIDS had been settled. HIV testing had become available for many of those who wanted it \u2013 and indeed often imposed on many who did not \u2013 and new treatments had become available. This chapter picks up at the point at which intense struggles of a different kind began, this time within networks of lawyers and activists in the UK concerned with ensuring employment equality for people with disabilities, amongst whom were counted those with diagnoses of HIV and AIDS. After a spirited campaign by disability rights activists, the UK's Disability Discrimination Act 1995 (DDA) came into force in 1996. For the first time, people who experienced disability discrimination in the workplace had a rights-based means of redress. For people with HIV \u2013 who often experienced racist and\/or homophobic harassment, dismissal, and 'outing' of various kinds associated with an HIV diagnosis \u2013 making a claim was often a significant route to badly needed financial compensation.\n\nHowever, these claims remained very difficult to make. Much of this was to do with the kinds of future that discrimination law could and could not imagine. In particular, a person could only go on to argue a discrimination claim if they had first established that their HIV constituted a disability for the purposes of the Act. Making this argument was fundamentally about time: it was about whether and how claimants, clinicians, lawyers, and Employment Tribunal members could find each other's understandings of the future effects of HIV realistic or credible within the relevant legal paradigms.\n\nJohn Lynch argues that 'AIDS was from the beginning of its discursive \"naming\" defined in temporal terms' (Lynch 2000: 247). Dominant discourses surrounding HIV\/AIDS by the mid-1990s had been temporally framed as both epidemiological crisis and individual moral failure. Indeed, AIDS in the UK during the 1980s and 1990s had its own distinct social chronology, which can be schematised as a 'dawning crisis', proceeding with a homophobic moral panic and then plateauing into crisis management (Lynch 2000: 249). The period before the DDA could be characterized as what Lynch understands as dawning crisis, or what Cindy Patton would term 'AIDS as emergency', which had its own effects in creating and sustaining a particular system of social control (Patton 1990, 107). Within this context, HIV activists concerned with the welfare of themselves and their friends, lovers, and colleagues had started out in the 1980s dealing with legal issues arising from the shockingly rapid decline in health that people with AIDS experienced before viable HIV tests and treatments became available. The context, as Douglas Crimp puts it, was one in which pragmatic hope was a necessity:\n\n> In practice, convincing the public and ourselves that people with AIDS can participate in life, even prosper, has been an uphill struggle. The facts are all too clear: people suffer terribly, and we're dying by the thousands. But I'm not talking about not dying; I don't deny the reality that sooner or later most people with AIDS die; it would be foolish not to address death. I'm talking about the business of living, or making choices, of _not_ being passive, helpless, dependent, the storm-tossed object of the ministrations of the kindly well.\n> \n> (Crimp and Bersani 1988: 145)\n\nEven with this reinforced hope, many AIDS organisations worked hard to get access to benefits and housing for people with significant health impediments on the assumption that they were dealing with a crisis and that any individual arrangements they made for people with AIDS (PWAs) were necessarily temporary, because \u2013 based on contemporary medical knowledge \u2013 PWAs were possibly going to die within a matter of months and would certainly live no longer than a few years. This paradigm also meant that employment-related activism worked with a short horizon, that of ensuring continued employment for PWAs if that was possible, and\/or extracting compensation (through settlements) for AIDS-related discrimination in order to provide badly needed financial support:\n\n> If you think, you've just had your diagnosis, you've got a CD4 count which is in double figures if you are lucky. You are already getting symptomatic illnesses. You've probably got diagnosed after your third bout of thrush, because nobody knew what it was the first couple of times.... An awful lot of people thought the gain was not worth a candle to challenge their employers. So in those cases, we would just try and extract maximum padding for them as it were.\n> \n> (health activist\/policy maker: 7)\n\nEmployment advice was largely voluntary and activist-based, aimed at dealing with pressing and often quite obvious health needs, and commonly involved non-legally qualified representation.\n\nAfter the very early stages of the crisis, as the aetiological link between HIV and AIDS was stabilised (Epstein 2009) and HIV tests eventually became available, this temporal paradigm of AIDS, the paradigm of imminent death, with its associated forms of representation and activism, began to shift. By no means (yet) considered a chronic condition, people were nevertheless being diagnosed earlier and treatments were coming online. Although these remained associated with hugely problematic side-effects \u2013 see the debates around AZT (Epstein 2009; Crimp and Bersani 1988) \u2013 and could be difficult to reach depending on where a person lived \u2013 see debates about aerosolized pentamidine in Canada (Silversides 2003) \u2013 people were surviving longer after diagnosis.\n\nLynch has analysed public health campaigns introduced in the UK in the 1990s, the period, incidentally, during which time the push for disability rights legislation was also in full force. As he demonstrates, one of the most important medico-cultural narratives about HIV focused on a new temporal sequence:\n\n> the first (stage) was the body's reaction to the initial HIV infection; followed by a latent period before the development of symptoms of the 'full-blown' condition; and finally inevitable death.\n> \n> (Lynch 2000: 250)\n\nThese campaigns introduced into public discourse the infected but asymptomatic person \u2013 who could be straight or queer \u2013 who unknowingly passed on the virus to others (Lynch 2000: 254). This was a different temporal frame from the suffering and terminally ill AIDS patient. Such campaigns relied on a representation of HIV-affected people as appearing 'normal' and potentially infectious. In one advert published in the mainstream press, a question was posed and answered, in white, on a black background: 'What is the difference between HIV and AIDS?' 'Time' (Lynch 2000: 255). As Lynch points out, this advert advances an essential closed, monologic, understanding of the temporal progression of HIV\/AIDS. The advert looks like a tombstone and effectively translates individual experiences of infection into what Lynch calls 'an abstracted realm of supra-human fate' (Lynch 2000, 256). Not only that, but within these narratives the Person With AIDs is understood through a fear-inducing matrix of 'past-transgression, present-agent of infection, future-death (the absence of future)' (Lynch 2000: 257).\n\nPublic health campaigns thus promoted fear of seemingly healthy but infectious people transmitting the virus. For their part, media outlets spread rumours about AIDS being more prevalent amongst Haitian populations in the United States or amongst African populations in London (Cohen 1999). It is not surprising, therefore, that many HIV-related employment discrimination issues were, as activists have put it, based on 'pure stigma':\n\n> It was sort of pure stigma related cases. That sort of fed into a lot of issues around whether, when to disclose your HIV status. If, for example, you were starting a new job and you knew that you would have to take time off for treatment or for hospital appointments or it would have some effect in the workplace should you disclose and take the risk the job offer would be withdrawn or you would be sacked.\n> \n> (legal activist: 2)\n\nAs treatments improved, disclosure issues changed, so that even though racialised and homophobic constructions of HIV still prevailed within popular culture, news stories, and employer attitudes, people with HIV had more control over whether and how to disclose to employers in the first place. PWAs with obvious symptoms still encountered outright hostility at work and immediate health crises, but more and more people with HIV diagnoses were facing difficult questions about whether to disclose to an employer in the first place, or how to deal with stigma relating to being 'out' as HIV-positive.\n\n## **1996\u20132005: the significance of likelihood**\n\nBy the time the DDA came into force in 1996, therefore, one distinct social and clinical shift had already happened in the lives of people concerned with the epidemic. This was the shift from very late diagnoses of what we now understand to be AIDS, in other words terminal diagnoses, to diagnoses of HIV infection (regarded as 'pre-AIDS'), and the beginnings of treatments which could delay progression from HIV to AIDS. In the midst of this, the DDA, which came into force in 1996, also promised a route for people with HIV\/AIDS to obtain some form of legal redress for HIV discrimination in the workplace. Yet, even with HIV recognised within the disability paradigm, between 1996 and 2005, the fact of a person's HIV diagnosis did not mean they were automatically considered to be disabled for the purposes of the DDA. In order to make a claim, a person had to prove that they had a disability for the purposes of the legislation. They had to show under section 1 of the DDA that they had a 'physical or mental impairment which ha[d] a substantial and long-term adverse effect on [their] ability to carry out normal day-to-day activities'. Often this issue would be determined by an Employment Tribunal at a preliminary hearing.\n\nOne lawyer spoke of the problems associated with the disability test more generally, which related to the trauma of having to 'prove' one's disability legally in this way:\n\n> When you are talking about someone's disability you are going to have to talk about what happens at home. What happens when they go to the doctor. It's a much more personal investigation of their circumstances. That is very difficult, often, for people to talk about, particularly in an open context, and that is what you are having to ask them to do. For many people, they didn't want to do that. And so they might pull out at that point.\n> \n> (lawyer\/policy maker: 5)\n\nGiven the social context, it might not have appeared too difficult to argue that HIV would have such an effect. As we have already seen, it was in very recent memory that people would be diagnosed with AIDS rather than HIV and in any case it was a matter of 'common sense' even to legal activists that HIV should be considered to be a disability. Yet despite all of this, people with HIV had two main problems fitting within the DDA's specific definition. The first was that if they had no symptoms, they would be unable to show an 'impairment' and would therefore not be covered at all by the DDA. This problem had been foreseen in the parliamentary debates preceding the enactment of the DDA in 1995 as affecting the majority of people with HIV. It could usually be countered by referring to even very minor symptoms associated with a person's sero-conversion illness (which people often at the time thought was a very bad cold). The second problem was that even if they had experienced symptoms, they then had to fit these symptoms within the provisions of the DDA that covered 'progressive conditions'.\n\nThrough case-law that developed over the period between 1996 and 2005, it became clear that people with HIV had to show that their condition was 'likely' in the future to have a substantial adverse effect on their ability to carry out day-to-day activities. 'Likelihood' meant more probable than not, importing, from the start, a medico-legal concept of the future into the definition of disability. This type of 'likelihood' was difficult to demonstrate, often requiring complex arguments in preliminary hearings and the commissioning of joint medical reports. Lawyers developed a field of expertise in creating draft questions for clinicians that would elicit the desired response, and they also spent a considerable amount of effort arguing, between parties, about which questions should be posed.\n\nBefore they were even written, expert medical reports, as material legal actors, had significant effects. Structured through negotiations between legal parties over which expert, how to instruct, and how to understand the features and prognosis of HIV generally and the treatments available, each report then had to address the question of _this_ HIV in _this_ claimant, as we will see below. Sometimes, representatives for employers would attempt to lead the clinical expert into making a statement about the legal status of the person's disability (or lack of disability), a question that was understood on normal legal principles to be firmly a matter for the Employment Tribunal itself. Claimants' representatives would resist these attempts, not always successfully:\n\n> there would be a wide range of good and bad chair people in [a] tribunal. Some would see that for what it was. Some would say, \"well, that's settled then isn't it and this expert said that this person doesn't meet the definition of disability. That is the end of the case.\" That was difficult.\n> \n> (legal activist: 9)\n\nClinical expertise, imported into legal proceedings for the purpose of providing evidence relating to a person's disability status, could sometimes have the effect of eliding constructed boundaries separating legal 'truth' from medical 'evidence' (Fassin and d'Halluin 2005), and claimants' representatives were on guard to ensure that this did not happen. In this way, through medical reports, and the tensions and discussions that produced and surrounded them, fields of clinical and legal knowledge, and the imagined overlap between them, were produced and fortified.\n\n## **Prognosis and uncertainty: contested technical imaginaries of time**\n\nAs we have seen, a key factor debated in medical reports about the question of likelihood was the individual's prognosis. Generally, prognosis refers to the mixture of statistics, survival rates, and other factors that go into predicting the course and outcome of a given condition. As a socio-technical invention, prognosis creates fields of power through the articulation of probabilities based on statistical measurements, clinical conjecture, and understandings of the life-course. As such, it interacts with other socio-temporal concepts such as 'decline', 'crisis', or 'recovery'. As Sara Lochlann Jain has argued, prognosis has a double effect, 'causing and evacuating the terror of a potential future' through a logic which is at once incessantly future-oriented and abstracted (Jain 2007: 81).\n\nWhen mobilised in legal settings, prognosis materialises diseases such as cancer, and in the present case, HIV. Yet in Jain's own study, prognosis was not necessarily enough to establish a legal claim. Cases relating to 'lost chances', where they were allowed, were only successful if the claimant suffered more than a 50 per cent loss of survival, bringing a claimant from more than a 50 per cent survival chance to less than 50 per cent. The staging indices usually used in prognosis (e.g., the size of a tumour) often did not provide enough specific information to allow this kind of determination (Jain 2007: 85\u201388).\n\nLegal activists in the realm of HIV discrimination were similarly beset with temporo-legal problems, oriented around a central concept of prognosis. In the context of clinical decisions around the treatment of HIV, prognosis is normally influenced by where a patient's test results fall on a normative scale, determined with reference to viral load and T-cell count (Persson 2004: 50). Viral load refers to the extent of HIV infection in the body, and is hence a clinical, and also cultural, measurement of disease. Clinicians normally view an undetectable viral load as indicating therapeutic success and 'exponential viral replication' as a decline in health (Persson 2004: 51). T-cell (or CD4) counts measure specialised immune cells that are targeted by HIV, so they indicate retained functionality in the immune system. In this way, technosci-entific practices of tracing, counting, and interpreting cells and viruses have a direct effect on the health constructs that are mobilised within legal spheres.\n\nFurthermore, in the late 1990s and early 2000s, the disability of HIVpositive claimants was legally established with reference to an individual's own prognosis, not an understanding of what this condition might usually do in a majority of cases. In other words, what activists reported was an articulation of scale and individuation, which prioritized a legal subject's personalized prognosis and lifespan over a population-level understanding of the progression of disease based on statistical likelihood:\n\n> [T]he definition of likely was more probably than not. That gave rise to real problems... Even if you can show that more than 50% of people with [HIV] go on to develop a substantial and long-term adverse effect from it... that didn't necessarily help you, because the tribunals and the courts were saying, \"well, we don't just want to know whether statistically that condition leads to that, is likely to lead to that. We want to know in the case of your client, does your client fall into the 51% who will go on to have substantial and long term effects or are they in the 49%?\"\n> \n> (legal activist: 6\u20137)\n\nLikelihood, 'more probably than not', was a universally applied legal invention, established through the interaction of definitions within the DDA, case-law, and clinical decisions, which instantiated the potentially disabled legal subject as _temporally_ distinct and distinguishable on the basis of his\/her prognosis. This temporalised individuation signalled a relationship, of sorts, between legal and clinical knowledges.\n\nOne way of understanding this relationship is through Catherine Waldby's concept of the 'shared technical imaginary'. Waldby analyses the new relationships of equivalence that emerged between quite distinct domains of biomedicine in the mapping of AIDs, and argues that one feature of this coming together, or equivalence, was a process of schematisation:\n\n> In the case of AIDS we have on the one hand a concern about T and B cells, lymph, tissues and glycoproteins, and on the other patterns of sexual practice and alliance. These diverse phenomena can be made to coalesce precisely through their transformation into a shared technical imaginary which is characterized by a maximum of schematisation. Wherever possible, pathological processes in both bodies and the population are quantified and mathematicised, rendered into units which can be graphed, plotted, counted. Such mathematicisation places the divergent spheres of clinical and epidemiological medicine into modular and compatible relations with one another, allowing a relationship to be expressed between each of the phenomena so rendered.\n> \n> (Waldby 2004: 97)\n\nIn the context of legal arguments about HIV as a disability in the late 1990s, prognosis functioned across legal and clinical fields, distributing knowledges about T-cells, viral load, decline, or recovery, with varying effects. This process relied on a schematic approach to the constructed, internalized, 'truths' of the body, creating a shared, or at least co-articulated, technical imaginary of time, linking the legal concept of likelihood with the realm of projected clinical outcomes. The resulting technical imaginary of time mobilised T-cell counts, viral load measurements, and pharmaceutical innovations as much as it drew on claimants' own (temporal) narratives of health and (dis)ability.\n\nThe temporalised effects of cells (pace Waldby and Cooper 2010), their functions and measurements, were quite distinct within the co-articulated technical imaginary of HIV time in the late 1990s. As non-human actors, viral load measurements and T-cell counts created a specific genre of HIV temporality, based on viral and cellular replication, through which the generative function of T-cells or the virus (respectively) was instead mapped back onto the future of the body-as-organism\/legal claimant, as well as the potential success of the case. On this logic, crudely expressed, as the number of T-cells grew, so the potential life time attributable to the human legal actor expanded, meaning that both a growing number and a high number of cells lengthened and improved a person's prognosis and made a finding of future impairment, and hence disability, less likely. Similarly, on this logic, a growing or high viral load, indicating exponential generation at the level of the virus, reduced and worsened prognosis, making a legal finding of disability more possible.\n\nIn this way, cellular processes, and the actions of clinicians in measuring them, influenced legal determinations of prognosis, evidencing a level of temporal co-articulation that accords with what Waldby identifies in her analysis of scientific knowledges relating to HIV. Indeed, Waldby perceives this type of co-articulation as being very smooth, a 'seamless reciprocity' between knowledge relating to individual symptoms and knowledge relating to epidemiology (Waldby 2004: 97\u201398). A degree of fit and exchange was established and maintained between clinical and legal understandings of HIV temporalities through various means: legal activists acquainted themselves with key medical terms and concepts in order to educate Employment Tribunals making decisions on the question of someone's disability; medical experts, carefully chosen and carefully instructed by lawyers, provided what they thought to be the most rigorous and appropriate assessments of a claimant's current condition and future medical outlook within the legal guidelines they had been given.\n\nYet, the machinery powering this 'seamless reciprocity' also broke down at times. Despite the apparent reliability of their own T-cell and viral load measurements for treating patients and\/or undertaking medical research, when faced with the task of making a prediction for legal purposes clinical experts often did not give the level of specificity that legal representatives wanted. Legal activists interpreted this as signifying an uncomfortable relationship with prediction, if not a stubborn unwillingness to give vital information about a claimant's future health:\n\n> It seems to me that the medical experts are always reluctant to really go out on a limb and give anything other than a vague, well, this could or this might happen.\n> \n> (legal activist: 13)\n\nThe reason for this, as a respondent from a slightly different health activist field speculated, could have been because the expert's role as a clinician involved communicating prognosis in terms of life-course to patients:\n\n> This was a thing I think that came up particularly in the context of cancer. I suspect it also came up in the context of HIV, although we didn't see as much of it at the [ _name of organisation_ ]. With the reluctance of physicians to say that it was more likely than not that it would progress. Certainly in the context of cancer, they just didn't want to say it. They didn't want to say that it's likely this is going to come back and it's going to get much worse. They didn't want their patients to think that that's what was going to happen, yet that was [what] they had to say in order for them to fall within the definition.\n> \n> (lawyer\/policy maker: 8)\n\nClinical experts resorted to what they thought was a more palatable or even more concrete and defensible temporal outlook: uncertainty. Professionally, what Fred Davis would term 'functional uncertainty' would consciously or not have served to deflect difficult discussions and emotional encounters between clinician and patient around prognosis (Davis 1960), and would not have equated directly with clinical uncertainty as such. Disclosing uncertainty has varying effects and motivations, amongst clinicians, signalling rigour and discretion in some circumstances, but sometimes also being feared as an admission of lack of knowledge (Gerrity _et al._ 1992), and any of these factors might have been at play in individual cases. However, as Tiago Moreira and colleagues argue, uncertainty can be also understood in biomedical terms as a 'discursive, interactive accomplishment' (Moreira _et al._ 2009: 667), constituent of \u2013 instead of undermining or reversing \u2013 biomedical knowledge production.\n\nGiven the potentially foundational effects of uncertainty in generating scientific knowledges, it becomes all the more difficult to imagine an encounter between legal and clinical fields about HIV prognosis in which likelihood can be articulated as entirely distinct from uncertainty. Yet a 'centrifrugal pull' of temporal orders was in motion in this period (Cooper 2013), through which the construct of biomedical uncertainty effectively destabilised the legal rationalities of prediction needed to establish a claimant's HIV as disability.\n\n## **Conclusion**\n\nIn this chapter, I have attempted to clear some conceptual space for thinking about how legal networks create plural ontologies of time. Multiple temporalities were at play in the case study: the individual lifespan; the recent temporal 'progression' of the AIDS and HIV public health crises; experiences of PWAs; various contested paradigms of the 'future' and 'prognosis'; temporalities of aetiology, diagnosis, the life-saving or life-preserving effects of anti-retroviral treatments, and disease itself; and temporalities of litigation and legal subjectivity. These temporalities, confabulated through the coming together (or not) of legal and clinical knowledges, objects, cells, tests, medical reports, HIV treatments, and legal resources, constituted what I term a 'likely story': the story of what happened in the late 1990s when 'uncertainty' about HIV prognosis met 'likelihood', the legal test for future impairment, and the gateway to beginning a disability discrimination claim.\n\nBy the mid-2000s, legally constructed temporalities of HIV, in perverse ways, mirrored the newfound (yet cautious) hope that positive people were experiencing with the development of 'combination therapy' or highly active anti-retroviral therapy (HAART), where these therapies were available. HAART suppressed the virus and also preserved the immune system. Similarly, within the logic of 'likely' future impairment, it was possible for HIV not to be debilitating, for legal purposes. Indeed, the problem was that HIV was not a disability unless proven so, and prognosis, despite the concrete nature of T-cell and viral load measurements, was often scientifically uncertain enough to disqualify people from legal claims.\n\nWhat is interesting, for me, is quite how much these ontologies of HIV \u2013 created through pharmaceutical innovations, new treatment regimes, medical reports, and legal processes \u2013 varied from temporalities of HIV being produced and circulated in health education networks in the UK around about the time period. In health education networks, as Lynch shows, the disease was debilitating: it was present without visible or obvious effects and prognosis was certain death. Entirely different temporal ontologies were deployed in one country for the same disease. If anything, this is evidence that our understandings of time cannot escape the networked interactions of human and non-human actors that create temporalities instead of inhabit them. Tracing and diagnosing such temporalities can contribute to our understanding, more broadly, of the variegated relationships, networks, and worlds conjured by encounters within law, science, and technology.\n\n## **Acknowledgements**\n\nI would like to thank Emilie Cloatre and Martyn Pickersgill for their wonderful editorial advice and suggestions. This research was funded by the UK Socio-Legal Studies Association and a Faculty Grant from the University of Kent, and was conducted whilst on sabbatical at the Institute for Feminist Legal Studies, Osgoode Hall Law School. Many thanks also to the following people for comments, engagement, and inspiration: contributors to the 'Temporalising Law' stream at the Canadian Law and Society conference _Law on the Edge_ July 2013 and audience, Davina Cooper, Mariana Valverde, Nick Piska, Maria Drakopoulou, Kate Bedford, Mairead Enright, Sonia Lawrence, and Donatella Alessandrini. The usual disclaimers apply.\n\n## **References**\n\nBeynon-Jones, S. M. (2012) 'Timing is everything: the demarcation of \"later\" abortions in Scotland'. _Social Studies of Science_ , 42(1): 53\u201374.\n\nCohen, C. J. (1999) _The Boundaries of Blackness: AIDS and the Breakdown of Black Politics_. Chicago, IL: University of Chicago Press.\n\nCooper, D. (2013) 'Time against time: normative temporalities and the failure of community labour in local exchange trading schemes'. _Time & Society_, 22(1): 31\u201354.\n\nCornell, D. (1990) 'Time, deconstruction, and the challenge to legal positivism: the call for judicial responsibility'. _Yale Journal of Law & the Humanities_, 2: 267.\n\nCrimp, D. and Bersani, L. (1988) _AIDS: Cultural Analysis, Cultural Activism_. Cambridge, MA: MIT Press.\n\nDavis, F. (1960) 'Uncertainty in medical prognosis clinical and functional'. _American Journal of Sociology_ , 66(1): 41\u201347.\n\nDouglas, S. (2011) 'Between constitutional mo(nu)ments: memorialising past, present and future at the district six museum and constitution hill'. _Law and Critique_ , 22(2): 171\u2013187.\n\nEngel, D. M. (1987) 'Law, time, and community'. _Law and Society Review_ , 21: 605.\n\nEpstein, S. (2009) _Impure Science: AIDS, Activism, and the Politics of Knowledge_. Berkeley, CA: University of California Press.\n\nFassin, D. and d'Halluin, E. (2005) 'The truth from the body: medical certificates as ultimate evidence for asylum seekers'. _American Anthropologist_ , 107(4): 597\u2013608.\n\nFaulkner, A., Lange, B. and Lawless, C. (2012) 'Introduction: material worlds: intersections of law, science, technology, and society'. _Journal of Law and Society_ , 39(1): 1\u201319.\n\nGerrity, M. S., Earp, J. A. L., DeVellis, R. F. and Light, D. W. (1992) 'Uncertainty and professional work: perceptions of physicians in clinical practice'. _American Journal of Sociology_ , 97(4): 1022\u20131051.\n\nGreenhouse, C. J. (1996) _A Moment's Notice: Time Politics Across Cultures_. Ithaca, NY: Cornell University Press.\n\nHenare, A. J. M., Holbraad, M. and Wastell, S. (eds) (2007) _Thinking Through Things: Theorising Artefacts Ethnographically_. London: Routledge.\n\nJain, S. L. (2007) 'Living in prognosis: toward an elegiac politics'. _Representations_ , 98(1): 77\u201392.\n\nKeenan, S. (2012) 'Subversive property: law and the production of spaces of belonging'. Unpublished.\n\nLatour, B. (1993) _We Have Never Been Modern_. Translated by Catherine Porter. Cambridge, MA: Harvard University Press.\n\n\u2014\u2014 (2004) 'Scientific objects and legal objectivity', in _Law, Anthropology, and the Constitution of the Social_ , Cambridge Studies in Law and Society. Cambridge; Cambridge University Press, pp. 73\u2013114.\n\n\u2014\u2014 (2010) _The Making of Law: An Ethnography of the Conseil d'\u00c9tat_. Cambridge, UK; Malden, MA: Polity.\n\nLynch, J. (2000) 'AIDSTimes representing AIDS in an age of anxiety'. _Time & Society_, 9(2\u20133): 247\u2013267.\n\nMawani, R. (forthcoming) 'Law as temporality: colonial politics and indian settlers'. _Irvine Law Review_.\n\nMelissaris, E. (2005) 'The chronology of the legal'. _McGill Law Journal_ , 50: 839\u2013861.\n\nMoreira, T., May, C. and Bond, J. (2009) 'Regulatory objectivity in action mild cognitive impairment and the collective production of uncertainty'. _Social Studies of Science_ , 39(5): 665\u2013690.\n\nPatton, C. (1990) _Inventing AIDS_. New York: Routledge.\n\nPersson, A. (2004) 'Incorporating pharmakon: HIV, medicine, and body shape change', _Body & Society_, 10(4): 45\u201367.\n\nPhilippopoulos-Mihalopoulos, A. (1999) 'The silence of the sirens', _Law and Critique_ , 10(2): 175\u2013197.\n\nPottage, A. (2004) 'Our original inheritance', in _Law, Anthropology, and the Constitution of the Social_ , Cambridge Studies in Law and Society. Cambridge: Cambridge University Press, pp. 249\u2013285.\n\n\u2014\u2014 (2012) 'The materiality of what?' _SSRN eLibrary_ (March). http:\/\/papers.ssrn.com\/sol3\/papers.cfm?abstract_id=2009206.\n\nPottage, A. and Mundy, M. (2004) _Law, Anthropology, and the Constitution of the Social: Making Persons and Things_ , Cambridge: Cambridge University Press.\n\nRooke, C., Cloatre, E. and Dingwall, R. (2012) 'The regulation of nicotine in the United Kingdom: how nicotine gum came to be a medicine, but not a drug'. _Journal of Law and Society_ , 39(1): 39\u201357.\n\nSilversides, A. (2003) _AIDS Activist: Michael Lynch and the Politics of Community_. Toronto: Between the Lines.\n\nVan Marle, K. (2003) 'Laws time, particularity and slowness', _South African Journal on Human Rights_ , 19: 239\u2013255.\n\nWaldby, C. (2004) _AIDS and the Body Politic: Biomedicine and Sexual Difference_. London: Routledge.\n\nWaldby, C. and Cooper, M. (2010) 'From reproductive work to regenerative labour the female body and the stem cell industries'. _Feminist Theory_ , 11(1): 3\u201322.\n\nZerubavel, E. (1985) _Hidden Rhythms: Schedules and Calendars in Social Life_. Berkeley, CA: University of California Press.\nChapter 13\n\n* * *\n\n# **Paper prototypes**\n\n_Alain Pottage_\n\n* * *\n\nThe techniques and institutions of patent law offer a ready-made intersection between law and science, one that has been quite extensively explored from various perspectives. Many of these explorations assume that intellectual property rights necessarily bring with them something like the powers of control and exclusion that lawyers ascribe to property. The assumption is that patents function as instruments for regulating access to scientific or technological knowledge. It is obvious that patents have effects, but in order to understand those effects one should be clear about what kind of object a patent is. The conventional understanding objectifies patents by construing them as instruments of broader social and economic forces. In fact, a patent (if one can use the noun form at all) is a complex assemblage of material, discursive, and cognitive elements: files, documents, indexing systems, bureaucratic practices and aesthetics, structural couplings to economic and scientific institutions and techniques, and so on. Accordingly, the better place to begin in developing a social-scientific account of the agency of patents is with the 'knowledge practices' that are immanent in this assemblage of elements (Riles, 2011). In other words, the agency of patent law should be construed not in terms of social forces of which law would be a dependent vehicle or instrument, but in terms of the emergent agency that articulates the composite 'object' of material and communicative elements and techniques. Of course, this quite considerably complicates the question of how law intersects with science in the medium of patent law, but it is better to acknowledge these complexities than it is to perpetuate the ideological narrative of ownership that has been retailed by more than two centuries of legal and economic commentary.\n\nIn this chapter, I focus on one element of the patent assemblage \u2013 the patent text or specification. I suggest that in classical patent jurisprudence the specification was understood in terms of a logic of prototyping. Patent lawyers did not talk about prototyping, but they imagined the invention as a kind of genetic template; as an immaterial causative principle that informed the form and functionality of the material artefact. This sense of the invention was probably an effect of the intimate historical connection between the formation of copyright and patent law and the emergence of large-scale manufacture and industrial production. Copyright law emerged when scribal publication was eclipsed by the diffusion of the printing press \u2013 'the first assembly line in the history of technology' (Ong, 2002: 118) \u2013 and the category of the invention in its modern sense was closely shaped by the formation of an ideology of manufacturing which construed manufactured machines or consumer artefacts as 'one of a multitude' (Babbage, 1851: 48). By stamping out potentially limitless copies of an original, the technology of manufacturing promoted the generalized sense that 'ideas' were the prime movers in the making of mechanical devices and literary artefacts (as Kant observed, even publishing was 'carried on in the manner of a factory [ _fabrikenm\u00e4ssig_ ]') (Kant, 1996a: 626). The sheer quantity of identical copies \u2013 even if quantity was potential and identity necessarily imperfect \u2013 affirmed the prestige of the original template or 'prototype' as the most vital agency in the manufacturing process.\n\nThe artefact itself was construed as the inessential embodiment of the invention. The art of patent discourse was to elicit the type from the tokens in which it was embodied, and, in the context of the infringement action, where the question was which of two claimants was the 'original and first inventor' of a new invention, to determine which of two types qualified as the 'prototype', or the first practicable expression of a novel idea. This particular take on prototyping took the invention as the endpoint of one process of prototyping, namely, the process of working out the practicability of an idea, and as the origin of another, namely, the process of manufacture. The process of eliciting the invention from its embodiment, or of anatomizing materiality to express the intangible 'type', took place in various settings, ranging from the office of the patent attorney, through the patent office, to the circuit or appellate courts. In each of these settings the inventive artefact was scrutinized to identify the novel 'principle' that it embodied.\n\nThe lawyer's sense of prototyping was essentially normative. According to patent doctrine, inventions had to be properly disclosed before they could be patented; the object was to fix the intangible idea in print, so that it could circulate to others 'skilled in the art' both as a notice of the existence of the invention and as a disclosure of the knowledge expressed in the invention. This theory of paper as the medium in which inventions were communicated followed from the basic normative premise of the US patent system, which is that the inventor is granted an exclusive right to exploit the invention for a limited period of time in return for disclosing the invention to the public in terms which would enable the skilled artisan to make and use the invention:\n\n> The law confers an exclusive patent-right on the inventor of any thing new and useful, as an encouragement and reward for his ingenuity, and for the expense and labor attending the invention. But this monopoly is granted for a limited term only, at the expiration of which the invention becomes the property of the public. Unless, therefore, such a specification was made, as would at all events enable other persons of competent skill to construct similar machines, the advantage to the public, which the act contemplates, would be entirely lost, and its principal object would be defeated.\n> \n> ( _Lowell v. Lewis_ 15 F.Cas. 1018, 1020 [1817])\n\nThe emergence of this theory was an effect of a quite profound transformation in the political or governmental diagram of modern Europe, which Mario Biagioli characterizes in terms of a relation between political representation and the representation of inventions (Biagioli _et al._ , 2011: 27\u201328). The theory of the printed specification identified inventions as prototypes in two senses: retrospective and prospective. The paper specification functioned retrospectively as the medium in which the invention was fixed as the endpoint of the process of conception and reduction to practice; in setting out a recipe from which other artisans could reproduce the invention, a properly prepared specification proved that the inventor truly possessed the new idea. The same criterion of reproducibility allowed the specification to function prospectively, as the medium in which the invention became an industrial prototype, or an effective genetic template.\n\nI begin with an account of the historical factors that allowed textual descriptions and perspectival drawings to be adopted as vehicles for recollecting and communicating mechanical designs. I then describe how patent jurisprudence settled on a definition of the invention which animated the patent text in ways which escaped the instrumental agency implied in the doctrinal theory of disclosure. Nineteenth-century patent lawyers in the United States defined the mechanical invention not as a spatial form but as a transitive or dynamic force. So the 'prototype' that was held in a patent text was not identical to the 'virtual' construct depicted in the drawings and descriptions; rather, it was the transitive 'mode of operation' that could be seen in the articulations of the machine once it was made. And a 'mode of operation' was an effect of the way that texts were put together with concepts, gestures, and material arrangements. That is, the invention as prototype was an artefact of knowledge practices that articulated this assemblage of elements. Indeed, in characterizing the mode of operation of a machine as a means abstracted from ends, or as a means in itself, patent lawyers found in their object the very mode of agency that articulated the discourse of patent law. I conclude by suggesting that the legal definition of the invention introduced a sense of prototyping that was more akin to postindustrial 'modulation' than to industrial 'moulding' (on this distinction, see Deleuze, 1980). Although the jurisprudential theory of disclosure was consonant with the classical sense of prototypes as points in a linear process of engineering, the definition of the mechanical invention as a 'means' allowed patent lawyers to reinvent inventions by recombining 'form' and 'matter', paper and materiality, or the visual and textual. It is precisely this kind of agency that should inform reflection on the social or economic effects of patents.\n\n## **Paper prototyping**\n\nIn his patent treatise of 1849, the patent lawyer George Ticknor Curtis drew an analogy between a transaction in a material thing and a transaction in a 'secret process':\n\n> Let us suppose that A., by the exertion of his inventive faculties, has ascertained, that by placing matter in certain positions to be operated upon by forces of nature, a result will be produced, in the shape of an instrument, wholly unknown before, and capable of being usefully applied to the wants of mankind. Let us suppose that B., seeing the result but wholly ignorant of the process by which it may be attained, desires to possess that instrument. Common gratitude would prompt him to return something valuable for it. How does it alter the case, if, instead of a single specimen of the instrument, we suppose A. to have retained in his recollection the process by which copies of that instrument may be indefinitely multiplied, and that it is the secret process of making the thing, the intellectual property and knowledge, which B. desires to possess?\n> \n> (Curtis, 1849: 3\u20134)\n\nIn one sense this merely restated the basic political justification of patents, which was that inventors were granted a patent in return for disclosing inventions that they might otherwise have kept secret. But the particular characterization of the invention as a process from which copies of a thing might be 'indefinitely multiplied' gets at the lawyer's sense of the invention as a prototype. The invention was not the material product but the recipe from which copies were reproduced.\n\nThe notion of 'recollection' is important because it expressed the principle that inventors had to be in possession of the invention that they purported to disclose, which in turn meant that they had to master the new principle well enough to be able to survey or traverse it intellectually, and well enough to be able to communicate it to other persons 'skilled in the art'. Towards the end of the nineteenth century, William Robinson made the same point by using the metaphor of 'perception' rather than 'recollection':\n\n> To create by accident without a recognition of the fact or nature of his own creation, and consequently without the power to repeat the same creative act, is not invention. [I]t is essential to the inventive act that the inventor should not only _con_ ceive, but should also _per_ ceive his original idea, and should do both so clearly as to make this idea an actual addition to his fund of knowledge, and to be able to communicate it to the public.\n> \n> (Robinson, 1890: 121\u2013122)\n\nOne could discover something accidentally, by recognizing the significance of a fortuitous event, but such a discovery became a patentable invention only if the inventor had understood it well enough to recollect it, and to present that recollection as a recipe from which it could be reproduced by 'the person skilled in the art'.\n\nThe invention became a prototype only through the agency of paper. First, the recollection of the invention in print proved that one kind of prototyping had resolved into an endpoint \u2013 that the inventor's experiments had matured into a complete 'perception' of the idea. Second, the genetic agency of the invention as an industrial prototype, as endpoint turned into original template, was an effect of its constitution in textual and graphical form. Paper was the medium in which the invention was constituted as the active principle protected by patent law. The agency or instrumentality of the invention was an effect of its representation in texts and drawings, or, more precisely, of the way that texts and drawings were drawn into the discursive practices that unfolded in the various institutions in which the invention had to be elicited and scrutinized. Paper was also the medium in which the invention was supposed to be anatomized, and formulated in terms that would make it intelligible and practicable for artisans or manufacturers. In that sense, paper was the medium in which the invention became an effective template or instrument (even if that template could be used freely only after the patent expired). A late nineteenth-century Supreme Court decision referred to the material thing or machine as an 'outward embodiment of the terms contained in the patent', meaning that the primary and most vital medium of the intangible invention was the text, and that the material artefact was merely the secondary or reflected embodiment of the representation formulated in text. Thus, the invention as prototype was 'prototyped' by the medium of paper and the discursive operations in which paper was mobilized.\n\nThe agency of paper might be formulated in more abstract, and perhaps more consequential terms. Media theorist Friedrich Kittler observes that the rise of engineering in Europe was made possible by the combining of texts and drawings in the plane of the printed page. To some extent, this observation merely recalls the effects of standardization that came with 'print culture' (Eisenstein, 1982); engineering knowledge circulated in the form of 'technical drawings, construction plans, and sketches, which looked the same in every printed copy simply because they were indestructible reproductions of a single original' (Kittler, 2010: 66). The ability to reproduce images as woodcuts predated the emergence of the printed press, so what is important is that the technical drawings that were joined with text were based on linear-perspectival geometry: the image and accompanying text 'correctly' reproduced an 'original' machine or structure because this original had been translated into an image by a draftsman who saw it through the frame of linear-perspectival geometry and because it was then decoded by a reader who visualized it using the same perspectival _dispositif_. Thus, '[t]he book became a medium in which technical innovations as such could take place. They could be stored, shared, and even advanced with the help of technical drawings in the text' (Kittler, 2010: 67). The emergence of this perspectival _dispositif_ might well have been coupled to the emergence of the governmental diagram that is described in Biagioli's story of the patent specification: just as the specification assured the transition from privileges to rights, so did the perspectival geometry implemented by the specification make the transition from 'guild secrets' to 'the knowledge of engineers' (Kittler, 2010: 67).\n\nGiven that the specifications lodged in the US Patent Office in the first half of the nineteenth century were manuscripts illustrated with non-perspectival drawings, it is not evident how this 'new' _dispositif_ of print and perspective came to shape the juridical-discursive prototyping of inventions. However, the problematic here is not to question _how_ this translation occurred, but rather to query the _effects_ of this _dispositif_. The theory that machines could be prototyped in paper was current in the nineteenth century, and it is also true that some scientists and engineers in the United States, notably the patrons of the mechanics' institutes that were established in the middle of the century, believed that patent specifications could be taken as a means of communicating innovation, even though access to the full specification meant paying Patent Office scribes and draftsmen to copy out the text and drawings at the then considerable cost of 20 cents per sheet. But the primary addressees of the patent specification were not engineers or artisans but lawyers, patent examiners, and judges. The construction of the invention as a prototype took place in the legal settings, where lawyers or patent examiners judged the adequacy of the specification by reference to the legal fiction of the 'person skilled in the art'. Accordingly, what really matters for our purposes is the discursive practice and competence of this community of interpreters. In a discussion focused on the US patent system in the latter part of the nineteenth century, Lisa Gitelman suggests that lawyers took the patent specification to be the transparent representation of an existent invention:\n\n> The patent system assumes a na\u00efve relation between words and things. According to law, any 'useful art, manufacture, engine, machine or device', may be protected by 'describing the said invention or discovery clearly, truly, and fully', and undergoing a successful examination by the patent office. Any technology can be described, and any invention can be neatly, if not uniquely, divided into individual claims, that is, bite-sized acts of objective signification, for which language is presumed to be clear, clean, natural, and free from ambiguity. No symbolic or figurative meanings pertain; the whole is exactly equal to the sum of its parts. In an age that imagined so many machines for language, the patent process offered a language for machines'.\n> \n> (Gitelman, 2000: 101)\n\nAnd, to the extent that the invention could not be captured in text, the lawyers' understanding was that 'drawings do what language cannot' (Gitelman, 2000: 106).\n\nPatent lawyers might have been somewhat more knowing about their own naivety. Writing in the late 1930s, Emerson Stringham, the pre-eminent expert on patent claim drafting, dismissed the idea that claims were indexed to the 'essential reality' of the invention:\n\n> The effect upon Western culture of the discovery that \"essential reality\" is a human concept having no foundation in the external world cannot yet be predicted. The discovery is only slowly filtering into our consciousness. When the dust has cleared we shall still have \"essential reality\" but we shall recognize it as something to be repeatedly picked and chosen and discarded, simply as seems convenient from time to time.\n> \n> (Stringham, 1939: 32)\n\nStringham approached the question of what the patent claims \u2013 and by implication the patent text as whole \u2013 might actually refer to by way of a reflection on the writings of Berkeley, Freud, Korzybski, and, ultimately, Pierce's differentiation of qualisigns, sinsigns, and legisigns: 'A patent claim never identifies anybody's \"invention\", but is merely a set of legisigns that, by virtue of conventional usage, help to select those events with respect to which the law will take some steps' (Stringham, 1939: 977). Despite his pre-eminence, Stringham was not in the usual run of patent lawyers, and the flourish of philosophy (and realism) was more characteristic of mid-twentieth-century than mid-nineteenth-century patent jurisprudence. But although most nineteenth-century patent lawyers in Europe and the US probably did believe that there really was such a thing as the invention waiting to be found in a material machine or artefact, there is one basic reason why the language of a specification could only have been understood and practised forensically, as a means of contesting or negotiating the shape and effect of the invention. In the age of mechanical invention, language always fell short of reality, not so much because the invention was intangible but because it was transitive.\n\n## **Transitive types**\n\nIn the introduction to his patent treatise, Curtis characterized the invention in terms of a distinction between a 'form' and a 'means':\n\n> [A]ll inventions are independent of form, except those whose entire essence, purpose, and characteristics begin and end in form alone, as is the case with all manufactures, the utility and advantage and proposed object of which depend on form. But where there is a purpose that does not begin and end in form alone, where the form or arrangement of matter is but the means to accomplish a result of a character which remains the same, through a certain range of variations of those means, the invention is independent of form and arrangement, to this extent, that it embraces every application of means which accomplishes the result without changing its nature and character.\n> \n> (Curtis, 1849: 17)\n\nInventions whose 'entire essence' consisted 'in form alone' were the kinds of consumer designs that were serially moulded by the industrial manufacturers that emerged in the US from the 1840s onwards. In one sense, the distinction between the mechanical invention and the artistic work or consumer design turned on an eminently practical point. To identify a mechanical invention with the form of a machine would have been to make machine patents almost worthless: any competitor could evade the inventor's patent by changing some incidental feature of the machine. Thus, in patent law the material form of the machine was deciphered as an equivocal and potentially deceptive expression of the embodied invention. According to an influential nineteenth-century decision:\n\n> the principles of two machines may be the same, although the form or proportions may be different. They may employ substantially the same power in the same way, though the external mechanism be apparently different. On the other hand, the principles of two machines may be very different, although their external structure may have great similarity in many respects.\n> \n> ( _Barrett v. Hall_ 2 F.Cas. 914, 923 [1818])\n\nConsequently, the mechanical invention could not be identified with the external form of a machine, but nor could it be construed as a kind of fixed inner form, as a virtual blueprint or spectral architecture.\n\nThis insight distinguished the form of the invention from the prevailing understanding of form as an effect of operations of moulding. This understanding has its origins in the old distinction between the liberal and the mechanical arts. In his history of the imprint [ _l'empreinte_ ] as the 'countermodel of the idea of art' (Didi-Huberman, 2008: 102), Georges Didi-Huberman describes how the Renaissance sense of artistic creativity emerged from a contrast between these two modes of production. The expansion of modes of imprinting in the Renaissance period \u2013 exemplified in Pisanello's portrait medals, new techniques of engraving, the refinement of the art of sculpture, and (last but not least) the advent of the printing press \u2013 was paralleled by the emergence of a discourse in which these techniques were characterized as 'mechanical' arts and subordinated to the 'liberal' arts:\n\n> The form obtained by an imprint was incompatible with the very notion or ideal of art because it was too directly derived from matter as it already existed and insufficiently derived from the _idea_ that was so dear to the classical theory of art. Being both borrowed and imprinted [ _\u00e0 la fois 'emprunt\u00e9e' (pr\u00e9lev\u00e9e) et empreint\u00e9e (physiquement marqu\u00e9e)_ ] form was transferred directly from matter to matter. In order to come into existence it did not have to be formed in the mind of the artist.\n> \n> (Didi-Huberman, 2008: 121)\n\nAn imprint did not involve artistic creativity. The theory was that although painters and sculptors quite commonly used techniques of imprinting they used them only as means or instruments of artistic expression. The true artist 'molded with his mind' (Didi-Huberman, 2008: 129), and the mind thoroughly controlled the means by which a style was made material. This carries over into copyright law in the form of the assumption that literary or artistic style or expression is fixed in the material form, pattern, or immanent architecture of a work, or in a distinctive arrangement of words or brushstrokes. Plainly, these different kinds of work are expressed in different kinds of materiality, ranging from the plastic materials of traditional sculpture, through the density of certain kinds of painting media, to the print-based combinatory of 'standardized elements or characters' (Kittler, 2002: 38) that underwrites uniqueness of literary style.\n\nBy contrast, Curtis's description of the invention as a 'means to accomplish a result', or as the means of applying a 'natural law, or power, or property of matter' (Curtis, 1849: 7\u20138), expressed the jurisprudential conception of the mechanical invention as its 'mode of operation' rather than its form or ultimate function:\n\n> It is obvious that, where the invention is in machinery, the mode of operation embodied in such machinery must constitute the essence of the means of producing the result. If any one think otherwise, let him test it by supposing the mode of operation to be taken away from the machine, and see what will remain. To enforce this truth, imagine, if possible, a machine without any mode of operation, and what is it? Clearly nothing but the wood and metal composing it. This shows that the mode of operation is the characterizing feature.\n> \n> (Argument of counsel in _Burr v. Duryee_ 68 U.S. 531, 554 [1863])\n\nThe cryptic proposition that a mechanical invention consisted in 'the essence of the means of producing the result' gets at the idea that the invention was the machine construed as a mode of functioning abstracted from its ultimate functions, or \u2013 with apologies to Kantians \u2013 as a 'means in itself'. An inventor could not have a patent for the ends of the machine \u2013 its ultimate function \u2013 because such a right would be too broad and indefinite: '[The patentee] cannot describe a machine which will perform a certain function, and then claim the function itself, and all other machines that may be invented to perform the same function'; it followed that the invention encompassed by the patent had to be defined not by the end but by the means, and the distinction between these two terms had to be drawn clearly enough to allow patents to be issued for alternative means to the same end.\n\nReturning to the Renaissance notion of moulded form, Curtis's point was that the 'type' of the mechanical invention was not like the 'form' of an artistic work. It was because the 'principle' of a machine consisted in the way that a machine elicited an effect from nature \u2013 in the 'application of the natural law, or power, or property of matter' (Curtis, 1849: 7\u20138) \u2013 that the invention had to be apprehended as a 'means', or as a transitive type. Thus, the material substance of a machine, to take the paradigm of invention, was not 'formed' in the way that one forms plaster to make a sculpture; rather, it was 'arrayed' in the way that one arrays experimental devices and instruments to generate or precipitate a particular effect. So the 'principle' of a machine \u2013 the 'thing' protected by a patent and in an infringement action \u2013 was the specific inflection of physical or mechanical forces that was elicited by a mechanical configuration.\n\nThere are two important points about this sense of the invention as an effect elicited from nature. First, the effect in question was not the ultimate end or function of the machine but the physical or kinetic process generated in and by the operation of the machine. Second, for the purposes of patent law, the machine designed by an inventor was just the means by which nature was instrumentalized:\n\n> [T]he peculiarity of the invention consists in the effect produced by the application of the natural law, as an agent; and this effect is not changed, by the use of different vehicles for the action of the agent, provided there is still the same agent, operating in substantially the same way, to produce substantially the same effect or result.\n> \n> (Curtis, 1849: 8)\n\nOf course, the machine was essential to the invention, but the intimate type of the invention was the 'machinic' trace of the encounter between mechanism and natural 'agent', and this trace was the thing reproduced by legitimate and illicit 'copiers'.\n\nThe form of the invention was revealed and defined by the periodic articulation of a mechanism rather than by the material or physical composition of the machine. How could one fix a mode of operation on paper? Forensically, a mechanical invention could be apprehended only by seeing or imagining a machine in operation. Writing in 1890, and taking the example of a machine for planing lumber, William Robinson suggested that what should be seen or imagined was the operation of the machinery abstracted from its immediate objects:\n\n> the planer [is not] changed in character or capability by the presence or the absence of the lumber; [it acts] with equal energy, and in precisely the same manner, whether [its] action terminates on the material [object], or issues only in the beating of the air.\n> \n> (Robinson, 1890: 210)\n\nI have described elsewhere the ways in which the doctrinal conception of the mechanical invention as a mode of operation was conditioned by media of representation (Pottage and Sherman, 2010; Pottage, 2011). Although the paper specification was the official medium for the recollection and communication of inventions, the process of fixing and communicating the invention in print actually involved a diversity of media \u2013 voice, gestures, sketches, scale models, and technical drawings. And, at least until the second half of the nineteenth century, the patent model was the form in which most inventors, attorneys, and judges recollected, scrutinized, and delimited the invention. The question of the limpidity or objectivity of the language or drawings in a specification was largely beside the point because most of the addressees of these descriptions simply could not see a machine on paper. In 1878, a former Commissioner of Patents observed that most judges could not read engineering drawings \u2013 'there is not to exceed one in five, at most, and probably not so large a proportion, of our best judges on the bench who get any just conception of a machine by the drawing' \u2013 and that for this reason the only way to reach an understanding of the 'principle' of a machine was by demonstrating a scale model.\n\nWith the question of prototyping in mind, this suggests that we need to take a more expansive view of the discursive media and techniques that were involved in eliciting a prototypical agency from the material form of the machine. One might go somewhat further than Gitelman, and suggest that the language of patent law was not so much a language for machines as it was language that (in the philosophical sense) 'machined' machines (Pottage, 2011). And this process of machining might be explored at different orders of scale, from the close texture of the rhetorical, gestural, and material manoeuvres that brought the invention into being for a jury in a courtroom, or for an examiner in the Patent Office, to the more abstract diagram.\n\nSo, if the invention did not consist in a material form, or in an immaterial form that could be radiographed by the specification, then what exactly was the 'type' that was reproduced in each exemplar of the invention, and which an infringer might be held to have 'lifted'? If the invention did not consist in form, and if lawful or unlawful reproduction of the invention could not be construed as transactions in form, then what exactly was being transacted? Although the mechanical invention could be defined only negatively (it was not the material form of the machine, not its ultimate function, and not the natural forces or materials that were mobilized in a machine), patent lawyers assumed that there was a 'principle' waiting to be found in the patented machine: 'However inadequate the term may be to express what it is used to convey, it is obvious that there is a characteristic, an essence, or purpose of every invention, which, in our law, has been termed by jurists its principle' (Curtis, 1849: 10). Speculatively, one might say that a transitive type could acquire determinacy because the 'mode of operation' of a machine is periodic and repeatable. If, to adapt a classic definition, a mechanism is an assembly of elements that is articulated in such a way that its configuration is maintained (or, rather, periodically restored) throughout the operation of the machine, then this periodicity is capable of definition. And, for the purposes of property law, reproducibility is as good as materiality.\n\n## **Prototyping machines**\n\nAlthough the transitive 'mode of operation' of a machine could be fixed as an observable trace, it is obvious, first, that this effect was contingent on the specific setting in which the demonstration of a machine or mechanical model was performed and witnessed, and, second, that depending on the strategy of demonstration, on the way that models, texts, and drawings were articulated together, and on the expectations of the judges and jurors, the same machine could yield a quite different 'principle'. Indeed, the point of litigation was to contest the nature of the 'type' embodied in a mechanical invention. Each side in an infringement action would mobilize available discursive or forensic resources that were available in such a way as to press its own argument; each side would anatomize and recompose the machines in question so as to reveal a mode of operation that supported its case. There were only a few variables to work with \u2013 patent jurisprudence reduced machines to a set of basic elements: 'shape, size, capacity, proportions, arrangement, or materials' (Robinson, 1890: 320) \u2013 but the invention was a legal rather than a mechanical artefact. In one prominent early nineteenth-century decision, the trial judge took the parties' lawyers to task for turning an infringement action into a lesson in mechanical science:\n\n> the question whether the principles are the same in two machines, is in reality, when all the facts are given, rather a matter of law, than of the opinion of mechanics; at least matter of law is necessarily mixed up with it, which mechanics could not be presumed to be acquainted with.\n> \n> ( _Barrett v. Hall_ 2, F. Cas. 914, 923, [1818])\n\nThe point was that lawyers should demonstrate mechanical form or draw on the 'opinion of mechanics' only for the purposes of revealing the inventive 'principle' of a machine, which was 'a matter of law'. So the machine, or the model that stood in for it, was actually two machines in one. In one dimension, the machine or model was a creature of mechanical ingenuity, an artefact embodying a principle that could be identified and explained by expert witnesses in the language of mechanical powers, structures and forces. In its alternative dimension, the model embodied the legal schema of the invention, a 'principle' that was accessible only to the legal imagination. Mechanical and legal re-engineering worked reciprocally; demonstrations switched between the two dimensions of the model, engineering mechanical form in such a way as to sustain a favourable interpretation of the doctrinal notion of invention, and engineering legal form so as to ask new questions of mechanical opinion.\n\nKeeping in mind the question of prototyping, and more especially the classical understanding of prototypes as instructional means that relayed design to manufacture and consumption, the most interesting effect of this mode of prototyping was that it turned machines into means in themselves. Patent law, which presented itself as a thoroughly instrumentalist technique, effectively turned the instrumentalist logic of the nineteenth-century inside out. From within a deeply entrenched understanding of machines and mechanical processes as means to ends, of industrial manufacturing as the means to achieve economy, efficiency, and precision, and of human competences as mere components or adjuncts to this mode of production, patent jurisprudence turned machines into instruments abstracted from their industrial causes or effects. Whereas engineers understood prototypes as designs that could be translated from a model or blueprint into a machine that would articulate the instrumental flow of an industrial process, the legal definition of the mechanical invention suspended or involuted this linear flow. The effect of construing machines as means in themselves was to locate the invention in what cyberneticians or systems theorists would call the 'blind spot' of each of the hallmark distinctions of the industrial ideology (means and ends, matter and form, inputs and outputs). In other words, and this was precisely what made the task of apprehending inventions so difficult, the invention took shape in the zone of indistinction of each of these cardinal distinctions, the zone in which they remained latent or contingent.\n\nOf course, the figure of the machine as a transitive type immediately re-entered the industrial economy; patents secured the investments made by 'projectors' and manufacturers in production runs, and they became bases of economic speculation in their own right. The peculiar definition of machines as means in themselves was essential to these industrial and economic functions. The definition of the machine as a 'mode of operation' enabled patent lawyers to identify patent infringement by reference to criteria that anticipated and defused the strategies that infringers might use to avoid liability for copying the 'type' embodied in a mechanical invention. So one might say that the virtue of locating the invention in a zone of indistinction was that it enabled the 'type' of the machine, its intangible substance, to be responsive to perceptions of the 'fairness' of certain kinds of industrial competition, and that in return the spirit of the machine, its inventive 'principle', took shape as a reflex of the articulation of a broader industrial _dispositif_.\n\n## **Conclusion**\n\nAlthough there is a longstanding historical sense of prototypes as ideal forms rather than material or medial agents, few people these days think of prototypes without having in mind the working sketches, experimental models, technical specifications, blueprints, and trial manufactures that are used to test mechanical hypotheses, to probe potential functionality or marketability, or to instruct mass production. Moreover, the recent critical understanding of prototyping as a contingent, recursive, or reflexive operation has been informed by the shift from these classical industrial prototypes to informatic prototyping; namely, rapid prototyping, direct writing, computer-aided design or manufacturing, and 3-D printing. Although some practitioners still understand these newer procedures instrumentally, as modes of 'instructional design', they mark the emergence of what one might call 'objectile' objects: things which no longer have the 'appearance of essence' that was generated by processes of industrial reproduction (see Deleuze, 1980: 26).\n\nIf in the case of the 'objectile' object the object becomes an event, then one might say that patent discourse reversed this move: the 'type' of the mechanical invention was an event fixed as an object. And what was fixed was not a specific intersection of form and matter, or an effect of moulding, but a transitive 'modulation' of the elements of the discursive _dispositif_ of patent law, and also of the machines to which a 'mode of operation' was ascribed. Some patent lawyers were alive to the philosophical implications of this mode of modulation. Discussing the prevalent sense of mechanical inventions as 'combinations' of known elements, Emerson Stringham drew on Whitehead's _Process and Reality_ to explain why a mechanical combination was a discursive artefact, or a product of judgment:\n\n> [For Whitehead] 'the production of novel togetherness' is the ultimate notion in the term 'concrescence'. These ultimate notions of 'production of novelty' and of 'concrete togetherness' are inexplicable either in terms of higher universals or in terms of the components participating in the concrescence. The analysis of the components abstracts from the concrescence. The sole appeal is to intuition.\n> \n> (Stringham, 1939: 1150).\n\nAnd, as in Deleuze's notion of postindustrial modulation, patent discourse machined the terms that the logic of moulding takes as determinate starting points: form and matter, means and ends, production and consumption. Even if transitive types were folded back (see Strathern, 2006) into the industrial economy, their mode of fabrication or prototyping suggests that patent law had created something like a postindustrial machine, the machine in question being the discursive or medial _dispositif_ of patent jurisprudence itself, more than the artefacts that were produced by that machine.\n\n E.g. section 2 of the US Patent Act of 1790.\n\n According to the dictionary definition, 'the mental operation by which objects or ideas are revived in the mind' ( _Oxford English Dictionary_ , online edition 'recollection').\n\n 'Paper' rather than 'print' because it was only in 1866 that the Patent Office began printing patent specifications.\n\n _Bischoff v. Wethered_ 76 U.S. 812 (1869).\n\n 'In the 1850s, a simple drawing of a washing machine attached to a patent application might look quite similar to a large cross-section of the Great Eastern steamship drawn to woo investors; both would show a non-perspective view, use blue colouring for steel and yellow for brass, and show light coming from the upper left' (Rankin, 2011: 67).\n\n See Patent Act 1793, section 11.\n\n For another example, see Kent (1945).\n\n Kant distinguished between literary works and other kinds of artistic 'manufacture' on the basis that whereas the material book was the medium through which the living 'speech' of the author was relayed, an engraving or sculpture was fully realized by its material form (Kant 1996b: 29\u201335). Nonetheless, this ongoing 'speech' presupposed the finite elements of a graphic combinatory.\n\n _Corning v. Burden_ 56 U.S. 252, 270 (1853).\n\n See Arguments before the Committee on Patents April 3rd 1878, 45th Congress 2d Session Mis Doc 50, pp. 99\u2013100.\n\n So, to take the leading example, Samuel Johnson's dictionary defines prototype as: 'the original of a copy; exemplar, archetype' (Johnson, 1766), at p. 413, s.v. 'prototype').\n\n## **References**\n\nBabbage, C. (1851) _The Exposition of 1851: or Views of the Industry, the Science and the Government of England_ , London: John Murray.\n\nBiagioli, M., Jaszi, P. and Woodmansee, M. (eds) (2011) _Making and Unmaking Intellectual Property. Creative Production in Legal and Cultural Perspective_ , Chicago, IL: University of Chicago Press.\n\nCurtis, G. (1849) _A Treatise on the Law of Patents for Useful Inventions_ , Boston: Little & Brown.\n\nDeleuze, G. (1980) _Le pli. Leibniz et le baroque_ , Paris: Minuit.\n\nDidi-Huberman, G. (2008) _La ressemblance par contact. Arch\u00e9ologie, anachronisme et modernit\u00e9 de l'empreinte_ , Paris: Minuit.\n\nGitelman, L. (2000) _Scripts, Grooves, and Writing Machines. Representing Technology in the Edison Era_ , Stanford, CA: Stanford University Press.\n\nKant, I. (1996a) 'On turning out books', in Mary J. Gregor (ed.), _Practical Philosophy_ , Cambridge: Cambridge University Press.\n\nKant, I. (1996b) 'On the wrongfulness of unauthorized publication of books', in Mary J. Gregor (ed.), _Practical Philosophy_ , Cambridge: Cambridge University Press.\n\nKent, C. (1945) 'General semantics and the patent system', _Journal of the Patent Office Society_ , 27: 37\u201347.\n\nKittler, F. A. (2000) _Optical Media_ , Cambridge: Polity Press.\n\nKittler, F. (2002) 'The perspective of print', _Configurations_ , 10(1): 37\u201350.\n\nOng, W. (2002) _Orality and Literacy. The Technologizing of the Word_ , London: Routledge.\n\nPottage, A. (2011) 'Law machines. Scale models, forensic materiality, and the making of modern patent law', _Social Studies of Science_ , 41: 621\u2013643.\n\nPottage, A. and Sherman, B. (2010) _Figures of Invention. A History of Modern Patent Law_ , Oxford: Oxford University Press.\n\nRiles, A. (2011) _Collateral Knowledge. Legal Reasoning in the Global Financial Markets_ , Chicago, IL: University of Chicago Press.\n\nRobinson, W. C. (1890) _The Law of Patents for Useful Inventions_ , Boston: Little, Brown & Co.\n\nStrathern, M. (2006) 'Cutting the network', _Journal of the Royal Anthropological Institute_ , 2: 517\u2013553.\n\nStringham, E. (1939) _Patent Claims. A Drafter's Manual_ , Madison, WI: Pacot Publications.\n* * *\n\n# Index\n\n* * *\n\n2000 Lisbon European Council\n\n2004 Department of Health survey: fertility clinics \u2013\n\n2007 Opinion of the Article 29 Working Party\n\nAarhus Convention (1998) 40n7\n\nAB 32\n\nabortions\n\nabsolutism\n\nAcademy of Medical Sciences (AMS) \u2013; review of current regulatory frameworks\n\nACP (Advisory Committee on Pesticides)\n\n_acquis communautaire_\n\nactor-network theory (ANT) _see_ ANT\n\nAdams, J. A.: 'it's normal to be normal' publication ; review of child and adolescent sexual assault victims\n\nadoption\n\nAdvisory Committee on Pesticides (ACP)\n\n_agencement_ ,\n\nAgency for Toxic Substances and Disease Registry (ATSDR): complaints against ; conference calls \u2013; formaldehyde minimum response levels ; forwarded results to FEMA ; level of concern , \u2013; nuisance symptoms ; toxicological investigation techniques\n\nagricultural commodities: financial derivatives\n\nAIDS: benefits and housing \u2013; credibility struggles between scientific teams ; crisis management ; dawning crisis ; defined in temporal terms ; emergence ; homophobic moral panic ; link with HIV ; media scaremongering ; paradigm of imminent death , ; pragmatic hope ; prognosis _see_ prognosis; schematisation ; social chronology ; _see also_ HIV; PWAs\n\nAlder, K.: American obsession with lie detectors ; polygraphs in America \u2013\n\nAlexander, Alana\n\nallergic sensitisation \u2013\n\n_American Journal of Obstetrics and Gynaecology_\n\nAmorim, Ant\u00f3nio ,\n\nAmsterdam Treaty\n\nANT (actor-network theory): invisibility of law ; socio-materiality\n\narbitrage\n\nArticle 11 TFEU\n\nArticle 3(3) of the Treaty on European Union (TEU)\n\narts: liberal ; mechanical\n\nasset markets\n\nasthma: bronchodilating inhalers ; defined by pharmaceutical industry ; environmental triggers \u2013; impact of formaldehyde exposure \u2013; pathophysiologies ; reconceptualisation of ; treatment of\n\nAtkinson, P.\n\nAtomic Energy Commission\n\nauctions\n\nAugmented Reality\n\nAZT\n\nBarbour, Virginia\n\nBarrett, Cleveland \u2013\n\nBartrip, B. ,\n\n_Beleno v. Texas Dept. of State Health Services_ (2009) \u2013\n\nbellwether trial \u2013\n\nBenrubi, G. I.: genital microtrauma caused by consensual sexual intercourse\n\nBiagioli, Mario: norms within science ; relation between political representation and representation of inventions ; technologies of integrity verification\n\nBiermann, F.\n\nBingaman, Senator Jeff\n\n_Biotech Products_ case \u2013\n\nBlackmore, Sue\n\nBlack-Scholes-Merton formula\n\nBlackwell\n\nblood pressure: apparatus for recording arterial blood pressure ; history of ; relationship to truth and deception ; relationship with fear ; systolic blood pressure deception test ,\n\nbloodspot collections , \u2013, \u2013\n\nboundary work\n\nBowyer, L.: injury interpretation\n\nBoyd, W.\n\nBoyle, Robert ,\n\nBrazil: challenge to EU Directive\n\nBrennan, P. A. W.: advocates routine use of colposcope \u2013; potential for the colposcope\n\nBretton Woods monetary management system\n\nBristol Centre for Reproductive Medicine (BCRM)\n\nBritish Fertility Society\n\nBritish Medical Association (BMA): work against quackery\n\nBritish Medical Association (BMA) House\n\n_British Medical Journal_\n\nbronchoconstriction\n\nbronchodilating inhalers\n\nBrown, C. R V.: evidence and image recordings ; study of injury caused by consensual and nonconsensual sexual intercourse \u2013; study on colposcopes \u2013; Toludine Blue \u2013\n\nBrown, M. A.\n\nBrownsword, R.\n\nBrundtland Report ,\n\nBryan, D.: value\n\nBSE crisis 17n1,\n\nBunn, G. C.: notion of the born criminal\n\nBush, George W.\n\nCalifornia: emissions targets ; lawsuits ; low-carbon fuel standard ; political settlement ; regulatory power struggle between State and county\n\nCalifornia Air Resources Board (CARB) ,\n\nCalifornia Environmental Quality Act (CEQA)\n\nCallon, Michael , , ; market formations\n\ncapital: commercial ; controls ; industrial ; speculative\n\ncarbon regulation \u2013,\n\nCare Quality Commission\n\nCBRC (cross-border reproductive care) \u2013; reproductive exile\n\nCenters for Disease Control and Prevention (CDC): conference calls \u2013\n\ncitizen participation: citizenship-as-participation ; citizenship-as-rights ; EU's strategies and policies on sustainable development ; Innovation Union ; _innovator citizens_ ; limited in EU SDS ; sustainable development ; technologies of participation _see_ technologies of participation; _see also_ publics\n\ncitizens: EU (European Union)\n\nClean Air Act (CAA) \u2013, ; epistemic jurisdiction\n\nclimate change , , ; bottom up regime formation ; climate federalism _see_ climate federalism; conflicts between regulatory authorities ; regulation at State and local level ; site of a complex global order\n\nclimate federalism: carbon regulation \u2013, ; decentralised approach ; GHG emission targets ; jurisdictional friction ; _Massachusetts v. EPA_ ; patchworks of regulation ; political settlement ; US obstructionism towards reducing GHGs\n\nClinton, Bill\n\nCode of Practice \u2013\n\nCodex Alimentarius Commission\n\nCollins, Harry ; separation of roles for experts and politics\n\ncolposcopes: access to images \u2013; benefits of ; circumscribed applications of ; controversy around ; cost of ; design of ; discretion over whether examinations are recorded , ; distinguishing between consensual and nonconsensual sexual intercourse ; dual role applications of , ; evidence and image recordings ; evidential benefit at expense of victim care ; examinations ; FNEs use in training ; forensic medical examinations ; forensic medical practitioners' support for \u2013; greater clarity than macroscopic observation ; improved visibility for genital examinations ; management of \u2013; recording decisions \u2013; requirement for therapeutic and evidential practices ; security of images , , ; study of \u2013; survivors' consent for examinations ; usefulness or appropriateness ; victims of routine application ; _see also_ images; rape; sexual assault\n\ncolposcopy controversy: evidential benefits of colposcope \u2013; injury interpretation \u2013; new quest for the rape injury \u2013; potential benefits identified by gynaecologists ; victims and the absence of injuries \u2013\n\nColumbia\n\ncommercial capital\n\nCommissioner for Research, Innovation, and Science\n\nCommittee on Publication Ethics (COPE) _see_ COPE\n\ncommodities: financial derivatives\n\ncomplex governance\n\nComte, Auguste , 20n6\n\nConseil d'\u00c9tat \u2013\n\ncontracts: financial derivatives \u2013\n\ncontractual outsourcing \u2013\n\nConvention on Human Rights (1950)\n\nconvictions: false eyewitness evidence ; nature of power and technoscience ; wrongful\n\nconvicts: barriers to research on ; definition 73n8; exclusion from democratic science and technology processes ; experience and expertise of \u2013; identification by anatomical features ; knowledge of technology ; marginalised groups , ; margins of society ; subordination of expertise\n\nCooper, Christopher: aggravation by water damaged trailer ; asthma and formaldehyde \u2013; bronchial sensitivity placed within the pharmaceuticalised definition of asthma ; medication \u2013; visit to National Jewish Hospital\n\nCOPE (Committee on Publication Ethics) ; award from Council of Science Editors (CSE) ; Britishness ; Code of Conduct ; convening of ; cosmopolitan membership ; Forum \u2013, __ , , ; funding ; innovative methods of deliberation ; large international network ; open-ended deliberations assessing misconduct ; open-ended publication of cases ; promotion of good practice in science ; virtual Forum \u2013; _see also_ misconduct hunting\n\ncoproduction: co-constitutional realms ; political centralisation and decentralisation ; regulatory science and jurisdictional architecture\n\ncopyright law ; emergence\n\nCouncil of Europe ,\n\nCouncil of Science Editors (CSE)\n\nCourtney, G. M.: critique of colposcopy ; rape without signs of injury ; sympathy to 'it's normal to be normal' publication\n\nCourt of Justice of the EU\n\nCox, Brian , ; politicians and science\n\ncriminal justice system: benefits of colposcopes , ; hegemony ; polygraphs ; problems and possibilities for errors ; recognition and power of experts ; social status and power of actors ; subordination of expertise of convicts ; truths , ; unexamined assumptions \u2013; _see also_ miscarriage of justice\n\ncriminals _see_ convicts\n\ncriminology: notion of the born criminal ; studied through biological lenses\n\nCrimp, Douglas\n\ncritical discourse theory\n\ncritical political economy _see_ feminist political economy\n\ncross-border reproductive care (CBRC) \u2013; reproductive exile\n\nCrown Office and Procurator Fiscal Service\n\nCurtis, George Tinkner: characterisation of invention ; material things and secret processes ; mechanical invention ; nature instrumentalised\n\nDahl, Robert\n\nDalton, M. E.: injury interpretation\n\ndata collection: Eurostat\n\n_Data Protection Act 1998_\n\ndata protection law: consent for retention ; consent for use in research \u2013; imperative for change\n\nData Sharing Review (2008)\n\n_Daubert Trilogy_\n\n_Daubert v. Merrell Dow Pharmaceuticals_ \u2013\n\nDavis, Fred\n\nDawson, A.: role of HFEA inspector\n\nDDA (Disability Discrimination Act) 1995: asymptomatic HIV carriers ; demonstration of likelihood \u2013; expert medical reports ; introduction ; legal redress for HIV discrimination \u2013; progressive conditions ; proof of disability\n\ndeath sentences\n\ndeception: relationship to blood pressure ; systolic blood pressure deception test ,\n\ndeceptive consciousness\n\ndecisionism: definition of , ; elected politicians ; role of science in \u2013\n\ndelegation: structures and accountability\n\nde Marchi, B.: high-profile controversies\n\ndemocracy: epistemic jurisdiction and ; nature of\n\ndemocratic experimentation\n\ndemocratic politics\n\nDemos 17n1\n\nDenmark\n\nDepartment of Defense\n\nDepartment of Energy (DOE) _see_ DOE\n\nDepartment Of Health: 'Report on the arm's-length bodies review'\n\nDepartment Of Health (Ireland)\n\nderivatives _see_ financial derivatives\n\ndetective fiction\n\nDidi-Huberman, Georges\n\ndigit preference theory\n\nDirective on the Legal Protection of Biotechnological Inventions\n\nDisability Discrimination Act 1995 (DDA) _see_ DDA\n\ndisplaced persons , 189n1; emergency housing units for Gulf Coast residents ; health problems of trailer residents ; subjected to sustained formaldehyde exposures\n\nDispute Settlement Body (DSP)\n\nDNA testing: Cleveland Barrett case \u2013; DNA wars ; ethnic profiling ; familial searching 77n10; flawed analysis of samples ; forensic analysis ; limitations ; markers and individuals \u2013; profiles ; random match probability 77n10; statistics \u2013, ; stratification ; STRs \u2013, 77n11; technological instrumentalism , \u2013; traditional technologies ; true individualisation , ; underlying principle\n\nDNA wars\n\n_Do Economists Make Markets?_\n\nDOE Counter Intelligence Office \u2013\n\nDOE (Department of Energy) \u2013; counterintelligence (CI) reports ; creation and purpose ; National Research Council report \u2013; new polygraph program ; NRC report challenges validity of polygraphs \u2013; polygraph testing of employees ; protests by employees against polygraph testing ; public hearings on polygraph testing ; restructuring \u2013; security issues\n\nDonor Gamete forum , \u2013\n\ndonors: anonymity \u2013, , , ; campaigns to support sperm donation ; extensive waiting times ; gametes ; identification by offspring of gamete donation ; recruitment ; register ; shortage , ; sperm , ; _see also_ gametes\n\ndynamic federalism\n\nEaster District of Louisiana\n\nEast Jefferson General Hospital\n\nEcuadorian Technical Commission\n\nelectronic tracking\n\nelite privilege: epistemic governance ; technocracy\n\nElsevier\n\nembryo donors _see_ donors\n\nEmployment Tribunals ,; disability tests\n\nenvironment: EU commitment to sustainable growth ; hazards \u2013; impact assessment of EU SDS ; limited citizen participation in governance ; public engagement in issues\n\nEnvironmental Health Network\n\nenvironmental impact report (EIR)\n\nEnvironmental Protection Agency (EPA) , , ; conference calls \u2013; delays releasing test results ; tests on unoccupied trailers\n\nEPA (Environmental Protection Agency) _see_ Environmental Protection Agency (EPA)\n\nepistemic governance: delegation of expert advisors \u2013; dilemmas \u2013; elite privilege ; focus on experts \u2013; independence of expertise ; normative principle of transparency \u2013; public\n\nengagement \u2013; public interest ; realm of ; scientific advisory systems ; uncertainty\n\nepistemic jurisdiction: Californian political settlement ; carbon regulation \u2013, ; definition ; EC food safety ; political centralisation ; questions for democracy\n\n'Equity and excellence: Liberating the NHS'\n\nespionage \u2013\n\nethanol regulation\n\nEU (European Union): Better Regulation agenda ; citizens ; Directive on the Legal Protection of Biotechnological Inventions ; Europe 2020 strategy \u2013; imaginaries ; industrial policy agenda ; legal-constitutional order ; low-carbon fuel standard model ; political identity ; politics of subsidiarity ; Resource-efficient Europe ; sustainable development _see_ sustainable development; _see also_ Innovation Union\n\nEurope 2020 strategy \u2013\n\nEuropean Association of Science Editors (EASE)\n\nEuropean Clinical Trials Directive\n\nEuropean Commission (EC) ; consolidation of epistemic jurisdiction ; flagship initiatives ; harmonizing social regulation \u2013; sustainable development progress reports\n\n_European Communities\u2014Measures Affecting the Approval and Marketing of Biotech Products_\n\nEuropean Community Treaty\n\nEuropean Convention for the Protection of Human Rights and Fundamental Freedoms 50n2\n\nEuropean Convention on Human Rights (ECHR): child's right to personal identity\n\nEuropean Council ,\n\nEuropean Court of Human Rights (ECtHR): data protection ; personal data of arrested persons ; retention of data \u2013\n\nEuropean Data Protection Regulation\n\nEuropean Food Safety Authority (EFSA) ; political legitimacy\n\nEuropean Medicines Agency (EMA)\n\nEuropean Parliament\n\nEuropean Parliament Committee report\n\nEuropean Society of Human Reproduction and Embryology (ESHRE) Task Force: survey\n\nEuropean Union (EU) _see_ EU\n\nEurostat\n\nEU SDS (EU Sustainable Development Strategy): complementary process ; data collection ; design of ; environmental impact assessment ; global promotion of sustainable development ; new approach to policy-making ; technocratic and expert-led governance ; unsustainable trends\n\nEU Sustainable Development Strategy (EU SDS) _see_ EU SDS (EU Sustainable Development Strategy)\n\nEU Tissues and Cells Directive (EUTCD) ,\n\nEvans, R.: separation of roles for experts and politics\n\nEvans, S.\n\nEwick, P.: authenticating site of science\n\nexchange rates , ,\n\nexpertise ; blurring of science and politics ; courtroom context ; delegated authority \u2013; EU SDS (EU Sustainable Development Strategy) ; influence on policymaking ; notion of ; ontologies of \u2013; role of experts in legal settings\n\nexperts: accountability ; committees ; Comtean view on ; role in relation to public and policymakers \u2013; science\/politics interface ; separate roles for politics and ; uncritical acceptance of authority of ; Weberian view on\n\nexperts-in-policymaking \u2013\n\neyewitnesses: false evidence\n\nEzrahi, Y.: tension between equality and order\/governance\n\nFabrication, Falsification and Plagiarism (FFP)\n\nFaculty of Forensic and Legal Medicine ,\n\nfamilial searching 77n10\n\nFarthing, Michael ,\n\nFaulkner, Alex\n\nFederal Communication Commission (FCC)\n\nFederal District Court: carbon regulation\n\nFederal Emergency Management Agency (FEMA) _see_ FEMA (Federal Emergency Management Agency)\n\nFederal Rules of Evidence (FRE)\n\nFEMA (Federal Emergency Management Agency) ; complaints to ; conference calls \u2013; press release on ATSDR report ; response to formaldehyde testing campaign ; _see also_ formaldehyde\n\nFEMA trailers ; chemical constitution of ; Christopher Cooper asthma case \u2013, , \u2013; exposure threshold ; formaldehyde content ; level of concern \u2013; nuisance symptoms ; safe levels of formaldehyde\n\nFEMA travel parks ; emergency housing units for Gulf Coast residents\n\nfeminist political economy: constructivist approaches ; nexus between economy and society \u2013; reconceptualisation of ; reproductive labor ; value-making processes\n\nfertility clinics \u2013\n\nfertility support groups\n\nfinancial capitalism\n\nfinancial derivatives: affecting current value-making processes ; beneficial role in the economy ; construction of value ; contracts \u2013; dangerous innovations ; decreasing attention to ; definition ; distinction between hedging and speculation ; entanglement of goods and production ; function ; futures ; impact on prices ; management of risks of global investing ; options ; over-the counter (OTC) derivatives ; price volatility ; productive power ; real and financial life of derivatives \u2013; realm of commodities ; regulation to reduce speculation ; relation to performativity ; speculative capital ; speculative role ; technical standards and rules ; trend in market prices \u2013; types of ; underlying asset , ; value of \u2013; worldwide value\n\n_Financial Derivatives and the Globalisation of Risks_\n\nfinancial markets: performativity ; social insurers\n\nfingerprints\n\nfloating exchange rates\n\nFMEs (Forensic Medical Examiners): clearer colposcopic images of genital area \u2013; colposcopic images banned in Scottish courtrooms ; decision on when to record images , ; forensic intervention of rape assaults ; objective evidence gatherer\/provider of therapeutic care ; opposition to use of colposcopes in courtroom in Scotland \u2013; peer review meetings ; potential for the colposcope ; use of colposcope in training\n\nFNEs (Forensic Nurse Examiners): clearer colposcopic images of genital area \u2013; decision on when to record images , ; forensic intervention of rape assaults ; objective evidence gatherer\/provider of therapeutic care ; peer review meetings ; potential for the colposcope ; use of colposcope in training\n\nFOIA (Freedom of Information Act) letter\n\nfood safety regime: agreements ; _Biotech Products_ case \u2013; Codex Alimentarius Commission standards ; legal disputes \u2013; regulatory harmonisation ; regulatory science ; risk assessment and a global regulatory jurisdiction \u2013; scientific justification , ; scientific validity\n\nFord, President Gerald\n\nforensic science: debunking assumptions of infallibility ; evidence and image recordings from colposcopes ; fingerprints ; scrutiny of ; solving wrongful convictions ; _see also_ DNA testing; polygraphs\n\nformaldehyde: allergic sensitisation \u2013; ATSDR findings ; Christopher Cooper asthma case \u2013, , \u2013; content in FEMA trailers ; impact on asthma \u2013; minimum response levels ; safe levels in FEMA trailers ; testing campaign ; trailer residents subjected to sustained exposure\n\nForster, G. E.: critique of colposcopy ; rape without signs of injury ; sympathy to 'it's normal to be normal' publication\n\nforums, online: would-be parents \u2013\n\nFoucault, Michel: panopticon's visual force ; silences ; subjugated knowledges\n\nframing\n\nFried, Jordan 193n10\n\nFrumkin, Dr Howard\n\nFrye, James Alfonse \u2013\n\n_Frye_ standard \u2013; difficulties applying the rule\n\n_Frye v. United States_ \u2013,\n\nfuel standards\n\nFumagalli, A.\n\nfundamental value\n\nfutures (financial derivatives)\n\ngametes (sperm): campaigns to support sperm donation ; donation , ; donors ; purchasing online ; sperm donors , ; _see also_ donors\n\nGarry, Dr Vincent: challenges ATSDR report findings\n\n_General Electric Co._ v. _Joiner_ (1997)\n\ngenetically modified organisms (GMOs) _see_ GMOs\n\ngenetic patrimony\n\ngenomes , ; patents\n\ngenotypes\n\n_Georgia v. Tennessee Copper Co._\n\nGHG emissions , , \u2013\n\nghost authorship\n\nGillette, Becky ; EPA testing\n\nGitelman, Lisa\n\n'Give a Toss' campaign\n\nGlobal Agreement on Tariffs and Trade (GATT)\n\nGlobal Warming Solutions Act\n\nGM crop trials 17n1\n\nGMOs (genetically modified organisms): crops ; dispute ; EC delayed approval of ; European process approach ; trade conflict ; US challenge to scientific basis of European actions ; US products approach\n\ngood research practice: norms 137n2\n\ngoods: performativity\n\nGothenburg European Council\n\nGrundmann, R.\n\nGulf Stream Inc.: compromised health of employees ; employees' complaints \u2013; working conditions for employees\n\nGuthrie cards , \u2013\n\nHales, Stephen\n\nHarper, K.: 'it's normal to be normal' publication ; review of child and adolescent sexual assault victims\n\nHarrison, Edward \u2013\n\nhazards 22n10; environmental \u2013\n\nHealth Research Authority , 125n1\n\nHealth Service Executive (Ireland)\n\nhedging: farming community ; firms and governments\n\nHFE Act 1990 ; opposition to\n\nHFEA (Human Fertilisation and Embryology Authority) ; consideration to abolish , ; consultation into egg and sperm donation policies ; guidance for overseas fertility treatment ; legal challenges ; licensing of sperm ; proactive approach to donor recruitment ; responses to consultations surveys ; support for lifting donor anonymity ; survey of clinics \u2013\n\nHigh Level Group of National Regulatory Experts 41n11\n\nhighly active anti-retroviral therapy (HAART)\n\nHind, D.\n\nHIV: asymptomatic affected people \u2013; complying with DDA's specific definition ; DDA act ; demonstration of likelihood \u2013; difficulties of claimants making claims \u2013; disability test \u2013; disclosure issues ; employment discrimination issues ; expert medical reports ; legal issues for activists ; link with AIDS ; prognosis _see_ prognosis; schematisation ; temporal progression \u2013; temporal sequence ; testing and treatments , ; _see also_ AIDS; PWAs\n\nHolmes, Oliver Wendell \u2013\n\nhomeostasis\n\nHorlick-Jones, T.: high-profile controversies\n\nHorton, Richard ,\n\nHouse Committee on Oversight and Government\n\nHouse of Commons Science and Technology Committee\n\nHouse Of Lords Science and Technology Committee on Genomic Medicine\n\nHouse Of Lords Select Committee on Science and Technology 17n1\n\nHulme, Mick ; subjective (Bayesian) assessment\n\nHuman Fertilisation and Embryology Act (HFE Act) , \u2013\n\nHuman Fertilisation and Embryology Authority (HFEA) _see_ HFEA\n\nHuman Rights Act 1998\n\nhuman tissue: collapsing the regulatory divide with personal data \u2013; consent ; crossing the regulatory divide with personal data \u2013; destruction of collections \u2013; individual autonomy and control over data \u2013; legal acts of legislation ; legal uncertainty about status of tissue samples ; privacy concerns ; regulatory divide with personal data \u2013; regulatory systems \u2013; samples ; storage on DNA databases\n\n_Human Tissue Act 1961_\n\n_Human Tissue Act 2004_\n\nHuman Tissue Authority\n\nHurricane Katrina\n\nHurricane Rita\n\nIgert, Jill 193n10\n\nIllinois State Police\n\nimages: colposcopic examinations ; importance in scientific and medical work 90n5; ownership of ; persuasive evidence 90n5; secure storage of ; security of , , ; survivors' consent for recording of ; survivors' fears of exposure ; _see also_ colposcopes\n\nimaginaries: EU (European Union) ; key repository of the nation-state ; nation-states ; salience of ; symbolism\n\nimplied volatility\n\nInce, Robin , ; politicians and science\n\nindustrial capital\n\ninfertility newsgroups\n\ninflation rates\n\nInformation Commissioner's Office: Code of Practice \u2013\n\nInnocence Project: creation of ; distinction between perpetrator and victim ; false eyewitness evidence ; reliance on technology to correct errors ; uncritical acceptance of authority of experts\n\ninnovation: citizen participation in European project ; citizen's virtue ; definition of ; symbolic of EU positive qualities ; videos \u2013\n\nInnovation Union: aims of ; Augmented Reality ; EU funded projects ; Europe 2020 flagship initiative ; Innovation Union Lab ; pledge to involve citizens in projects ; promise of higher quality of life ; Tales from the Future ; website ,\n\n_Innovation Union Act_ \u2013\n\n_Innovation Union Act_ \u2013\n\n_innovator citizens_\n\nIntelligence Authorization Act\n\nintentional ignorance \u2013\n\ninterest rates\n\nIntergovernmental Panel on Climate Change (IPCC) _see_ IPCC\n\nInternational Committee of Medical Journal Editors (ICMJE)\n\nInternational Office of Epizootics 177n2\n\nInternet sperm providers\n\nINUK (Infertility Network UK): survey\n\ninventions ; becoming a prototype through paper ; copies indefinitely multiplied ; effect elicited from nature ; forms and means , ; mechanical \u2013, , \u2013; perception ; prospective ; recollection , ; representation of ; retrospective ; _see also_ paper prototyping; patents; prototypes\n\nIPCC (Intergovernmental Panel on Climate Change) , ; self-governance and peer review\n\nIrwin, A.: farm workers and pesticides\n\nIVF (in vitro fertilisation)\n\nJain, Sara Lochlann ,\n\nJardine, Professor Lisa\n\nJasanoff, Sheila: coproduction of technology and social order \u2013; epistemic disagreement ; legal writing about science\n\nJefferson Pulmonary Associates\n\njurisdictional architecture: period of rapid change\n\nKant, Immanuel\n\nKeeler, Leonarde ; avoided mass production of polygraphs ; department stores and banks \u2013; employee screening ; patent of an apparatus for recording arterial blood pressure ; patent ran out ; popularising the polygraph ; sales and training ; strategies for commercialising the polygraph\n\nKennedy, Donald\n\nKennedy, Ian\n\nKittler, Friedrich\n\nknowledge-based economies: 2000 Lisbon European Council strategy\n\nknowledge claims: policymaking\n\nknowledge-for-governance _see_ epistemic governance\n\nKnudson, S.: 'it's normal to be normal' publication ; review of child and adolescent sexual assault victims\n\n_Kumho Tire Co. v. Carmichael_\n\nKyoto Protocol ,\n\nlaboratories ; t nce\n\nlabour: productive\n\n_The Lancet_\n\nLarson, John\n\nLatour, Bruno: coproduction of technology and social order ; legal predictability ; legal temporalities \u2013; Royal Society debates ; temporality\n\nlaw: functioning of ; role of materiality \u2013; social studies of \u2013\n\nlaw of value: Marxist\n\n_LD and ED (Guardian Ad Litem) v. British Columbia Women's Hospital and Health Centre and British Columbia Children's Hospital_ (2010)\n\nLeather, Suzi ; warning against overseas clinics\n\nLee, B. ; financial derivatives \u2013\n\nlegal temporalities: analysis by Bruno Latour \u2013\n\nlegitimacy: normative judgements of better or worse decisions\n\nLeo, Richard \u2013\n\n_Leviathan and the Air-Pump_\n\nLewis, J.\n\nLezaun, Javier ; ontology and epistemology\n\nliberal arts\n\nliberal democracy: public knowledge and governmental evaluation \u2013\n\nlie detectors _see_ polygraphs\n\nlife sciences regulation: decentred approach ; decontextualisation of materials and data ; multiple authorities ; reflexive system of governance \u2013; research , \u2013; research ethics committees ; sector-specific legal regimes , ; struggle of law to keep pace in ; thickets and divides in \u2013; _see also_ human tissue; personal data; thickets\n\nLincoln, C.: concerns about the cost of colposcopes\n\nLipuma, E. ; financial derivatives \u2013\n\nLisbon Strategy\n\nLittlefield, M. M.: mind reading ; significant lying curve\n\nLombrosso, Cesare \u2013\n\nLow-Carbon Fuel Standard (LCFS)\n\nLuhmann, Nikolas\n\nLynch, John ; public health campaigns ; temporal progression of HIV\/AIDS \u2013\n\nMackenzie, D. ; performativity \u2013\n\nmacroeconomic stability\n\nManhattan Project\n\nmarginalised groups \u2013\n\nMarres, N.: public involvement in policymaking\n\nMarston, William Moulton: examination of James Alfonse Frye \u2013; experimental protocols ; first version of the lie detector ; systolic blood pressure symptoms of deception\n\nMarx, Karl ,\n\n_Massachusetts v. EPA_\n\nmateriality: attention to matter within legal scholarship ; entanglement with other kinds of objects ; law and \u2013; textual\n\nMauna Lao laboratory \u2013\n\nMay, Peter: hazards 21n10; policies-without-publics \u2013\n\nMcCarthyite period \u2013\n\nMcGoey, Linsey\n\nMcLean, Professor Sheila\n\nmechanical arts\n\nmechanical inventions \u2013, , \u2013\n\nMedical and Healthcare products Regulatory Agency (MHRA)\n\n_Medical Ethics_\n\nmedical journals: growth of ; past records ; quackery \u2013; statistics\n\nmedicine: professionalisation of \u2013\n\nMerton, Robert ; performativity\n\nMezzadra, S.\n\nMillennium Ecosystem Assessment: power dynamics\n\nMiller, Brad\n\nMirowski, P. ; measurement of commodities\n\nmiscarriage of justice: unexamined assumptions \u2013\n\nmisconduct hunting: COPE Forum \u2013; open-ended deliberations at COPE ; professionalisation of studies into ; regulatory organisations ; self-regulation \u2013; _see also_ COPE; research integrity; technologies of integrity verification\n\nmock crimes\n\nMol, A.: ontology\n\nMontgomery, K.: institutional logics\n\nMoreira, Tiago\n\nMorris, Andrew\n\nMosso, Angelo ,\n\nMouffe, Chantal\n\nMundy, Martha: distinctions between persons and things\n\nMuniesa, F.\n\nMunsterberg, Hugo\n\nMurphy, Michelle , 192n8\n\nNational Cancer Institute 194n12\n\nNational Defense Authorization Act\n\nNational Donation Strategy Group\n\nNational Gamete Donation Trust (NGDT)\n\nNational Jewish Hospital\n\nNational Reform Programmes 43n17\n\nNational Research Council (NRC): report \u2013\n\nNational Science Foundation\n\nNational Security Agency\n\nNational Sustainable Development Strategies\n\nNational Toxins Campaign Fund \u2013\n\nnation-states: imagery of ; imaginaries\n\nneo-classical economists: price system \u2013\n\nNew Mexico\n\nNik-Khah, E.\n\nnormative principle of transparency: plurality of knowledge\n\nNorvell, M. K.: genital microtrauma caused by consensual sexual intercourse\n\nnucleotides\n\nNutt, David\n\nOccupational Safety and Health Administration 194n12\n\noffenders: definition 73n8\n\nOffice of Science and Technology Policy\n\nOliver, A. L.: institutional logics\n\nonline forums: control by moderators ; Donor Gamete forum , \u2013; fertility support groups ; Repromed ; virtual patient communities ; would-be parents \u2013,\n\nontological gerrymandering 192n8\n\nontology: connections between emotion, the body and lying ; definition\n\noptions (financial derivatives)\n\nOstrom, Elinor\n\nover-the counter (OTC) derivatives\n\nPacheco, Dr. Karen A. ,\n\nP\u00e1ez, Pedro ,\n\npaper: retrospective functioning ; theory of\n\npaper prototyping: emergence of _dispositif_ ; inventions formulated for artisans or manufacturers ; lawyers' interpretation of invention as prototype ; recollection of inventions\n\nParry, Bronwyn\n\nPatel, H. C.: critique of colposcopy ; rape without signs of injury ; sympathy to 'it's normal to be normal' publication\n\npatents: copies indefinitely multiplied ; doctrine ; genetic template ; invention ; justification of ; legal rights \u2013; life sciences ; machining ; system ; understanding of ; _see also_ inventions; paper prototyping; prototypes\n\npathophysiologies\n\nPattberg, P.\n\nPatton, Cindy\n\npension funds\n\n_The People of California v. San Bernardino County_ \u2013\n\npeople with AIDS (PWAs) _see_ PWAs\n\nPercival, Thomas\n\nperformativity pieces: _agencement_\n\nperformativity thesis: auctions ; Black-Scholes-Merton formula ; emergence of ; exceeding the sphere of human beliefs ; incorporation into financial markets ; relation to financial derivatives\n\npersonal data: anonymity \u2013; arrested persons , ; collapsing the regulatory divide with human tissue \u2013; consent ; consent for use in research \u2013; crossing the regulatory divide with human tissue \u2013; European data protection regime \u2013; individual autonomy and control over data \u2013; privacy concerns ; processing of ; regulatory divide with human tissue \u2013; regulatory systems \u2013; storage on DNA databases\n\npesticides\n\nphantomtic ontology\n\npharmaceutical industry \u2013\n\nphysiology: standardisation of measurements\n\nPickersgill, M. D.: reshaping social and legal ontologies ; scientific knowledge\n\nPielke, R. A.: boundary between factual claims and value choices\n\nplagiarism\n\n_PLoS Medicine_\n\nPolice Doctor _see_ FMEs (Forensic Medical Examiners); FNEs (Forensic Nurse Examiners)\n\npolicies-without-publics \u2013\n\npolicy domains: without publics \u2013\n\npolicymakers: democratic politics\n\npolicymaking: agenda set by science ; expert judgements ; individual studies ; knowledge claims ; problems identified by science \u2013; public involvement in ; science-based ; _see also_ populism\n\npolitical centralisation: coproduction ; dependence on regulatory science ; epistemic jurisdiction\n\npolitical decentralisation: coproduction ; dependence on regulatory science ; epistemic jurisdiction\n\npolitical economy _see_ feminist political economy\n\npolitics: decisionism in ; democratic ; pitted against science ; regulation of environmental hazards in USA ; separate roles for experts and\n\npolygraphs: admissibility \u2013; admissibility practices open to reinterpretation ; admissible evidence for criminal trials ; alternative to police methods of extracting confessions \u2013; American obsession with ; deterrent to espionage ; distrust in American workplaces, criminal procedure and governance \u2013; early stage of development ; exclusionary at Federal level ; expertise of examiners ; false positives , ; _Fryev. United States_ \u2013; government security policies against communism \u2013; histories ; innovation ; introduction to department stores and banks \u2013; legal status in the USA \u2013; ontological connections ; popularising ; prior-stipulation mechanism at state level ; scholarly literature on ; sex offenders ; struggle to gain scientific and legal legitimacy ; Supreme Court accepts as expert evidence \u2013; surveillance \u2013; surveillance technique in industry ; technical developments leading to increased reliability ; testing of employees in DOE ; training courses ; usage ; use during McCarthyite period \u2013; validity challenged by NRC report ; _see also_ Keeler, Leonarde\n\npopulism: concept of ; policies-without-publics \u2013\n\nPottage, Alain: common patrimony of humanity ; distinctions between persons and things ; functioning of law \u2013; genetic patrimony ; legal temporalities\n\nPresidential Commission\n\nPresidential Decision Directive\n\nPreston, Patrick \"Rick\" Edward ; FEMA's formaldehyde testing issue ; stipulated ATSDR results remain confidential\n\nprice index\n\nprice volatility: financial derivatives\n\nprint _see_ paper\n\nprinting press\n\nprior-stipulation rule ,\n\nproduction: performativity\n\nproductive labour\n\nprognosis: co-articulation ; determinant of likelihood ; double effect of ; function across legal and clinical fields ; legal claims ; likelihood \u2013; medical experts' reluctance to predict \u2013; patient's test results ; scale and individuation ; schematisation ; T-cell count , \u2013; uncertainty ; viral loads , \u2013\n\nprotection of endangered species\n\nprototypes: invention ; machines _see_ prototyping machines; machining ; paper prototyping _see_ paper prototyping; process of manufacturing ; vital agency in manufacturing process ; _see also_ inventions; patents\n\nprototyping machines: instrument abstracted from their industrial courses ; legal schema of invention ; machine as mechanical ingenuity ; transitive type embodied ,\n\nPublic Bodies Act 2011\n\npublic health campaigns: asymptomatic HIV carriers \u2013\n\npublic participation _see_ citizen participation\n\npublics: accountability of experts ; definition 21n9; engaged in policymaking ; _innovator citizens_ ; liberal democracy and \u2013; limited participation in EU data analysis ; public engagement around science-based policymaking \u2013; _see also_ citizen participation\n\nPullin, Scott 190n1\n\nPurchasing Power Parity\n\nPWAs (people with AIDS): financial support ; hostility at work ; temporary benefits and housing arrangements \u2013\n\nquackery: BMA's work against ; campaigners against ; characterisation of ; dilemmas for medical reformers ; petition to Parliament ; professionalisation of medicine \u2013\n\nquangos: abolition of ; HFEA _see_ HFEA (Human Fertilisation and Embryology Authority); transferring the functions of ; unelected\n\nracial categorisation\n\nRafferty, M.: value\n\nrandom match probability 77n10,\n\nrape: absence of signs of injury ; criminal prosecutions ; evidence provided by colposcopes ; forensic intervention ; invasive genital examination of colposcope ; 'it's normal to be normal' publication ; rape injury \u2013; Rape Trauma Syndrome ; recording and documenting injuries ; survivors' consent for recording of colposcopic images ; terminology of victims 88n2; _see also_ colposcopes; colposcopy controversy; sexual assault\n\nRape Trauma Syndrome\n\nregimes of imperceptibility\n\nRegional Monetary Agreement (RMA) ; exchange rates ; interaction of governments and markets\n\nregulation: financial derivatives\n\nregulatory capture\n\nregulatory harmonisation\n\nregulatory landscape: apt metaphor for lawyers\n\nregulatory science ; coproductionist framework ; definition ; focal point of conflict ; global and international levels ; institutions\n\nRenda, Frank: concerns about ATSDR report ; removal from post\n\nRenewable Energy Directive\n\nreproductive care: cross-border \u2013\n\nreproductive medicine\n\nreproductive travel\n\nRepromed ; advice on backstreet abortions ; Egg and Sperm Donation forum , ; forum statistics ; posts on ; reproductive medicine\n\nRepublic of Ecuador\n\nRepublic of Ireland: abortion\n\nresearch integrity: high-tech integrity verification tools ; regulatory agenda ; regulatory capture ; regulatory technological fixes ; technological verification ; technologies of integrity verification \u2013; _see also_ misconduct hunting\n\nResource-efficient Europe\n\nrespiratory illnesses\n\nRevilla, J.: 'it's normal to be normal' publication ; review of child and adolescent sexual assault victims\n\nRiles, Annelise\n\nRva-Rocci, Scipione\n\nRoberts, Justice John\n\nRobinson, William\n\nRooke, Catriona: legal things\n\nRound-up Ready soy\n\nRoyal Society 17n1, , ; conflicting testimonies\n\nRules Enabling Act\n\nSaint-Simon, Henri 20n6\n\nSan Bernardino County \u2013\n\n_S and Marper v. United Kingdom_ (2008) \u2013,\n\nSanitary and Phytosanitary (SPS) Agreement , , ,\n\nSarewitz, D. ,\n\nSchaffer, S.: connection between experimental knowledge and legal knowledge \u2013; scientific experimentation\n\nSchrader, Astrid\n\nSchwarzenegger, Governor Arnold\n\nscience: decisionism \u2013; definition of ; deviances in ; experimentation ; good research practice ; indeterminancy ; institutional logics ; laboratories ; management of uncertainty \u2013; politicisation of ; professionalisation of medicine \u2013; publishing ; quackery \u2013; regulation of research ; regulatory , ; research integrity _see_ research integrity; review of current regulatory frameworks ; role in making of law and public policy \u2013; scientific publication ; sound \u2013; value-free ; virtual global networks ; _see also_ climate change; climate federalism; COPE; Royal Society; science-in-policymaking\n\nscience and technology studies (STS) _see_ STS\n\nscience-based policymaking _see_ epistemic governance\n\nscience-in-policymaking ; climate science and policy ; environmental hazards \u2013; excess of objectivity \u2013; impact of high political stakes \u2013; limiting the role of scientism ; politicisation of ; probabilistic judgements ; STS insights on\n\nscientific governance _see_ epistemic governance\n\nscientific indeterminacy\n\nScientific Revolution\n\nscientism\n\nSecretariat of International Plant Protection Convention 177n2\n\nself-plagiarism , \u2013\n\nSenate Energy and National Resources Committee\n\nsexual assault: absence of signs of injury ; evidence provided by colposcopes ; invasive genital examination of colposcope ; 'it's normal to be normal' publication ; review of child and adolescent victims ; survivors' consent for recording of colposcopic images ; terminology of victims 88n2; _see also_ colposcopes; colposcopy controversy; rape\n\nShapin, S.: connection between experimental knowledge and legal knowledge \u2013; scientific experimentation\n\nShapiro, B. J.\n\nshared technical imaginary\n\nShea, Dan 190n1\n\nshort tandem repeats (STRs) \u2013, 77n11\n\nsick building syndrome\n\nSierra Club ; press release\n\nSilbey, S.: authenticating site of science\n\nSiu, L.\n\nSlaughter, Laura: application of the colposcope ; evidence and image recordings ; study of injury caused by consensual and nonconsensual sexual intercourse \u2013; study on colposcopes \u2013; Toludine Blue \u2013\n\nSmith, Dr Kenneth: asthma medication \u2013; disposition for FEMA trailer formaldehyde litigation \u2013; translation of affidavit\n\nSmith, Richard\n\nSocial Studies of Finance (SSF) _see_ SSF\n\nsocial studies of law \u2013\n\nsocio-legal scholars: multiple forms of disputing\n\nSouza, Kevin 193n10\n\nSovacool, B. K.\n\nspeculation: capital ; distinct from hedging ; role of financial derivatives\n\nsperm clinics ,\n\nsperm donors ,\n\nsperm (gametes) _see_ gametes\n\nSSF (Social Studies of Finance): contribution ; emergence of ; explaining economic activities ; second order economy\n\nState of Massachusetts\n\n_State of New Mexico v. Shane Glen Harrison_\n\nstatistics: Eurostat data collection ; medical journals ; secrecy of methods\n\nStehr, N.\n\nStern Review 87n1\n\nStevens, Justice John Paul \u2013,\n\nStewart, Paul\n\nstock prices\n\nStrange, S.\n\nstratification\n\nStringham, Emerson \u2013; discursive artefact\n\nSTRs (short tandem repeats) \u2013, 77n11\n\nSTS (science and technology studies): causal relationship between products and adverse health outcomes ; challenging categories and concepts in law _;_ challenging the unproblematic and obvious \u2013; controversies ; coproduction ; differences between risk and uncertainty ; distinction between perpetrator and victim ; DNA wars ; ethical, legal, and social implications (ELSI) ; exchanging methods and information with sociolegal scholars \u2013; exclusion of convicts from scholarship \u2013; false eyewitness evidence ; focus on usual suspects ; forensic science ; importance of images in scientific and medical work 90n5; instances of _situated, dis\/empowerment_ ; multiple forms of disputing ; regulatory science ; rendering explicit the distribution of power and political agency ; salience of imaginaries ; scholarship ; science-based policy-making ; science-in-policy \u2013; scientific knowledge ; social studies of \u2013; studies of law and science ; technological instrumentalism , \u2013; unexamined assumptions \u2013, \u2013\n\nSubcommittee on Investigations and Oversight\n\nsubjective (Bayesian) assessment\n\nsubmissions, statistical analysis of\n\nsupply and demand\n\nsurrogacy \u2013\n\nsurveillance: department stores and banks \u2013; deterrent to criminal activity ; DOE (Department of Energy) employees , ; employee confessions ; employee screening ; polygraph testing \u2013\n\nsurveys: HFEA (Human Fertilisation and Embryology Authority) consultations ; overseas fertility treatment \u2013\n\nsustainable development: Article 3(3) TEU ; Article 11 TFEU ; Brundtland Report ; citizen participation ; cross-cutting ; data collection ; different formulations of ; EU (European Union) \u2013; limitations ; limits of participation in the EU governance \u2013; long-term objective of the EU ; needs of the world's poor ; policy documents ; prioritisation of growth ; progress reports ; technoscientific innovation , \u2013\n\nsustainable growth: EU commitment ; Resource-efficient Europe\n\nsystolic blood pressure deception test ,\n\nTales from the Future\n\nT-cell count , \u2013\n\ntechnocracy: assumption of elite privilege ; definition of , ; normative case against ; sound science\n\ntechnologies of integrity verification \u2013; 'computerised the detective's role' ; statistics ; surveillance in laboratories ; textual similarities software\n\ntechnologies of participation: active\/passive participation ; citizen regulation ; _ex ante_ , ; _ex post_ ; techniques and practices \u2013\n\ntechnoscience: governance of ; risks and ethics about ; sustainable development ; sustainable growth \u2013; tool of regulation\n\nTeixeira, W. R.: colposcopic study\n\ntemporality: definition ; legal \u2013; material worlds ; natural or social ; systems theory ; _see also_ HIV\n\ntextual similarity algorithms ,\n\nthickets: collapsing the regulatory divide between personal data and human tissue \u2013; conflation of private and public interest , \u2013; consent and protection of individual concerns \u2013; continuum of personal data and human tissue ; crossing the regulatory divide between human tissue and personal data \u2013; definition ; European data protection regime \u2013; evolution of ; focus on objects of regulatory concern ; growth of ; Guthrie cards , \u2013; human tissue ; individual autonomy and control over data \u2013; personal data ; reflexive system of governance \u2013; regulatory burden ; root cause of many problems ; sector-specific legal regimes , ; upstream considerations \u2013; widening regulatory divide in personal data and human tissue \u2013; _see also_ life sciences regulation\n\nthird degree policing , ,\n\nThomas, Ronald: polygraphy and detective fiction\n\nThompson, G. I.: genital microtrauma caused by consensual sexual intercourse\n\nThrift, Nigel 191n7\n\ntime: segmented understandings of ; social horizons of \u2013; _see also_ HIV; temporality\n\nToludine Blue \u2013\n\ntrailers _see_ FEMA trailers\n\ntrans-science\n\ntravel parks _see_ FEMA travel parks\n\nTreaty on the Functioning of the European Union (TFEU)\n\nUK Research Integrity Office\n\nuncertainty: clinical ; governance of futures \u2013; management of \u2013; production and perpetuation of\n\nunderlying assets ,\n\nunexamined assumptions: challenging the unproblematic and obvious ; miscarriage of justice \u2013; technological instrumentalism \u2013\n\nUnited States of America (USA): admissibility of polygraphs \u2013; challenge to EU Directive ; climate federalism \u2013; security \u2013; third degree policing , , ; use of polygraphs during McCarthyite period \u2013; _see also_ DOE (Department of Energy)\n\nun-knowing: ATSDRFEMA trailer health consultation \u2013; Christopher Cooper asthma case \u2013, ; definition ; formaldehyde testing campaign exposure ; pharmaceutical logics \u2013; secrecy of ATSDR test results\n\nvalue: construction of ; contestability ; financial derivatives , \u2013; fundamental , ; law of ; measurement of , ; real and financial ; social ; value-making processes\n\nv-CJD (variant Creutzfeldt-Jakob disease)\n\nviral loads , \u2013\n\nvirtual global networks: scientists and experts\n\nvirtual patient communities\n\nin vitro fertilisation (IVF)\n\nVollmer, August: removal of third degree policing\n\nwages for housework initiative\n\nWakley, Thomas\n\nWaldby, Catherine ; co-articulation ; shared technical imaginary\n\nWeber, Max\n\nWeber, S.: mock crimes ; scientific work on fear\n\nWeinberg, Alvin\n\nWeinel, M.: separation of roles for experts and politics\n\nWellcome Trust Scottish Health Informatics Programme 125n1\n\nWhitmarsh, I.: pharmaceutical intervention\n\nWilliams, Dawson\n\nWoolgar, S.: ontology and epistemology\n\nWorld Association of Medical Editors (WAME)\n\nWorld Bank: environmental impact assessment\n\nWorld Trade Organization (WTO) _see_ WTO (World Trade Organization)\n\nwould-be parents: challenges faced by ; difficulties accessing required treatment ; managing own treatment ; obtaining sperm, eggs or embryos from within the EU ; online forums \u2013; opposition to HFE Act ; overseas treatment , \u2013; Repromed ; survey on overseas fertility treatment \u2013; virtual patient communities\n\nWTO Dispute Settlement Body\n\nWTO (World Trade Organization): adjudicating GMOs \u2013; challenge to EU Directive ; food safety agreements ; review of risk assessment processes ; risk assessment protocols\n\nWynne, B.: science-based arguments for policy\n\nYeung, K. \n","meta":{"redpajama_set_name":"RedPajamaBook"}}
+{"text":" \n# Practical Deep Learning for Cloud, Mobile, and Edge\n\nReal-World AI and Computer-Vision Projects Using Python, Keras, and TensorFlow\n\nAnirudh Koul, Siddha Ganju, and Meher Kasam\n\n# Practical Deep Learning for Cloud, Mobile, and Edge\n\nby Anirudh  Koul, Siddha  Ganju, and Meher  Kasam\n\nCopyright \u00a9 2020 Anirudh Koul, Siddha Ganju, Meher Kasam. All rights reserved.\n\nPrinted in the United States of America.\n\nPublished by O'Reilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA 95472.\n\nO'Reilly books may be purchased for educational, business, or sales promotional use. Online editions are also available for most titles ( _http:\/\/oreilly.com_). For more information, contact our corporate\/institutional sales department: 800-998-9938 or _corporate@oreilly.com_.\n\n  * Acquisitions Editor: Rachel Roumeliotis \n  * Development Editor: Nicole Tache \n  * Production Editor: Christopher Faucher \n  * Copyeditor: Octal Publishing, LLC \n  * Proofreader: Christina Edwards \n  * Indexer: Judith McConville \n  * Interior Designer: David Futato \n  * Cover Designer: Karen Montgomery \n  * Illustrator: Rebecca Demarest\n\n  * October 2019: First Edition\n\n# Revision History for the First Edition\n\n  * 2019-10-14: First Release\n\nSee _http:\/\/oreilly.com\/catalog\/errata.csp?isbn=9781492034865_ for release details.\n\nThe O'Reilly logo is a registered trademark of O'Reilly Media, Inc. _Practical Deep Learning for Cloud, Mobile, and Edge_ , the cover image, and related trade dress are trademarks of O'Reilly Media, Inc.\n\nThe views expressed in this work are those of the authors, and do not represent the publisher's views. While the publisher and the authors have used good faith efforts to ensure that the information and instructions contained in this work are accurate, the publisher and the authors disclaim all responsibility for errors or omissions, including without limitation responsibility for damages resulting from the use of or reliance on this work. Use of the information and instructions contained in this work is at your own risk. If any code samples or other technology this work contains or describes is subject to open source licenses or the intellectual property rights of others, it is your responsibility to ensure that your use thereof complies with such licenses and\/or rights.\n\n978-1-492-03486-5\n\n[LSI]\n\n# Preface\n\nWe are experiencing a renaissance of artificial intelligence, and everyone and their neighbor wants to be a part of this movement. That's quite likely why you are browsing through this book. There are tons of books about deep learning out there. So you might ask us, very reasonably, why does this book even exist? We'll get to that in just a second.\n\nDuring our own deep learning journeys since 2013 (while building products at companies including Microsoft, NVIDIA, Amazon, and Square), we witnessed dramatic shifts in this landscape. Constantly evolving research was a given and a lack of mature tooling was a reality of life.\n\nWhile growing and learning from the community, we noticed a lack of clear guidance on how to convert research to an end product for everyday users. After all, the end user is somewhere in front of a web browser, a smartphone, or an edge device. This often involved countless hours of hacking and experimentation, extensively searching through blogs, GitHub issue threads, and Stack Overflow answers, and emailing authors of packages to get esoteric knowledge, as well as occasional \"Aha!\" moments. Even the books on the market tended to focus more on theory or how to use a specific tool. The best we could hope to learn from the available books was to build a toy example.\n\nTo fill this gap between theory and practice, we originally started giving talks on taking artificial intelligence from research to the end user with a particular focus on practical applications. The talks were structured to showcase motivating examples, as well as different levels of complexity based on skill level (from a hobbyist to a Google-scale engineer) and effort involved in deploying deep learning in production. We discovered that beginners and experts alike found value in these talks.\n\nOver time, the landscape thankfully became accessible to beginners and more tooling became available. Great online material like Fast.ai and DeepLearning.ai made understanding how to train AI models easier than ever. Books also cornered the market on teaching fundamentals using deep learning frameworks such as TensorFlow and PyTorch. But even with all of this, the wide chasm between theory and production remained largely unaddressed. And we wanted to bridge this gap. Thus, the book you are now reading.\n\nUsing approachable language as well as ready-to-run fun projects in computer vision, the book starts off with simple classifiers assuming no knowledge of machine learning and AI, gradually building to add complexity, improve accuracy and speed, scale to millions of users, deploy on a wide variety of hardware and software, and eventually culminate in using reinforcement learning to build a miniature self-driving car.\n\nNearly every chapter begins with a motivating example, establishes the questions upfront that one might ask through the process of building a solution, and discusses multiple approaches for solving problems, each with varying levels of complexity and effort involved. If you are seeking a quick solution, you might end up just reading a few pages of a chapter and be done. Someone wanting to gain a deeper understanding of the subject should read the entire chapter. Of course, everyone should peruse the case studies included in these chapters for two reasons\u2014they are fun to read and they showcase how people in the industry are using the concepts discussed in the chapter to build real products.\n\nWe also discuss many of the practical concerns faced by deep learning practitioners and industry professionals in building real-world applications using the cloud, browsers, mobile, and edge devices. We compiled a number of practical \"tips and tricks,\" as well as life lessons in this book to encourage our readers to build applications that can make someone's day just a little bit better.\n\n# To the Backend\/Frontend\/Mobile Software Developer\n\nYou are quite likely a proficient programmer already. Even if Python is an unfamiliar language to you, we expect that you will be able to pick it up easily and get started in no time. Best of all, we don't expect you to have any background in machine learning and AI; that's what we are here for! We believe that you will gain value from the book's focus on the following areas:\n\n  * How to build user-facing AI products\n\n  * How to train models quickly\n\n  * How to minimize the code and effort required in prototyping\n\n  * How to make models more performant and energy efficient\n\n  * How to operationalize and scale, and estimate the costs involved\n\n  * Discovering how AI is applied in the industry with 40+ case studies and real-world examples\n\n  * Developing a broad-spectrum knowledge of deep learning\n\n  * Developing a generalized skill set that can be applied on new frameworks (e.g., PyTorch), domains (e.g., healthcare, robotics), input modalities (e.g., video, audio, text), and tasks (e.g., image segmentation, one-shot learning)\n\n# To the Data Scientist\n\nYou might already be proficient at machine learning and potentially know how to train deep learning models. Good news! You can further enrich your skill set and deepen your knowledge in the field in order to build real products. This book will help inform your everyday work and beyond by covering how to:\n\n  * Speed up your training, including on multinode clusters\n\n  * Build an intuition for developing and debugging models, including hyperparameter tuning, thus dramatically improving model accuracy\n\n  * Understand how your model works, uncover bias in the data, and automatically determine the best hyperparameters as well as model architecture using AutoML\n\n  * Learn tips and tricks used by other data scientists, including gathering data quickly, tracking your experiments in an organized manner, sharing your models with the world, and being up to date on the best available models for your task\n\n  * Use tools to deploy and scale your best model to real users, even automatically (without involving a DevOps team)\n\n# To the Student\n\nThis is a great time to be considering a career in AI\u2014it's turning out to be the next revolution in technology after the internet and smartphones. A lot of strides have been made, and a lot remains to be discovered. We hope that this book can serve as your first step in whetting your appetite for a career in AI and, even better, developing deeper theoretical knowledge. And the best part is that you don't have to spend a lot of money to buy expensive hardware. In fact, you can train on powerful hardware entirely for free from your web browser (thank you, Google Colab!). With this book, we hope you will:\n\n  * Aspire to a career in AI by developing a portfolio of interesting projects\n\n  * Learn from industry practices to help prepare for internships and job opportunities\n\n  * Unleash your creativity by building fun applications like an autonomous car\n\n  * Become an AI for Good champion by using your creativity to solve the most pressing problems faced by humanity\n\n# To the Teacher\n\nWe believe that this book can nicely supplement your coursework with fun, real-world projects. We've covered every step of the deep learning pipeline in detail, along with techniques on how to execute each step effectively and efficiently. Each of the projects we present in the book can make for great collaborative or individual work in the classroom throughout the semester. Eventually, we will be releasing PowerPoint Presentation Slides on _http:\/\/PracticalDeepLearning.ai_ that can accompany coursework.\n\n# To the Robotics Enthusiast\n\nRobotics is exciting. If you're a robotics enthusiast, we don't really need to convince you that adding intelligence to robots is the way to go. Increasingly capable hardware platforms such as Raspberry Pi, NVIDIA Jetson Nano, Google Coral, Intel Movidius, PYNQ-Z2, and others are helping drive innovation in the robotics space. As we grow towards Industry 4.0, some of these platforms will become more and more relevant and ubiquitous. With this book, you will:\n\n  * Learn how to build and train AI, and then bring it to the edge\n\n  * Benchmark and compare edge devices on performance, size, power, battery, and costs\n\n  * Understand how to choose the optimal AI algorithm and device for a given scenario\n\n  * Learn how other makers are building creative robots and machines\n\n  * Learn how to build further progress in the field and showcase your work\n\n# What to Expect in Each Chapter\n\nChapter 1, _Exploring the Landscape of Artificial Intelligence_\n\nWe take a tour of this evolving landscape, from the 1950s to today, analyze the ingredients that make for a perfect deep learning recipe, get familiar with common AI terminology and datasets, and take a peek into the world of responsible AI.\n\nChapter 2, _What's in the Picture: Image Classification with Keras_\n\nWe delve into the world of image classification in a mere five lines of Keras code. We then learn what neural networks are paying attention to while making predictions by overlaying heatmaps on videos. Bonus: we hear the motivating personal journey of Fran\u00e7ois Chollet, the creator of Keras, illustrating the impact a single individual can have.\n\nChapter 3, _Cats Versus Dogs: Transfer Learning in 30 Lines with Keras_\n\nWe use transfer learning to reuse a previously trained network on a new custom classification task to get near state-of-the-art accuracy in a matter of minutes. We then slice and dice the results to understand how well it is classifying. Along the way, we build a common machine learning pipeline, which is repurposed throughout the book. Bonus: we hear from Jeremy Howard, cofounder of fast.ai, on how hundreds of thousands of students use transfer learning to jumpstart their AI journey.\n\nChapter 4, _Building a Reverse Image Search Engine: Understanding Embeddings_\n\nLike Google Reverse Image Search, we explore how one can use embeddings\u2014a contextual representation of an image to find similar images in under ten lines. And then the fun starts when we explore different strategies and algorithms to speed this up at scale, from thousands to several million images, and making them searchable in microseconds.\n\nChapter 5, _From Novice to Master Predictor: Maximizing Convolutional Neural Network Accuracy_\n\nWe explore strategies to maximize the accuracy that our classifier can achieve, with the help of a range of tools including TensorBoard, the What-If Tool, tf-explain, TensorFlow Datasets, AutoKeras, and AutoAugment. Along the way, we conduct experiments to develop an intuition of what parameters might or might not work for your AI task.\n\nChapter 6, _Maximizing Speed and Performance of TensorFlow: A Handy Checklist_\n\nWe take the speed of training and inference into hyperdrive by going through a checklist of 30 tricks to reduce as many inefficiencies as possible and maximize the value of your current hardware.\n\nChapter 7, _Practical Tools, Tips, and Tricks_\n\nWe diversify our practical skills in a variety of topics and tools, ranging from installation, data collection, experiment management, visualizations, and keeping track of state-of-the-art research all the way to exploring further avenues for building the theoretical foundations of deep learning.\n\nChapter 8, _Cloud APIs for Computer Vision: Up and Running in 15 Minutes_\n\nWork smart, not hard. We utilize the power of cloud AI platforms from Google, Microsoft, Amazon, IBM, and Clarifai in under 15 minutes. For tasks not solved with existing APIs, we then use custom classification services to train classifiers without coding. And then we pit them against each other in an open benchmark\u2014you might be surprised who won.\n\nChapter 9, _Scalable Inference Serving on Cloud with TensorFlow Serving and KubeFlow_\n\nWe take our custom trained model to the cloud\/on-premises to scalably serve from hundreds to millions of requests. We explore Flask, Google Cloud ML Engine, TensorFlow Serving, and KubeFlow, showcasing the effort, scenario, and cost-benefit analysis.\n\nChapter 10, _AI in the Browser with TensorFlow.js and ml5.js_\n\nEvery single individual who uses a computer or a smartphone uniformly has access to one software program\u2014their browser. Reach all those users with browser-based deep learning libraries including TensorFlow.js and ml5.js. Guest author Zaid Alyafeai walks us through techniques and tasks such as body pose estimation, generative adversarial networks (GANs), image-to-image translation with Pix2Pix, and more, running not on a server but in the browser itself. Bonus: hear from key contributors to TensorFlow.js and ml5.js on how the projects incubated.\n\nChapter 11, _Real-Time Object Classification on iOS with Core ML_\n\nWe explore the landscape of deep learning on mobile, with a sharp focus on the Apple ecosystem with Core ML. We benchmark models on different iPhones, investigate strategies to reduce app size and energy impact, and look into dynamic model deployment, training on device, and how professional apps are built.\n\nChapter 12, _Not Hotdog on iOS with Core ML and Create ML_\n\n_Silicon Valley_ 's Not Hotdog app (from HBO) is considered the \"Hello World\" of mobile AI, so we pay tribute by building a real-time version in not one, not two, but three different ways.\n\nChapter 13, _Shazam for Food: Developing Android Apps with TensorFlow Lite and ML Kit_\n\nWe bring AI to Android with the help of TensorFlow Lite. We then look at cross-platform development using ML Kit (which is built on top of TensorFlow Lite) and Fritz to explore the end-to-end development life cycle for building a self-improving AI app. Along the way we look at model versioning, A\/B testing, measuring success, dynamic updates, model optimization, and other topics. Bonus: we get to hear about the rich experience of Pete Warden (technical lead for Mobile and Embedded TensorFlow) in bringing AI to edge devices.\n\nChapter 14, _Building the Purrfect Cat Locator App with TensorFlow Object Detection API_\n\nWe explore four different methods for locating the position of objects within images. We take a look at the evolution of object detection over the years, and analyze the tradeoffs between speed and accuracy. This builds the base for case studies such as crowd counting, face detection, and autonomous cars.\n\nChapter 15, _Becoming a Maker: Exploring Embedded AI at the Edge_\n\nGuest author Sam Sterckval brings deep learning to low-power devices as he showcases a range of AI-capable edge devices with varying processing power and cost including Raspberry Pi, NVIDIA Jetson Nano, Google Coral, Intel Movidius, and PYNQ-Z2 FPGA, opening the doors for robotics and maker projects. Bonus: hear from the NVIDIA Jetson Nano team on how people are building creative robots quickly from their open source recipe book.\n\nChapter 16, _Simulating a Self-Driving Car Using End-to-End Deep Learning with Keras_\n\nUsing the photorealistic simulation environment of Microsoft AirSim, guest authors Aditya Sharma and Mitchell Spryn guide us in training a virtual car by driving it first within the environment and then teaching an AI model to replicate its behavior. Along the way, this chapter covers a number of concepts that are applicable in the autonomous car industry.\n\nChapter 17, _Building an Autonomous Car in Under an Hour: Reinforcement Learning with AWS DeepRacer_\n\nMoving from the virtual to the physical world, guest author Sunil Mallya showcases how AWS DeepRacer, a miniature car, can be assembled, trained, and raced in under an hour. And with the help of reinforcement learning, the car learns to drive on its own, penalizing its mistakes and maximizing success. We learn how to apply this knowledge to races from the Olympics of AI Driving to RoboRace (using full-sized autonomous cars). Bonus: hear from Anima Anandkumar (NVIDIA) and Chris Anderson (founder of DIY Robocars) on where the self-driving automotive industry is headed.\n\n# Conventions Used in This Book\n\nThe following typographical conventions are used in this book:\n\n_Italic_\n\nIndicates new terms, URLs, email addresses, filenames, and file extensions.\n\n`Constant width`\n\nUsed for program listings, as well as within paragraphs to refer to program elements such as variable or function names, databases, data types, environment variables, statements, and keywords.\n\n**`Constant width bold`**\n\nShows commands or other text that should be typed literally by the user.\n\n_`Constant width italic`_\n\nShows text that should be replaced with user-supplied values or by values determined by context.\n\n###### Tip\n\nThis element signifies a tip or suggestion.\n\n###### Note\n\nThis element signifies a general note.\n\n###### Warning\n\nThis element indicates a warning or caution.\n\n# Using Code Examples\n\nSupplemental material (code examples, exercises, etc.) is available for download at _http:\/\/PracticalDeepLearning.ai_. If you have a technical question or a problem using the code examples, please send email to _PracticalDLBook@gmail.com_.\n\nThis book is here to help you get your job done. In general, if example code is offered with this book, you may use it in your programs and documentation. You do not need to contact us for permission unless you're reproducing a significant portion of the code. For example, writing a program that uses several chunks of code from this book does not require permission. Selling or distributing examples from O'Reilly books does require permission. Answering a question by citing this book and quoting example code does not require permission. Incorporating a significant amount of example code from this book into your product's documentation does require permission.\n\nWe appreciate, but generally do not require, attribution. An attribution usually includes the title, author, publisher, and ISBN. For example: \" _Practical Deep Learning for Cloud, Mobile, and Edge_ by Anirudh Koul, Siddha Ganju, and Meher Kasam (O'Reilly). Copyright 2020 Anirudh Koul, Siddha Ganju, Meher Kasam, 978-1-492-03486-5.\"\n\nIf you feel your use of code examples falls outside fair use or the permission given above, feel free to contact us at _permissions@oreilly.com_.\n\n# O'Reilly Online Learning\n\n###### Note\n\nFor more than 40 years, _O'Reilly Media_ has provided technology and business training, knowledge, and insight to help companies succeed.\n\nOur unique network of experts and innovators share their knowledge and expertise through books, articles, conferences, and our online learning platform. O'Reilly's online learning platform gives you on-demand access to live training courses, in-depth learning paths, interactive coding environments, and a vast collection of text and video from O'Reilly and 200+ other publishers. For more information, please visit _http:\/\/oreilly.com_.\n\n# How to Contact Us\n\nPlease address comments and questions concerning this book to the publisher:\n\n  * O'Reilly Media, Inc.\n  * 1005 Gravenstein Highway North\n  * Sebastopol, CA 95472\n  * 800-998-9938 (in the United States or Canada)\n  * 707-829-0515 (international or local)\n  * 707-829-0104 (fax)\n\nO'Reilly has a web page for this book, where we list errata, examples, and any additional information. You can access this page at _https:\/\/oreil.ly\/practical-deep-learning_. The authors have a website for this book as well: _http:\/\/PracticalDeepLearning.ai_.\n\nEmail _bookquestions@oreilly.com_ to comment or ask technical questions about this book; email _PracticalDLBook@gmail.com_ to contact the authors about this book.\n\nFor more information about our books, courses, conferences, and news, see our website at _http:\/\/www.oreilly.com_.\n\nFind us on Facebook: _http:\/\/facebook.com\/oreilly_\n\nFollow us on Twitter: _http:\/\/twitter.com\/oreillymedia_\n\nWatch us on YouTube: _http:\/\/www.youtube.com\/oreillymedia_\n\n# Acknowledgments\n\n## Group Acknowledgments\n\nWe'd like to thank the following people for their immense help throughout our journey in writing this book. Without them, this book would not be possible.\n\nThis book came to life because of our development editor Nicole Tach\u00e9's efforts. She rooted for us throughout our journey and provided important guidance at each step of the process. She helped us prioritize the right material (believe it or not, the book was going to be even larger!) and ensured that we were on track. She was reader number one for every single draft that we had written, so our goal first and foremost was ensuring that she was able to follow the content, despite her being new to AI. We're immensely grateful for her support.\n\nWe also want to thank the rest of the O'Reilly team including our production editor Christopher Faucher who worked tireless hours on a tight schedule to ensure that this book made it to the printing press on time. We are also grateful to our copy editor Bob Russell who really impressed us with his lightning-fast edits and his attention to detail. He made us realize the importance of paying attention to English grammar in school (though a few years too late, we're afraid). We also want to acknowledge Rachel Roumeliotis (VP of Content Strategy) and Olivia MacDonald (Managing Editor for Development) for believing in the project and for offering their continued support.\n\nHuge thanks are in order for our guest authors who brought in their technical expertise to share their passion for this field with our readers. Aditya Sharma and Mitchell Spryn (from Microsoft) showed us that our love for playing video racing games can be put to good use to train autonomous cars by driving them in a simulated environment (with AirSim). Sunil Mallya (from Amazon) helped bring this knowledge to the physical world by demonstrating that all it takes is one hour to assemble and get a miniature autonomous car (AWS DeepRacer) to navigate its way around a track using reinforcement learning. Sam Sterckval (from Edgise) summarized the vast variety of embedded AI hardware available in the market, so we can get a leg up on our next robotics project. And finally, Zaid Alyafeai (from King Fahd University) demonstrated that browsers are no less capable of running serious interactive AI models (with the help of TensorFlow.js and ml5js).\n\nThe book is in its current shape because of timely feedback from our amazing technical reviewers, who worked tirelessly on our drafts, pointed out any technical inaccuracies they came across, and gave us suggestions on better conveying our ideas. Due to their feedback (and the ever-changing APIs of TensorFlow), we ended up doing a rewrite of a majority of the book from the original prerelease. We thank Margaret Maynard-Reid (Google Developer Expert for Machine Learning, you might have read her work while reading TensorFlow documentation), Paco Nathan (35+ years industry veteran at Derwin Inc., who introduced Anirudh to the world of public speaking), Andy Petrella (CEO and Founder at Kensu and creator of SparkNotebook whose technical insights stood up to his reputation), and Nikhita Koul (Senior Data Scientist at Adobe who read and suggested improvements after every iteration, effectively reading a few thousand pages, thus making the content significantly more approachable) for their detailed reviews of each chapter. Additionally, we also had a lot of help from reviewers with expertise in specific topics be it AI in the browser, mobile development, or autonomous cars. The chapter-wise reviewer list (in alphabetical order) is as follows:\n\n  * Chapter 1: Dharini Chandrasekaran, Sherin Thomas\n\n  * Chapter 2: Anuj Sharma, Charles Kozierok, Manoj Parihar, Pankesh Bamotra, Pranav Kant\n\n  * Chapter 3: Anuj Sharma, Charles Kozierok, Manoj Parihar, Pankesh Bamotra, Pranav Kant\n\n  * Chapter 4: Anuj Sharma, Manoj Parihar, Pankesh Bamotra, Pranav Kant\n\n  * Chapter 6: Gabriel Ibagon, Jiri Simsa, Max Katz, Pankesh Bamotra\n\n  * Chapter 7: Pankesh Bamotra\n\n  * Chapter 8: Deepesh Aggarwal\n\n  * Chapter 9: Pankesh Bamotra\n\n  * Chapter 10: Brett Burley, Laurent Denoue, Manraj Singh\n\n  * Chapter 11: David Apgar, James Webb\n\n  * Chapter 12: David Apgar\n\n  * Chapter 13: Jesse Wilson, Salman Gadit\n\n  * Chapter 14: Akshit Arora, Pranav Kant, Rohit Taneja, Ronay Ak\n\n  * Chapter 15: Geertrui Van Lommel, Joke Decubber, Jolien De Cock, Marianne Van Lommel, Sam Hendrickx\n\n  * Chapter 16: Dario Salischiker, Kurt Niebuhr, Matthew Chan, Praveen Palanisamy\n\n  * Chapter 17: Kirtesh Garg, Larry Pizette, Pierre Dumas, Ricardo Sueiras, Segolene Dessertine-panhard, Sri Elaprolu, Tatsuya Arai\n\nWe have short excerpts throughout the book from creators who gave us a little peek into their world, and how and why they built the project for which they are most known. We are grateful to Fran\u00e7ois Chollet, Jeremy Howard, Pete Warden, Anima Anandkumar, Chris Anderson, Shanqing Cai, Daniel Smilkov, Cristobal Valenzuela, Daniel Shiffman, Hart Woolery, Dan Abdinoor, Chitoku Yato, John Welsh, and Danny Atsmo.\n\n## Personal Acknowledgments\n\n> \"I would like to thank my family\u2014Arbind, Saroj, and Nikhita who gave me the support, resources, time, and freedom to pursue my passions. To all the hackers and researchers at Microsoft, Aira, and Yahoo who stood with me in turning ideas to prototypes to products, it's not the successes but the hiccups which taught us a lot during our journey together. Our trials and tribulations provided glorious material for this book, enough to exceed our original estimate by an extra 250 pages! To my academic families at Carnegie Mellon, Dalhousie, and Thapar University, you taught me more than just academics (unlike what my GPA might suggest). And to the blind and low-vision community, you inspired me daily to work in the AI field by demonstrating that armed with the right tools, people are truly limitless.\"\n> \n> Anirudh\n\n> \"My grandfather, an author himself, once told me, 'Authoring a book is harder than I thought and more rewarding than I could have ever imagined.' I am eternally grateful to my grandparents and family, Mom, Dad, and Shriya for advocating seeking out knowledge and helping me become the person I am today. To my wonderful collaborators and mentors from Carnegie Mellon University, CERN, NASA FDL, Deep Vision, NITH, and NVIDIA who were with me throughout my journey, I am indebted to them for teaching and helping develop a scientific temperament. To my friends, who I hope still remember me, as I've been pretty invisible as of late, I would like to say a big thanks for being incredibly patient. I hope to see you all around. To my friends who selflessly reviewed chapters of the book and acted as a sounding board, a huge thank you\u2014without you, the book would not have taken shape.\"\n> \n> Siddha\n\n> \"I am indebted to my parents Rajagopal and Lakshmi for their unending love and support starting from the very beginning and their strong will to provide me with a good life and education. I am grateful to my professors from UF and VNIT who taught me well and made me glad that I majored in CS. I am thankful to my incredibly supportive partner Julia Tanner who, for nearly two years, had to endure nights and weekends of unending Skype calls with my coauthors, as well as my several terrible jokes (some of which unfortunately made it into this book). I'd also like to acknowledge the role of my wonderful manager Joel Kustka in supporting me during the process of writing this book. A shout out to my friends who were incredibly understanding when I couldn't hang out with them as often as they would have liked me to.\"\n> \n> Meher\n\nLast but not least, thank you to the makers of Grammarly, which empowered people with mediocre English grades to become published authors!\n\n# Chapter 1. Exploring the Landscape of Artificial Intelligence\n\nFollowing are the words from Dr. May Carson's (Figure 1-1) seminal paper on the changing role of artificial intelligence (AI) in human life in the twenty-first century:\n\n> Artificial Intelligence has often been termed as the electricity of the 21st century. Today, artificial intelligent programs will have the power to drive all forms of industry (including health), design medical devices and build new types of products and services, including robots and automobiles. As AI is advancing, organizations are already working to ensure those artificial intelligence programs can do their job and, importantly, avoid mistakes or dangerous accidents. Organizations need AI, but they also recognize that not everything they can do with AI is a good idea.\n> \n> We have had extensive studies of what it takes to operate artificial intelligence using these techniques and policies. The main conclusion is that the amount of money spent on AI programs per person, per year versus the amount used to research, build and produce them is roughly equal. That seems obvious, but it's not entirely true.\n> \n> First, AI systems need support and maintenance to help with their functions. In order to be truly reliable, they need people to have the skills to run them and to help them perform some of their tasks. It's essential that AI organizations provide workers to do the complex tasks needed by those services. It's also important to understand the people who are doing those jobs, especially once AI is more complex than humans. For example, people will most often work in jobs requiring advanced knowledge but are not necessarily skilled in working with systems that need to be built and maintained.\n\n###### Figure 1-1. Dr. May Carson\n\n# An Apology\n\nWe now have to come clean and admit that everything in this chapter up to now was entirely fake. Literally everything! All of the text (other than the first italicized sentence, which was written by us as a seed) was generated using the GPT-2 model (built by Adam King) on the website _TalkToTransformer.com_. The name of the author was generated using the \"Nado Name Generator\" on the website _Onitools.moe_. At least the picture of the author must be real, right? Nope, the picture was generated from the website _ThisPersonDoesNotExist.com_ which shows us new pictures of nonexistent people each time we reload the page using the magic of Generative Adversarial Networks (GANs).\n\nAlthough we feel ambivalent, to say the least, about starting this entire book on a dishonest note, we thought it was important to showcase the state-of-the-art of AI when you, our reader, least expected it. It is, frankly, mind-boggling and amazing and terrifying at the same time to see what AI is already capable of. The fact that it can create sentences out of thin air that are more intelligent and eloquent than some world leaders is...let's just say big league.\n\nThat being said, one thing AI can't appropriate from us just yet is the ability to be fun. We're hoping that those first three fake paragraphs will be the driest in this entire book. After all, we don't want to be known as \"the authors more boring than a machine.\"\n\n# The Real Introduction\n\nRecall that time you saw a magic show during which a trick dazzled you enough to think, \"How the heck did they do that?!\" Have you ever wondered the same about an AI application that made the news? In this book, we want to equip you with the knowledge and tools to not only deconstruct but also build a similar one.\n\nThrough accessible, step-by-step explanations, we dissect real-world applications that use AI and showcase how you would go about creating them on a wide variety of platforms\u2014from the cloud to the browser to smartphones to edge AI devices, and finally landing on the ultimate challenge currently in AI: autonomous cars.\n\nIn most chapters, we begin with a motivating problem and then build an end-to-end solution one step at a time. In the earlier portions of the book, we develop the necessary skills to build the brains of the AI. But that's only half the battle. The true value of building AI is in creating usable applications. And we're not talking about toy prototypes here. We want you to construct software that can be used in the real world by real people for improving their lives. Hence, the word \"Practical\" in the book title. To that effect, we discuss various options that are available to us and choose the appropriate options based on performance, energy consumption, scalability, reliability, and privacy trade-offs.\n\nIn this first chapter, we take a step back to appreciate this moment in AI history. We explore the meaning of AI, specifically in the context of deep learning and the sequence of events that led to deep learning becoming one of the most groundbreaking areas of technological progress in the early twenty-first century. We also examine the core components underlying a complete deep learning solution, to set us up for the subsequent chapters in which we actually get our hands dirty.\n\nSo our journey begins here, with a very fundamental question.\n\n# What Is AI?\n\nThroughout this book, we use the terms \"artificial intelligence,\" \"machine learning,\" and \"deep learning\" frequently, sometimes interchangeably. But in the strictest technical terms, they mean different things. Here's a synopsis of each (see also Figure 1-2):\n\nAI\n\nThis gives machines the capabilities to mimic human behavior. IBM's Deep Blue is a recognizable example of AI.\n\nMachine learning\n\nThis is the branch of AI in which machines use statistical techniques to learn from previous information and experiences. The goal is for the machine to take action in the future based on learning observations from the past. If you watched IBM's Watson take on Ken Jennings and Brad Rutter on _Jeopardy!_ , you saw machine learning in action. More relatably, the next time a spam email doesn't reach your inbox, you can thank machine learning.\n\nDeep learning\n\nThis is a subfield of machine learning in which deep, multilayered neural networks are used to make predictions, especially excelling in computer vision, speech recognition, natural language understanding, and so on.\n\n###### Figure 1-2. The relationship between AI, machine learning, and deep learning\n\nThroughout this book, we primarily focus on deep learning.\n\n## Motivating Examples\n\nLet's cut to the chase. What compelled us to write this book? Why did you spend your hard-earned money buying this book? Our motivation was simple: to get more people involved in the world of AI. The fact that you're reading this book means that our job is already halfway done.\n\nHowever, to really pique your interest, let's take a look at some stellar examples that demonstrate what AI is already capable of doing:\n\n  * \"DeepMind's AI agents conquer human pros at StarCraft II\": _The Verge_ , 2019\n\n  * \"AI-Generated Art Sells for Nearly Half a Million Dollars at Christie's\": _AdWeek_ , 2018\n\n  * \"AI Beats Radiologists in Detecting Lung Cancer\": _American Journal of Managed Care_ , 2019\n\n  * \"Boston Dynamics Atlas Robot Can Do Parkour\": _ExtremeTech_ , 2018\n\n  * \"Facebook, Carnegie Mellon build first AI that beats pros in 6-player poker\": _ai.facebook.com_ , 2019\n\n  * \"Blind users can now explore photos by touch with Microsoft's Seeing AI\": _TechCrunch_ , 2019\n\n  * \"IBM's Watson supercomputer defeats humans in final Jeopardy match\": _VentureBeat_ , 2011\n\n  * \"Google's ML-Jam challenges musicians to improvise and collaborate with AI\": _VentureBeat_ , 2019\n\n  * \"Mastering the Game of Go without Human Knowledge\": _Nature_ , 2017\n\n  * \"Chinese AI Beats Doctors in Diagnosing Brain Tumors\": _Popular Mechanics_ , 2018\n\n  * \"Two new planets discovered using artificial intelligence\": _Phys.org_ , 2019\n\n  * \"Nvidia's latest AI software turns rough doodles into realistic landscapes\": _The Verge_ , 2019\n\nThese applications of AI serve as our North Star. The level of these achievements is the equivalent of a gold-medal-winning Olympic performance. However, applications solving a host of problems in the real world is the equivalent of completing a 5K race. Developing these applications doesn't require years of training, yet doing so provides the developer immense satisfaction when crossing the finish line. We are here to coach you through that 5K.\n\nThroughout this book, we intentionally prioritize breadth. The field of AI is changing so quickly that we can only hope to equip you with the proper mindset and array of tools. In addition to tackling individual problems, we will look at how different, seemingly unrelated problems have fundamental overlaps that we can use to our advantage. As an example, sound recognition uses Convolutional Neural Networks (CNNs), which are also the basis for modern computer vision. We touch upon practical aspects of multiple areas so you will be able to go from 0 to 80 quickly to tackle real-world problems. If we've generated enough interest that you decide you then want to go from 80 to 95, we'd consider our goal achieved. As the oft-used phrase goes, we want to \"democratize AI.\"\n\nIt's important to note that much of the progress in AI happened in just the past few years\u2014it's difficult to overstate that. To illustrate how far we've come along, take this example: five years ago, you needed a Ph.D. just to get your foot in the door of the industry. Five years later, you don't even need a Ph.D. to write an entire book on the subject. (Seriously, check our profiles!)\n\nAlthough modern applications of deep learning seem pretty amazing, they did not get there all on their own. They stood on the shoulders of many giants of the industry who have been pushing the limits for decades. Indeed, we can't fully appreciate the significance of this time without looking at the entire history.\n\n# A Brief History of AI\n\nLet's go back in time a little bit: our whole universe was in a hot dense state. Then nearly 14 billion years ago expansion started, wait...okay, we don't have to go back that far (but now we have the song stuck in your head for the rest of the day, right?). It was really just 70 years ago when the first seeds of AI were planted. Alan Turing, in his 1950 paper, \"Computing Machinery and Intelligence,\" first asked the question \"Can machines think?\" This really gets into a larger philosophical debate of sentience and what it means to be human. Does it mean to possess the ability to compose a concerto and know that you've composed it? Turing found that framework rather restrictive and instead proposed a test: if a human cannot distinguish a machine from another human, does it really matter? An AI that can mimic a human is, in essence, human.\n\n## Exciting Beginnings\n\nThe term \"artificial intelligence\" was coined by John McCarthy in 1956 at the Dartmouth Summer Research Project. Physical computers weren't even really a thing back then, so it's remarkable that they were able to discuss futuristic areas such as language simulation, self-improving learning machines, abstractions on sensory data, and more. Much of it was theoretical, of course. This was the first time that AI became a field of research rather than a single project.\n\nThe paper \"Perceptron: A Perceiving and Recognizing Automaton\" in 1957 by Frank Rosenblatt laid the foundation for deep neural networks. He postulated that it should be feasible to construct an electronic or electromechanical system that will learn to recognize similarities between patterns of optical, electrical, or tonal information. This system would function similar to the human brain. Rather than using a rule-based model (which was standard for the algorithms at the time), he proposed using statistical models to make predictions.\n\n###### Note\n\nThroughout this book, we repeat the phrase _neural network_. What is a neural network? It is a simplified model of the human brain. Much like the brain, it has _neurons_ that activate when encountering something familiar. The different neurons are connected via connections (corresponding to synapses in our brain) that help information flow from one neuron to another.\n\nIn Figure 1-3, we can see an example of the simplest neural network: a perceptron. Mathematically, the perceptron can be expressed as follows:\n\n_output = f(x 1, x2, x3) = x1 w1 \\+ x2 w2 \\+ x3 w3 \\+ b_\n\n###### Figure 1-3. An example of a perceptron\n\nIn 1965, Ivakhnenko and Lapa published the first working neural network in their paper \"Group Method of Data Handling\u2014A Rival Method of Stochastic Approximation.\" There is some controversy in this area, but Ivakhnenko is regarded by some as the father of deep learning.\n\nAround this time, bold predictions were made about what machines would be capable of doing. Machine translation, speech recognition, and more would be performed better than humans. Governments around the world were excited and began opening up their wallets to fund these projects. This gold rush started in the late 1950s and was alive and well into the mid-1970s.\n\n## The Cold and Dark Days\n\nWith millions of dollars invested, the first systems were put into practice. It turned out that a lot of the original prophecies were unrealistic. Speech recognition worked only if it was spoken in a certain way, and even then, only for a limited set of words. Language translation turned out to be heavily erroneous and much more expensive than what it would cost a human to do. Perceptrons (essentially single-layer neural networks) quickly hit a cap for making reliable predictions. This limited their usefulness for most problems in the real world. This is because they are linear functions, whereas problems in the real world often require a nonlinear classifier for accurate predictions. Imagine trying to fit a line to a curve!\n\nSo what happens when you over-promise and under-deliver? You lose funding. The Defense Advanced Research Project Agency, commonly known as DARPA (yeah, those people; the ones who built the ARPANET, which later became the internet), funded a lot of the original projects in the United States. However, the lack of results over nearly two decades increasingly frustrated the agency. It was easier to land a man on the moon than to get a usable speech recognizer!\n\nSimilarly, across the pond, the Lighthill Report was published in 1974, which said, \"The general-purpose robot is a mirage.\" Imagine being a Brit in 1974 watching the bigwigs in computer science debating on the BBC as to whether AI is a waste of resources. As a consequence, AI research was decimated in the United Kingdom and subsequently across the world, destroying many careers in the process. This phase of lost faith in AI lasted about two decades and came to be known as the \"AI Winter.\" If only Ned Stark had been around back then to warn them.\n\n## A Glimmer of Hope\n\nEven during those freezing days, there was some groundbreaking work done in this field. Sure, perceptrons\u2014being linear functions\u2014had limited capabilities. How could one fix that? By chaining them in a network, such that the output of one (or more) perceptron is the input to one (or more) perceptron. In other words, a multilayer neural network, as illustrated in Figure 1-4. The higher the number of layers, the more the nonlinearity it would learn, resulting in better predictions. There is just one issue: how does one train it? Enter Geoffrey Hinton and friends. They published a technique called _backpropagation_ in 1986 in the paper \"Learning representations by back-propagating errors.\" How does it work? Make a prediction, see how far off the prediction is from reality, and propagate back the magnitude of the error into the network so it can learn to fix it. You repeat this process until the error becomes insignificant. A simple yet powerful concept. We use the term backpropagation repeatedly throughout this book.\n\n###### Figure 1-4. An example multilayer neural network (image source)\n\nIn 1989, George Cybenko provided the first proof of the _Universal Approximation Theorem_ , which states that a neural network with a single hidden layer is theoretically capable of modeling any problem. This was remarkable because it meant that neural networks could (at least in theory) outdo any machine learning approach. Heck, it could even mimic the human brain. But all of this was only on paper. The size of this network would quickly pose limitations in the real world. This could be overcome somewhat by using multiple hidden layers and training the network with...wait for it...backpropagation!\n\nOn the more practical side of things, a team at Carnegie Mellon University built the first-ever autonomous vehicle, NavLab 1, in 1986 (Figure 1-5). It initially used a single-layer neural network to control the angle of the steering wheel. This eventually led to NavLab 5 in 1995. During a demonstration, a car drove all but 50 of the 2,850-mile journey from Pittsburgh to San Diego on its own. NavLab got its driver's license before many Tesla engineers were even born!\n\n###### Figure 1-5. The autonomous NavLab 1 from 1986 in all its glory (image source)\n\nAnother standout example from the 1980s was at the United States Postal Service (USPS). The service needed to sort postal mail automatically according to the postal codes (ZIP codes) they were addressed to. Because a lot of the mail has always been handwritten, optical character recognition (OCR) could not be used. To solve this problem, Yann LeCun et al. used handwritten data from the National Institute of Standards and Technology (NIST) to demonstrate that neural networks were capable of recognizing these handwritten digits in their paper \"Backpropagation Applied to Handwritten Zip Code Recognition.\" The agency's network, LeNet, became what the USPS used for decades to automatically scan and sort the mail. This was remarkable because it was the first convolutional neural network that really worked in the wild. Eventually, in the 1990s, banks would use an evolved version of the model called LeNet-5 to read handwritten numbers on checks. This laid the foundation for modern computer vision.\n\n###### Note\n\nThose of you who have read about the MNIST dataset might have already noticed a connection to the NIST mention we just made. That is because the MNIST dataset essentially consists of a subset of images from the original NIST dataset that had some modifications (the \"M\" in \"MNIST\") applied to them to ease the train and test process for the neural network. Modifications, some of which you can see in Figure 1-6, included resizing them to 28 x 28 pixels, centering the digit in that area, antialiasing, and so on.\n\n###### Figure 1-6. A sample of handwritten digits from the MNIST dataset\n\nA few others kept their research going, including J\u00fcrgen Schmidhuber, who proposed networks like the Long Short-Term Memory (LSTM) with promising applications for text and speech.\n\nAt that point, even though the theories were becoming sufficiently advanced, results could not be demonstrated in practice. The main reason was that it was too computationally expensive for the hardware back then and scaling them for larger tasks was a challenge. Even if by some miracle the hardware was available, the data to realize its full potential was certainly not easy to come by. After all, the internet was still in its dial-up phase. Support Vector Machines (SVMs), a machine learning technique introduced for classification problems in 1995, were faster and provided reasonably good results on smaller amounts of data, and thus had become the norm.\n\nAs a result, AI and deep learning's reputation was poor. Graduate students were warned against doing deep learning research because this is the field \"where smart scientists would see their careers end.\" People and companies working in the field would use alternative words like informatics, cognitive systems, intelligent agents, machine learning, and others to dissociate themselves from the AI name. It's a bit like when the U.S. Department of War was rebranded as the Department of Defense to be more palatable to the people.\n\n## How Deep Learning Became a Thing\n\nLuckily for us, the 2000s brought high-speed internet, smartphone cameras, video games, and photo-sharing sites like Flickr and Creative Commons (bringing the ability to legally reuse other people's photos). People in massive numbers were able to quickly take photos with a device in their pockets and then instantly upload. The data lake was filling up, and gradually there were ample opportunities to take a dip. The 14 million-image ImageNet dataset was born from this happy confluence and some tremendous work by (then Princeton's) Fei-Fei Li and company.\n\nDuring the same decade, PC and console gaming became really serious. Gamers demanded better and better graphics from their video games. This, in turn, pushed Graphics Processing Unit (GPU) manufacturers such as NVIDIA to keep improving their hardware. The key thing to remember here is that GPUs are damn good at matrix operations. Why is that the case? Because the math demands it! In computer graphics, common tasks such as moving objects, rotating them, changing their shape, adjusting their lighting, and so on all use matrix operations. And GPUs specialize in doing them. And you know what else needs a lot of matrix calculations? Neural networks. It's one big happy coincidence.\n\nWith ImageNet ready, the annual ImageNet Large Scale Visual Recognition Challenge (ILSVRC) was set up in 2010 to openly challenge researchers to come up with better techniques for classifying this data. A subset of 1,000 categories consisting of approximately 1.2 million images was available to push the boundaries of research. The state-of-the-art computer-vision techniques like Scale-Invariant Feature Transform (SIFT) + SVM yielded a 28% (in 2010) and a 25% (2011) top-5 error rate (i.e., if one of the top five guesses ranked by probability matches, it's considered accurate).\n\nAnd then came 2012, with an entry on the leaderboard that nearly halved the error rate down to 16%. Alex Krizhevsky, Ilya Sutskever (who eventually founded OpenAI), and Geoffrey Hinton from the University of Toronto submitted that entry. Aptly called AlexNet, it was a CNN that was inspired by LeNet-5. Even at just eight layers, AlexNet had a massive 60 million parameters and 650,000 neurons, resulting in a 240 MB model. It was trained over one week using two NVIDIA GPUs. This single event took everyone by surprise, proving the potential of CNNs that snowballed into the modern deep learning era.\n\nFigure 1-7 quantifies the progress that CNNs have made in the past decade. We saw a 40% year-on-year decrease in classification error rate among ImageNet LSVRC\u2013winning entries since the arrival of deep learning in 2012. As CNNs grew deeper, the error continued to decrease.\n\n###### Figure 1-7. Evolution of winning entries at ImageNet LSVRC\n\nKeep in mind we are vastly simplifying the history of AI, and we are surely glossing over some of the details. Essentially, it was a confluence of data, GPUs, and better techniques that led to this modern era of deep learning. And the progress kept expanding further into newer territories. As Table 1-1 highlights, what was in the realm of science fiction is already a reality.\n\nTable 1-1. A highlight reel of the modern deep learning era 2012 | Neural network from Google Brain team starts recognizing cats after watching YouTube videos  \n---|---  \n2013 |\n\n  * Researchers begin tinkering with deep learning on a variety of tasks\n\n  * word2vec brings context to words and phrases, getting one step closer to understanding meanings\n\n  * Error rate for speech recognition went down 25%\n\n2014 |\n\n  * GANs invented\n\n  * Skype translates speech in real time\n\n  * Eugene Goostman, a chatbot, passes the Turing Test\n\n  * Sequence-to-sequence learning with neural networks invented\n\n  * Image captioning translates images to sentences\n\n2015 |\n\n  * Microsoft ResNet beats humans in image accuracy, trains 1,000-layer network\n\n  * Baidu's Deep Speech 2 does end-to-end speech recognition\n\n  * Gmail launches Smart Reply\n\n  * YOLO (You Only Look Once) does object detection in real time\n\n  * Visual Question Answering allows asking questions based on images\n\n2016 |\n\n  * AlphaGo wins against professional human Go players\n\n  * Google WaveNets help generate realistic audio\n\n  * Microsoft achieves human parity in conversational speech recognition\n\n2017 |\n\n  * AlphaGo Zero learns to play Go itself in 3 days\n\n  * Capsule Nets fix flaws in CNNs\n\n  * Tensor Processing Units (TPUs) introduced\n\n  * California allows sale of autonomous cars\n\n  * Pix2Pix allows generating images from sketches\n\n2018 |\n\n  * AI designs AI better than humans with Neural Architecture Search\n\n  * Google Duplex demo makes restaurant reservations on our behalf\n\n  * Deepfakes swap one face for another in videos\n\n  * Google's BERT succeeds humans in language understanding tasks\n\n  * DawnBench and MLPerf established to benchmark AI training\n\n2019 |\n\n  * OpenAI Five crushes Dota2 world champions\n\n  * StyleGan generates photorealistic images\n\n  * OpenAI GPT-2 generates realistic text passages\n\n  * Fujitsu trains ImageNet in 75 seconds\n\n  * Microsoft invests $1 billion in OpenAI\n\n  * AI by the Allen Institute passes 12th-grade science test with 80% score\n\nHopefully, you now have a historical context of AI and deep learning and have an understanding of why this moment in time is significant. It's important to recognize the rapid rate at which progress is happening in this area. But as we have seen so far, this was not always the case.\n\nThe original estimate for achieving real-world computer vision was \"one summer\" back in the 1960s, according to two of the field's pioneers. They were off by only half a century! It's not easy being a futurist. A study by Alexander Wissner-Gross observed that it took 18 years on average between when an algorithm was proposed and the time it led to a breakthrough. On the other hand, that gap was a mere three years on average between when a dataset was made available and the breakthrough it helped achieve! Look at any of the breakthroughs in the past decade. The dataset that enabled that breakthrough was very likely made available just a few years prior.\n\nData was clearly the limiting factor. This shows the crucial role that a good dataset can play for deep learning. However, data is not the only factor. Let's look at the other pillars that make up the foundation of the perfect deep learning solution.\n\n# Recipe for the Perfect Deep Learning Solution\n\nBefore Gordon Ramsay starts cooking, he ensures he has all of the ingredients ready to go. The same goes for solving a problem using deep learning (Figure 1-8).\n\n###### Figure 1-8. Ingredients for the perfect deep learning solution\n\nAnd here's your deep learning _mise en place_!\n\n    _Dataset + Model + Framework + Hardware = Deep Learning Solution_\n\nLet's look into each of these in a little more detail.\n\n## Datasets\n\nJust like Pac-Man is hungry for dots, deep learning is hungry for data\u2014lots and lots of data. It needs this amount of data to spot meaningful patterns that can help make robust predictions. Traditional machine learning was the norm in the 1980s and 1990s because it would function even with few hundreds to thousands of examples. In contrast, Deep Neural Networks (DNNs), when built from scratch, would need orders more data for typical prediction tasks. The upside here is far better predictions.\n\nIn this century, we are having a data explosion with quintillions of bytes of data being created every single day\u2014images, text, videos, sensor data, and more. But to make effective use of this data, we need labels. To build a sentiment classifier to know whether an Amazon review is positive or negative, we need thousands of sentences and an assigned emotion for each. To train a face segmentation system for a Snapchat lens, we need the precise location of eyes, lips, nose, and so forth on thousands of images. To train a self-driving car, we need video segments labeled with the human driver's reactions on controls such as the brakes, accelerator, steering wheel, and so forth. These labels act as teachers to our AI and are far more valuable than unlabeled data alone.\n\nGetting labels can be pricey. It's no wonder that there is an entire industry around crowdsourcing labeling tasks among thousands of workers. Each label might cost from a few cents to dollars, depending on the time spent by the workers to assign it. For example, during the development of the Microsoft COCO (Common Objects in Context) dataset, it took roughly three seconds to label the name of each object in an image, approximately 30 seconds to place a bounding box around each object, and 79 seconds to draw the outlines for each object. Repeat that hundreds of thousands of times and you can begin to fathom the costs around some of the larger datasets. Some labeling companies like Appen and Scale AI are already valued at more than a billion dollars each.\n\nWe might not have a million dollars in our bank account. But luckily for us, two good things happened in this deep learning revolution:\n\n  * Gigantic labeled datasets have been generously made public by major companies and universities.\n\n  * A technique called _transfer learning,_ which allows us to tune our models to datasets with even hundreds of examples\u2014as long as our model was originally trained on a larger dataset similar to our current set. We use this repeatedly in the book, including in Chapter 5 where we experiment and prove even a few tens of examples can get us decent performance with this technique. Transfer learning busts the myth that big data is necessary for training a good model. Welcome to the world of _tiny data_!\n\nTable 1-2 showcases some of the popular datasets out there today for a variety of deep learning tasks.\n\nTable 1-2. A diverse range of public datasets **Data type** | **Name** | **Details**  \n---|---|---  \nImage |\n\nOpen Images V4\n\n(from Google) |   * Nine million images in 19,700 categories\n\n  * 1.74 Million images with 600 categories (bounding boxes)\n\n| Microsoft COCO |\n\n  * 330,000 images with 80 object categories\n\n  * Contains bounding boxes, segmentation, and five captions per image\n\nVideo | YouTube-8M |\n\n  * 6.1 million videos, 3,862 classes, 2.6 billion audio-visual features\n\n  * 3.0 labels\/video\n\n  * 1.53 TB of randomly sampled videos\n\nVideo, images | BDD100K\n\n(from UC Berkeley) |   * 100,000 driving videos over 1,100 hours\n\n  * 100,000 images with bounding boxes for 10 categories\n\n  * 100,000 images with lane markings\n\n  * 100,000 images with drivable-area segmentation\n\n  * 10,000 images with pixel-level instance segmentation\n\nWaymo Open Dataset | 3,000 driving scenes totaling 16.7 hours of video data, 600,000 frames, approximately 25 million 3D bounding boxes, and 22 million 2D bounding boxes  \nText | SQuAD | 150,000 Question and Answer snippets from Wikipedia  \n| Yelp Reviews | Five million Yelp reviews  \nSatellite data | Landsat Data | Several million satellite images (100 nautical mile width and height), along with eight spectral bands (15- to 60-meter spatial resolution)\n\nAudio\n\n| Google AudioSet | 2,084,320 10-second sound clips from YouTube with 632 categories  \nLibriSpeech | 1,000 hours of read English speech\n\n## Model Architecture\n\nAt a high level, a model is just a function. It takes in one or more inputs and gives an output. The input might be in the form of text, images, audio, video, and more. The output is a prediction. A good model is one whose predictions reliably match the expected reality. The model's accuracy on a dataset is a major determining factor as to whether it's suitable for use in a real-world application. For many people, this is all they really need to know about deep learning models. But it's when we peek into the inner workings of a model that it becomes really interesting (Figure 1-9).\n\n###### Figure 1-9. A black box view of a deep learning model\n\nInside the model is a graph that consists of nodes and edges. Nodes represent mathematical operations, whereas edges represent how the data flows from one node to another. In other words, if the output of one node can become the input to one or more nodes, the connections between those nodes are represented by edges. The structure of this graph determines the potential for accuracy, its speed, how much resources it consumes (memory, compute, and energy), and the type of input it's capable of processing.\n\nThe layout of the nodes and edges is known as the _architecture_ of the model. Essentially, it's a blueprint. Now, the blueprint is only half the picture. We still need the actual building. Training is the process that utilizes this blueprint to construct that building. We train a model by repeatedly 1) feeding it input data, 2) getting outputs from it, 3) monitoring how far these predictions are from the expected reality (i.e., the labels associated with the data), and then, 4) propagating the magnitude of error back to the model so that it can progressively learn to correct itself. This training process is performed iteratively until we are satisfied with the accuracy of the predictions.\n\nThe result from this training is a set of numbers (also known as weights) that is assigned to each of the nodes. These weights are necessary parameters for the nodes in the graph to operate on the input given to them. Before the training begins, we usually assign random numbers as weights. The goal of the training process is essentially to gradually tune the values of each set of these weights until they, in conjunction with their corresponding nodes, produce satisfactory predictions.\n\nTo understand weights a little better, let's examine the following dataset with two inputs and one output:\n\nTable 1-3. Example dataset **input 1** | **input 2** | **output**  \n---|---|---  \n1 | 6 | 20  \n2 | 5 | 19  \n3 | 4 | 18  \n4 | 3 | 17  \n5 | 2 | 16  \n6 | 1 | 15\n\nUsing linear algebra (or guesswork in our minds), we can deduce that the equation governing this dataset is:\n\n_output = f(input 1, input2_) = 2 x _input 1_ \\+ 3 x _input 2_\n\nIn this case, the weights for this mathematical operation are 2 and 3. A deep neural network has millions of such weight parameters.\n\nDepending on the types of nodes used, different themes of model architectures will be better suited for different kinds of input data. For example, CNNs are used for image and audio, whereas Recurrent Neural Networks (RNNs) and LSTM are often used in text processing.\n\nIn general, training one of these models from scratch can take a pretty significant amount of time, potentially weeks. Luckily for us, many researchers have already done the difficult work of training them on a generic dataset (like ImageNet) and have made them available for everyone to use. What's even better is that we can take these available models and tune them to our specific dataset. This process is called transfer learning and accounts for the vast majority of needs by practitioners.\n\nCompared to training from scratch, transfer learning provides a two-fold advantage: significantly reduced training time a (few minutes to hours instead of weeks), and it can work with a substantially smaller dataset (hundreds to thousands of data samples instead of millions). Table 1-4 shows some famous examples of model architectures.\n\nTable 1-4. Example model architectures over the years **Task** | **Example model architectures**  \n---|---  \nImage classification | ResNet-152 (2015), MobileNet (2017)  \nText classification | BERT (2018), XLNet (2019)  \nImage segmentation | U-Net (2015), DeepLabV3 (2018)  \nImage translation | Pix2Pix (2017)  \nObject detection | YOLO9000 (2016), Mask R-CNN (2017)  \nSpeech generation | WaveNet (2016)\n\nEach one of the models from Table 1-4 has a published accuracy metric on reference datasets (e.g., ImageNet for classification, MS COCO for detection). Additionally, these architectures have their own characteristic resource requirements (model size in megabytes, computation requirements in floating-point operations, or FLOPS).\n\nWe explore transfer learning in-depth in the upcoming chapters. Now, let's look at the kinds of deep learning frameworks and services that are available to us.\n\n###### Note\n\nWhen Kaiming He et al. came up with the 152-layer ResNet architecture in 2015\u2014a feat of its day considering the previous largest GoogLeNet model consisted of 22 layers\u2014there was just one question on everyone's mind: \"Why not 153 layers?\" The reason, as it turns out, was that Kaiming ran out of GPU memory!\n\n## Frameworks\n\nThere are several deep learning libraries out there that help us train our models. Additionally, there are frameworks that specialize in using those trained models to make predictions (or _inference_ ), optimizing for where the application resides.\n\nHistorically, as is the case with software generally, many libraries have come and gone\u2014Torch (2002), Theano (2007), Caffe (2013), Microsoft Cognitive Toolkit (2015), Caffe2 (2017)\u2014and the landscape has been evolving rapidly. Learnings from each have made the other libraries easier to pick up, driven interest, and improved productivity for beginners and experts alike. Table 1-5 looks at some of the popular ones.\n\nTable 1-5. Popular deep learning frameworks **Framework** | **Best suited for** | **Typical target platform**  \n---|---|---  \nTensorFlow (including Keras) | Training | Desktops, servers  \nPyTorch | Training | Desktops, servers  \nMXNet | Training | Desktops, servers  \nTensorFlow Serving | Inference | Servers  \nTensorFlow Lite | Inference | Mobile and embedded devices  \nTensorFlow.js | Inference | Browsers  \nml5.js | Inference | Browsers  \nCore ML | Inference | Apple devices  \nXnor AI2GO | Inference | Embedded devices\n\n### TensorFlow\n\nIn 2011, Google Brain developed the DNN library DistBelief for internal research and engineering. It helped train Inception (2014's winning entry to the ImageNet Large Scale Visual Recognition Challenge) as well as helped improve the quality of speech recognition within Google products. Heavily tied into Google's infrastructure, it was not easy to configure and to share code with external machine learning enthusiasts. Realizing the limitations, Google began working on a second-generation distributed machine learning framework, which promised to be general-purpose, scalable, highly performant, and portable to many hardware platforms. And the best part, it was open source. Google called it TensorFlow and announced its release on November 2015.\n\nTensorFlow delivered on a lot of these aforementioned promises, developing an end-to-end ecosystem from development to deployment, and it gained a massive following in the process. With more than 100,000 stars on GitHub, it shows no signs of stopping. However, as adoption gained, users of the library rightly criticized it for not being easy enough to use. As the joke went, TensorFlow was a library by Google engineers, of Google engineers, for Google engineers, and if you were smart enough to use TensorFlow, you were smart enough to get hired there.\n\nBut Google was not alone here. Let's be honest. Even as late as 2015, it was a given that working with deep learning libraries would inevitably be an unpleasant experience. Forget even working on these; installing some of these frameworks made people want to pull their hair out. (Caffe users out there\u2014does this ring a bell?)\n\n### Keras\n\nAs an answer to the hardships faced by deep learning practitioners, Fran\u00e7ois Chollet released the open source framework Keras in March 2015, and the world hasn't been the same since. This solution suddenly made deep learning accessible to beginners. Keras provided an intuitive and easy-to-use interface for coding, which would then use other deep learning libraries as the backend computational framework. Starting with Theano as its first backend, Keras encouraged rapid prototyping and reduced the number of lines of code. Eventually, this abstraction expanded to other frameworks including Cognitive Toolkit, MXNet, PlaidML, and, yes, TensorFlow.\n\n### PyTorch\n\nIn parallel, PyTorch started at Facebook early in 2016, where engineers had the benefit of observing TensorFlow's limitations. PyTorch supported native Python constructs and Python debugging right off the bat, making it flexible and easier to use, quickly becoming a favorite among AI researchers. It is the second-largest end-to-end deep learning system. Facebook additionally built Caffe2 to take PyTorch models and deploy them to production to serve more than a billion users. Whereas PyTorch drove research, Caffe2 was primarily used in production. In 2018, Caffe2 was absorbed into PyTorch to make a full framework.\n\n### A continuously evolving landscape\n\nHad this story ended with the ease of Keras and PyTorch, this book would not have the word \"TensorFlow\" in the subtitle. The TensorFlow team recognized that if it truly wanted to broaden the tool's reach and democratize AI, it needed to make the tool easier. So it was welcome news when Keras was officially included as part of TensorFlow, offering the best of both worlds. This allowed developers to use Keras for defining the model and training it, and core TensorFlow for its high-performance data pipeline, including distributed training and ecosystem to deploy. It was a match made in heaven! And to top it all, TensorFlow 2.0 (released in 2019) included support for native Python constructs and eager execution, as we saw in PyTorch.\n\nWith so many competing frameworks available, the question of portability inevitability arises. Imagine a new research paper published with the state-of-the-art model being made public in PyTorch. If we didn't work in PyTorch, we would be locked out of the research and would have to reimplement and train it. Developers like to be able to share models freely and not be restricted to a specific ecosystem. Organically, many developers wrote libraries to convert model formats from one library to another. It was a simple solution, except that it led to a combinatorial explosion of conversion tools that lacked official support and sufficient quality due to the sheer number of them. To address this issue, the Open Neural Network Exchange (ONNX) was championed by Microsoft and Facebook, along with major players in the industry. ONNX provided a specification for a common model format that was readable and writable by a number of popular libraries officially. Additionally, it provided converters for libraries that did not natively support this format. This allowed developers to train in one framework and do inferences in a different framework.\n\nApart from these frameworks, there are several Graphical User Interface (GUI) systems that make code-free training possible. Using transfer learning, they generate trained models quickly in several formats useful for inference. With point-and-click interfaces, even your grandma can now train a neural network!\n\nTable 1-6. Popular GUI-based model training tools **Service** | **Platform**  \n---|---  \nMicrosoft CustomVision.AI | Web-based  \nGoogle AutoML | Web-based  \nClarifai | Web-based  \nIBM Visual Recognition | Web-based  \nApple Create ML | macOS  \nNVIDIA DIGITS | Desktop  \nRunway ML | Desktop\n\nSo why did we choose TensorFlow and Keras as the primary frameworks for this book? Considering the sheer amount of material available, including documentation, Stack Overflow answers, online courses, the vast community of contributors, platform and device support, industry adoption, and, yes, open jobs available (approximately three times as many TensorFlow-related roles compared to PyTorch in the United States), TensorFlow and Keras currently dominate the landscape when it comes to frameworks. It made sense for us to select this combination. That said, the techniques discussed in the book are generalizable to other libraries, as well. Picking up a new framework shouldn't take you too long. So, if you really want to move to a company that uses PyTorch exclusively, don't hesitate to apply there.\n\n## Hardware\n\nIn 1848, when James W. Marshall discovered gold in California, the news spread like wildfire across the United States. Hundreds of thousands of people stormed to the state to begin mining for riches. This was known as the _California Gold Rush_. Early movers were able to extract a decent chunk, but the latecomers were not nearly as lucky. But the rush did not stop for many years. Can you guess who made the most money throughout this period? The shovel makers!\n\nCloud and hardware companies are the shovel makers of the twenty-first century. Don't believe us? Look at the stock performance of Microsoft and NVIDIA in the past decade. The only difference between 1849 and now is the mind-bogglingly large amount of shovel choices available to us.\n\nGiven the variety of hardware available, it is important to make the correct choices for the constraints imposed by resource, latency, budget, privacy, and legal requirements of the application.\n\nDepending on how your application interacts with the user, the inference phase usually has a user waiting at the other end for a response. This imposes restrictions on the type of hardware that can be used as well as the location of the hardware. For example, a Snapchat lens cannot run on the cloud due to the network latency issues. Additionally, it needs to run in close to real time to provide a good user experience (UX), thus setting a minimum requirement on the number of frames processed per second (typically >15 fps). On the other hand, a photo uploaded to an image library such as Google Photos does not need immediate image categorization done on it. A few seconds or few minutes of latency is acceptable.\n\nGoing to the other extreme, training takes a lot more time; anywhere between minutes to hours to days. Depending on our training scenario, the real value of better hardware is enabling faster experimentation and more iterations. For anything more serious than basic neural networks, better hardware can make a mountain of difference. Typically, GPUs would speed things up by 10 to 15 times compared to CPUs, and at a much higher performance per watt, reducing the wait time for our experiment to finish from a week to a few hours. This can be the difference in watching a documentary about the Grand Canyon (two hours) versus actually making the trip to visit the Grand Canyon (four days).\n\nFollowing are a few fundamental hardware categories to choose from and how they are typically characterized (see also Figure 1-10):\n\nCentral Processing Unit (CPU)\n\nCheap, flexible, slow. For example, Intel Core i9-9900K.\n\nGPU\n\nHigh throughput, great for batching to utilize parallel processing, expensive. For example, NVIDIA GeForce RTX 2080 Ti.\n\nField-Programmable Gate Array (FPGA)\n\nFast, low power, reprogrammable for custom solutions, expensive. Known companies include Xilinx, Lattice Semiconductor, Altera (Intel). Because of the ability to run in seconds and configurability to any AI model, Microsoft Bing runs the majority of its AI on FPGAs.\n\nApplication-Specific Integrated Circuit (ASIC)\n\nCustom-made chip. Extremely expensive to design, but inexpensive when built for scale. Just like in the pharmaceutical industry, the first item costs the most due to the R&D effort that goes into designing and making it. Producing massive quantities is rather inexpensive. Specific examples include the following:\n\nTensor Processing Unit (TPU)\n\nASIC specializing in operations for neural networks, available on Google Cloud only.\n\nEdge TPU\n\nSmaller than a US penny, accelerates inference on the edge.\n\nNeural Processing Unit (NPU)\n\nOften used by smartphone manufacturers, this is a dedicated chip for accelerating neural network inference.\n\n###### Figure 1-10. Comparison of different types of hardware relative to flexibility, performance, and cost\n\nLet's look at a few scenarios for which each one would be used:\n\n  * Getting started with training \u2192 CPU\n\n  * Training large networks \u2192 GPUs and TPUs\n\n  * Inference on smartphones \u2192 Mobile CPU, GPU, Digital Signal Processor (DSP), NPU\n\n  * Wearables (e.g., smart glasses, smartwatches) \u2192 Edge TPU, NPUs\n\n  * Embedded AI projects (e.g., flood surveying drone, autonomous wheelchair) \u2192 Accelerators like Google Coral, Intel Movidius with Raspberry Pi, or GPUs like NVIDIA Jetson Nano, all the way down to $15 microcontrollers (MCUs) for wake word detection in smart speakers\n\nAs we go through the book, we will closely explore many of these.\n\n# Responsible AI\n\nSo far, we have explored the power and the potential of AI. It shows great promise to enhance our abilities, to make us more productive, to give us superpowers.\n\nBut with great power comes great responsibility.\n\nAs much as AI can help humanity, it also has equal potential to harm us when not designed with thought and care (either intentionally or unintentionally). The AI is not to blame; rather, it's the AI's designers.\n\nConsider some real incidents that made the news in the past few years.\n\n  * \"____ can allegedly determine whether you're a terrorist just by analyzing your face\" (Figure 1-11): _Computer World_ , 2016\n\n  * \"AI is sending people to jail\u2014and getting it wrong\": _MIT Tech Review_ , 2019\n\n  * \"____ supercomputer recommended 'unsafe and incorrect' cancer treatments, internal documents show\": _STAT News_ , 2018\n\n  * \"____ built an AI tool to hire people but had to shut it down because it was discriminating against women\": _Business Insider_ , 2018\n\n  * \"____ AI study: Major object recognition systems favor people with more money\": _VentureBeat_ , 2019\n\n  * \"____ labeled black people 'gorillas'\" _USA Today_ , 2015. \"Two years later, ____ solves 'racist algorithm' problem by purging 'gorilla' label from image classifier\": _Boing Boing_ , 2018\n\n  * \"____ silences its new A.I. bot Tay, after Twitter users teach it racism\": _TechCrunch_ , 2016\n\n  * \"AI Mistakes Bus-Side Ad for Famous CEO, Charges Her With Jaywalking\": _Caixin Global_ , 2018\n\n  * \"____ to drop Pentagon AI contract after employee objections to the 'business of war'\": _Washington Post_ , 2018\n\n  * \"Self-driving ____ death car 'spotted pedestrian six seconds before mowing down and killing her'\": _The Sun_ , 2018\n\n###### Figure 1-11. Startup claiming to classify people based on their facial structure\n\nCan you fill in the blanks here? We'll give you some options\u2014Amazon, Microsoft, Google, IBM, and Uber. Go ahead and fill them out. We'll wait.\n\nThere's a reason we kept them blank. It's to recognize that it's not a problem belonging to a specific individual or a company. This is everyone's problem. And although these things happened in the past, and might not reflect the current state, we can learn from them and try not to make the same mistakes. The silver lining here is that everyone learned from these mistakes.\n\nWe, as developers, designers, architects, and leaders of AI, have the responsibility to think beyond just the technical problem at face value. Following are just a handful of topics that are relevant to any problem we solve (AI or otherwise). They must not take a backseat.\n\n## Bias\n\nOften in our everyday work, we bring in our own biases, knowingly or unknowingly. This is the result of a multitude of factors including our environment, upbringing, cultural norms, and even our inherent nature. After all, AI and the datasets that power them were not created in a vacuum\u2014they were created by human beings with their own biases. Computers don't magically create bias on their own, they reflect and amplify existing ones.\n\nTake the example from the early days of the YouTube app when the developers noticed that roughly 10% of uploaded videos were upside-down. Maybe if that number had been lower, say 1%, it could have been brushed off as user error. But 10% was too high a number to be ignored. Do you know who happens to make up 10% of the population? Left-handed people! These users were holding their phones in the opposite orientation as their right-handed peers. But the engineers at YouTube had not accounted for that case during the development and testing of their mobile app, so YouTube uploaded videos to its server in the same orientation for both left-handed and right-handed users.\n\nThis problem could have been caught much earlier if the developers had even a single left-handed person on the team. This simple example demonstrates the importance of diversity. Handedness is just one small attribute that defines an individual. Numerous other factors, often outside their control, often come into play. Factors such as gender, skin tone, economic status, disability, country of origin, speech patterns, or even something as trivial as hair length can determine life-changing outcomes for someone, including how an algorithm treats them.\n\nGoogle's machine learning glossary lists several forms of bias that can affect a machine learning pipeline. The following are just some of them:\n\nSelection bias\n\nThe dataset is not representative of the distribution of the real-world problem and is skewed toward a subset of categories. For example, in many virtual assistants and smart home speakers, some spoken accents are overrepresented, whereas other accents have no data at all in the training dataset, resulting in a poor UX for large chunks of the world's population.\n\nSelection bias can also happen because of co-occurrence of concepts. For example, Google Translate, when used to translate the sentences \"She is a doctor. He is a nurse\" into a gender-neutral language such as Turkish and then back, switches the genders, as demonstrated in Figure 1-12. This is likely because the dataset contains a large sample of co-occurrences of male pronouns and the word \"doctor,\" and female pronouns and the word \"nurse.\"\n\n###### Figure 1-12. Google Translate reflecting the underlying bias in data (as of September 2019)\n\nImplicit bias\n\nThis type of bias creeps in because of implicit assumptions that we all make when we see something. Consider the highlighted portion in Figure 1-13. Anyone shown it might assume with a high amount of certainty that those stripes belong to a zebra. In fact, given how much ImageNet-trained networks are biased toward textures, most of them will classify the full image as a zebra. Except that we know that the image is of a sofa upholstered in a zebra-like fabric.\n\n###### Figure 1-13. Zebra sofa by Glen Edelson (image source)\n\nReporting bias\n\nSometimes the loudest voices in the room are the most extreme ones and dominate the conversation. One good look at Twitter might make it seem as if the world is ending, whereas most people are busy leading mundane lives. Unfortunately, boring does not sell.\n\nIn-group\/out-group bias\n\nAn annotator from East Asia might look at a picture of the Statue of Liberty and give it tags like \"America\" or \"United States,\" whereas someone from the US might look at the same picture and assign more granular tags such as \"New York\" or \"Liberty Island.\" It's human nature to see one's own groups with nuance while seeing other groups as more homogenous, and that reflects in our datasets, as well.\n\n## Accountability and Explainability\n\nImagine, in the late 1800s, Mr. Karl Benz told you that he invented this four-wheeled device that could transport you quicker than anything else in existence. Except he had no idea how it worked. All he knew was that it consumed a highly flammable liquid that exploded several times inside it to propel it forward. What caused it to move? What caused it to stop? What stopped it from burning the person sitting inside it? He had no answers. If this was the origin story of the car, you'd probably not want to get into that contraption.\n\nThis is precisely what is happening with AI right now. Previously, with traditional machine learning, data scientists had to manually pick features (predictive variables) from data, from which a machine learning model then would learn. This manual selection process, although cumbersome and restrictive, gave them more control and insight into how the prediction came about. However, with deep learning, these features are automatically selected. Data scientists are able to build models by providing lots of data, and these models somehow end up making predictions reliably\u2014most of the time. But the data scientist doesn't know exactly how the model works, what features it learned, under what circumstances the model works, and, more importantly, the circumstances under which it doesn't work. This approach might be acceptable when Netflix is recommending TV shows to us based on what we've already watched (although we're fairly certain they have the line `recommendations.append(\"Stranger Things\")` in their code somewhere). But AI does a lot more than just recommend movies these days. Police and judicial systems are beginning to rely on algorithms to decide whether someone poses a risk to society and whether they should be detained before their trial. The lives and freedoms of many people are at stake. We simply must not outsource important decision making to an unaccountable black box. Thankfully, there's momentum to change that with investments in _Explainable AI_ , wherein the model would be able to not just provide predictions but also account for the factors that caused it to make a certain prediction, and reveal areas of limitations.\n\nAdditionally, cities (such as New York) are beginning to make their algorithms accountable to the public by recognizing that the public has a right to know what algorithms they use for vital decision making and how they work, allowing reviews and audits by experts, improving expertise in government agencies to better evaluate each system they add, and by providing mechanisms to dispute a decision made by an algorithm.\n\n## Reproducibility\n\nResearch performed in the scientific field gains wide acceptance by the community only when it's reproducible; that is, anyone studying the research should be able to replicate the conditions of the test and obtain the same results. Unless we can reproduce a model's past results, we cannot hold it accountable when using it in the future. In the absence of reproducibility, research is vulnerable to _p-hacking_ \u2014tweaking the parameters of an experiment until the desired results are obtained. It's vital for researchers to extensively document their experimental conditions, including the dataset(s), benchmarks, and algorithms, and declare the hypothesis they will be testing prior to performing an experiment. Trust in institutions is at an all-time low and research that is not grounded in reality, yet sensationalized by the media, can erode that trust even more. Traditionally, replicating a research paper was considered a dark art because many implementation details are left out. The uplifting news is that researchers are now gradually beginning to use publicly available benchmarks (as opposed to their privately constructed datasets) and open sourcing the code they used for their research. Members of the community can piggyback on this code, prove it works, and make it better, thereby leading to newer innovations rapidly.\n\n## Robustness\n\nThere's an entire area of research on one-pixel attacks on CNNs. Essentially, the objective is to find and modify a single pixel in an image to make a CNN predict something entirely different. For example, changing a single pixel in a picture of an apple might result in a CNN classifying it as a dog. A lot of other factors can influence predictions, such as noise, lighting conditions, camera angle, and more that would not have affected a human's ability to make a similar call. This is particularly relevant for self-driving cars, where it would be possible for a bad actor on a street to modify the input the car sees in order to manipulate it into doing bad things. In fact, Tencent's Keen Security Lab was able to exploit a vulnerability in Tesla's AutoPilot by strategically placing small stickers on the road, which led it to change lanes and drive into the oncoming lane. Robust AI that is capable of withstanding noise, slight deviations, and intentional manipulation is necessary if we are to be able to trust it.\n\n## Privacy\n\nIn the pursuit of building better and better AI, businesses need to collect lots of data. Unfortunately, sometimes they overstep their bounds and collect information overzealously beyond what is necessary for the task at hand. A business might believe that it is using the data it collects only for good. But what if it is acquired by a company that does not have the same ethical boundaries for data use? The consumer's information could be used for purposes beyond the originally intended goals. Additionally, all that data collected in one place makes it an attractive target for hackers, who steal personal information and sell it on the black market to criminal enterprises. Moreover, governments are already overreaching in an attempt to track each and every individual.\n\nAll of this is at odds to the universally recognized human right of privacy. What consumers desire is having transparency into what data is being collected about them, who has access to it, how it's being used, and mechanisms to opt out of the data collection process, as well to delete data that was already collected on them.\n\nAs developers, we want to be aware of all the data we are collecting, and ask ourselves whether a piece of data is even necessary to be collected in the first place. To minimize the data we collect, we could implement privacy-aware machine learning techniques such as Federated Learning (used in Google Keyboard) that allow us to train networks on the users' devices without having to send any of the Personally Identifiable Information (PII) to a server.\n\nIt turns out that in many of the aforementioned headlines at the beginning of this section, it was the bad PR fallout that brought mainstream awareness of these topics, introduced accountability, and caused an industry-wide shift in mindset to prevent repeats in the future. We must continue to hold ourselves, academics, industry leaders, and politicians accountable at every misstep and act swiftly to fix the wrongs. Every decision we make and every action we take has the potential to set a precedent for decades to come. As AI becomes ubiquitous, we need to come together to ask the tough questions and find answers for them if we want to minimize the potential harm while reaping the maximum benefits.\n\n# Summary\n\nThis chapter explored the landscape of the exciting world of AI and deep learning. We traced the timeline of AI from its humble origins, periods of great promise, through the dark AI winters, and up to its present-day resurgence. Along the way, we answered the question of why it's different this time. We then looked at the necessary ingredients to build a deep learning solution, including datasets, model architectures, frameworks, and hardware. This sets us up for further exploration in the upcoming chapters. We hope you enjoy the rest of the book. It's time to dig in!\n\n# Frequently Asked Questions\n\n  1. I'm just getting started. Do I need to spend a lot of money on buying powerful hardware?\n\nLuckily for you, you can get started even with your web browser. All of our scripts are available online, and can be run on free GPUs courtesy of the kind people at Google Colab (Figure 1-14), who are generously making powerful GPUs available for free (up to 12 hours at a time). This should get you started. As you become better at it by performing more experiments (especially in a professional capacity or on large datasets), you might want to get a GPU either by renting one on the cloud (Microsoft Azure, Amazon Web Services (AWS), Google Cloud Platform (GCP), and others) or purchasing the hardware. Watch out for those electricity bills, though!\n\n###### Figure 1-14. Screenshot of a notebook on GitHub running on Colab inside Chrome\n\n  2. Colab is great, but I already have a powerful computer that I purchased for playing <insert name of video game>. How should I set up my environment?\n\nThe ideal setup involves Linux, but Windows and macOS work, too. For most chapters, you need the following:\n\n    * Python 3 and PIP\n\n    * `tensorflow` or `tensorflow-gpu` PIP package (version 2 or greater)\n\n    * Pillow\n\nWe like keeping things clean and self-contained, so we recommend using Python virtual environments. You should use the virtual environment whenever you install a package or run a script or a notebook.\n\nIf you do not have a GPU, you are done with the setup.\n\nIf you have an NVIDIA GPU, you would want to install the appropriate drivers, then CUDA, then cuDNN, then `tensorflow-gpu` package. If you're using Ubuntu, there's an easier solution than installing these packages manually, which can be tedious and error prone even for the best of us: simply install the entire environment with just one line using Lambda Stack.\n\nAlternatively, you could install all of your packages using Anaconda Distribution, which works equally well for Windows, Mac, and Linux.\n\n  3. Where will I find the code used in this book?\n\nYou'll find ready-to-run examples at _http:\/\/PracticalDeepLearning.ai_.\n\n  4. What are the minimal prerequisites to be able to read this book?\n\nA Ph.D. in areas including Calculus, Statistical Analysis, Variational Autoencoders, Operations Research, and so on...are definitely _not_ necessary to be able to read this book (had you a little nervous, didn't we?). Some basic coding skills, familiarity with Python, a healthy amount of curiosity, and a sense of humor should go a long way in the process of absorbing the material. Although a beginner-level understanding of mobile development (with Swift and\/or Kotlin) will help, we've designed the examples to be self-sufficient and easy enough to be deployed by someone who has never written a mobile app previously.\n\n  5. What frameworks will we be using?\n\nKeras + TensorFlow for training. And chapter by chapter, we explore different inference frameworks.\n\n  6. Will I be an expert when I finish this book?\n\nIf you follow along, you'll have the know-how on a wide variety of topics all the way from training to inference, to maximizing performance. Even though this book primarily focuses on computer vision, you can bring the same know-how to other areas such as text, audio, and so on and get up to speed very quickly.\n\n  7. Who is the cat from earlier in the chapter?\n\nThat is Meher's cat, Vader. He will be making multiple cameos throughout this book. And don't worry, he has already signed a model release form.\n\n  8. Can I contact you?\n\nSure. Drop us an email at PracticalDLBook@gmail.com with any questions, corrections, or whatever, or tweet to us @PracticalDLBook.\n\n If you're reading a pirated copy, consider us disappointed in you.\n\n Robert Geirhos et al.\n\n# Chapter 2. What's in the Picture: Image Classification with Keras\n\nIf you have skimmed through deep learning literature, you might have come across a barrage of academic explanations laced with intimidating mathematics. Don't worry. We will ease you into practical deep learning with an example of classifying images with just a few lines of code.\n\nIn this chapter, we take a closer look at the Keras framework, discuss its place in the deep learning landscape, and then use it to classify a few images using existing state-of-the-art classifiers. We visually investigate how these classifiers operate by using _heatmaps_. With these heatmaps, we make a fun project in which we classify objects in videos.\n\nRecall from the \"Recipe for the Perfect Deep Learning Solution\" that we need four ingredients to create our deep learning recipe: hardware, dataset, framework, and model. Let's see how each of these comes into play in this chapter:\n\n  * We begin with the easy one: _hardware_. Even an inexpensive laptop would suffice for what we we're doing in this chapter. Alternatively, you can run the code in this chapter by opening the GitHub notebook (see _http:\/\/PracticalDeepLearning.ai_) in Colab. This is just a matter of a few mouse clicks.\n\n  * Because we won't be training a neural network just yet, we don't need a _dataset_ (other than a handful of sample photos to test with).\n\n  * Next, we come to the _framework_. This chapter's title has Keras in it, so that is what we will be using for now. In fact, we use Keras for our training needs throughout a good part of the book.\n\n  * One way to approach a deep learning problem is to obtain a dataset, write the code to train it, spend a lot of time and energy (both human and electrical) in training that model, and then use it for making predictions. But we are not gluttons for punishment. So, we will use a _pretrained model_ instead. After all, the research community has already spent blood, sweat, and tears training and publishing many of the standard models that are now publicly available. We will be reusing one of the more famous models called ResNet-50, the little sibling of ResNet-152 that won the ILSVRC in 2015.\n\nYou will get hands-on with some code in this chapter. As we all know, the best way to learn is by doing. You might be wondering, though, what's the theory behind this? That comes in later chapters, in which we delve deeper into the nuts and bolts of CNNs using this chapter as a foundation.\n\n# Introducing Keras\n\nAs Chapter 1 discussed, Keras started in 2015 as an easy-to-use abstraction layer over other libraries, making rapid prototyping possible. This made the learning curve a lot less steep for beginners of deep learning. At the same time, it made deep learning experts more productive by helping them rapidly iterate on experiments. In fact, the majority of the winning teams on _Kaggle.com_ (which hosts data science competitions) have used Keras. Eventually, in 2017, the full implementation of Keras was available directly in TensorFlow, thereby combining the high scalability, performance, and vast ecosystem of TensorFlow with the ease of Keras. On the web, we often see the TensorFlow version of Keras referred to as `tf.keras`.\n\nIn this chapter and Chapter 3, we write all of the code exclusively in Keras. That includes boilerplate functions such as file reading, image manipulation (augmentation), and so on. We do this primarily for ease of learning. From Chapter 5 onward, we begin to gradually use more of the native performant TensorFlow functions directly for more configurability and control.\n\n##### From the Creator's Desk\n\nBy Fran\u00e7ois Chollet, creator of Keras, AI researcher, and author of _Deep Learning with Python_\n\nI originally started Keras for my own use. At the time, in late 2014 and early 2015, there weren't any good options for deep learning frameworks with solid usability and strong support for both RNNs and convnets. Back then, I wasn't deliberately trying to democratize deep learning, I was just building what I wanted for myself. But as time went by, I saw lots of people pick up deep learning through Keras, and use it to solve many different problems I didn't even know existed. To me, that has been really fascinating. It made me realize that deep learning can be deployed, in transformative ways, to far more domains than what machine learning researchers are aware of. There are so many people out there that could benefit from using these technologies in their work. Because of that, I've grown to care a lot about making deep learning accessible to as many people as possible. That's the only way we're going to deploy AI to the full extent of its potential\u2014by making it broadly available. Today, in TensorFlow 2.0, the Keras API consolidates the power of deep learning in a spectrum of really productive and enjoyable workflows, suited to a variety of user profiles, from research to applications, including deployment. I'm looking forward to seeing what you'll build with it!\n\n# Predicting an Image's Category\n\nIn layperson's terms, image classification answers the question: \"what object does this image contain?\" More specifically, \"This image contains _X_ object with what probability,\" where _X_ is from a predefined list of categories of objects. If the probability is higher than a minimum threshold, the image is likely to contain one or more instances of _X_.\n\nA simple image classification pipeline would consist of the following steps:\n\n  1. Load an image.\n\n  2. Resize it to a predefined size such as 224 x 224 pixels.\n\n  3. Scale the values of the pixel to the range [0,1] or [\u20131,1], a.k.a. normalization.\n\n  4. Select a pretrained model.\n\n  5. Run the pretrained model on the image to get a list of category predictions and their respective probabilities.\n\n  6. Display a few of the highest probability categories.\n\n###### Tip\n\nThe GitHub link is provided on the website _http:\/\/PracticalDeepLearning.ai_. Navigate to `code\/chapter-2` where you will find the Jupyter notebook `1-predict-class.ipynb` that has all the steps detailed.\n\nWe begin by importing all of the necessary modules from the Keras and Python packages:\n\n    import tensorflow as tf\n    from tf.keras.applications.resnet50 import preprocess_input, decode_predictions\n    from tf.keras.preprocessing import image\n    import numpy as np\n    import matplotlib.pyplot as plt\n\nNext, we load and display the image that we want to classify (see Figure 2-1):\n\n    img_path = \"..\/..\/sample-images\/cat.jpg\"\n    img = image.load_img(img_path, target_size=(224, 224))\n    plt.imshow(img)\n    plt.show()\n\n###### Figure 2-1. Plot showing the contents of the input file\n\nYup, it's a cat (although the filename kind of gave it away). And that's what our model should ideally be predicting.\n\n##### A Brief Refresher on Images\n\nBefore we dive into how images are processed, it would be good to take a look at how images store information. At the most basic level, an image is a collection of pixels that are laid out in a rectangular grid. Depending on the type of image, each pixel can consist of 1 to 4 parts (also known as _components_ or _channels_ ). For the images we're using, these components represent the intensities of the colors red, green, and blue (RGB). They are typically 8 bits in length, so their values range between 0 and 255 (i.e., 28 \u2013 1).\n\nBefore feeding any image to Keras, we want to convert it to a standard format. This is because pretrained models expect the input to be of a specific size. The standardization in our case involves resizing the image to 224 x 224 pixels.\n\nMost deep learning models expect a batch of images as input. But what do we do when we have just one image? We create a batch of one image, of course! That essentially involves making an array consisting of that one object. Another way to look at this is to expand the number of dimensions from three (representing the three channels of the image) to four (the extra one for the length of the array itself).\n\nIf that is not clear, consider this scenario: for a batch of 64 images of size 224 x 224 pixels, each containing three channels (RGB), the object representing that batch would have a shape 64 x 224 x 224 x 3. In the code that follows, where we'd be using only one 224 x 224 x 3 image, we'd create a batch of just that image by expanding the dimensions from three to four. The shape of this newly created batch would be 1 x 224 x 224 x 3:\n\n    img_array = image.img_to_array(img)\n    img_batch = np.expand_dims(img_array, axis=0) # Increase the number of dimensions\n\nIn machine learning, models perform best when they are fed with data within a consistent range. Ranges typically include [0,1] and [\u20131,1]. Given that image pixel values are between 0 and 255, running the `preprocess_input` function from Keras on input images will normalize each pixel to a standard range. _Normalization_ or _feature scaling_ is one of the core steps in preprocessing images to make them suitable for deep learning.\n\nNow comes the model. We will be using a _Convolutional Neural Network_ (CNN) called ResNet-50. The very first question we should ask is, \"Where will I find the model?\" Of course, we could hunt for it on the internet to find something that is compatible with our deep learning framework (Keras). _But ain't nobody got time for that!_ Luckily, Keras loves to make things easy and provides it to us in a single function call. After we call this function for the first time, the model will be downloaded from a remote server and cached locally:\n\n    model = tf.keras.applications.resnet50.ResNet50()\n\nWhen predicting with this model, the results include probability predictions for each class. Keras also provides the `decode_predictions` function, which tells us the probability of each category of objects contained in the image.\n\nNow, let's see the entire code in one handy function:\n\n    def classify(img_path):\n        img = image.load_img(img_path, target_size=(224, 224))\n        model = tf.keras.applications.resnet50.ResNet50()\n        img_array = image.img_to_array(img)\n        img_batch = np.expand_dims(img_array, axis=0)\n        img_preprocessed = preprocess_input(img_batch)\n        prediction = model.predict(img_preprocessed)\n        print(decode_predictions(prediction, top=3)[0])\n\n    classify(\"..\/..\/sample-images\/cat.jpg\")\n\n    [('n02123045', 'tabby', 0.50009364),\n     ('n02124075', 'Egyptian_cat', 0.21690978),\n     ('n02123159', 'tiger_cat', 0.2061722)]\n\nThe predicted categories for this image are various types of felines. Why doesn't it simply predict the word \"cat,\" instead? The short answer is that the ResNet-50 model was trained on a granular dataset with many categories and does not include the more general \"cat.\" We investigate this dataset in more detail a little later, but first, let's load another sample image (see Figure 2-2):\n\n    img_path = '..\/..\/sample-images\/dog.jpg'\n    img = image.load_img(img_path, target_size=(224, 224))\n    plt.imshow(img)\n    plt.show()\n\n###### Figure 2-2. Plot showing the contents of the file dog.jpg\n\nAnd, again, we run our handy function from earlier:\n\n    classify(\"..\/..\/sample-images\/dog.jpg\")\n\n    [(u'n02113186', u'Cardigan', 0.809839),\n     (u'n02113023', u'Pembroke', 0.17665945),\n     (u'n02110806', u'basenji', 0.0042166105)]\n\nAs expected, we get different breeds of canines (and not just the \"dog\" category). If you are unfamiliar with the Corgi breed of dogs, the word \"corgi\" literally means \"dwarf dog\" in Welsh. The Cardigan and Pembroke are subbreeds of the Corgi family, which happen to look pretty similar to one another. It's no wonder our model thinks that way, too.\n\nNotice the predicted probability of each category. Usually, the prediction with the highest probability is considered the answer. Alternatively, any value over a predefined threshold can be considered as the answer, too. In the dog example, if we set a threshold of 0.5, Cardigan would be our answer.\n\n###### Figure 2-3. Running the notebook on Google Colab using the browser\n\n###### Tip\n\nYou can follow along with the code in this chapter and execute it interactively without any installations in the browser itself with Google Colab. Simply find the \"Run on Colab\" link at the top of each notebook on GitHub that you'd like to experiment with. Then, click the \"Run Cell\" button; this should execute the code within that cell, as shown in Figure 2-3.\n\n# Investigating the Model\n\nWe got the predictions from our model, great! But what factors led to those predictions? There are a few questions that we need to ask here:\n\n  * What dataset was the model trained on?\n\n  * Are there other models that I can use? How good are they? Where can I get them?\n\n  * Why does my model predict what it predicts?\n\nWe look into the answers to each of these questions in this section.\n\n## ImageNet Dataset\n\nLet's investigate the ImageNet dataset on which ResNet-50 was trained. ImageNet, as the name suggests, is a network of images; that is, a dataset of images organized as a network, as demonstrated in Figure 2-4. It is arranged in a hierarchical manner (like the WordNet hierarchy) such that the parent node encompasses a collection of images of all different varieties possible within that parent. For example, within the \"animal\" parent node, there are fish, birds, mammals, invertebrates, and so on. Each category has multiple subcategories, and these have subsubcategories, and so forth. For example, the category \"American water spaniel\" is eight levels from the root. The dog category contains 189 total subcategories in five hierarchical levels.\n\nVisually, we developed the tree diagram shown in Figure 2-5 to help you to understand the wide variety of high-level entities that the ImageNet dataset contains. This treemap also shows the relative percentage of different categories that make up the ImageNet dataset.\n\n###### Figure 2-4. The categories and subcategories in the ImageNet dataset\n\n###### Figure 2-5. Treemap of ImageNet and its classes\n\nThe ImageNet dataset was the basis for the famous ILSVRC that started in 2010 to benchmark progress in computer vision and challenge researchers to innovate on tasks including object classification. Recall from Chapter 1 that the ImageNet challenge saw submissions that drastically improved in accuracy each year. When it started out, the error rate was nearly 30%. And now, it is 2.2%, already better than how an average human would perform at this task. This dataset and challenge are considered the single biggest reasons for the recent advancements in computer vision.\n\nWait, AI has better-than-human accuracy? If the dataset was created by humans, won't humans have 100% accuracy? Well, the dataset was created by experts, with each image verified by multiple people. Then Stanford researcher (and now of Tesla fame) Andrej Karpathy attempted to figure out how much a normal human would fare on ImageNet-1000. Turns out he achieved an accuracy of 94.9%, well short of the 100% we all expected. Andrej painstakingly spent a week going over 1,500 images, spending approximately one minute per image in tagging it. How did he misclassify 5.1% of the images? The reasons are a bit subtle:\n\nFine-grained recognition\n\nFor many people, it is really tough to distinguish a Siberian husky from a Alaskan Malamute. Someone who is really familiar with dog breeds would be able to tell them apart because they look for finer-level details that distinguish both breeds. It turns out that neural networks are capable of learning those finer-level details much more easily than humans.\n\nCategory unawareness\n\nNot everyone is aware of all the 120 breeds of dogs and most certainly not each one of the 1,000 classes. But the AI is. After all, it was trained on it.\n\n###### Note\n\nSimilar to ImageNet, speech datasets like Switchboard report a 5.1% error rate for speech transcription (coincidentally the same number as ImageNet). It's clear that humans have a limit, and AI is gradually beating us.\n\nOne of the other key reasons for this fast pace of improvement was that researchers were openly sharing models trained on datasets like ImageNet. In the next section, we learn about model reuse in more detail.\n\n## Model Zoos\n\nA model zoo is a place where organizations or individuals can publicly upload models that they have built for others to reuse and improve upon. These models can be trained using any framework (e.g., Keras, TensorFlow, MXNet), for any task (classification, detection, etc.), or trained on any dataset (e.g., ImageNet, Street View House Numbers (SVHN)).\n\nThe tradition of model zoos started with Caffe, one of the first deep learning frameworks, developed at the University of California, Berkeley. Training a deep learning model from scratch on a multimillion-image database requires weeks of training time and lots of GPU computational energy, making it a difficult task. The research community recognized this as a bottleneck, and the organizations that participated in the ImageNet competition open sourced their trained models on Caffe's website. Other frameworks soon followed suit.\n\nWhen starting out on a new deep learning project, it's a good idea to first explore whether there's already a model that performs a similar task and was trained on a similar dataset.\n\nThe model zoo in Keras is a collection of various architectures trained using the Keras framework on the ImageNet dataset. We tabulate their details in Table 2-1.\n\nTable 2-1. Architectural details of select pretrained ImageNet models **Model** | **Size** | **Top-1 accuracy** | **Top-5 accuracy** | **Parameters** | **Depth**  \n---|---|---|---|---|---  \nVGG16 | 528 MB | 0.713 | 0.901 | 138,357,544 | 23  \nVGG19 | 549 MB | 0.713 | 0.9 | 143,667,240 | 26  \nResNet-50 | 98 MB | 0.749 | 0.921 | 25,636,712 | 50  \nResNet-101 | 171 MB | 0.764 | 0.928 | 44,707,176 | 101  \nResNet-152 | 232 MB | 0.766 | 0.931 | 60,419,944 | 152  \nInceptionV3 | 92 MB | 0.779 | 0.937 | 23,851,784 | 159  \nInceptionResNetV2 | 215 MB | 0.803 | 0.953 | 55,873,736 | 572  \nNASNetMobile | 23 MB | 0.744 | 0.919 | 5,326,716 | \u2014  \nNASNetLarge | 343 MB | 0.825 | 0.96 | 88,949,818 | \u2014  \nMobileNet | 16 MB | 0.704 | 0.895 | 4,253,864 | 88  \nMobileNetV2 | 14 MB | 0.713 | 0.901 | 3,538,984 | 88\n\nThe column \"Top-1 accuracy\" indicates how many times the best guess was the correct answer, and the column \"Top-5 accuracy\" indicates how many times at least one out of five guesses were correct. The \"Depth\" of the network indicates how many layers are present in the network. The \"Parameters\" column indicates the size of the model; that is, how many individual weights the model has: the more parameters, the \"heavier\" the model is, and the slower it is to make predictions. In this book, we often use ResNet-50 (the most common architecture cited in research papers for high accuracy) and MobileNet (for a good balance between speed, size, and accuracy).\n\n## Class Activation Maps\n\nImage saliency, usually famous in UX research, is trying to answer the question \"What part of the image are users paying attention to?\" This is facilitated with the help of eye-tracking studies and represented in heatmaps. For example, big, bold fonts or people's faces usually get more attention than backgrounds. It's easy to guess how useful these heatmaps would be to designers and advertisers, who can then adapt their content to maximize users' attention. Taking inspiration from this human version of saliency, wouldn't it be great to learn which part of the image the neural network is paying attention to? That's precisely what we will be experimenting with.\n\nIn our experiment, we will be overlaying a _class activation map_ (or colloquially a _heatmap_ ) on top of a video in order to understand what the network pays attention to. The heatmap tells us something like \"In this picture, these pixels were responsible for the prediction of the class `dog` where \"dog\" was the category with the highest probability. The \"hot\" pixels are represented with warmer colors such as red, orange, and yellow, whereas the \"cold\" pixels are represented using blue. The \"hotter\" a pixel is, the higher the signal it provides toward the prediction. Figure 2-6 gives us a clearer picture. (If you're reading the print version, refer to the book's GitHub for the original color image.)\n\n###### Figure 2-6. Original image of a dog and its generated heatmap\n\nIn the GitHub repository (see _http:\/\/PracticalDeepLearning.ai_), navigate to _code\/chapter-2_. There, you'll find a handy Jupyter notebook, _2-class-activation-map-on-video.ipynb,_ which describes the following steps:\n\nFirst, we need to install `keras-vis` using `pip`:\n\n    $ pip install keras-vis --user\n\nWe then run the visualization script on a single image to generate the heatmap for it:\n\n    $ python visualization.py --process image --path ..\/sample-images\/dog.jpg\n\nWe should see a newly created file called _dog-output.jpg_ that shows a side-by-side view of the original image and its heatmap. As we can see from Figure 2-6, the right half of the image indicates the \"areas of heat\" along with the correct prediction of a \"Cardigan\" (i.e., Welsh Corgi).\n\nNext, we want to visualize the heatmap for frames in a video. For that, we need `FFmpeg`, an open source multimedia framework. You can find the download binary as well as the installation instructions for your operating system at _https:\/\/www.ffmpeg.org_.\n\nWe use `ffmpeg` to split up a video into individual frames (at 25 frames per second) and then run our visualization script on each of those frames. We must first create a directory to store these frames and pass its name as part of the `ffmpeg` command:\n\n    $ mkdir kitchen\n    $ ffmpeg -i video\/kitchen-input.mov -vf fps=25 kitchen\/thumb%04d.jpg -hide_banner\n\nWe then run the visualization script with the path of the directory containing the frames from the previous step:\n\n    $ python visualization.py --process video --path kitchen\/\n\nWe should see a newly created _kitchen-output_ directory that contains all of the heatmaps for the frames from the input directory.\n\nFinally, compile a video from those frames using `ffmpeg`:\n\n    $ ffmpeg -framerate 25 -i kitchen-output\/result-%04d.jpg kitchen-output.mp4\n\nPerfect! The result is the original video side by side with a copy of the heatmap overlaid on it. This is a useful tool, in particular, to discover whether the model has learned the correct features or if it picked up stray artifacts during its training.\n\nImagine generating heatmaps to analyze the strong points and shortfalls of our trained model or a pretrained model.\n\nYou should try this experiment out on your own by shooting a video with your smartphone camera and running the aforementioned scripts on the file. Don't forget to post your videos on Twitter, tagging @PracticalDLBook!\n\n###### Tip\n\nHeatmaps are a great way to visually detect bias in the data. The quality of a model's predictions depends heavily on the data on which it was trained. If the data is biased, that will reflect in the predictions. A great example of this is (although probably an urban legend) one in which the US Army wanted to use neural networks to detect enemy tanks camouflaged in trees. The researchers who were building the model took photographs\u201450% containing camouflaged tanks and 50% with just trees. Model training yielded 100% accuracy. A cause for celebration? That sadly wasn't the case when the US Army tested it. The model had performed very poorly\u2014no better than random guesses. Investigation revealed that photos with the tanks were taken on cloudy (overcast) days and those without the tanks on clear, sunny days. And the neural network model began looking for the sky instead of the tank. If the researchers had visualized the model using heatmaps, they would have caught that issue pretty early.\n\nAs we collect data, we must be vigilant at the outset of potential bias that can pollute our model's learning. For example, when collecting images to build a food classifier, we should verify that the other artifacts such as plates and utensils are not being learned as food. Otherwise, the presence of chopsticks might get our food classified as chow mein. Another term to define this is _co-occurrence_. Food very frequently co-occurs with cutlery. So watch out for these artifacts seeping into your classifier's training.\n\n# Summary\n\nIn this chapter, we got a glimpse of the deep learning universe using Keras. It's an easy-to-use yet powerful framework that we use in the next several chapters. We observed that there is often no need to collect millions of images and use powerful GPUs to train a custom model because we can use a pretrained model to predict the category of an image. By diving deeper into datasets like ImageNet, we learned the kinds of categories these pretrained models can predict. We also learned about finding these models in model zoos that exist for most frameworks.\n\nIn Chapter 3, we explore how we can tweak an existing pretrained model to make predictions on classes of input for which it was not originally intended. As with the current chapter, our approach is geared toward obtaining output without needing millions of images and lots of hardware resources to train a classifier.\n\n \"Artificial Intelligence as a Positive and Negative Factor in Global Risk\" by Eliezer Yudkowsky in _Global Catastrophic Risks_ (Oxford University Press).\n\n# Chapter 3. Cats Versus Dogs: Transfer Learning in 30 Lines with Keras\n\nImagine that we want to learn how to play the melodica, a wind instrument in the form of a handheld keyboard. Without a musical background, and the melodica being our very first instrument, it might take us a few months to become proficient at playing it. In contrast, if we were already skilled at playing another instrument, such as the piano, it might just take a few days, given how similar the two instruments are. Taking the learnings from one task and fine tuning them on a similar task is something we often do in real life (as illustrated in Figure 3-1). The more similar the two tasks are, the easier it is to adapt the learning from one task to the other.\n\nWe can apply this phenomenon from real life to the world of deep learning. Starting a deep learning project can be relatively quick when using a pretrained model, which reuses the knowledge that it learned during its training, and adapt it to the task at hand. This process is known as _transfer learning_.\n\nIn this chapter, we use transfer learning to modify existing models by training our own classifier in minutes using Keras. By the end of this chapter, we will have several tools in our arsenal to create high-accuracy image classifiers for any task.\n\n###### Figure 3-1. Transfer learning in real life\n\n# Adapting Pretrained Models to New Tasks\n\nBefore we discuss the process of transfer learning, let's quickly take a step back and review the primary reasons for the boom in deep learning:\n\n  * Availability of bigger and better-quality datasets like ImageNet\n\n  * Better compute available; i.e., faster and cheaper GPUs\n\n  * Better algorithms (model architecture, optimizer, and training procedure)\n\n  * Availability of pretrained models that have taken months to train but can be quickly reused\n\nThe last point is probably one of the biggest reasons for the widespread adoption of deep learning by the masses. If every training task took a month, not more than a handful of researchers with deep pockets would be working in this area. Thanks to transfer learning, the underappreciated hero of training models, we can now modify an existing model to suit our task in as little as a few minutes.\n\nFor example, we saw in Chapter 2 that the pretrained ResNet-50 model, which is trained on ImageNet, can predict feline and canine breeds, among thousands of other categories. So, if we just want to classify between the high-level \"cat\" and \"dog\" categories (and not the lower-level breeds), we can begin with the ResNet-50 model and quickly retrain this model to classify cats and dogs. All we need to do is show it a dataset with these two categories during training, which should take anywhere between a few minutes to a few hours. In comparison, if we had to train a cat versus dog model without a pretrained model, it could take several hours to days.\n\n##### From the Creator's Desk\n\nBy Jeremy Howard, cofounder of fast.ai and former chief scientist at Kaggle\n\nHundreds of thousands of students have studied deep learning through fast.ai. Our goal is to get them up and running as quickly as possible, solving real problems quickly. So what's the first thing we teach? It's transfer learning!\n\nThousands of students have now shared their success stories on our forum ( _http:\/\/forums.fast.ai_) describing how with as few as 30 images they have created 100% accurate image classifiers. We've also heard from students that have broken academic records in many domains and created commercially valuable models using this simple technique.\n\nFive years ago, I created Enlitic, the first company to focus on deep learning for medicine. As an initial proof of concept, I decided to develop a lung tumor classifier from CT scans. You can probably guess what technique we used...yes, it was transfer learning! In our open source fast.ai library we make transfer learning trivially easy\u2014it's just three lines of code, and the most important best practices are built in.\n\n## A Shallow Dive into Convolutional Neural Networks\n\nWe have been using the term \"model\" to refer to the part of AI that makes our predictions. In deep learning for computer vision, that model is usually a special type of neural network called a CNN. Although we explore CNNs in greater detail later in the book, we look at them very briefly in relation to training them via transfer learning here.\n\nIn machine learning, we need to convert data into a set of discernible features and then add a classification algorithm to classify them. It's the same with CNNs. They consist of two parts: convolutional layers and fully connected layers. The job of the convolutional layers is to take the large number of pixels of an image and convert them into a much smaller representation; that is, features. The fully connected layers convert these features into probabilities. A fully connected layer is really a neural network with hidden layers, as we saw in Chapter 1. In summary, the convolutional layers act as feature extractors, whereas the fully connected layers act as classifiers. Figure 3-2 shows a high-level overview of a CNN.\n\n###### Figure 3-2. A high-level overview of a CNN\n\nImagine that we want to detect a human face. We might want to use a CNN to classify an image and determine whether it contains a face. Such a CNN would be composed of several layers connected one after another. These layers represent mathematical operations. The output of one layer is the input to the next. The first (or the lowermost) layer is the input layer, where the input image is fed. The last (or the topmost) layer is the output layer, which gives the predictions.\n\nThe way it works is the image is fed into the CNN and passes through a series of layers, with each performing a mathematical operation and passing the result to the subsequent layer. The resulting output is a list of object classes and their probabilities. For example, categories like ball\u201465%, grass\u201420%, and so on. If the output for an image contains a \"face\" class with a 70% probability, we conclude that there is a 70% likelihood that the image contains a human face.\n\n###### Note\n\nAn intuitive (and overly simplified) way to look at CNNs is to see them as a series of filters. As the word filter implies, each layer acts as a sieve of information, letting something \"pass through\" only if it recognizes it. (If you have heard of high-pass and low-pass filters in electronics, this might seem familiar.) We say that the layer was \"activated\" for that information. Each layer is activated for visual patterns resembling parts of cats, dogs, cars, and so forth. If a layer does not recognize information (due to what it learned while training), its output is close to zero. CNNs are the \"bouncers\" of the deep learning world!\n\nIn the facial detection example, lower-level layers (Figure 3-3, a; layers that are closer to the input image) are \"activated\" for simpler shapes; for example, edges and curves. Because these layers activate only for basic shapes, they can be easily reused for a different purpose than face recognition such as detecting a car (every image is composed of edges and curves, after all). Middle-level layers (Figure 3-3 b) are activated for more complex shapes such as eyes, noses, and lips. These layers are not nearly as reusable as the lower-level layers. They might not be as useful for detecting a car, but might still be useful for detecting animals. And higher-level layers (Figure 3-3 c) are activated for even more complex shapes; for example, most of the human face. These layers tend to be more task-specific and thus the least reusable across other image classification problems.\n\n###### Figure 3-3. (a) Lower-level activations, followed by (b) midlevel activations and (c) upper-layer activations (image source: Convolutional Deep Belief Networks for Scalable Unsupervised Learning of Hierarchical Representations, Lee et al., ICML 2009)\n\nThe complexity and power of what a layer can recognize increases as we approach the final layers. Conversely, the reusability of a layer decreases as we get closer to the output. This will become apparent very soon when we look at what these layers learn.\n\n## Transfer Learning\n\nIf we want to transfer knowledge from one model to another, we want to reuse more of the _generic_ layers (closer to the input) and fewer of the _task-specific_ layers (closer to the output). In other words, we want to remove the last few layers (typically the fully connected layers) so that we can utilize the more generic ones, and add layers that are geared toward our specific classification task. Once training begins, the generic layers (which form the majority of our new model) are kept frozen (i.e., they are unmodifiable), whereas the newly added task-specific layers are allowed to be modified. This is how transfer learning helps quickly train new models. Figure 3-4 illustrates this process for a pretrained model trained for task X adapted to task Y.\n\n###### Figure 3-4. An overview of transfer learning\n\n## Fine Tuning\n\nBasic transfer learning gets us only so far. We usually add only two to three fully connected layers after the generic layers to make the new classifier model. If we want higher accuracy, we must allow more layers to be trained. This means unfreezing some of the layers that would have otherwise been frozen in transfer learning. This is known as _fine tuning_. Figure 3-5 shows an example where some convolutional layers near the head\/top are unfrozen and trained for the task at hand.\n\n###### Figure 3-5. Fine tuning a convolutional neural network\n\nIt's obvious that compared to basic transfer learning, more layers are tweaked to our dataset during fine tuning. Because a higher number of layers have adapted to our task compared to transfer learning, we can achieve greater accuracy for our task. The decision on how many layers to fine tune is dependent on the amount of data at hand as well as the similarity of the target task to the original dataset on which the pretrained model was trained.\n\nWe often hear data scientists saying, \"I fine tuned the model,\" which means that they took a pretrained model, removed task-specific layers and added new ones, froze the lower layers, and trained the upper part of the network on the new dataset they had.\n\n###### Note\n\nIn daily lingo, transfer learning and fine tuning are used interchangeably. When spoken, transfer learning is used more as a general concept, whereas fine tuning is referred to as its implementation.\n\n## How Much to Fine Tune\n\nHow many layers of a CNN should we fine tune? This can be guided by the following two factors:\n\nHow much data do we have?\n\nIf we have a couple hundred labeled images, it would be difficult to train and test a freshly defined model from scratch (i.e., define a model architecture with random seed weights) because we need a lot more data. The danger of training with such a small amount of data is that these powerful networks might potentially memorize it, leading to undesirable overfitting (which we explore later in the chapter). Instead, we will borrow a pretrained network and fine tune the last few layers. But if we had a million labeled images, it would be feasible to fine tune all layers of the network and, if necessary, train from scratch. So, the amount of task-specific data dictates whether we can fine tune, and how much.\n\nHow similar is the data?\n\nIf the task-specific data is similar to the data used for the pretrained network, we can fine tune the last few layers. But if our task is identifying different bones in an X-ray image and we want to start out from an ImageNet trained network, the high dissimilarity between regular ImageNet images and X-ray images would require nearly all layers to be trained.\n\nTo summarize, Table 3-1 offers an easy-to-follow cheat sheet.\n\nTable 3-1. Cheatsheet for when and how to fine tune | **High similarity among datasets** | **Low similarity among datasets**  \n---|---|---  \n**Large amount of training data** | Fine tune all layers | Train from scratch, or fine tune all layers  \n**Small amount of training data** | Fine tune last few layers | Tough luck! Train on a smaller network with heavy data augmentation or somehow get more data\n\nEnough theory, let's see it in action.\n\n# Building a Custom Classifier in Keras with Transfer Learning\n\nAs promised, it's time to build our state-of-the-art classifier in 30 lines or less. At a high level, we will use the following steps:\n\n  1. Organize the data. Download labeled images of cats and dogs and then divide the images into training and validation folders.\n\n  2. Build the data pipeline. Define a pipeline for reading data, including preprocessing the images (e.g., resizing) and grouping multiple images together into batches.\n\n  3. Augment the data. In the absence of a ton of training images, make small changes (augmentation) like rotation, zooming, and so on to increase variation in training data.\n\n  4. Define the model. Take a pretrained model, remove the last few task-specific layers, and append a new classifier layer. Freeze the weights of original layers (i.e., make them unmodifiable). Select an optimizer algorithm and a metric to track (like accuracy).\n\n  5. Train and test. Train for a few iterations until our validation accuracy is high. Save the model to eventually load as part of any application for predictions.\n\nThis will all make sense pretty soon. Let's explore this process in detail.\n\n##### Solving the World's Most Pressing Computer-Vision Problem\n\nIn early 2014, Microsoft Research was figuring out how to solve the world's most pressing problem at the time: \"Differentiating cats and dogs.\" (Where else would we have gotten the idea for this chapter?) Keep in mind that it was a much more difficult computer-vision problem back then. To facilitate this effort, Microsoft Research released the Asirra (Animal Species Image Recognition for Restricting Access) dataset. The motivation behind the Asirra dataset was to develop a sufficiently challenging CAPTCHA system. More than three million images, labeled by animal shelters throughout the United States, were provided by _Petfinder.com_ to Microsoft Research. When this problem was initially introduced, the highest possible accuracy attained was around 80%. By using deep learning, in just a few weeks, it went to 98%! This (now relatively easy) task shows the power of deep learning.\n\n# Organize the Data\n\nIt's essential to understand the distinction between train, validation, and test data. Let's look at a real-world analogy of a student preparing for standardized exams (e.g., SAT in the US, the Gaokao in China, JEE in India, CSAT in Korea, etc.). The in-class instruction and homework assignments are analogous to the training process. The quizzes, midterms, and other tests in school are the equivalent to the validation\u2014the student is able to take them frequently, assess performance, and make improvements in their study plan. They're ultimately optimizing for their performance in the final standardized exam for which they get only one chance. The final exam is equivalent to the test set\u2014the student does not get an opportunity to improve here (ignoring the ability to retake the test). This is their one shot at showing what they have learned.\n\nSimilarly, our aim is to give the best predictions in the real world. To enable this, we divide our data into three parts: train, validation, and test. A typical distribution would be 80% for train, 10% for validation, and 10% for test. Note that we randomly divide our data into these three sets in order to ensure the least amount of _bias_ that might creep in unknowingly. The final accuracy of the model is determined by the accuracy on the _test set_ , much like the student's score is determined only on their performance on the standardized exam.\n\nThe model learns from the training data and uses the validation set to evaluate its performance. Machine learning practitioners take this performance as feedback to find opportunities to improve their models on a continuous basis, similar to how students improve their preparation with the help of quizzes. There are several knobs that we can tune to improve performance; for example, the number of layers to train.\n\nIn many research competitions (including _Kaggle.com_ ), contestants receive a test set that is separate from the data they can use for building the model. This ensures uniformity across the competition when it comes to reporting accuracy. It is up to the contestants to divide the available data into training and validation sets. Similarly, during our experiments in this book, we will continue to divide data in these two sets, keeping in mind that a test dataset is still essential to report real-world numbers.\n\nSo why even use a validation set? Data is sometimes difficult to obtain, so why not use all the available samples for training, and then report accuracy on them? Sure, when the model begins to learn, it will gradually give higher accuracy predictions on the training dataset (called training accuracy). But because they are so powerful, deep neural networks can potentially memorize the training data, even resulting in 100% accuracy on the training data sometimes. However, its real-world performance will be quite poor. It's like if the student knew the questions that would be on the exam before taking it. This is why a validation set, not used to train the model, gives a realistic assessment of the model performance. Even though we might assign 10-15% of the data as a validation set, it will go a long way in guiding us on how good our model really is.\n\nFor the training process, we need to store our dataset in the proper folder structure. We'll divide the images into two sets: training and validation. For an image file, Keras will automatically assign the name of the _class_ (category) based on its parent folder name. Figure 3-6 depicts the ideal structure to recreate.\n\n###### Figure 3-6. Example directory structure of the training and validation data for different classes\n\nThe following sequence of commands can help download the data and achieve this directory structure:\n\n    $ wget https:\/\/www.kaggle.com\/c\/dogs-vs-cats-redux-kernels-edition\/ \n    download\/train.zip \n    $ unzip train.zip\n    $ mv train data\n    $ cd data\n    $ mkdir train val\n    $ mkdir train\/cat train\/dog\n    $ mkdir val\/cat val\/dog\n\nThe 25,000 files within the data folder are prefixed with \"cat\" and \"dog.\" Now, move the files into their respective directories. To keep our initial experiment short, we pick 250 random files per class and place them in training and validation folders. We can increase\/decrease this number anytime, to experiment with a trade-off between accuracy and speed:\n\n    $ ls | grep cat | sort -R | head -250 | xargs -I {} mv {} train\/cat\/\n    $ ls | grep dog | sort -R | head -250 | xargs -I {} mv {} train\/dog\/\n    $ ls | grep cat | sort -R | head -250 | xargs -I {} mv {} val\/cat\/\n    $ ls | grep dog | sort -R | head -250 | xargs -I {} mv {} val\/dog\/\n\n# Build the Data Pipeline\n\nTo start off with our Python program, we begin by importing the necessary packages:\n\n    import tensorflow as tf\n    from tf.keras.preprocessing.image import ImageDataGenerator\n    from tf.keras.models import Model\n    from tf.keras.layers import Input, Flatten, Dense, Dropout,\n    GlobalAveragePooling2D\n    from tf.keras.applications.mobilenet import MobileNet, preprocess_input\n    import math\n\nPlace the following lines of configuration right after the import statements, which we can modify based on our dataset:\n\n    TRAIN_DATA_DIR = 'data\/train_data\/'\n    VALIDATION_DATA_DIR = 'data\/val_data\/'\n    TRAIN_SAMPLES = 500\n    VALIDATION_SAMPLES = 500\n    NUM_CLASSES = 2\n    IMG_WIDTH, IMG_HEIGHT = 224, 224\n    BATCH_SIZE = 64\n\n## Number of Classes\n\nWith two classes to distinguish between, we can treat this problem as one of the following:\n\n  * A binary classification task\n\n  * A multiclass classification task\n\n### Binary classification\n\nAs a binary classification task, it's important to note that \"cat versus dog is really \"cat versus not cat.\" A dog would be classified as a \"not cat\" much like a desk or a ball would. For a given image, the model will give a single probability value corresponding to the \"cat\" class\u2014hence the probability of \"not cat\" is 1 - _P(cat)_. If the probability is higher than 0.5, we predict it as \"cat\"; otherwise, \"not cat.\" To keep things simple, we assume that it's guaranteed that the test set would contain only images of either cats or dogs. Because \"cat versus not cat\" is a binary classification task, we set the number of classes to 1; that is, \"cat.\" Anything that cannot be classified as \"cat\" will be classified as \"not cat.\"\n\n###### Tip\n\nKeras processes the input data in the alphabetical order of the folder names. Because \"cat\" comes before \"dog\" alphabetically, our first class for prediction is \"cat.\" For a multiclass task, we can apply the same concept and infer each class identifier (index) based on the folder sort order. Note that the class index starts at 0 for the first class.\n\n### Multiclass classification\n\nIn a hypothetical world that had only cats and dogs and nothing else, a \"not cat\" would always be a dog. So the label \"not cat\" could simply be replaced with the label \"dog.\" However, in the real world, we have more than two types of objects. As explained before, a ball or a sofa would also be classified as \"dog,\" which would be incorrect. Hence, for a real-world scenario, treating this as a multiclass classification task instead of a binary classification task is far more useful. As a multiclass classification task, we predict separate probability values for each class, and the highest one is our winner. In the case of \"cat versus dog,\" we set the number of classes to two. To keep our code reusable for future tasks, we will treat this as a multiclassification task.\n\n## Batch Size\n\nAt a high level, the training process includes the following steps:\n\n  1. Make predictions on images ( _forward pass_ ).\n\n  2. Determine which predictions were incorrect and propagate back the difference between the prediction and the true value ( _backpropagation_ ).\n\n  3. Rinse and repeat until the predictions become sufficiently accurate.\n\nIt's quite likely that the initial iteration would have close to 0% accuracy. Repeating the process several times, however, can yield a highly accurate model (>90%).\n\nThe batch size defines how many images are seen by the model at a time. It's important that each batch has a good variety of images from different classes in order to prevent large fluctuations in the accuracy metric between iterations. A sufficiently large batch size would be necessary for that. However, it's important not to set the batch size too large; a batch that is too large might not fit in GPU memory, resulting in an \"out of memory\" crash. Usually, batch sizes are set as powers of 2. A good number to start with is 64 for most problems, and we can play with the number by increasing or decreasing it.\n\n# Data Augmentation\n\nUsually, when we hear deep learning, we associate it with millions of images. So, 500 images like what we have might be a low number for real-world training. While these deep neural networks are powerful, a little too powerful for small quantities of data, the danger of a limited set of training images is that the neural network might memorize our training data, and show great prediction performance on the training set, but bad accuracy on the validation set. In other words, the model has overtrained and does not generalize on previously unseen images. And we definitely don't want that.\n\n###### Tip\n\nOften, when we attempt to train a neural network on a small amount of data, the result is a model that performs extremely well on the training data itself but makes rather poor predictions on data that it has not seen before. Such a model would be described as an _overfitted_ model and the problem itself is known as _overfitting_.\n\nFigure 3-7 illustrates this phenomenon for a distribution of points close to a sine curve (with little noise). The dots represent the training data visible to our network, and the crosses represent the testing data that was not seen during training. On one extreme (underfitting), an unsophisticated model, such as a linear predictor, will not be able to represent the underlying distribution well and a high error rate on both the training data and the test data will result. On the other extreme (overfitting), a powerful model (such as a deep neural network) might have the capacity to memorize the training data, which would result in a really low error on the training data, but still a high error on the testing data. What we want is the happy middle where the training error and the testing error are both modestly low, which ideally ensures that our model will perform just as well in the real world as it does during training.\n\n###### Figure 3-7. Underfitting, overfitting, and ideal fitting for points close to a sine curve\n\nWith great power comes great responsibility. It's our responsibility to ensure that our powerful deep neural network does not overfit on our data. Overfitting is common when we have little training data. We can reduce this likelihood in a few different ways:\n\n  * Somehow get more data\n\n  * Heavily augment existing data\n\n  * Fine tune fewer layers\n\nThere are often situations for which there's not enough data available. Perhaps we're working on a niche problem and data is difficult to come by. But there are a few ways that we can artificially augment our dataset for classification:\n\nRotation\n\nIn our example, we might want to rotate the 500 images randomly by 20 degrees in either direction, yielding up to 20,000 possible unique images.\n\nRandom Shift\n\nShift the images slightly to the left, or to the right.\n\nZoom\n\nZoom in and out slightly of the image.\n\nBy combining rotation, shifting, and zooming, the program can generate an almost infinite number of unique images. This important step is called _data augmentation_. Data augmentation is useful not only for adding more data, but also for training more robust models for real-world scenarios. For example, not all images have the cat properly centered in the middle or at a perfect 0-degree angle. Keras provides the `ImageDataGenerator` function that augments the data while it is being loaded from the directory. To illustrate what data augmentations of images look like, Figure 3-8 showcases example augmentations generated by the imgaug library for a sample image. (Note that we will not be using imgaug for our actual training.)\n\n###### Figure 3-8. Possible image augmentations generated from a single image\n\nColored images usually have three channels: red, green, and blue. Each channel has an intensity value ranging from 0 to 255. To normalize it (i.e., scale down the value to between 0 and 1), we use the `preprocess_input` function (which, among other things, divides each pixel by 255):\n\n    train_datagen = ImageDataGenerator(preprocessing_function=preprocess_input,\n                                       rotation_range=20,\n                                       width_shift_range=0.2,\n                                       height_shift_range=0.2,\n                                       zoom_range=0.2)\n    val_datagen = ImageDataGenerator(preprocessing_function=preprocess_input)\n\n###### Tip\n\nSometimes knowing the label of a training image can be useful in determining appropriate ways of augmenting it. For example, when training a digit recognizer, you might be okay with augmentation by flipping vertically for an image of the digit \"8,\" but not for \"6\" and \"9.\"\n\nUnlike our training set, we don't want to augment our validation set. The reason is that with dynamic augmentation, the validation set would keep changing in each iteration, and the resulting accuracy metric would be inconsistent and difficult to compare across other iterations.\n\nIt's time to load the data from its directories. Training one image at a time can be pretty inefficient, so we can batch them into groups. To introduce more randomness during the training process, we'll keep shuffling the images in each batch. To bring reproducibility during multiple runs of the same program, we'll give the random number generator a seed value:\n\n    train_generator = train_datagen.flow_from_directory(\n                            TRAIN_DATA_DIR,\n                            target_size=(IMG_WIDTH, IMG_HEIGHT),\n                            batch_size=BATCH_SIZE,\n                            shuffle=True,\n                            seed=12345,\n                            class_mode='categorical')\n    validation_generator = val_datagen.flow_from_directory(\n                            VALIDATION_DATA_DIR,\n                            target_size=(IMG_WIDTH, IMG_HEIGHT),\n                            batch_size=BATCH_SIZE,\n                            shuffle=False,\n                            class_mode='categorical')\n\n# Model Definition\n\nNow that the data is taken care of, we come to the most crucial component of our training process: the model. In the code that follows, we reuse a CNN previously trained on the ImageNet dataset (MobileNet in our case), throw away the last few layers, called fully connected layers (i.e., ImageNet-specific classifier layers), and replace them with our own classifier suited to the task at hand.\n\nFor transfer learning, we \"freeze\" the weights of the original model; that is, set those layers as unmodifiable, so only the layers of the new classifier (that we'll add) can be modified. We use MobileNet here to keep things fast, but this method will work just as well for any neural network. The following lines include a few terms such as `Dense`, `Dropout`, and so on. Although we won't explore them in this chapter, you can find explanations in Appendix A.\n\n    def model_maker():\n        base_model = MobileNet(include_top=False, input_shape =\n    (IMG_WIDTH,IMG_HEIGHT,3))\n        for layer in base_model.layers[:]:\n            layer.trainable = False _# Freeze the layers_\n        input = Input(shape=(IMG_WIDTH, IMG_HEIGHT, 3))\n        custom_model = base_model(input)\n        custom_model = GlobalAveragePooling2D()(custom_model)\n        custom_model = Dense(64, activation='relu')(custom_model)\n        custom_model = Dropout(0.5)(custom_model)\n        predictions = Dense(NUM_CLASSES, activation='softmax')(custom_model)\n        return Model(inputs=input, outputs=predictions)\n\n# Train the Model\n\n## Set Training Parameters\n\nWith both the data and model ready, all we have left to do is train the model. This is also known as _fitting the model to the data_. For training a model, we need to select and modify a few different training parameters.\n\nLoss function\n\nThe `loss` function is the penalty we impose on the model for incorrect predictions during the training process. It is the value of this function that we seek to _minimize_. For example, in a task to predict house prices, the `loss` function could be the root-mean-square error.\n\nOptimizer\n\nThis is an algorithm that helps minimize the `loss` function. We use Adam, one of the fastest optimizers out there.\n\nLearning rate\n\nLearning is incremental. The learning rate tells the optimizer how big of a step to take toward the solution; in other words, where the loss is minimum. Take too big of a step, and we end up wildly swinging and overshooting our target. Take too small a step, and it can take a really long time before eventually arriving at the target loss value. It is important to set an optimal learning rate to ensure that we reach our learning goal in a reasonable amount of time. In our example, we set the learning rate at 0.001.\n\nMetric\n\nChoose a metric to judge the performance of the trained model. Accuracy is a good explainable metric, especially when the classes are not imbalanced (i.e., roughly equal amounts of data for each class). Note that this metric is not related to the `loss` function and is mainly used for reporting and not as feedback for the model.\n\nIn the following piece of code, we create the custom model using the `model_maker` function that we wrote earlier. We use the parameters described here to customize this model further for our task of cats versus dogs:\n\n    model = model_maker()\n    model.compile(loss='categorical_crossentropy',\n                  optimizer= tf.train.Adam(lr=0.001),\n                  metrics=['acc'])\n    num_steps = math.ceil(float(TRAIN_SAMPLES)\/BATCH_SIZE)              \n    model.fit_generator(train_generator,\n                        steps_per_epoch = num_steps,\n                        epochs=10,\n                        validation_data = validation_generator,\n                        validation_steps = num_steps)\n\n###### Note\n\nYou might have noticed the term epoch in the preceding code. One epoch represents a full training step where the network has gone over the entire dataset. One epoch may consist of several minibatches.\n\n## Start Training\n\nRun this program and let the magic begin. If you don't have a GPU, brew a cup of coffee while you wait\u2014it might take 5 to 10 minutes. Or why wait, when you can run the notebooks of this chapter (posted on GitHub) on Colab with a GPU runtime for free?\n\nWhen complete, notice that there are four statistics: `loss` and `acc` on both the training and validation data. We are rooting for `val_acc:`\n\n    > Epoch 1\/100 7\/7 [====] - 5s - \n    loss: 0.6888 - acc: 0.6756 - val_loss: 0.2786 - val_acc: 0.9018\n    > Epoch 2\/100 7\/7 [====] - 5s - \n    loss: 0.2915 - acc: 0.9019 - val_loss: 0.2022 - val_acc: 0.9220\n    > Epoch 3\/100 7\/7 [====] - 4s - \n    loss: 0.1851 - acc: 0.9158 - val_loss: 0.1356 - val_acc: 0.9427\n    > Epoch 4\/100 7\/7 [====] - 4s - \n    loss: 0.1509 - acc: 0.9341 - val_loss: 0.1451 - val_acc: 0.9404\n    > Epoch 5\/100 7\/7 [====] - 4s - \n    loss: 0.1455 - acc: 0.9464 - val_loss: 0.1637 - val_acc: 0.9381\n    > Epoch 6\/100 7\/7 [====] - 4s - \n    loss: 0.1366 - acc: 0.9431 - val_loss: 0.2319 - val_acc: 0.9151\n    > Epoch 7\/100 7\/7 [====] - 4s - \n    loss: 0.0983 - acc: 0.9606 - val_loss: 0.1420 - val_acc: 0.9495\n    > Epoch 8\/100 7\/7 [====] - 4s - \n    loss: 0.0841 - acc: 0.9731 - val_loss: 0.1423 - val_acc: 0.9518\n    > Epoch 9\/100 7\/7 [====] - 4s - \n    loss: 0.0714 - acc: 0.9839 - val_loss: 0.1564 - val_acc: 0.9509\n    > Epoch 10\/100 7\/7 [====] - 5s - \n    loss: 0.0848 - acc: 0.9677 - val_loss: 0.0882 - val_acc: 0.9702\n\nAll it took was 5 seconds in the very first epoch to reach 90% accuracy on the validation set, with just 500 training images. Not bad! And by the 10th step, we observe about 97% _validation accuracy_. That's the power of transfer learning.\n\nLet us take a moment to appreciate what happened here. With just 500 images, we were able to reach a high level of accuracy in a matter of a few seconds and with very little code. In contrast, if we did not have a model previously trained on ImageNet, getting an accurate model might have needed training time anywhere between a couple of hours to a few days, and tons more data.\n\nThat's all the code we need to train a state-of-the-art classifier on any problem. Place data into folders with the name of the class, and change the corresponding values in the configuration variables. In case our task has more than two classes, we should use `categorical_crossentropy` as the `loss` function and replace the `activation` function in the last layer with `softmax`. Table 3-2 illustrates this.\n\nTable 3-2. Deciding the loss and activation type based on the task **Classification type** | **Class mode** | **Loss** | **Activation on the last layer**  \n---|---|---|---  \n1 or 2 classes | binary | binary_crossentropy | sigmoid  \nMulticlass, single label | categorical | categorical_crossentropy | softmax  \nMulticlass, multilabel | categorical | binary_crossentropy | sigmoid\n\nBefore we forget, save the model that you just trained so that we can use it later:\n\n    model.save('model.h5')\n\n# Test the Model\n\nNow that we have a trained model, we might eventually want to use it later for our application. We can now load this model anytime and classify an image. `load_model`, as its name suggests, loads the model:\n\n    from tf.keras.models import load_model\n    model = load_model('model.h5')\n\nNow let's try loading our original sample images and see what results we get:\n\n    img_path = '..\/..\/sample_images\/dog.jpg'\n    img = image.load_img(img_path, target_size=(224,224))\n    img_array = image.img_to_array(img)\n    expanded_img_array = np.expand_dims(img_array, axis=0)\n    preprocessed_img = preprocess_input(expanded_img_array) _# Preprocess the image_\n    prediction = model.predict(preprocessed_img)\n    print(prediction)\n    print(validation_generator.class_indices)\n    [[0.9967706]]\n    {'dog': 1, 'cat': 0}\n\nPrinting the value of the probability, we see that it is 0.996. This is the probability of the given image belonging to the class \"1,\" which is a dog. Because the probability is greater than 0.5, the image is predicted as a dog.\n\nThat's all that we need to train our own classifiers. Throughout this book, you can expect to reuse this code for training with minimal modifications. You can also reuse this code in your own projects. Play with the number of epochs and images, and observe how it affects the accuracy. Also, we should play with any other data we can find online. It doesn't get easier than that!\n\n# Analyzing the Results\n\nWith our trained model, we can analyze how it's performing on the validation dataset. Beyond the more straightforward accuracy metrics, looking at the actual images of mispredictions should give an intuition as to whether the example was truly challenging or whether our model is not sophisticated enough.\n\nThere are three questions that we want to answer for each category (cat, dog):\n\n  * Which images are we most confident about being a cat\/dog?\n\n  * Which images are we least confident about being a cat\/dog?\n\n  * Which images have incorrect predictions in spite of being highly confident?\n\nBefore we get to that, let's get predictions over the entire validation dataset. First, we set the pipeline configuration correctly:\n\n    _# VARIABLES_\n    IMG_WIDTH, IMG_HEIGHT = 224, 224\n    VALIDATION_DATA_DIR = 'data\/val_data\/'\n    VALIDATION_BATCH_SIZE = 64\n\n    _# DATA GENERATORS_\n    validation_datagen = ImageDataGenerator(\n            preprocessing_function=preprocess_input)\n    validation_generator = validation_datagen.flow_from_directory(\n            VALIDATION_DATA_DIR,\n            target_size=(IMG_WIDTH, IMG_HEIGHT),\n            batch_size=VALIDATION_BATCH_SIZE,\n            shuffle=False,\n            class_mode='categorical')\n    ground_truth = validation_generator.classes\n\nThen, we get the predictions:\n\n    predictions = model.predict_generator(validation_generator)\n\nTo make our analysis easier, we make a dictionary storing the image index to the prediction and ground truth (the expected prediction) for each image:\n\n    _# prediction_table is a dict with index, prediction, ground truth_\n    prediction_table = {}\n    for index, val in enumerate(predictions):\n        _# get argmax index_\n        index_of_highest_probability = np.argmax(val)\n        value_of_highest_probability = val[index_of_highest_probability]\n        prediction_table[index] = [value_of_highest_probability,\n    index_of_highest_probability, ground_truth[index]]\n    assert len(predictions) == len(ground_truth) == len(prediction_table)\n\nFor the next two code blocks, we provide boilerplate code, which we reuse regularly throughout the book.\n\nThe following is the signature of the helper function we'll use to find the images with the highest\/lowest probability value for a given category. Additionally, we will be using another helper function, \\- `display`(), to output the images as a grid on-screen:\n\n    def display(sorted_indices, message):\n        similar_image_paths = []\n        distances = []\n        for name, value in sorted_indices:\n            [probability, predicted_index, gt] = value\n            similar_image_paths.append(VALIDATION_DATA_DIR + fnames[name])\n            distances.append(probability)\n        plot_images(similar_image_paths, distances, message)\n\nThis function is defined the book's Github website (see _http:\/\/PracticalDeepLearning.ai_), at _code\/chapter-3_ ).\n\nNow the fun starts! Which images are we most confident contain dogs? Let's find images with the highest prediction probability (i.e., closest to 1.0; see Figure 3-9) with the predicted class dog (i.e., 1):\n\n    _# Most confident predictions of 'dog'_\n    indices = get_images_with_sorted_probabilities(prediction_table,\n    get_highest_probability=True, label=1, number_of_items=10,\n    only_false_predictions=False)\n    message = 'Images with the highest probability of containing dogs'\n    display(indices[:10], message)\n\n###### Figure 3-9. Images with the highest probability of containing dogs\n\nThese images are indeed very dog-like. One of the reasons the probability is so high may be the fact that the images contain multiple dogs, as well as clear, unambiguous views. Now let's try to find which images we are least confident contain dogs (see Figure 3-10):\n\n    _# Least confident predictions of 'dog'_\n    indices = get_images_with_sorted_probabilities(prediction_table,\n    get_highest_probability=False, label=1, number_of_items=10,\n    only_false_predictions=False)\n    message = 'Images with the lowest probability of containing dogs'\n    display(indices[:10], message)\n\n###### Figure 3-10. Images with the lowest probability of containing dogs\n\nTo repeat, these are the images our classifier is most unsure of containing a dog. Most of these predictions are at the tipping point (i.e., 0.5 probability) to be the majority prediction. Keep in mind the probability of being a cat is just slightly smaller, around 0.49. Compared to the previous set of images, the animals appearing in these images are often smaller and less clear. And these images often result in mispredictions\u2014only 2 of the 10 images were correctly predicted. One possible way to do better here is to train with a larger set of images.\n\nIf you are concerned about these misclassifications, worry not. A simple trick to improve the classification accuracy is to have a higher threshold for accepting a classifier's results, say 0.75. If the classifier is unsure of an image category, its results are withheld. In Chapter 5, we look at how to find an optimal threshold.\n\nSpeaking of mispredictions, they are obviously expected when the classifier has low confidence (i.e., near 0.5 probability for a two-class problem). But what we don't want is to mispredict when our classifier is really sure of its predictions. Let's check which images the classifier is confident contain dogs in spite of them being cats (see Figure 3-11):\n\n    _# Incorrect predictions of 'dog'_\n    indices = get_images_with_sorted_probabilities(prediction_table,\n    get_highest_probability=True, label=1, number_of_items=10,\n    only_false_predictions=True)\n    message = 'Images of cats with the highest probability of containing dogs'\n    display(indices[:10], message)\n\n###### Figure 3-11. Images of cats with the highest probability of containing dogs\n\nHmm...turns out half of these images contain both cats and dogs, and our classifier is correctly predicting the dog category because they are bigger in size in these images. Thus, it's not the classifier but the data that is incorrect here. This often happens in large datasets. The remaining half often contains unclear and relatively smaller objects (but ideally we want lower confidence for these difficult-to-identify images).\n\nRepeating the same set of questions for the cat class, which images are more cat-like (see Figure 3-12)?\n\n    _# Most confident predictions of 'cat'_\n    indices = get_images_with_sorted_probabilities(prediction_table,\n    get_highest_probability=True, label=0, number_of_items=10,\n    only_false_predictions=False)\n    message = 'Images with the highest probability of containing cats'\n    display(indices[:10], message)\n\n###### Figure 3-12. Images with the highest probability of containing cats\n\nInterestingly, many of these have multiple cats. This affirms our previous hypothesis that multiple clear, unambiguous views of cats can give higher probabilities. On the other hand, which images are we most unsure about containing cats (see Figure 3-13)?\n\n    _# Least confident predictions of 'cat'_\n    indices = get_images_with_sorted_probabilities(prediction_table,\n    get_highest_probability=False, label=0, number_of_items=10,\n    only_false_predictions=False)\n    message = 'Images with the lowest probability of containing cats'\n    display(indices[:10], message)\n\n###### Figure 3-13. Images with the lowest probability of containing cats\n\nAs seen previously, the key object size is small, and some of the images are quite unclear, meaning that there is too much contrast in some cases or the object is too bright, something not in line with most of the training images. For example, the camera flash in the eighth (dog.6680) and tenth (dog.1625) images in Figure 3-13 makes the dog difficult to recognize. The sixth image contains a dog in front of a sofa of the same color. Two images contain cages.\n\nLastly, which images is our classifier mistakenly sure of containing cats (see Figure 3-14)?\n\n    _# Incorrect predictions of 'cat'_\n    indices = get_images_with_sorted_probabilities(prediction_table,\n    get_highest_probability=True, label=0, number_of_items=10,\n    only_false_predictions=True)\n    message = 'Images of dogs with the highest probability of containing cats'\n    display(indices[:10], message)\n\n###### Figure 3-14. Images of dogs with the highest probability of containing cats\n\nThese mispredictions are what we want to reduce. Some of them are clearly wrong, whereas others are understandably confusing images. The sixth image (dog.4334) in Figure 3-14 seems to be incorrectly labeled as a dog. The seventh and tenth images are difficult to distinguish against the background. The first and tenth lack enough texture within them to give the classifier enough identification power. And some of the dogs are too small, like the second and fourth.\n\nGoing over the various analyses, we can summarize that mispredictions can be caused by low illumination, unclear, difficult-to-distinguish backgrounds, lack of texture, and smaller occupied area with regard to the image.\n\nAnalyzing our predictions is a great way to understand what our model has learned and what it's bad at, and highlights opportunities to enhance its predictive power. Increasing the size of the training examples and more robust augmentation will help in improving the classification. It's also important to note that showing real-world images to our model (images that look similar to the scenario where our app will be used) will help improve its accuracy drastically. In Chapter 5, we make the classifier more robust.\n\n# Further Reading\n\nTo help understand neural networks and CNNs better, our website features a learning guide which includes recommended resources like video lectures, blogs, and, more interestingly, interactive visual tools which allow you to play with different scenarios in the browser without the need to install any packages. If you're a first-time learner of deep learning, we highly recommend this guide in order to strengthen your foundational knowledge. It covers the theory that you will need to build the intuition to solve future problems. We use Google's TensorFlow Playground (Figure 3-15) for neural networks and Andrej Karpathy's ConvNetJS (Figure 3-16) for CNNs.\n\n###### Figure 3-15. Building a neural network in TensorFlow Playground\n\n###### Figure 3-16. Defining a CNN and visualizing the output of each layer during training in ConvNetJS\n\nWe additionally have a short guide in Appendix A, which summarizes convolutional neural networks, as a ready reference.\n\n# Summary\n\nIn this chapter, we introduced the concept of transfer learning. We reused a pretrained model to build our own cats versus dogs classifier in under 30 lines of code and with barely 500 images, reaching state-of-the-art accuracy in a few minutes. By writing this code, we also debunk the myth that we need millions of images and powerful GPUs to train our classifier (though they help).\n\nHopefully, with these skills, you might be able to finally answer the age-old question of who let the dogs out.\n\nIn the next couple of chapters, we use this learning to understand CNNs in more depth and take the model accuracy to the next level.\n\n# Chapter 4. Building a Reverse Image Search Engine: Understanding Embeddings\n\nBob just bought a new home and is looking to fill it up with some fancy modern furniture. He's flipping endlessly through furniture catalogs and visiting furniture showrooms, but hasn't yet landed on something he likes. Then one day, he spots the sofa of his dreams\u2014a unique L-shaped white contemporary sofa in an office reception. The good news is that he knows what he wants. The bad news is that he doesn't know where to buy it from. The brand and model number is not written on the sofa. Asking the office manager doesn't help either. So, he takes a few photos from different angles to ask around in local furniture shops, but tough luck: no one knows this particular brand. And searching on the internet with keywords like \"white L-shaped,\" \"modern sofa\" gives him thousands of results, but not the one he's looking for.\n\nAlice hears Bob's frustration and asks, \"Why don't you try reverse image search?\" Bob uploads his images on Google and Bing's Reverse Image Search and quickly spots a similar-looking image on an online shopping website. Taking this more perfect image from the website, he does a few more reverse image searches and finds other websites offering the same sofa at cheaper prices. After a few minutes of being online, Bob has officially ordered his dream sofa!\n\n_Reverse image search_ (or as it is more technically known,  _instance retrieval_ ) enables developers and researchers to build scenarios beyond simple keyword search. From discovering visually similar objects on Pinterest to recommending similar songs on Spotify to camera-based product search on Amazon, a similar class of technology under the hood is used. Sites like TinEye alert photographers on copyright infringement when their photographs are posted without consent on the internet. Even face recognition in several security systems uses a similar concept to ascertain the identity of the person.\n\nThe best part is, with the right knowledge, you can build a working replica of many of these products in a few hours. So let's dig right in!\n\nHere's what we're doing in this chapter:\n\n  1. Performing feature extraction and similarity search on Caltech101 and Caltech256 datasets\n\n  2. Learning how to scale to large datasets (up to billions of images)\n\n  3. Making the system more accurate and optimized\n\n  4. Analyzing case studies to see how these concepts are used in mainstream products\n\n# Image Similarity\n\nThe first and foremost question is: given two images, are they similar or not?\n\nThere are several approaches to this problem. One approach is to compare patches of areas between two images. Although this can help find exact or near-exact images (that might have been cropped), even a slight rotation would result in dissimilarity. By storing the hashes of the patches, duplicates of an image can be found. One use case for this approach would be the identification of plagiarism in photographs.\n\nAnother naive approach is to calculate the histogram of RGB values and compare their similarities. This might help find near-similar images captured in the same environment without much change in the contents. For example, in Figure 4-1, this technique is used in image deduplication software aimed at finding bursts of photographs on your hard disk, so you can select the best one and delete the rest. Of course, there is an increasing possibility of false positives as your dataset grows. Another downside to this approach is that small changes to the color, hue, or white balance would make recognition more difficult.\n\n###### Figure 4-1. RGB histogram-based \"Similar Image Detector\" program\n\nA more robust traditional computer vision-based approach is to find visual features near edges using algorithms like Scale-Invariant Feature Transform (SIFT), Speeded Up Robust Features (SURF), and Oriented FAST and Rotated BRIEF (ORB) and then compare the number of similar features that are common between the two photos. This helps you go from a generic image-level understanding to a relatively robust object-level understanding. Although this is great for images with rigid objects that have less variation like the printed sides of a box of cereal, which almost always look the same, it's less helpful for comparing deformable objects like humans and animals, which can exhibit different poses. As an example, you can see the features being displayed on the camera-based product search experience on the Amazon app. The app displays these features in the form of blue dots (Figure 4-2). When it sees a sufficient number of features, it sends them to the Amazon servers to retrieve product information.\n\n###### Figure 4-2. Product scanner in Amazon app with visual features highlighted\n\nGoing deeper, another approach is to find the category (e.g., sofa) of an image using deep learning and then find other images within the same category. This is equivalent to extracting metadata from an image so that it can then be indexed and used in a typical text query-based search. This can be easily scaled by using the metadata in open source search engines like ElasticSearch. Many ecommerce sites show recommendations based on tags extracted from an image while performing a query-based search internally. As you would expect, by extracting the tags, we lose certain information like color, pose, relationships between objects in the scene, and so on. Additionally, a major disadvantage of this approach is that it requires enormous volumes of labeled data to train the classifier for extracting these labels on new images. And every time a new category needs to be added, the model needs to be retrained.\n\nBecause our aim is to search among millions of images, what we ideally need is a way to summarize the information contained in the millions of pixels in an image into a smaller representation (of say a few thousand dimensions), and have this summarized representation be close together for similar objects and further away for dissimilar items.\n\nLuckily, deep neural networks come to the rescue. As we saw in Chapter 2 and Chapter 3, the CNNs take an image input and convert it into feature vectors of a thousand dimensions, which then act as input to a classifier that outputs the top identities to which the image might belong (say dog or cat). The _feature vectors_ (also called _embeddings_ or _bottleneck features_ ) are essentially a collection of a few thousand floating-point values. Going through the convolution and pooling layers in a CNN is basically an act of reduction, to filter the information contained in the image to its most important and salient constituents, which in turn form the bottleneck features. Training the CNN molds these values in such a way that items belonging to the same class have small Euclidean distance between them (or simply the square root of the sum of squares of the difference between corresponding values) and items from different classes are separated by larger distances. This is an important property that helps solve so many problems where a classifier can't be used, especially in unsupervised problems because of a lack of adequate labeled data.\n\n###### Tip\n\nAn ideal way to find similar images would be to use _transfer learning_. For example, pass the images through a pretrained convolutional neural network like ResNet-50, extract the features, and then use a metric to calculate the error rate like the Euclidean distance.\n\nEnough talk, let's code!\n\n# Feature Extraction\n\nAn image is worth a thousand ~~words~~ features.\n\nIn this section, we play with and understand the concepts of feature extraction, primarily with the Caltech 101 dataset (131 MB, approximately 9,000 images), and then eventually with Caltech 256 (1.2 GB, approximately 30,000 images). Caltech 101, as the name suggests, consists of roughly 9,000 images in 101 categories, with about 40 to 800 images per category. It's important to note that there is a 102nd category called \"BACKGROUND_Google\" consisting of random images not contained in the first 101 categories, which needs to be deleted before we begin experimenting. Remember that all of the code we are writing is also available in the GitHub repository.\n\nLet's download the dataset:\n\n    $ wget\n    http:\/\/www.vision.caltech.edu\/Image_Datasets\/Caltech101\/\n    101_ObjectCategories.tar.gz\n    $ tar -xvf 101_ObjectCategories.tar.gz\n    $ mv 101_ObjectCategories caltech101\n    $ rm -rf caltech101\/BACKGROUND_Google\n\nNow, import all of the necessary modules:\n\n    import numpy as np\n    from numpy.linalg import norm\n    import pickle\n    from tqdm import tqdm, tqdm_notebook\n    import os\n    import time\n    from tf.keras.preprocessing import image\n    from tf.keras.applications.resnet50 import ResNet50, preprocess_input\n\nLoad the ResNet-50 model without the top classification layers, so we get only the _bottleneck features._ Then define a function that takes an image path, loads the image, resizes it to proper dimensions supported by ResNet-50, extracts the features, and then normalizes them:\n\n    model = ResNet50(weights='imagenet', include_top=False,\n                     input_shape=(224, 224, 3))\n    def extract_features(img_path, model):\n        input_shape = (224, 224, 3)\n        img = image.load_img(img_path, target_size=(\n            input_shape[0], input_shape[1]))\n        img_array = image.img_to_array(img)\n        expanded_img_array = np.expand_dims(img_array, axis=0)\n        preprocessed_img = preprocess_input(expanded_img_array)\n        features = model.predict(preprocessed_img)\n        flattened_features = features.flatten()\n        normalized_features = flattened_features \/ norm(flattened_features)\n        return normalized_features\n\n###### Tip\n\nThe function defined in the previous example is the `key` function that we use for almost every feature extraction need in Keras.\n\nThat's it! Let's see the feature-length that the model generates:\n\n    features = extract_features('..\/..\/sample_images\/cat.jpg', model)\n    print(len(features))\n\nannoy\n\n    > 2048\n\nThe ResNet-50 model generated 2,048 features from the provided image. Each feature is a floating-point value between 0 and 1.\n\n###### Tip\n\nIf your model is trained or fine tuned on a dataset that is not similar to ImageNet, redefine the \"preprocess_input(img)\" step accordingly. The mean values used in the function are particular to the ImageNet dataset. Each model in Keras has its own preprocessing function so make sure you are using the right one.\n\nNow it's time to extract features for the entire dataset. First, we get all the filenames with this handy function, which recursively looks for all the image files (defined by their extensions) under a directory:\n\n    extensions = ['.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG']\n    def get_file_list(root_dir):\n        file_list = []\n        counter = 1\n        for root, directories, filenames in os.walk(root_dir):\n            for filename in filenames:\n                if any(ext in filename for ext in extensions):\n                    file_list.append(os.path.join(root, filename))\n                    counter += 1\n        return file_list\n\nThen, we provide the path to our dataset and call the function:\n\n    _# path to the datasets_\n    root_dir = '..\/..\/datasets\/caltech101'\n    filenames = sorted(get_file_list(root_dir))\n\nWe now define a variable that will store all of the features, go through all filenames in the dataset, extract their features, and append them to the previously defined variable:\n\n    feature_list = []\n    for i in tqdm_notebook(range(len(filenames))):\n        feature_list.append(extract_features(filenames[i], model))\n\nOn a CPU, this should take under an hour. On a GPU, only a few minutes.\n\n###### Tip\n\nTo get a better sense of time, use the super handy tool `tqdm`, which shows a progress meter (Figure 4-3) along with the speed per iteration as well as the time that has passed and expected finishing time. In Python, wrap an iterable with `tqdm;` for example, `tqdm(range(10))`. Its Jupyter Notebook variant is `tqdm_notebook`.\n\n###### Figure 4-3. Progress bar shown with `tqdm_notebook`\n\nFinally, write these features to a pickle file so that we can use them in the future without having to recalculate them:\n\n    pickle.dump(feature_list, open('data\/features-caltech101-resnet.pickle', 'wb'))\n    pickle.dump(filenames, open('data\/filenames-caltech101.pickle','wb'))\n\nThat's all folks! We're done with the feature extraction part.\n\n# Similarity Search\n\nGiven a photograph, our aim is to find another photo in our dataset similar to the current one. We begin by loading the precomputed features:\n\n    filenames = pickle.load(open('data\/filenames-caltech101.pickle', 'rb'))\n    feature_list = pickle.load(open('data\/features-caltech101-resnet.pickle', 'rb'))\n\nWe'll use Python's machine learning library `scikit-learn` for finding _nearest neighbors_ of the query features; that is, features that represent a query image. We train a nearest-neighbor model using the brute-force algorithm to find the nearest five neighbors based on Euclidean distance (to install `scikit-learn` on your system, use `pip3 install sklearn)`:\n\n    from sklearn.neighbors import NearestNeighbors\n    neighbors = NearestNeighbors(n_neighbors=5, algorithm='brute',\n    metric='euclidean').fit(feature_list)\n    distances, indices = neighbors.kneighbors([feature_list[0]])\n\nNow you have both the indices and distances of the nearest five neighbors of the very first query feature (which represents the first image). Notice the quick execution of the first step\u2014the training step. Unlike training most machine learning models, which can take from several minutes to hours on large datasets, instantiating the nearest-neighbor model is instantaneous because at training time there isn't much processing. This is also called _lazy learning_ because all the processing is deferred to classification or inference time.\n\nNow that we know the indices, let's see the actual image behind that feature. First, we pick an image to query, located at say, index = 0:\n\n    import matplotlib.pyplot as plt\n    import matplotlib.image as mpimg\n    %matplotlib inline       # Show the plots as a cell within the Jupyter Notebooks\n    plt.imshow(mpimg.imread(filenames[0]))\n\nFigure 4-4 shows the result.\n\n###### Figure 4-4. The query image from the Caltech-101 dataset\n\nNow, let's examine the nearest neighbors by plotting the first result.\n\n    plt.imshow(mpimg.imread(filenames[indices[0]]))\n\nFigure 4-5 shows that result.\n\n###### Figure 4-5. The nearest neighbor to our query image\n\nWait, isn't that a duplicate? Actually, the nearest index will be the image itself because that is what is being queried:\n\n    for i in range(5):\n        print(distances[0][i])\n\n    0.0\n    0.8285478\n    0.849847\n    0.8529018\n\nThis is also confirmed by the fact that the distance of the first result is zero. Now let's plot the real first nearest neighbor:\n\n    plt.imshow(mpimg.imread(filenames[indices[1]]))\n\nTake a look at the result this time in Figure 4-6.\n\n###### Figure 4-6. The second nearest neighbor of the queried image\n\nThis definitely looks like a similar image. It captured a similar concept, has the same image category (faces), same gender, and similar background with pillars and vegetation. In fact, it's the same person!\n\nWe would probably use this functionality regularly, so we have already built a helper function `plot_images()` that visualizes several query images with their nearest neighbors. Now let's call this function to visualize the nearest neighbors of six random images. Also, note that every time you run the following piece of code, the displayed images will be different (Figure 4-7) because the displayed images are indexed by a random integer.\n\n    for i in range(6):\n        random_image_index = random.randint(0,num_images)\n        distances, indices = neighbors.kneighbors([featureList[random_image_index]])\n        _# don't take the first closest image as it will be the same image_\n        similar_image_paths = [filenames[random_image_index]] +\n                              [filenames[indices[0][i]] for i in range(1,4)]\n        plot_images(similar_image_paths, distances[0])\n\n###### Figure 4-7. Nearest neighbor for different images returns similar-looking images\n\n# Visualizing Image Clusters with t-SNE\n\nLet's step up the game by visualizing the entire dataset!\n\nTo do this, we need to reduce the dimensions of the feature vectors because it's not possible to plot a 2,048-dimension vector (the feature-length) in two dimensions (the paper). The t-distributed stochastic neighbor embedding (t-SNE) algorithm reduces the high-dimensional feature vector to 2D, providing a bird's-eye view of the dataset, which is helpful in recognizing clusters and nearby images. t-SNE is difficult to scale to large datasets, so it is a good idea to reduce the dimensionality using Principal Component Analysis (PCA) and then call t-SNE:\n\n    _# Perform PCA over the features_\n    num_feature_dimensions=100      _# Set the number of features_\n    pca = PCA(n_components = num_feature_dimensions)\n    pca.fit(featureList)\n    feature_list_compressed = pca.transform(featureList)\n\n    _# For speed and clarity, we'll analyze about first half of the dataset._\n    selected_features = feature_list_compressed[:4000]\n    selected_class_ids = class_ids[:4000]\n    selected_filenames = filenames[:4000]\n\n    tsne_results =\n    TSNE(n_components=2,verbose=1,metric='euclidean')\n        .fit_transform(selected_features)\n\n    _# Plot a scatter plot from the generated t-SNE results_\n    colormap = plt.cm.get_cmap('coolwarm')\n    scatter_plot = plt.scatter(tsne_results[:,0],tsne_results[:,1], c =\n                   selected_class_ids, cmap=colormap)\n    plt.colorbar(scatter_plot)\n    plt.show()\n\nWe discuss PCA in more detail in later sections. In order to scale to larger dimensions, use Uniform Manifold Approximation and Projection (UMAP).\n\nFigure 4-8 shows clusters of similar classes, and how they are spread close to one another.\n\n###### Figure 4-8. t-SNE visualizing clusters of image features, where each cluster represents one object class in the same color\n\nEach color in Figure 4-8 indicates a different class. To make it even more clear, we can use another helper function, `plot_images_in_2d()`, to plot the images in these clusters, as demonstrated in Figure 4-9.\n\n###### Figure 4-9. t-SNE visualization showing image clusters; similar images are in the same cluster\n\nNeat! There is a clearly demarcated cluster of human faces, flowers, vintage cars, ships, bikes, and a somewhat spread-out cluster of land and marine animals. There are lots of images on top of one another, which makes Figure 4-9 a tad bit confusing, so let's try to plot the t-SNE as clear tiles with the helper function `tsne_to_grid_plotter_manual()`, the results of which you can see in Figure 4-10.\n\n    tsne_to_grid_plotter_manual(tsne_results[:,0], tsne_results[:,1],\n                                selected_filenames)\n\n###### Figure 4-10. t-SNE visualization with tiled images; similar images are close together\n\nThis is definitely much clearer. We can see similar images are colocated within the clusters of human faces, chairs, bikes, airplanes, ships, laptops, animals, watches, flowers, tilted minarets, vintage cars, anchor signs, and cameras, all close to their own kind. Birds of a feather indeed do flock together!\n\n###### Tip\n\n2D clusters are great, but visualizing them in 3D would look stellar. It would be even better if they could be rotated, zoomed in and out, and manipulated using the mouse without any coding. And bonus points if the data could be searched interactively, revealing its neighbors. The TensorFlow Embedding projector does all this and more in a browser-based GUI tool. The preloaded embeddings from image and text datasets are helpful in getting a better intuition of the power of embeddings. And, as Figure 4-11 shows, it's reassuring to see deep learning figure out that John Lennon, Led Zeppelin, and Eric Clapton happen to be used in a similar context to the Beatles in the English language.\n\n###### Figure 4-11. TensorFlow Embedding projector showing a 3D representation of 10,000 common English words and highlighting words related to \"Beatles\"\n\n# Improving the Speed of Similarity Search\n\nThere are several opportunities to improve the speed of the similarity search step. For similarity search, we can make use of two strategies: either reduce the feature-length, or use a better algorithm to search among the features. Let's examine each of these strategies individually.\n\n## Length of Feature Vectors\n\nIdeally, we would expect that the smaller the amount of data in which to search, the faster the search should be. Recall that the ResNet-50 model gives 2,048 features. With each feature being a 32-bit floating-point, each image is represented by an 8 KB feature vector. For a million images, that equates to nearly 8 GB. Imagine how slow it would be to search among 8 GB worth of features. To give us a better picture of our scenario, Table 4-1 gives the feature-lengths that we get from different models.\n\nTable 4-1. Top 1% accuracy and feature-lengths for different CNN models **Model** | **Bottleneck feature-length** | **Top-1% accuracy on ImageNet**  \n---|---|---  \nVGG16 | 512 | 71.5%  \nVGG19 | 512 | 72.7%  \nMobileNet | 1024 | 66.5%  \nInceptionV3 | 2048 | 78.8%  \nResNet-50 | 2048 | 75.9%  \nXception | 2048 | 79.0%\n\n###### Note\n\nUnder the hood, many models available in `tf.keras.applications` yield several thousand features. For example, InceptionV3 yields features in the shape of 1 x 5 x 5 x 2048, which translates to 2,048 feature maps of 5 x 5 convolutions, resulting in a total of 51,200 features. Hence, it becomes essential to reduce this large vector by using an average or max-pooling layer. The pooling layer will condense each convolution (e.g., 5 x 5 layer) into a single value. This can be defined during model instantiation as follows:\n\n    model = InceptionV3(weights='imagenet', include_top=False,\n    input_shape = (224,224,3), pooling='max')\n\nFor models that yield a large number of features, you will usually find that all code examples make use of this pooling option. Table 4-2 shows the before and after effect of max pooling on the number of features in different models.\n\nTable 4-2. Number of features before and after pooling for different models **Model** | **# features before pooling** | **# features after pooling**  \n---|---|---  \nResNet-50 | [1,1,1,2048] = 2048 | 2048  \nInceptionV3 | [1,5,5,2048] = 51200 | 2048  \nMobileNet | [1,7,7,1024] = 50176 | 1024\n\nAs we can see, almost all the models generate a large number of features. Imagine how much faster the search would be if we could reduce to a mere 100 features (a whopping reduction of 10 to 20 times!) without compromising the quality of the results. Apart from just the size, this is an even bigger improvement for big data scenarios, for which the data can be loaded into RAM all at once instead of periodically loading parts of it, thus giving an even bigger speedup. PCA will help us make this happen.\n\n## Reducing Feature-Length with PCA\n\nPCA is a statistical procedure that questions whether features representing the data are equally important. Are some of the features redundant enough that we can get similar classification results even after removing those features? PCA is considered one of the go-to techniques for dimensionality reduction. Note that it does not eliminate redundant features; rather, it generates a new set of features that are a linear combination of the input features. These linear features are orthogonal to one another, which is why all the redundant features are absent. These features are known as _principal components._\n\nPerforming PCA is pretty simple. Using the `scikit-learn` library, execute the following:\n\n    import sklearn.decomposition.PCA as PCA\n    num_feature_dimensions=100\n    pca = PCA(n_components = num_feature_dimensions)\n    pca.fit(feature_list)\n    feature_list_compressed = pca.transform(feature_list)\n\nPCA can also tell us the relative importance of each feature. The very first dimension has the most variance and the variance keeps on decreasing as we go on:\n\n    # Explain the importance of first 20 features\n    print(pca.explained_variance_ratio_[0:20])\n\n    [ 0.07320023  0.05273142   0.04310822 0.03494248  0.02166119  0.0205037\n      0.01974325  0.01739547   0.01611573 0.01548918  0.01450421  0.01311005\n      0.01200541  0.0113084    0.01103872 0.00990405  0.00973481  0.00929487\n      0.00915592  0.0089256 ]\n\nHmm, why did we pick 100 dimensions from the original 2,048? Why not 200? PCA is representing our original feature vector but in reduced dimensions. Each new dimension has diminishing returns in representing the original vector (i.e., the new dimension might not explain the data much) and takes up valuable space. We can balance between how well the original data is explained versus how much we want to reduce it. Let's visualize the importance of say the first 200 dimensions.\n\n    pca = PCA(200)\n    pca.fit(feature_list)\n    matplotlib.style.use('seaborn')\n    plt.plot(range(1,201),pca.explained_variance_ratio_,'o--', markersize=4)\n    plt.title ('Variance for each PCA dimension')\n    plt.xlabel('PCA Dimensions')\n    plt.ylabel('Variance')\n    plt.grid(True)\n    plt.show()\n\nFigure 4-12 presents the results.\n\n###### Figure 4-12. Variance for each PCA dimension\n\nThe individual variance will tell us how important the newly added features are. For example, after the first 100 dimensions, the additional dimensions don't add much variance (almost equal to 0) and can be neglected. Without even checking the accuracy it is safe to assume that the PCA with 100 dimensions will be a robust model. Another way to look at this is to visualize how much of the original data is explained by the limited number of features by finding the cumulative variance (see Figure 4-13).\n\n    plt.plot(range(1,201),pca.explained_variance_ratio_.cumsum(),'o--', markersize=4)\n    plt.title ('Cumulative Variance with each PCA dimension')\n    plt.xlabel('PCA Dimensions')\n    plt.ylabel('Variance')\n    plt.grid(True)\n    plt.show()\n\n###### Figure 4-13. Cumulative variance with each PCA dimension\n\nAs expected, adding 100 dimensions (from 100 to 200) adds only 0.1 variance and begins to gradually plateau. For reference, using the full 2,048 features would result in a cumulative variance of 1.\n\nThe number of dimensions in PCA is an important parameter that we can tune to the problem at hand. One way to directly justify a good threshold is to find a good balance between the number of features and its effect on accuracy versus speed:\n\n    pca_dimensions = [1,2,3,4,5,10,20,50,75,100,150,200]\n    pca_accuracy = []\n    pca_time = []\n\n    for dimensions in pca_dimensions:\n        _# Perform PCA_\n        pca = PCA(n_components = dimensions)\n        pca.fit(feature_list)\n        feature_list_compressed = pca.transform(feature_list[:])\n        _# Calculate accuracy over the compressed features_\n        accuracy, time_taken = accuracy_calculator(feature_list_compressed[:])\n        pca_time.append(time_taken)\n        pca_accuracy.append(accuracy)\n        print(\"For PCA Dimensions = \", dimensions, \",\\tAccuracy = \",accuracy,\"%\",\n    \",\\tTime = \", pca_time[-1])\n\nWe visualize these results using the graph in Figure 4-14 and see that after a certain number of dimensions an increase in dimensions does not lead to higher accuracy:\n\n    plt.plot(pca_time, pca_accuracy,'o--', markersize=4)\n    for label, x, y in zip(pca_dimensions, pca_time,pca_accuracy):\n        plt.annotate(label, xy=(x, y), ha='right', va='bottom')\n    plt.title ('Test Time vs Accuracy for each PCA dimension')\n    plt.xlabel('Test Time')\n    plt.ylabel('Accuracy')\n    plt.grid(True)\n    plt.show()\n\n###### Figure 4-14. Test time versus accuracy for each PCA dimension\n\nAs is visible in the graph, there is little improvement in accuracy after increasing beyond a feature-length of 100 dimensions. With almost 20 times fewer dimensions (100) than the original (2,048), this offers drastically higher speed and less time on almost any search algorithm, while achieving similar (and sometimes slightly better) accuracy. Hence, 100 would be an ideal feature-length for this dataset. This also means that the first 100 dimensions contain the most information about the dataset.\n\nThere are a number of benefits to using this reduced representation, like efficient use of computational resources, noise removal, better generalization due to fewer dimensions, and improved performance for machine learning algorithms learning on this data. By reducing the distance calculation to the most important features, we can also improve the result accuracy slightly. This is because previously all the 2,048 features were contributing equally in the distance calculation, whereas now, only the most important 100 features get their say. But, more importantly, it saves us from the _curse of dimensionality_. It's observed that as the number of dimensions increases, the ratio of the Euclidean distance between the two closest points and the two furthest points tends to become 1. In very high-dimensional space, the majority of points from a real-world dataset seem to be a similar distance away from one another, and the Euclidean distance metric begins to fail in discerning similar versus dissimilar items. PCA helps bring sanity back.\n\nYou can also experiment with different distances like Minkowski distance, Manhattan distance, Jaccardian distance, and weighted Euclidean distance (where the weight is the contribution of each feature as explained in `pca.explained_variance_ratio_`).\n\nNow, let's turn our minds toward using this reduced set of features to make our search even faster.\n\n# Scaling Similarity Search with Approximate Nearest Neighbors\n\nWhat do we want? Nearest neighbors. What is our baseline? Brute-force search. Although convenient to implement in two lines, it goes over each element and hence scales linearly with data size (number of items as well as the number of dimensions). Having PCA take our feature vector from a length of 2,048 to 100 will not only yield a 20-times reduction in data size, but also result in an increase in speed of 20 times when using brute force. PCA does pay off!\n\nLet's assume similarity searching a small collection of 10,000 images, now represented with 100 feature-length vectors, takes approximately 1 ms. Even though this looks fast for 10,000 items, in a real production system with larger data, perhaps 10 million items, this will take more than a second to search. Our system might not be able to fulfill more than one query per second per CPU core. If you receive 100 requests per second from users, even running on multiple CPU cores of the machine (and loading the search index per thread), you would need multiple machines to be able to serve the traffic. In other words, an inefficient algorithm means money, lots of money, spent on hardware.\n\nBrute force is our baseline for every comparison. As in most algorithmic approaches, brute force is the slowest approach. Now that we have our baseline set, we will explore approximate nearest-neighbor algorithms. Instead of guaranteeing the correct result as with the brute-force approach, approximation algorithms _generally_ get the correct result because they are...well, approximations. Most of the algorithms offer some form of tuning to balance between correctness and speed. It is possible to evaluate the quality of the results by comparing against the results of the brute-force baseline.\n\n## Approximate Nearest-Neighbor Benchmark\n\nThere are several approximate nearest-neighbor (ANN) libraries out there, including well-known ones like Spotify's Annoy, FLANN, Facebook's Faiss, Yahoo's NGT, and NMSLIB. Benchmarking each of them would be a tedious task (assuming you get past installing some of them). Luckily, the good folks at _ann-benchmarks.com_ (Martin Aumueller, Erik Bernhardsson, and Alec Faitfull) have done the legwork for us in the form of reproducible benchmarks on 19 libraries on large public datasets. We'll pick the comparisons on a dataset of feature embeddings representing words (instead of images) called GloVe. This 350 MB dataset consists of 400,000 feature vectors representing words in 100 dimensions. Figure 4-15 showcases their raw performance when tuned for correctness. Performance is measured in the library's ability to respond to queries each second. Recall that a measure of correctness is the fraction of top- _n_ closest items returned with respect to the real top- _n_ closest items. This ground truth is measured by brute-force search.\n\n###### Figure 4-15. Comparison of ANN libraries (data from ann-benchmarks.com)\n\nThe strongest performers on this dataset return close to several thousand queries per second at the acceptable 0.8 recall. To put this in perspective, our brute-force search performs under 1 query per second. At the fastest, some of these libraries (like NGT) can return north of 15,000 results per second (albeit at a low recall, making it impractical for usage).\n\n## Which Library Should I Use?\n\nIt goes without saying that the library you use will end up depending heavily on your scenario. Each library presents a trade-off between search speed, accuracy, size of index, memory consumption, hardware use (CPU\/GPU), and ease of setup. Table 4-3 presents a synopsis of different scenarios and recommendations as to which library might be work best for each scenario.\n\nTable 4-3. ANN library recommendations **Scenario** | **Recommendation**  \n---|---  \nI want to experiment quickly in Python without too much setup but I also care about fast speed. | Use Annoy or NMSLIB  \nI have a large dataset (up to 10 million entries or several thousand dimensions) and care utmost about speed. | Use NGT  \nI have a ridiculously large dataset (100 million-plus entries) and have a cluster of GPUs, too. | Use Faiss  \nI want to set a ground-truth baseline with 100% correctness. Then immediately move to a faster library, impress my boss with the orders of magnitude speedup, and get a bonus. | Use brute-force approach\n\nWe offer much more detailed examples in code of several libraries on the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_), but for our purposes here, we'll showcase our go-to library, Annoy, in detail and compare it with brute-force search on a synthetic dataset. Additionally, we briefly touch on Faiss and NGT.\n\n## Creating a Synthetic Dataset\n\nTo make an apples-to-apples comparison between different libraries, we first create a million-item dataset composed of random floating-point values with mean 0 and variance 1. Additionally, we pick a random feature vector as our query to find the nearest neighbors:\n\n    num_items = 1000000\n    num_dimensions = 100\n    dataset = np.random.randn(num_items, num_dimensions)\n    dataset \/= np.linalg.norm(dataset, axis=1).reshape(-1, 1)\n\n    random_index = random.randint(0,num_items)\n    query = dataset[random_index]\n\n## Brute Force\n\nFirst, we calculate the time for searching with the brute-force algorithm. It goes through the entire data serially, calculating the distance between the query and current item one at a time. We use the `timeit` command for calculating the time. First, we create the search index to retrieve the five nearest neighbors and then search with a query:\n\n    neighbors = NearestNeighbors(n_neighbors=5, algorithm='brute',\n    metric='euclidean').fit(dataset)\n    %timeit distances, indices = neighbors.kneighbors([query])\n\n    > 177 ms \u00b1 136 \u03bcs per loop (mean \u00b1 std. dev. of 7 runs, 1000 loops each)\n\n###### Tip\n\nThe `timeit` command is a handy tool. To benchmark the time of a single operation, prefix it with this command. Compared to the time command, which runs a statement for one time, `timeit` runs the subsequent line multiple times to give more precise aggregated statistics along with the standard deviation. By default, it turns off garbage collection, making independent timings more comparable. That said, this might not reflect timings in real production loads where garbage collection is turned on.\n\n## Annoy\n\nAnnoy (Approximate Nearest Neighbors Oh Yeah) is a C++ library with Python bindings for searching nearest neighbors. Synonymous with speed, it was released by Spotify and is used in production to serve its music recommendations. In contrast to its name, it's actually fun and easy to use.\n\nTo use Annoy, we install it using `pip`:\n\n    $ pip install annoy\n\nIt's fairly straightforward to use. First, we build a search index with two hyperparameters: the number of dimensions of the dataset and the number of trees:\n\n    from annoy import AnnoyIndex\n    annoy_index = AnnoyIndex(num_dimensions) _# Length of item vector that will be\n    indexed_\n    for i in range(num_items):\n        annoy_index.add_item(i, dataset[i])\n    annoy_index.build(40) _# 40 trees_\n\nNow let's find out the time it takes to search the five nearest neighbors of one image:\n\n    %timeit indexes=t.get_nns_by_vector(query, 5, include_distances=True)\n\n    > 34.9 \u03bcs \u00b1 165 ns per loop (mean \u00b1 std. dev. of 7 runs, 10000 loops each)\n\nNow that is blazing fast! To put this in perspective, even for our million-item dataset, this can serve almost 28,000 requests on a single CPU core. Considering most CPUs have multiple cores, it should be able to handle more than 100,000 requests on a single system. The best part is that it lets you share the same index in memory between multiple processes. Thus, the biggest index can be equivalent to the size of your overall RAM, making it possible to serve multiple requests on a single system.\n\nOther benefits include that it generates a modestly sized index. Moreover, it decouples creating indexes from loading them, so you can create an index on one machine, pass it around, and then on your serving machine load it in memory and serve it.\n\n###### Tip\n\nWondering about how many trees to use? More trees give higher precision, but larger indexes. Usually, no more than 50 trees are required to attain the highest precision.\n\n## NGT\n\nYahoo Japan's Neighborhood Graph and Tree (NGT) library currently leads most benchmarks and is best suited for large datasets (in millions of items) with large dimensions (in several thousands). Although the library has existed since 2016, its real entry into the industry benchmark scene happened in 2018 with the implementation of the ONNG algorithm (short for Optimization of indexing based on _k_ -Nearest Neighbor Graph for proximity). Considering multiple threads might be running NGT on a server, it can place the index in shared memory with the help of memory mapped files, helping to reduce memory usage as well as increase load time.\n\n## Faiss\n\nFaiss is Facebook's efficient similarity search library. It can scale to billions of vectors in RAM on a single server by storing a compressed representation of the vectors (compact quantization codes) instead of the original values. It's especially suited for dense vectors. It shines particularly on machines with GPUs by storing the index on GPU memory (VRAM). This works on both single-GPU and multi-GPU setups. It provides the ability to configure performance based on search time, accuracy, memory usage, and indexing time. It's one of the fastest known implementations of ANN search on GPU. Hey, if it's good enough for Facebook, it's good enough for most of us (as long as we have enough data).\n\nWhile showing the entire process is beyond the scope of this book, we recommend installing Faiss using Anaconda or using its Docker containers to quickly get started.\n\n# Improving Accuracy with Fine Tuning\n\nMany of the pretrained models were trained on the ImageNet dataset. Therefore, they provide an incredible starting point for similarity computations in most situations. That said, if you tuned these models to adapt to your specific problem, they would perform even more accurately at finding similar images.\n\nIn this portion of the chapter, we identify the worst-performing categories, visualize them with t-SNE, fine tune, and then see how their t-SNE graph changes.\n\nWhat is a good metric to check whether you are indeed getting similar images?\n\nPainful option 1\n\nGo through the entire dataset one image at a time, and manually score whether the returned images indeed look similar.\n\nHappier option 2\n\nSimply calculate accuracy. That is, for an image belonging to category _X_ , are the similar images belonging to the same category? We will refer to this similarity accuracy.\n\nSo, what are our worst-performing categories? And why are they the worst? To answer this, we have predefined a helper function `worst_classes`. For every image in the dataset, it finds the nearest neighbors using the brute-force algorithm and then returns six classes with the least accuracy. To see the effects of fine tuning, we run our analysis on a more difficult dataset: Caltech-256. Calling this function unveils the least-accurate classes:\n\n    names_of_worst_classes_before_finetuning, accuracy_per_class_before_finetuning =\n    worst_classes(feature_list[:])\n\n        Accuracy is 56.54\n        Top 6 incorrect classifications\n        059.drinking-straw    Accuracy: 11.76%\n        135.mailbox             Accuracy: 16.03%\n        108.hot-dog             Accuracy: 16.72%\n        163.playing-card      Accuracy: 17.29%\n        195.soda-can            Accuracy: 19.68%\n        125.knife         Accuracy: 20.53%\n\nTo see why they are performing so poorly on certain classes, we've plotted a t-SNE graph to visualize the embeddings in 2D space, which you can see in Figure 4-16. To prevent overcrowding on our plot, we use only 50 items from each of the 6 classes.\n\n###### Tip\n\nTo enhance the visibility of the graph we can define different markers and different colors for each class. Matplotlib provides a wide variety of markers and colors.\n\n    markers = [ \"^\", \".\",\"s\", \"o\",\"x\", \"P\" ]\n    colors = ['red', 'blue', 'fuchsia', 'green', \n    'purple', 'orange']\n\n###### Figure 4-16. t-SNE visualization of feature vectors of least-accurate classes before fine tuning\n\nAah, these feature vectors are all over the place and on top of one another. Using these feature vectors in other applications such as classification might not be a good idea because it would be difficult to find a clean plane of separation between them. No wonder they performed so poorly in this nearest neighbor\u2013based classification test.\n\nWhat do you think will be the result if we repeat these steps with the fine-tuned model? We reckon something interesting; let's take a look at Figure 4-17 to see.\n\n###### Figure 4-17. t-SNE visualization of feature vectors of least-accurate classes after fine tuning\n\nThis is so much cleaner. With just a little bit of fine tuning as shown in Chapter 3, the embeddings begin to group together. Compare the noisy\/scattered embeddings of the pretrained models against those of the fine-tuned model. A machine learning classifier would be able to find a plane of separation between these classes with much more ease, hence yielding better classification accuracy as well as more similar images when not using a classifier. And, remember, these were the classes with the highest misclassifications; imagine how nicely the classes with originally higher accuracy would be after fine tuning.\n\nPreviously, the pretrained embeddings achieved 56% accuracy. The new embeddings after fine tuning deliver a whopping 87% accuracy! A little magic goes a long way.\n\nThe one limitation for fine tuning is the requirement of labeled data, which is not always present. So depending on your use case, you might need to label some amount of data.\n\nThere's a small unconventional training trick involved, though, which we discuss in the next section.\n\n## Fine Tuning Without Fully Connected Layers\n\nAs we already know, a neural network comprises three parts:\n\n  * Convolutional layers, which end up generating the feature vectors\n\n  * Fully connected layers\n\n  * The final classifier layer\n\nFine tuning, as the name suggests, involves tweaking a neural network lightly to adapt to a new dataset. It usually involves stripping off the fully connected layers (top layers), substituting them with new ones, and then training this newly composed neural network using this dataset. Training in this manner will cause two things:\n\n  * The weights in all the newly added fully connected layers will be significantly affected.\n\n  * The weights in the convolutional layers will be only slightly changed.\n\nThe fully connected layers do a lot of the heavy lifting to get maximum classification accuracy. As a result, the majority of the network that generates the feature vectors will change insignificantly. Thus, the feature vectors, despite fine tuning, will show little change.\n\nOur aim is for similar-looking objects to have closer feature vectors, which fine tuning as described earlier fails to accomplish. By forcing all of the task-specific learning to happen in the convolutional layers, we can see much better results. How do we achieve that? _By removing all of the fully connected layers and placing a classifier layer directly after the convolutional layers (which generate the feature vectors)._ This model is optimized for similarity search rather than classification.\n\nTo compare the process of fine tuning a model optimized for classification tasks as opposed to similarity search, let's recall how we fine tuned our model in Chapter 3 for classification:\n\n    from tf.keras.applications.resnet50 import ResNet50\n    model = ResNet50(weights='imagenet', include_top=False,\n    input_shape = (224,224,3))\n    input = Input(shape=(224, 224, 3))\n    x = model(input)\n    x = GlobalAveragePooling2D()(x)\n    x = Dense(64, activation='relu')(x)\n    x = Dropout(0.5)(x)\n    x = Dense(NUM_CLASSES, activation='softmax')(x)\n    model_classification_optimized = Model(inputs=input, outputs=x)\n\nAnd here's how we fine tune our model for similarity search. Note the missing hidden dense layer in the middle:\n\n    from tf.keras.applications.resnet50 import ResNet50\n    model = ResNet50(weights='imagenet', include_top=False,\n    input_shape = (224,224,3))\n    input = Input(shape=(224, 224, 3))\n    x = model(input)\n    x = GlobalAveragePooling2D()(x)\n    _# No dense or dropout layers_\n    x = Dense(NUM_CLASSES, activation='softmax')(x)\n    model_similarity_optimized = Model(inputs=input, outputs=x)\n\nAfter fine tuning, to use the `model_similarity_optimized` for extracting features instead of giving probabilities for classes, simply `pop` (i.e., remove) the last layer:\n\n    model_similarity_optimized.layers.pop()\n    model = Model(model_similarity_optimized.input,\n    model_similarity_optimized.layers[-1].output)\n\nThe key thing to appreciate here is if you used the regular fine-tuning process, we would get lower similarity accuracy than `model_similarity_optimized`. Obviously, we would want to use `model_classification_optimized` for classification scenarios and `model_similarity_optimized` for extracting embeddings for similarity search.\n\nWith all this knowledge, you can now make both a fast and accurate similarity system for any scenario you are working on. It's time to see how the giants in the AI industry build their products.\n\n# Siamese Networks for One-Shot Face Verification\n\nA face verification system is usually trying to ascertain\u2014given two images of faces\u2014whether the two images are of the same person. This is a high-precision binary classifier that needs to robustly work with different lighting, clothing, hairstyles, backgrounds, and facial expressions. To make things more challenging, although there might be images of many people in, for instance an employee database, there might be only a handful of images of the same person available. Similarly, signature identification in banks and product identification on Amazon suffer the same challenge of limited images per item.\n\nHow would you go about training such a classifier? Picking embeddings from a model like ResNet pretrained on ImageNet might not discern these fine facial attributes. One approach is to put each person as a separate class and then train like we usually train a regular network. Two key issues arise:\n\n  * If we had a million individuals, training for a million categories is not feasible.\n\n  * Training with a few images per class will lead to overtraining.\n\nAnother thought: instead of teaching different categories, we could teach a network to directly compare and decide whether a pair of images are similar or dissimilar by giving guidance on their similarity during training. And this is the key idea behind Siamese networks. Take a model, feed in two images, extract two embeddings, and then calculate the distance between the two embeddings. If the distance is under a threshold, consider them similar, else not. By feeding a pair of images with the associated label, similar or dissimilar, and training the network end to end, the embeddings begin to capture the fine-grained representation of the inputs. This approach, shown in Figure 4-18, of directly optimizing for the distance metric is called _metric learning_.\n\n###### Figure 4-18. A Siamese network for signature verification; note that the same CNN was used for both input images\n\nWe could extend this idea and even feed three images. Pick one anchor image, pick another positive sample (of the same category), and another negative sample (of a different category). Let's now train this network to directly optimize for the distance between similar items to be minimized and the distance between dissimilar items to be maximized. This loss function that helps us achieve this is called a _triplet loss_ function. In the previous case with a pair of images, the loss function is called a _contrastive loss_ function. The triplet loss function tends to give better results.\n\nAfter the network is trained, we need only one reference image of a face for deciding at test time whether the person is the same. This methodology opens the doors for _one-shot learning_. Other common uses include signature and logo recognition. One remarkably creative application by Saket Maheshwary and Hemant Misra is to use a Siamese network for matching r\u00e9sum\u00e9s with job applicants by calculating the semantic similarity between the two.\n\n# Case Studies\n\nLet's look at a few interesting examples that show how what we have learned so far is applied in the industry.\n\n## Flickr\n\nFlickr is one of the largest photo-sharing websites, especially popular among professional photographers. To help photographers find inspiration as well as showcase content the users might find interesting, Flickr produced a similarity search feature based on the same semantic meaning. As demonstrated in Figure 4-19, exploring a desert pattern leads to several similarly patterned results. Under the hood, Flickr adopted an ANN algorithm called Locally Optimized Product Quantization (LOPQ), which has been open sourced in Python as well as Spark implementations.\n\n###### Figure 4-19. Similar patterns of a desert photo (image source)\n\n## Pinterest\n\nPinterest is an application used widely for its visual search capabilities, more specifically in its features called Similar Pins and Related Pins. Other companies like Baidu and Alibaba have launched similar visual search systems. Also, Zappos, Google Shopping, and like.com are using computer vision for recommendation.\n\nWithin Pinterest \"women's fashion\" is one of the most popular themes of pins and the Similar Looks feature (Figure 4-20) helps people discover similar products. Additionally, Pinterest also reports that its Related Pins feature increased its repin rate. Not every pin on Pinterest has associated metadata, which makes recommendation a difficult cold-start problem due to lack of context. Pinterest developers solved this cold-start problem by using the visual features for generating the related pins. Additionally, Pinterest implements an incremental fingerprinting service that generates new digital signatures if either a new image is uploaded or if there is feature evolution (due to improvements or modifications in the underlying models by the engineers).\n\n###### Figure 4-20. The Similar Looks feature of the Pinterest application (image source: Pinterest blog)\n\n## Celebrity Doppelgangers\n\nWebsite applications like _Celebslike.me_ , which went viral in 2015, look for the nearest neighbor among celebrities, as shown in Figure 4-21. A similar viral approach was taken by the Google Arts & Culture app in 2018, which shows the nearest existing portrait to your face. Twins or not is another application with a similar aim.\n\n###### Figure 4-21. Testing our friend Pete Warden's photo (technical lead for mobile and embedded TensorFlow at Google) on the celebslike.me website\n\n## Spotify\n\nSpotify uses nearest neighbors for recommending music and creating automatic playlists and radio stations based on the current set of songs being played. Usually, collaborative filtering techniques, which are employed for recommending content like movies on Netflix, are content agnostic; that is, the recommendation happens because large groups of users with similar tastes are watching similar movies or listening to similar songs. This presents a problem for new and not yet popular content because users will keep getting recommendations for existing popular content. This is also referred to as the aforementioned cold-start problem. The solution is to use the latent understanding of the content. Similar to images, we can create feature vectors out of music using MFCC features (Mel Frequency Cepstral Coefficients), which in turn generates a 2D spectrogram that can be thought of as an image and can be used to generate features. Songs are divided into three-second fragments, and their spectrograms are used to generate features. These features are then averaged together to represent the complete song. Figure 4-22 shows artists whose songs are projected in specific areas. We can discern hip-hop (upper left), rock (upper right), pop (lower left), and electronic music (lower right). As already discussed, Spotify uses Annoy in the background.\n\n###### Figure 4-22. t-SNE visualization of the distribution of predicted usage patterns, using latent factors predicted from audio (image source: \"Deep content-based music recommendation\" by Aaron van den Oord, Sander Dieleman, Benjamin Schrauwen, NIPS 2013)\n\n## Image Captioning\n\nImage captioning is the science of translating an image into a sentence (as illustrated in Figure 4-23). Going beyond just object tagging, this requires a deeper visual understanding of the entire image and relationships between objects. To train these models, an open source dataset called MS COCO was released in 2014, which consists of more than 300,000 images along with object categories, sentence descriptions, visual question-answer pairs, and object segmentations. It serves as a benchmark for a yearly competition to see progress in image captioning, object detection, and segmentation.\n\n###### Figure 4-23. Image captioning feature in Seeing AI: the Talking Camera App for the blind community\n\nA common strategy applied in the first year of the challenge (2015) was to append a language model (LSTM\/RNN) with a CNN in such a way that the output of a CNN feature vector is taken as the input to the language model (LSTM\/RNN). This combined model was trained jointly in an end-to-end manner, leading to very impressive results that stunned the world. Although every research lab was trying to beat one another, it was later found that doing a simple nearest-neighbor search could yield state-of-the-art results. For a given image, find similar images based on similarity of the embeddings. Then, note the common words in the captions of the similar images, and print the caption containing the most common words. In short, a lazy approach would still beat the state-of-the-art one, and this exposed a critical bias in the dataset.\n\nThis bias has been coined the _Giraffe-Tree_ problem by Larry Zitnick. Do an image search for \"giraffe\" on a search engine. Look closely: in addition to giraffe, is there grass in almost every image? Chances are you can describe the majority of these images as \"A giraffe standing in a grass field.\" Similarly, if a query image like the photo on the far left in Figure 4-24 contains a giraffe and a tree, almost all similar images (right) can be described as \"a giraffe standing in the grass, next to a tree.\" Even without a deeper understanding of the image, one would arrive at the correct caption using a simple nearest-neighbor search. This shows that to measure the real intelligence of a system, we need more semantically novel\/original images in the test set.\n\n###### Figure 4-24. The Giraffe-Tree problem (image source: Measuring Machine Intelligence Through Visual Question Answering, C. Lawrence Zitnick, Aishwarya Agrawal, Stanislaw Antol, Margaret Mitchell, Dhruv Batra, Devi Parikh)\n\nIn short, don't underestimate a simple nearest-neighbor approach!\n\n# Summary\n\nNow we are at the end of a successful expedition where we explored locating similar images with the help of embeddings. We took this one level further by exploring how to scale searches from a few thousand to a few billion documents with the help of ANN algorithms and libraries including Annoy, NGT, and Faiss. We also learned that fine tuning the model to your dataset can improve the accuracy and representative power of embeddings in a supervised setting. To top it all off, we looked at how to use Siamese networks, which use the power of embeddings to do one-shot learning, such as for face verification systems. We finally examined how nearest-neighbor approaches are used in various use cases across the industry. Nearest neighbors are a simple yet powerful tool to have in your toolkit.\n\n# Chapter 5. From Novice to Master Predictor: Maximizing Convolutional Neural Network Accuracy\n\nIn Chapter 1, we looked at the importance of responsible AI development. One of the aspects we discussed was the importance of robustness of our models. Users can trust what we build only if they can be assured that the AI they encounter on a day-to-day basis is accurate and reliable. Obviously, the context of the application matters a lot. It would be okay for a food classifier to misclassify pasta as bread on occasion. But it would be dangerous for a self-driving car to misclassify a pedestrian as a street lane. The main goal of this chapter is thus a rather important one\u2014to build more accurate models.\n\nIn this chapter, you will develop an intuition for recognizing opportunities to improve your model's accuracy the next time you begin training one. We first look at the tools that will ensure that you won't be going in blind. After that, for a good chunk of this chapter, we take a very experimental approach by setting up a baseline, isolating individual parameters to tweak, and observing their effect on model performance and training speed. A lot of the code we use in this chapter is all aggregated in a single Jupyter Notebook, along with an actionable checklist with interactive examples. It is meant to be highly reusable should you choose to incorporate it in your next training script.\n\nWe explore several questions that tend to come up during model training:\n\n  * I am unsure whether to use transfer learning or building from scratch to train my own network. What is the preferred approach for my scenario?\n\n  * What is the least amount of data that I can supply to my training pipeline to get acceptable results?\n\n  * I want to ensure that the model is learning the correct thing and not picking up spurious correlations. How can I get visibility into that?\n\n  * How can I ensure that I (or someone else) will obtain the same results from my experiments every single time they are run? In other words, how do I ensure reproducibility of my experiments?\n\n  * Does changing the aspect ratio of the input images have an impact on the predictions?\n\n  * Does reducing input image size have a significant effect on prediction results?\n\n  * If I use transfer learning, what percentage of layers should I fine tune to achieve my preferred balance of training time versus accuracy?\n\n  * Alternatively, if I were to train from scratch, how many layers should I have in my model?\n\n  * What is the appropriate \"learning rate\" to supply during model training?\n\n  * There are too many things to remember. Is there a way to automate all of this work?\n\nWe will try to answer these questions one by one in the form of experiments on a few datasets. Ideally, you should be able to look at the results, read the takeaways, and gain some insight into the concept that the experiment was testing. If you're feeling more adventurous, you can choose to perform the experiments yourself using the Jupyter Notebook.\n\n# Tools of the Trade\n\nOne of the main priorities of this chapter is to reduce the code and effort involved during experimentation while trying to gain insights into the process in order to reach high accuracy. An arsenal of tools exists that can assist us in making this journey more pleasant:\n\nTensorFlow Datasets\n\nQuick and easy access to around 100 datasets in a performant manner. All well-known datasets are available starting from the smallest MNIST (a few megabytes) to the largest MS COCO, ImageNet, and Open Images (several hundred gigabytes). Additionally, medical datasets like the Colorectal Histology and Diabetic Retinopathy are also available.\n\nTensorBoard\n\nClose to 20 easy-to-use methods to visualize many aspects of training, including visualizing the graph, tracking experiments, and inspecting the images, text, and audio data that pass through the network during training.\n\nWhat-If Tool\n\nRun experiments in parallel on separate models and tease out differences in them by comparing their performance on specific data points. Edit individual data points to see how that affects the model training.\n\ntf-explain\n\nAnalyze decisions made by the network to identify bias and inaccuracies in the dataset. Additionally, use heatmaps to visualize what parts of the image the network activated on.\n\nKeras Tuner\n\nA library built for `tf.keras` that enables automatic tuning of hyperparameters in TensorFlow 2.0.\n\nAutoKeras\n\nAutomates Neural Architecture Search (NAS) across different tasks like image, text, and audio classification and image detection.\n\nAutoAugment\n\nUtilizes reinforcement learning to improve the amount and diversity of data in an existing training dataset, thereby increasing accuracy.\n\nLet's now explore these tools in greater detail.\n\n## TensorFlow Datasets\n\nTensorFlow Datasets is a collection of nearly 100 ready-to-use datasets that can quickly help build high-performance input data pipelines for training TensorFlow models. Instead of downloading and manipulating data sets manually and then figuring out how to read their labels, TensorFlow Datasets standardizes the data format so that it's easy to swap one dataset with another, often with just a single line of code change. As you will see later on, doing things like breaking the dataset down into training, validation, and testing is also a matter of a single line of code. We will additionally be exploring TensorFlow Datasets from a performance point of view in the next chapter.\n\nYou can list all of the available datasets by using the following command (in the interest of conserving space, only a small subset of the full output is shown in this example):\n\n    import tensorflow_datasets as tfds\n    print(tfds.list_builders())\n\n    ['amazon_us_reviews', 'bair_robot_pushing_small', 'bigearthnet', 'caltech101',\n    'cats_vs_dogs', 'celeb_a', 'imagenet2012',... , 'open_images_v4',\n    'oxford_flowers102', 'stanford_dogs','voc2007', 'wikipedia', 'wmt_translate',\n    'xnli']\n\nLet's see how simple it is to load a dataset. We will plug this into a full working pipeline later:\n\n    # Import necessary packages\n    import tensorflow_datasets as tfds\n\n    # Downloading and loading the dataset\n    dataset = tfds.load(name=\"cats_vs_dogs\", split=tfds.Split.TRAIN)\n\n    # Building a performance data pipeline\n    dataset = dataset.map(preprocess).cache().repeat().shuffle(1024).batch(32).\n    prefetch(tf.data.experimental.AUTOTUNE)\n\n    model.fit(dataset, ...)\n\n###### Tip\n\n`tfds` generates a lot of progress bars, and they take up a lot of screen space\u2014using `tfds.disable_progress_bar()` might be a good idea.\n\n## TensorBoard\n\nTensorBoard is a one-stop-shop for all of your visualization needs, offering close to 20 tools to understand, inspect, and improve your model's training.\n\nTraditionally, to track experiment progress, we save the values of loss and accuracy per epoch and then, when done, plot it using `matplotlib`. The downside with that approach is that it's not real time. Our usual options are to watch for the training progress in text. Additionally, after the training is done, we need to write additional code to make the graph in `matplotlib`. TensorBoard solves these and more pressing issues by offering a real-time dashboard (Figure 5-1) that helps us visualize all logs (such as train\/validation accuracy and loss) to assist in understanding the progression of training. Another benefit it offers is the ability to compare our current experiment's progress with the previous experiment, so we can see how a change in parameters affected our overall accuracy.\n\n###### Figure 5-1. TensorBoard default view showcasing real-time training metrics (the lightly shaded lines represent the accuracy from the previous run)\n\nTo enable TensorBoard to visualize our training and models, we need to log information about our training with the help of summary writer:\n\n    summary_writer = tf.summary.FileWriter('.\/logs')\n\nTo follow our training in real time, we need to load TensorBoard before the model training begins. We can load TensorBoard by using the following commands:\n\n    # Get TensorBoard to run\n    %load_ext tensorboard\n\n    # Start TensorBoard\n    %tensorboard --logdir .\/log\n\nAs more TensorFlow components need a visual user interface, they reuse TensorBoard by becoming embeddable plug-ins within it. You'll notice the Inactive drop-down menu on TensorBoard; that's where you can see all the different profiles or tools that TensorFlow offers. Table 5-1 showcases a handful of the wide variety of tools available.\n\nTable 5-1. Plugins for TensorBoard **TensorBoard plug-in name** | **Description**  \n---|---  \nDefault Scalar | Visualize scalar values such as classification accuracy.  \nCustom Scalar | Visualize user-defined custom metrics. For example, different weights for different classes, which might not be a readily available metric.  \nImage | View the output from each layer by clicking the Images tab.  \nAudio | Visualize audio data.  \nDebugging tools | Allows debugging visually and setting conditional breakpoints (e.g., tensor contains Nan or Infinity).  \nGraphs | Shows the model architecture graphically.  \nHistograms | Show the changes in the weight distribution in the layers of a model as the training progresses. This is especially useful for checking the effect of compressing a model with quantization.  \nProjector | Visualize projections using t-SNE, PCA, and others.  \nText | Visualize text data.  \nPR curves | Plot precision-recall curves.  \nProfile | Benchmark speed of all operations and layers in a model.  \nBeholder | Visualize the gradients and activations of a model in real time during training. It allows seeing them filter by filter, and allows them to be exported as images or even as a video.  \nWhat-If Tool | For investigating the model by slicing and dicing the data and checking its performance. Especially helpful for discovering bias.  \nHParams | Find out which params and at what values are the most important, allow logging of the entire parameter server (discussed in detail in this chapter).  \nMesh | Visualize 3D data (including point clouds).\n\nIt should be noted that TensorBoard is not TensorFlow specific, and can be used with other frameworks like PyTorch, scikit-learn, and more, depending on the plugin used. To make a plugin work, we need to write the specific metadata that we want to visualize. For example, TensorBoard embeds the TensorFlow Projector tool within to cluster images, text, or audio using t-SNE (which we examined in detail in Chapter 4). Apart from calling TensorBoard, we need to write the metadata like the feature embeddings of our image, so that TensorFlow Projector can use it to do clustering, as demonstrated in Figure 5-2.\n\n###### Figure 5-2. TensorFlow Embedding Projector showcasing data in clusters (can be run as a TensorBoard plugin)\n\n## What-If Tool\n\nWhat if we could inspect our AI model's predictions with the help of visualizations? What if we could find the best threshold for our model to maximize precision and recall? What if we could slice and dice the data along with the predictions our model made to see what it's great at and where there are opportunities to improve? What if we could compare two models to figure out which is indeed better? What if we could do all this and more, with a few clicks in the browser? Sounds appealing for sure! The What-If Tool (Figure 5-3 and Figure 5-4) from Google's People + AI Research (PAIR) initiative helps open up the black box of AI models to enable model and data explainability.\n\n###### Figure 5-3. What-If Tool's datapoint editor makes it possible to filter and visualize data according to annotations of the dataset and labels from the classifier\n\n###### Figure 5-4. PR curves in the Performance and Fairness section of the What-If Tool help to interactively select the optimal threshold to maximize precision and recall\n\nTo use the What-If Tool, we need the dataset and a model. As we just saw, TensorFlow Datasets makes downloading and loading the data (in the `tfrecord` format) relatively easy. All we need to do is to locate the data file. Additionally, we want to save the model in the same directory:\n\n    # Save model for What If Tool\n    tf.saved_model.save(model, \"\/tmp\/model\/1\/\")\n\nIt's best to perform the following lines of code in a local system rather than a Colab notebook because the integration between Colab and the What-If Tool is still evolving.\n\nLet's start TensorBoard:\n\n    $ mkdir tensorboard\n    $ tensorboard --logdir .\/log --alsologtostderr\n\nNow, in a new terminal, let's make a directory for all of our What-If Tool experiments:\n\n    $ mkdir what-if-stuff\n\nMove the trained model and TFRecord data here. The overall directory structure looks something like this:\n\n    $ tree .\n    \u251c\u2500\u2500 colo\n    \u2502   \u2514\u2500\u2500 model\n    \u2502         \u2514\u2500\u2500 1\n    \u2502         \u251c\u2500\u2500 assets\n    \u2502         \u251c\u2500\u2500 saved_model.pb\n    \u2502         \u2514\u2500\u2500 variables\n\nWe'll serve the model using Docker within the newly created directory:\n\n    $ sudo docker run -p 8500:8500 \\\n    --mount type=bind,source=\/home\/{ _your_username_ }\/what-if-stuff\/colo\/model\/,\n     target=\/models\/colo \\\n    -e MODEL_NAME=colo -t tensorflow\/serving\n\nA word of caution: the port must be `8500` and all parameters must be spelled exactly as shown in the preceding example.\n\nNext, at the far right, click the settings button (the gray gear icon) and add the values listed in Table 5-2.\n\nTable 5-2. Configurations for the What-If Tool **Parameter** | **Value**  \n---|---  \nInference address |\n\n`ip_addr:8500`\n\nModel name |\n\n`\/models\/colo`\n\nModel type | Classification  \nPath to examples | _\/home\/{ _`your_username`_ }\/what_if_stuff\/colo\/models\/colo.tfrec_ (Note: this must be an absolute path)\n\nWe can now open the What-If Tool in the browser within TensorBoard, as depicted in Figure 5-5.\n\n###### Figure 5-5. Setup window for the What-If Tool\n\nThe What-If Tool can also be used to visualize datasets according to different bins, as shown in Figure 5-6. We can also use the tool to determine the better performing model out of multiple models on the same dataset using the `set_compare_estimator_and_feature_spec` function.\n\n    from witwidget.notebook.visualization import WitConfigBuilder\n\n    _# features are the test examples that we want to load into the tool_\n    models = [model2, model3, model4]\n    config_builder =\n    WitConfigBuilder(test_examples).set_estimator_and_feature_spec(model1, features)\n\n    for each_model in models:\n        config_builder =\n     config_builder.set_compare_estimator_and_feature_spec(each_model, features)\n\n###### Figure 5-6. The What-If tool enables using multiple metrics, data visualization, and many more things under the sun\n\nNow, we can load TensorBoard, and then, in the Visualize section, choose the model we want to compare, as shown in Figure 5-7. This tool has many features to explore!\n\n###### Figure 5-7. Choose the model to compare using the What-If Tool\n\n## tf-explain\n\nDeep learning models have traditionally been black boxes, and up until now, we usually learn about their performance by watching the class probabilities and validation accuracies. To make these models more interpretable and explainable, heatmaps come to the rescue. By showing the area of an image that leads to the prediction with higher intensity, heatmaps can help visualize their learning. For example, an animal often seen in surroundings with snow might be getting high-accuracy predictions, but if the dataset has only that animal with snow as the background, the model might just be paying attention to the snow as the distinctive pattern instead of the animal. Such a dataset demonstrates bias, making the predictions not too robust when the classifier is put in the real world (and potentially dangerous!). Heatmaps can be especially useful to explore such bias, as often spurious correlations can seep in if the dataset is not carefully curated.\n\n`tf-explain` (by Raphael Meudec) helps understand the results and inner workings of a neural network with the help of such visualizations, removing the veil on bias in datasets. We can add multiple types of callbacks while training or use its core API to generate TensorFlow events that can later be loaded into TensorBoard. For inference, all we need to do is pass an image, its ImageNet object ID along with a model into tf-explain's functions. You must supply the object ID because `tf.explain` needs to know what is activated for that particular class. A few different visualization approaches are available with `tf.explain`:\n\nGrad CAM\n\nThe Gradient-weighted Class Activation Mapping (Grad CAM) visualizes how parts of the image affect the neural network's output by looking into the activation maps. A heatmap (illustrated in Figure 5-8) is generated based on the gradients of the object ID from the last convolutional layer. Grad CAM is largely a broad-spectrum heatmap generator given that it is robust to noise and can be used on an array of CNN models.\n\nOcclusion Sensitivity\n\nOccludes a part of the image (using a small square patch placed randomly) to establish how robust the network is. If the prediction is still correct, on average, the network is robust. The area in the image that is the warmest (i.e., red) has the most effect on the prediction when occluded.\n\nActivations\n\nVisualizes the activations for the convolutional layers.\n\n###### Figure 5-8. Visualizations on images using MobileNet and tf-explain\n\nAs demonstrated in the code example that follows, such visualizations can be built with very little code. By taking a video, generating individual frames, and running tf-explain with Grad CAM and joining them together, we can build a detailed understanding of how these neural networks would react to moving camera angles.\n\n    from tf_explain.core.grad_cam import GradCAM\n    From tf.keras.applications.MobileNet import MobileNet\n\n    model = MobileNet(weights='imagenet', include_top=True)\n\n    # Set Grad CAM System\n    explainer = GradCAM()\n\n    # Image Processing\n    IMAGE_PATH = 'dog.jpg'\n    dog_index = 263\n    img = tf.keras.preprocessing.image.load_img(IMAGE_PATH, target_size=(224, 224))\n    img = tf.keras.preprocessing.image.img_to_array(img)\n    data = ([img], None)\n\n    # Passing the image through Grad CAM\n    grid = explainer.explain(data, model, 'conv1', index)\n    name = IMAGE_PATH.split(\".jpg\")[0]\n    explainer.save(grid, '\/tmp', name + '_grad_cam.png')\n\n# Common Techniques for Machine Learning Experimentation\n\nThe first few chapters focused on training the model. The following sections, however, contain a few more things to keep in the back of your mind while running your training experiments.\n\n## Data Inspection\n\nData inspection's first biggest hurdle is determining the structure of the data. TensorFlow Datasets has made this step relatively easy because all of the available datasets are in the same format and structure and can be used in a performant way. All we need to do is load the dataset into the What-If Tool and use the various options already present to inspect the data. As an example, on the SMILE dataset, we can visualize the dataset according to its annotations, such as images of people wearing eyeglasses and those without eyeglasses, as illustrated in Figure 5-9. We observe that a wider distribution of the dataset has images of people wearing no eyeglasses, thus uncovering bias in the data due to an unbalanced dataset. This can be solved by modifying the weights of the metrics accordingly, through the tool.\n\n###### Figure 5-9. Slicing and dividing the data based on predictions and real categories\n\n## Breaking the Data: Train, Validation, Test\n\nSplitting a dataset into train, validation, and test is pretty important because we want to report the results on an unseen dataset by the classifier (i.e., the test dataset). TensorFlow Datasets makes it easy to download, load, and split the dataset into these three parts. Some datasets already come with three default splits. Alternatively, the data can be split by percentages. The following code showcases using a default split:\n\n    dataset_name = \"cats_vs_dogs\"\n    train, info_train = tfds.load(dataset_name, split=tfds.Split.TRAIN,\n                        with_info=True)\n\nThe cats-and-dogs dataset in `tfds` has only the train split predefined. Similar to this, some datasets in TensorFlow Datasets do not have a `validation` split. For those datasets, we take a small percentage of samples from the predefined `training` set and treat it as the `validation` set. To top it all off, splitting the dataset using the `weighted_splits` takes care of randomizing and shuffling data between the splits:\n\n    # Load the dataset\n    dataset_name = \"cats_vs_dogs\"\n\n    # Dividing data into train (80), val (10) and test (10)\n    split_train, split_val, split_test = tfds.Split.TRAIN.subsplit(weighted=[80, 10,\n                                         10])\n    train, info_train = tfds.load(dataset_name, split=split_train , with_info=True)\n    val, info_val = tfds.load(dataset_name, split=split_val, with_info=True)\n    test, info_test = tfds.load(dataset_name, split=split_test, with_info=True)\n\n## Early Stopping\n\nEarly stopping helps to avoid overtraining of the network by keeping a lookout for the number of epochs that show limited improvement. Assuming a model is set to train for 1,000 epochs and reaches 90% accuracy at the 10th epoch and stops improving any further for the next 10 epochs, it might be a waste of resources to train any further. If the number of epochs exceeds a predefined threshold called `patience`, training is stopped even if there might still be more epochs left to train. In other words, early stopping decides the point at which the training would no longer be useful and stops training. We can change the metric using the `monitor` parameter and add early stopping to our list of callbacks for the model:\n\n    # Define Early Stopping callback\n    earlystop_callback = tf.keras.callbacks.EarlyStopping(monitor='val_acc',\n    \t\t\t\t\t min_delta=0.0001, patience=10)\n\n    # Add to the training model\n    model.fit_generator(... callbacks=[earlystop_callback])\n\n## Reproducible Experiments\n\nTrain a network once. Then, train it again, without changing any code or parameters. You might notice that the accuracies in two subsequent runs came out slightly different, even if no change was made in code. This is due to random variables. To make experiments reproducible across runs, we want to control this randomization. Initialization of weights of models, randomized shuffling of data, and so on all utilize randomization algorithms. We know that random number generators can be made reproducible by initializing a seed and that's exactly what we will do. Various frameworks have their own ways of setting a random seed, some of which are shown here:\n\n    # Seed for Tensorflow\n    tf.random.set_seed(1234)\n\n    # Seed for Numpy\n    import numpy as np\n    np.random.seed(1234)\n\n    # Seed for Keras\n    seed = 1234\n    fit(train_data, augment=True, seed=seed)\n    flow_from_dataframe(train_dataframe, shuffle=True, seed=seed)\n\n###### Note\n\nIt is necessary to set a seed in all the frameworks and subframeworks that are being used, as seeds are not transferable between frameworks.\n\n# End-to-End Deep Learning Example Pipeline\n\nLet's combine several tools and build a skeletal backbone, which will serve as our pipeline in which we will add and remove parameters, layers, functionality, and various other addons to really understand what is happening. Following the code on the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_), you can interactively run this code for more than 100 datasets in your browser with Colab. Additionally, you can modify it for most classification tasks.\n\n## Basic Transfer Learning Pipeline\n\nFirst, let's build this end-to-end example for transfer learning.\n\n    # Import necessary packages\n    import tensorflow as tf\n    import tensorflow_datasets as tfds\n\n    # tfds makes a lot of progress bars, which takes up a lot of screen space, hence\n    # disabling them\n    tfds.disable_progress_bar()\n\n    tf.random.set_seed(1234)\n\n    # Variables\n    BATCH_SIZE = 32\n    NUM_EPOCHS= 20\n    IMG_H = IMG_W = 224\n    IMG_SIZE = 224\n    LOG_DIR = '.\/log'\n    SHUFFLE_BUFFER_SIZE = 1024\n    IMG_CHANNELS = 3\n\n    dataset_name = \"oxford_flowers102\"\n\n    def preprocess(ds):\n      x = tf.image.resize_with_pad(ds['image'], IMG_SIZE, IMG_SIZE)\n      x = tf.cast(x, tf.float32)\n      x = (x\/127.5) - 1\n      return x, ds['label']\n\n    def augmentation(image,label):\n      image = tf.image.random_brightness(image, .1)\n      image = tf.image.random_contrast(image, lower=0.0, upper=1.0)\n      image = tf.image.random_flip_left_right(image)\n      return image, label\n\n    def get_dataset(dataset_name):\n      split_train, split_val = tfds.Split.TRAIN.subsplit(weighted=[9,1])\n      train, info_train = tfds.load(dataset_name, split=split_train , with_info=True)\n      val, info_val = tfds.load(dataset_name, split=split_val, with_info=True)\n      NUM_CLASSES = info_train.features['label'].num_classes\n      assert NUM_CLASSES >= info_val.features['label'].num_classes\n      NUM_EXAMPLES = info_train.splits['train'].num_examples * 0.9\n      IMG_H, IMG_W, IMG_CHANNELS = info_train.features['image'].shape\n      train = train.map(preprocess).cache().\n              repeat().shuffle(SHUFFLE_BUFFER_SIZE).batch(BATCH_SIZE)\n      train = train.map(augmentation)\n      train = train.prefetch(tf.data.experimental.AUTOTUNE)\n      val = val.map(preprocess).cache().repeat().batch(BATCH_SIZE)\n      val = val.prefetch(tf.data.experimental.AUTOTUNE)\n      return train, info_train, val, info_val, IMG_H, IMG_W, IMG_CHANNELS,\n             NUM_CLASSES, NUM_EXAMPLES\n\n    train, info_train, val, info_val, IMG_H, IMG_W, IMG_CHANNELS, NUM_CLASSES,\n    NUM_EXAMPLES = get_dataset(dataset_name)\n\n    # Allow TensorBoard callbacks\n    tensorboard_callback = tf.keras.callbacks.TensorBoard(LOG_DIR,\n                                                          histogram_freq=1,\n                                                          write_graph=True,\n                                                          write_grads=True,\n                                                          batch_size=BATCH_SIZE,\n                                                          write_images=True)\n\n    def transfer_learn(train, val, unfreeze_percentage, learning_rate):\n       mobile_net = tf.keras.applications.ResNet50(input_shape=(IMG_SIZE, IMG_SIZE,\n                    IMG_CHANNELS), include_top=False)\n       mobile_net.trainable=False\n       # Unfreeze some of the layers according to the dataset being used\n       num_layers = len(mobile_net.layers)\n       for layer_index in range(int(num_layers - unfreeze_percentage*num_layers),\n                                 num_layers ):\n       \t\tmobile_net.layers[layer_index].trainable = True\n       model_with_transfer_learning = tf.keras.Sequential([mobile_net,\n                              tf.keras.layers.GlobalAveragePooling2D(),\n                              tf.keras.layers.Flatten(),\n                              tf.keras.layers.Dense(64),\n                              tf.keras.layers.Dropout(0.3),\n                              tf.keras.layers.Dense(NUM_CLASSES, \n                                                    activation='softmax')],)\n      model_with_transfer_learning.compile(\n                   optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),\n                       loss='sparse_categorical_crossentropy',\n                       metrics=[\"accuracy\"])\n      model_with_transfer_learning.summary()\n      earlystop_callback = tf.keras.callbacks.EarlyStopping(\n                                       monitor='val_accuracy', \n                                       min_delta=0.0001, \n                                       patience=5)\n      model_with_transfer_learning.fit(train,\n                                       epochs=NUM_EPOCHS,\n                                       steps_per_epoch=int(NUM_EXAMPLES\/BATCH_SIZE),\n                                       validation_data=val,\n                                       validation_steps=1,\n                                       validation_freq=1,\n                                       callbacks=[tensorboard_callback,\n                                                  earlystop_callback])\n      return model_with_transfer_learning\n\n    # Start TensorBoard\n    %tensorboard --logdir .\/log\n\n    # Select the last % layers to be trained while using the transfer learning\n    # technique. These layers are the closest to the output layers.\n    unfreeze_percentage = .33\n    learning_rate = 0.001\n\n    model = transfer_learn(train, val, unfreeze_percentage, learning_rate)\n\n## Basic Custom Network Pipeline\n\nApart from transfer learning on state-of-the-art models, we can also experiment and develop better intuitions by building our own custom network. Only the model needs to be swapped in the previously defined transfer learning code:\n\n    def create_model():\n      model = tf.keras.Sequential([\n         tf.keras.layers.Conv2D(32, (3, 3), activation='relu',\n              input_shape=(IMG_SIZE, IMG_SIZE, IMG_CHANNELS)),\n         tf.keras.layers.MaxPool2D(pool_size=(2, 2)),\n         tf.keras.layers.Conv2D(32, (3, 3), activation='relu'),\n         tf.keras.layers.MaxPool2D(pool_size=(2, 2)),\n         tf.keras.layers.Conv2D(32, (3, 3), activation='relu'),\n         tf.keras.layers.MaxPool2D(pool_size=(2, 2)),\n         tf.keras.layers.Dropout(rate=0.3),\n         tf.keras.layers.Flatten(),\n         tf.keras.layers.Dense(128, activation='relu'),\n         tf.keras.layers.Dropout(rate=0.3),\n         tf.keras.layers.Dense(NUM_CLASSES, activation='softmax')\n      ])\n      return model\n\n    def scratch(train, val, learning_rate):\n      model = create_model()\n      model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),\n                     loss='sparse_categorical_crossentropy',\n                     metrics=['accuracy'])\n\n      earlystop_callback = tf.keras.callbacks.EarlyStopping(\n                     monitor='val_accuracy', \n                     min_delta=0.0001, \n                     patience=5)\n\n      model.fit(train,\n               epochs=NUM_EPOCHS,\n               steps_per_epoch=int(NUM_EXAMPLES\/BATCH_SIZE),\n               validation_data=val, \n               validation_steps=1,\n               validation_freq=1,\n               callbacks=[tensorboard_callback, earlystop_callback])\n      return model\n\nNow, it's time to use our pipeline for various experiments.\n\n# How Hyperparameters Affect Accuracy\n\nIn this section, we aim to modify various parameters of a deep learning pipeline one at a time\u2014from the number of layers fine-tuned, to the choice of the activation function used\u2014and see its effect primarily on validation accuracy. Additionally, when relevant, we also observe its effect on the speed of training and time to reach the best accuracy (i.e., convergence).\n\nOur experimentation setup is as follows:\n\n  * To reduce experimentation time, we have used a faster architecture\u2014MobileNet\u2014in this chapter.\n\n  * We reduced the input image resolution to 128 x 128 pixels to further speed up training. In general, we would recommend using a higher resolution (at least 224 x 224) for production systems.\n\n  * Early stopping is applied to stop experiments if they don't increase in accuracy for 10 consecutive epochs.\n\n  * For training with transfer learning, we generally unfreeze the last 33% of the layers.\n\n  * Learning rate is set to 0.001 with Adam optimizer.\n\n  * We're mostly using the Oxford Flowers 102 dataset for testing, unless otherwise stated. We chose this dataset because it is reasonably difficult to train on due to the large number of classes it contains (102) and the similarities between many of the classes that force networks to develop a fine-grained understanding of features in order to do well.\n\n  * To make apples-to-apples comparisons, we take the maximum accuracy value in a particular experiment and normalize all other accuracy values within that experiment with respect to this maximum value.\n\nBased on these and other experiments, we have compiled a checklist of actionable tips to implement in your next model training adventure. These are available on the book's GitHub (see _http:\/\/PracticalDeepLearning.ai_) along with interactive visualizations. If you have more tips, feel free to tweet them @PracticalDLBook or submit a pull request.\n\n## Transfer Learning Versus Training from Scratch\n\nExperimental setup\n\nTrain two models: one using transfer learning, and one from scratch on the same dataset.\n\nDatasets used\n\nOxford Flowers 102, Colorectal Histology\n\nArchitectures used\n\nPretrained MobileNet, Custom model\n\nFigure 5-10 shows the results.\n\n###### Figure 5-10. Comparing transfer learning versus training a custom model on different datasets\n\nHere are the key takeaways:\n\n  * Transfer learning leads to a quicker rise in accuracy during training by reusing previously learned features.\n\n  * Although it is expected that transfer learning (based on pretrained models on ImageNet) would work when the target dataset is also of natural imagery, the patterns learned in the early layers by a network work surprisingly well for datasets beyond ImageNet. That does not necessarily mean that it will yield the best results, but it can get close. When the images match more real-world images that the model was pretrained on, we get relatively quick improvement in accuracy.\n\n## Effect of Number of Layers Fine-Tuned in Transfer Learning\n\nExperimental setup\n\nVary the percentage of trainable layers from 0 to 100%\n\nDataset used\n\nOxford Flowers 102\n\nArchitecture used\n\nPretrained MobileNet\n\nFigure 5-11 shows the results.\n\n###### Figure 5-11. Effect of % layers fine-tuned on model accuracy\n\nHere are the key takeaways:\n\n  * The higher the number of layers fine-tuned, the fewer epochs it took to reach convergence and the higher the accuracy.\n\n  * The higher the number of layers fine-tuned, the more time it took per epoch for training, due to more computation and updates involved.\n\n  * For a dataset that required fine-grained understanding of images, making more layers task specific by unfreezing them was the key to a better model.\n\n## Effect of Data Size on Transfer Learning\n\nExperimental setup\n\nAdd one image per class at a time\n\nDataset used\n\nCats versus dogs\n\nArchitecture used\n\nPretrained MobileNet\n\nFigure 5-12 shows the results.\n\n###### Figure 5-12. Effect of the amount of data per category on model accuracy\n\nHere are the key takeaways:\n\n  * Even with only three images in each class, the model was able to predict with close to 90% accuracy. This shows how powerful transfer learning can be in reducing data requirements.\n\n  * Because ImageNet has several cats and dogs, pretrained networks on ImageNet suited our dataset much more easily. More difficult datasets like Oxford Flowers 102 might require a much higher number of images to achieve similar accuracies.\n\n## Effect of Learning Rate\n\nExperimental setup\n\nVary the learning rate between .1, .01, .001, and .0001\n\nDataset used\n\nOxford Flowers 102\n\nArchitecture used\n\nPretrained MobileNet\n\nFigure 5-13 shows the results.\n\n###### Figure 5-13. Effect of learning rate on model accuracy and speed of convergence\n\nHere are the key takeaways:\n\n  * Too high of a learning rate, and the model might never converge.\n\n  * Too low a learning rate results in a long time taken to convergence.\n\n  * Striking the right balance is crucial in training quickly.\n\n## Effect of Optimizers\n\nExperimental setup\n\nExperiment with available optimizers including AdaDelta, AdaGrad, Adam, Gradient Descent, Momentum, and RMSProp\n\nDataset used\n\nOxford Flowers 102\n\nArchitecture used\n\nPretrained MobileNet\n\nFigure 5-14 shows the results.\n\n###### Figure 5-14. Effect of different optimizers on the speed of convergence\n\nHere are the key takeaways:\n\n  * Adam is a great choice for faster convergence to high accuracy.\n\n  * RMSProp is usually better for RNN tasks.\n\n## Effect of Batch Size\n\nExperimental setup\n\nVary batch sizes in powers of two\n\nDataset used\n\nOxford Flowers 102\n\nArchitecture used\n\nPretrained\n\nFigure 5-15 shows the results.\n\n###### Figure 5-15. Effect of batch size on accuracy and speed of convergence\n\nHere are the key takeaways:\n\n  * The higher the batch size, the more the instability in results from epoch to epoch, with bigger rises and drops. But higher accuracy also leads to more efficient GPU utilization, so faster speed per epoch.\n\n  * Too low a batch size slows the rise in accuracy.\n\n  * 16\/32\/64 are good to start batch sizes with.\n\n## Effect of Resizing\n\nExperimental setup\n\nChange image size to 128x128, 224x224\n\nDataset used\n\nOxford Flowers 102\n\nArchitecture used\n\nPretrained\n\nFigure 5-16 shows the results.\n\n###### Figure 5-16. Effect of image size on accuracy\n\nHere are the key takeaways:\n\n  * Even with a third of the pixels, there wasn't a significant difference in validation accuracies. On the one hand, this shows the robustness of CNNs. It might partly be because the Oxford Flowers 102 dataset has close-ups of flowers visible. For datasets in which the objects have much smaller portions in an image, the results might be lower.\n\n## Effect of Change in Aspect Ratio on Transfer Learning\n\nExperimental Setup\n\nTake images of various aspect ratios (width:height ratio) and resize them to a square (1:1 aspect ratio).\n\nDataset Used\n\nCats vs. Dogs\n\nArchitecture Used\n\nPretrained\n\nFigure 5-17 shows the results.\n\n###### Figure 5-17. Distribution of aspect ratio and corresponding accuracies in images\n\nHere are the key takeaways:\n\n  * Most common aspect ratio is 4:3; that is, 1.33, whereas our neural networks are generally trained at 1:1 ratio.\n\n  * Neural networks are relatively robust to minor modifications in aspect ratio brought upon by resizing to a square shape. Even up to 2.0 ratio gives decent results.\n\n# Tools to Automate Tuning for Maximum Accuracy\n\nAs we have seen since the rise of the nineteenth century, automation has always led to an increase in productivity. In this section, we investigate tools that can help us automate the search for the best model.\n\n## Keras Tuner\n\nWith so many potential combinations of hyperparameters to tune, coming up with the best model can be a tedious process. Often two or more parameters might have correlated effects on the overall speed of convergence as well as validation accuracy, so tuning one at a time might not lead to the best model. And if curiosity gets the best of us, we might want to experiment on all the hyperparameters together.\n\nKeras Tuner comes in to automate this hyperparameter search. We define a search algorithm, the potential values that each parameter can take (e.g., discrete values or a range), our target object to maximize (e.g., validation accuracy), and sit back to see the program start training. Keras Tuner conducts multiple experiments changing the parameters on our behalf, storing the metadata of the best model. The following code example adapted from Keras Tuner documentation showcases searching through the different model architectures (varying in the number of layers between 2 and 10) as well as varying the learning rate (between 0.1 and 0.001):\n\n    from tensorflow import keras\n    from tensorflow.keras import layers\n    import numpy as np\n\n    from kerastuner.engine.hypermodel import HyperModel\n    from kerastuner.engine.hyperparameters import HyperParameters\n\n    # Input data\n    (x, y), (val_x, val_y) = keras.datasets.mnist.load_data()\n    x = x.astype('float32') \/ 255.\n    val_x = val_x.astype('float32') \/ 255.\n\n    # Defining hyper parameters\n    hp = HyperParameters()\n    hp.Choice('learning_rate', [0.1, 0.001])\n    hp.Int('num_layers', 2, 10)\n\n    # Defining model with expandable number of layers\n    def build_model(hp):\n        model = keras.Sequential()\n        model.add(layers.Flatten(input_shape=(28, 28)))\n        for _ in range(hp.get('num_layers')):\n            model.add(layers.Dense(32, activation='relu'))\n        model.add(layers.Dense(10, activation='softmax'))\n        model.compile(\n            optimizer=keras.optimizers.Adam(hp.get('learning_rate')),\n            loss='sparse_categorical_crossentropy',\n            metrics=['accuracy'])\n        return model\n\n    hypermodel = RandomSearch(\n                 build_model,\n                 max_trials=20, # Number of combinations allowed\n                 hyperparameters=hp,\n                 allow_new_entries=False,\n                 objective='val_accuracy')\n\n    hypermodel.search(x=x,\n                 y=y,\n                 epochs=5,\n                 validation_data=(val_x, val_y))\n\n    # Show summary of overall best model\n    hypermodel.results_summary()\n\nEach experiment will show values like this:\n\n     > Hp values:\n      |-learning_rate: 0.001\n      |-num_layers: 6\n    \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n    \u2502 Name         \u2502 Best model \u2502 Current model \u2502\n    \u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n    \u2502 accuracy     \u2502 0.9911     \u2502 0.9911        \u2502\n    \u2502 loss         \u2502 0.0292     \u2502 0.0292        \u2502\n    \u2502 val_loss     \u2502 0.227      \u2502 0.227         \u2502\n    \u2502 val_accuracy \u2502 0.9406     \u2502 0.9406        \u2502\n    \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n\nOn the experiment end, the result summary gives a snapshot of the experiments conducted so far, and saves more metadata.\n\n    Hypertuning complete - results in .\/untitled_project\n    [Results summary]\n     |-Results in .\/untitled_project\n     |-Ran 20 trials\n     |-Ran 20 executions (1 per trial)\n     |-Best val_accuracy: 0.9406\n\nAnother big benefit is the ability to track experiments online in real time and get notifications on their progress by visiting _http:\/\/keras-tuner.appspot.com_, getting an API key (from Google App Engine), and entering the following line in our Python program along with the real API key:\n\n    tuner.enable_cloud(api_key=api_key)\n\nDue to the potentially large combinatorial space, random search is preferred to grid search as a more practical way to get to a good solution on a limited experimentation budget. But there are faster ways, including Hyperband (Lisha Li et al.), whose implementation is also available in Keras Tuner.\n\nFor computer-vision problems, Keras Tuner includes ready-to-use tunable applications like HyperResNet.\n\n## AutoAugment\n\nAnother example hyperparameter are augmentations. Which augmentations to use? How much magnitude to augment? Would combining one too many make matters worse? Instead of leaving these decisions to humans, we can let AI decide. AutoAugment utilizes reinforcement learning to come up with the best combination of augmentations (like translation, rotation, shearing) and the probabilities and magnitudes to apply, to maximize the validation accuracy. (The method was applied by Ekin D. Cubuk et al. to come up with the new state-of-the-art ImageNet validation numbers.) By learning the best combination of augmentation parameters on ImageNet, we can readily apply it to our problem.\n\nApplying the prelearned augmentation strategy from ImageNet is pretty simple:\n\n    from PIL import Image\n    from autoaugment import ImageNetPolicy\n    img = Image.open(\"cat.jpg\")\n    policy = ImageNetPolicy()\n    imgs = [policy(img) for _ in range(8) ]\n\nFigure 5-18 displays the results.\n\n###### Figure 5-18. Output of augmentation strategies learned by reinforcement learning on the ImageNet dataset\n\n## AutoKeras\n\nWith AI automating more and more jobs, it is no surprise it can finally automate designing AI architectures, too. NAS approaches utilize reinforcement learning to join together mini-architectural blocks until they are able to maximize the objective function; in other words, our validation accuracy. The current state-of-the-art networks are all based on NAS, leaving human-designed architectures in the dust. Research in this area started showing promising results in 2017, with a bigger focus on making train faster in 2018. And now with AutoKeras (Haifeng Jin et al.), we can also apply this state-of-the-art technique on our particular datasets in a relatively accessible manner.\n\nGenerating new model architectures with AutoKeras is a matter of supplying our images and associated labels as well as a time limit by which to finish running the jobs. Internally, it implements several optimization algorithms, including a Bayesian optimization approach to search for an optimal architecture:\n\n    !pip3 install autokeras\n    !pip3 install graphviz\n    from keras.datasets import mnist\n    from autokeras.image.image_supervised import ImageClassifier\n\n    (x_train, y_train), (x_test, y_test) = mnist.load_data()\n    x_train = x_train.reshape(x_train.shape + (1,))\n    x_test = x_test.reshape(x_test.shape + (1,))\n\n    clf = ImageClassifier(path=\".\",verbose=True, augment=False)\n    clf.fit(x_train, y_train, time_limit= 30 * 60) # 30 minutes\n    clf.final_fit(x_train, y_train, x_test, y_test, retrain=True)\n    y = clf.evaluate(x_test, y_test)\n    print(y)\n\n    _# Save the model as a pickle file_\n    clf.export_autokeras_model(\"model.pkl\")\n\n    visualize('.')\n\nPost-training, we are all eager to learn how the new model architecture looks. Unlike most of the cleaner-looking images we generally get to see, this will look pretty obfuscated to understand or print out. But what we do find faith in is that it yields high accuracy.\n\n# Summary\n\nIn this chapter, we saw a range of tools and techniques to help investigate opportunities to improve our CNN accuracy. Building a case for iterative experimentation, you learned how tuning hyperparameters can bring about optimal performance. And with so many hyperparameters to choose from, we then looked at automated approaches, including AutoKeras, AutoAugment, and Keras Tuner. Best of all, the core code for this chapter combining multiple tools in a single Colab file is available online on the book's GitHub (see _http:\/\/PracticalDeepLearning.ai_) and can easily be tuned to more than 100 datasets with a single line change and run online in the browser. Additionally, we compiled a checklist of actionable tips along with interactive experiments hosted online to help give your model a little extra edge. We hope that the material covered in this chapter will help make your models more robust, reduce bias, make them more explainable, and ultimately contribute to the responsible development of AI.\n\n# Chapter 6. Maximizing Speed and Performance of TensorFlow: A Handy Checklist\n\nLife is all about making do with what we have, and optimization is the name of the game.\n\nIt's not about having everything\u2014it's about using your resources wisely. Maybe we really want to buy that Ferrari, but our budget allows for a Toyota. You know what, though? With the right kinds of performance tuning, we can make that bad boy race at NASCAR!\n\nLet's look at this in terms of the deep learning world. Google, with its engineering might and TPU pods capable of boiling the ocean, set a speed record by training ImageNet in just about 30 minutes! And yet, just a few months later, a ragtag team of three researchers (Andrew Shaw, Yaroslav Bulatov, and Jeremy Howard), with $40 in their pockets using a public cloud, were able to train ImageNet in only 18 minutes!\n\nThe lesson we can draw from these examples is that the amount of resources that you have is not nearly as important as using them to their maximum potential. It's all about doing more with less. In that spirit, this chapter is meant to serve as a handy checklist of potential performance optimizations that we can make when building all stages of the deep learning pipelines, and will be useful throughout the book. Specifically, we will discuss optimizations related to data preparation, data reading, data augmentation, training, and finally inference.\n\nAnd the story starts and ends with two words...\n\n# GPU Starvation\n\nA commonly asked question by AI practitioners is, \"Why is my training so slow?\" The answer more often than not is GPU starvation.\n\nGPUs are the lifelines of deep learning. They can also be the most expensive component in a computer system. In light of that, we want to fully utilize them. This means that a GPU should not need to wait for data to be available from other components for processing. Rather, when the GPU is ready to process, the preprocessed data should already be available at its doorstep and ready to go. Yet, the reality is that the CPU, memory, and storage are frequently the performance bottlenecks, resulting in suboptimal utilization of the GPU. In other words, we want the GPU to be the bottleneck, not the other way round.\n\nBuying expensive GPUs for thousands of dollars can be worthwhile, but only if the GPU is the bottleneck to begin with. Otherwise, we might as well burn the cash.\n\nTo illustrate this better, consider Figure 6-1. In a deep learning pipeline, the CPU and GPU work in collaboration, passing data to each other. The CPU reads the data, performs preprocessing steps including augmentations, and then passes it on to the GPU for training. Their collaboration is like a relay race, except one of the relay runners is an Olympic athlete, waiting for a high school track runner to pass the baton. The more time the GPU stays idle, the more wasted resources.\n\n###### Figure 6-1. GPU starvation, while waiting for CPU to finish preparing the data\n\nA large portion of this chapter is devoted to reducing the idle time of the GPU and the CPU.\n\nA logical question to ask is: how do we know whether the GPU is starving? Two handy tools can help us answer this question:\n\n`nvidia-smi`\n\nThis command shows GPU statistics including utilization.\n\nTensorFlow Profiler + TensorBoard\n\nThis visualizes program execution interactively in a timeline within TensorBoard.\n\n## nvidia-smi\n\nShort for NVIDIA System Management Interface program, `nvidia-smi` provides detailed statistics about our precious GPUs, including memory, utilization, temperature, power wattage, and more. It's a geek's dream come true.\n\nLet's take it for a test drive:\n\n    $ nvidia-smi\n\nFigure 6-2 shows the result.\n\n###### Figure 6-2. Terminal output of `nvidia-smi` highlighting the GPU utilization\n\nWhile training a network, the key figure we are interested in is the GPU utilization, defined in the documentation as the percent of time over the past second during which _one or more_ kernels was executing on the GPU. Fifty-one percent is frankly not that great. But this is utilization at the moment in time when `nvidia-smi` is called. How do we continuously monitor these numbers? To better understand the GPU usage, we can refresh the utilization metrics every half a second with the `watch` command (it's worth memorizing this command):\n\n    $ watch -n .5 nvidia-smi\n\n###### Note\n\nAlthough GPU utilization is a good proxy for measuring the efficiency of our pipeline, it does not alone measure how well we're using the GPU, because the work could still be using a small fraction of the GPU's resources.\n\nBecause staring at the terminal screen with the number jumping around is not the most optimal way to analyze, we can instead poll the GPU utilization every second and dump that into a file. Run this for about 30 seconds while any GPU-related process is running on our system and stop it by pressing Ctrl+C:\n\n    $ nvidia-smi --query-gpu=utilization.gpu --format=csv,noheader,nounits -f\n    gpu_utilization.csv -l 1\n\nNow, calculate the median GPU utilization from the file generated:\n\n    $ sort -n gpu_utilization.csv | grep -v '^0$' | datamash median 1\n\n###### Tip\n\nDatamash is a handy command-line tool that performs basic numeric, textual, and statistical operations on textual data files. You can find instructions to install it at _https:\/\/www.gnu.org\/software\/datamash\/_.\n\n`nvidia-smi` is the most convenient way to check our GPU utilization on the command line. Could we get a deeper analysis? It turns out, for advanced users, TensorFlow provides a powerful set of tools.\n\n## TensorFlow Profiler + TensorBoard\n\nTensorFlow ships with `tfprof` (Figure 6-3), the TensorFlow profiler to help analyze and understand the training process at a much deeper level, such as generating a detailed model analysis report for each operation in our model. But the command line can be a bit daunting to navigate. Luckily, TensorBoard, a suite of browser-based visualization tools for TensorFlow, includes a plugin for the profiler that lets us interactively debug the network with a few mouse clicks. This includes Trace Viewer, a feature that shows events in a timeline. It helps investigate how resources are being used precisely at a given period of time and spot inefficiencies.\n\n###### Note\n\nAs of this writing, TensorBoard is fully supported only in Google Chrome and might not show the profile view in other browsers, like Firefox.\n\n###### Figure 6-3. Profiler's timeline in TensorBoard shows an idle GPU while the CPU is processing as well as CPU idling while the GPU is processing\n\nTensorBoard, by default, has the profiler enabled. Activating TensorBoard involves a simple callback function:\n\n    tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=\"\/tmp\",\n                                                          profile_batch=7)\n\n    model.fit(train_data,\n              steps_per_epoch=10,\n              epochs=2, \n              callbacks=[ **tensorboard_callback**])\n\nWhen initializing the callback, unless `profile_batch` is explicitly specified, it profiles the second batch. Why the second batch? Because the first batch is usually slower than the rest due to some initialization overhead.\n\n###### Note\n\nIt bears reiterating that profiling using TensorBoard is best suited for power users of TensorFlow. If you are just starting out, you are better off using `nvidia-smi`. (Although `nvidia-smi` is a far more capable than just providing GPU utilization info, which is typically how most practitioners use it.) For users wanting even deeper access to their hardware utilization metrics, NVIDIA Nsight is a great tool.\n\nAlright. With these tools at our disposal, we know that our program needs some tuning and has room for efficiency improvements. We look at those areas one by one in the next few sections.\n\n# How to Use This Checklist\n\nIn business, an oft-quoted piece of advice is \"You can't improve what you can't measure.\" This applies to deep learning pipelines, as well. Tuning performance is like a science experiment. You set up a baseline run, tune a knob, measure the effect, and iterate in the direction of improvement. The items on the following checklist are our knobs\u2014some are quick and easy, whereas others are more involved.\n\nTo use this checklist effectively, do the following:\n\n  1. Isolate the part of the pipeline that you want to improve.\n\n  2. Find a relevant point on the checklist.\n\n  3. Implement, experiment, and observe if runtime is reduced. If not reduced, ignore change.\n\n  4. Repeat steps 1 through 3 until the checklist is exhausted.\n\nSome of the improvements might be minute, some more drastic. But the cumulative effect of all these changes should hopefully result in faster, more efficient execution and best of all, more bang for the buck for your hardware. Let's look at each area of the deep learning pipeline step by step, including data preparation, data reading, data augmentation, training, and, finally, inference.\n\n# Performance Checklist\n\n## Data Preparation\n\n  * \"Store as TFRecords\"\n\n  * \"Reduce Size of Input Data\"\n\n  * \"Use TensorFlow Datasets\"\n\n## Data Reading\n\n  * \"Use tf.data\"\n\n  * \"Prefetch Data\"\n\n  * \"Parallelize CPU Processing\"\n\n  * \"Parallelize I\/O and Processing\"\n\n  * \"Enable Nondeterministic Ordering\"\n\n  * \"Cache Data\"\n\n  * \"Turn on Experimental Optimizations\"\n\n  * \"Autotune Parameter Values\"\n\n## Data Augmentation\n\n  * \"Use GPU for Augmentation\"\n\n## Training\n\n  * \"Use Automatic Mixed Precision\"\n\n  * \"Use Larger Batch Size\"\n\n  * \"Use Multiples of Eight\"\n\n  * \"Find the Optimal Learning Rate\"\n\n  * \"Use tf.function\"\n\n  * \"Overtrain, and Then Generalize\"\n\n    * \"Use progressive sampling\"\n\n    * \"Use progressive augmentation\"\n\n    * \"Use progressive resizing\"\n\n  * \"Install an Optimized Stack for the Hardware\"\n\n  * \"Optimize the Number of Parallel CPU Threads\"\n\n  * \"Use Better Hardware\"\n\n  * \"Distribute Training\"\n\n  * \"Examine Industry Benchmarks\"\n\n## Inference\n\n  * \"Use an Efficient Model\"\n\n  * \"Quantize the Model\"\n\n  * \"Prune the Model\"\n\n  * \"Use Fused Operations\"\n\n  * \"Enable GPU Persistence\"\n\n###### Note\n\nA printable version of this checklist is available at <http:\/\/PracticalDeepLearning.ai>. Feel free to use it as a reference next time you train or deploy a model. Or even better, spread the cheer by sharing with your friends, colleagues, and more importantly, your manager.\n\n# Data Preparation\n\nThere are a few optimizations that we can make even before we do any kind of training, and they have to do with how we prepare our data.\n\n## Store as TFRecords\n\nImage datasets typically consist of thousands of tiny files, each file measuring a few kilobytes. And our training pipeline must read each file individually. Doing this thousands of times has significant overhead, causing a slowdown of the training process. That problem is even more severe in the case of spinning hard drives, for which the magnetic head needs to seek to the beginning of each file. This problem is further exacerbated when the files are stored on a remote storage service like the cloud. And there lies our first hurdle!\n\nTo speed up the reads, one idea is to combine thousands of files into a handful of larger files. And that's exactly what TFRecord does. It stores data in efficient Protocol Buffer (protobuf) objects, making them quicker to read. Let's see how to create TFRecord files:\n\n    _# Create TFRecord files_\n\n    import tensorflow as tf\n    from PIL import Image\n    import numpy as np\n    import io\n\n    cat = \"cat.jpg\"\n    img_name_to_labels = {'cat' : 0}\n    img_in_string = open(cat, 'rb').read()\n    label_for_img = img_name_to_labels['cat']\n\n    def getTFRecord(img, label):\n     feature = {\n        'label': _int64_feature(label),\n        'image_raw': _bytes_feature(img),\n     }\n     return tf.train.Example(features=tf.train.Features(feature=feature))\n\n    with tf.compat.v1.python_io.TFRecordWriter('img.tfrecord') as writer:\n      for filename, label in img_name_to_labels.items():\n        image_string = open(filename, 'rb').read()\n        tf_example = getTFRecord(image_string, label)\n        writer.write(tf_example.SerializeToString())\n\nNow, let's take a look at reading these TFRecord files:\n\n    # Reading TFRecord files\n\n    dataset = tf.data.TFRecordDataset('img.tfrecord')\n    ground_truth_info = {\n      'label': tf.compat.v1.FixedLenFeature([], tf.int64),\n      'image_raw': tf.compat.v1.FixedLenFeature([], tf.string),\n    }\n\n    def map_operation(read_data):\n      return tf.compat.v1.parse_single_example(read_data, ground_truth_info)\n\n    imgs = dataset.map(map_operation)\n\n    for image_features in imgs:\n      image_raw = image_features['image_raw'].numpy()\n      label = image_features['label'].numpy()\n      image = Image.open(io.BytesIO(image_raw))\n      image.show()\n      print(label)\n\nSo, why not join all of the data in a single file, like say for ImageNet? Although reading thousands of tiny files harms performance due to the overhead involved, reading gigantic files is an equally bad idea. They reduce our ability to make parallel reads and parallel network calls. The sweet spot to shard (divide) a large dataset in TFRecord files lies at around 100 MB.\n\n## Reduce Size of Input Data\n\nImage datasets with large images need to be resized before passing through to the GPU. This means the following:\n\n  * Repeated CPU cycles at every iteration\n\n  * Repeated I\/O bandwidth being consumed at a larger rate than needed in our data pipeline\n\nOne good strategy to save compute cycles is to perform common preprocessing steps once on the entire dataset (like resizing) and then saving the results in TFRecord files for all future runs.\n\n## Use TensorFlow Datasets\n\nFor commonly used public datasets, from MNIST (11 MB) to CIFAR-100 (160 MB) all the way to MS COCO (38 GB) and Google Open Images (565 GB), it's quite an effort to download the data (often spread across multiple zipped files). Imagine your frustration if after downloading 95% of the file slowly, the connection becomes spotty and breaks. This is not unusual because these files are typically hosted on university servers, or are downloaded from various sources like Flickr (as is the case with ImageNet 2012, which gives us the URLs from which to download 150 GB-plus of images). A broken connection might mean having to start all over again.\n\nIf you think that was tedious, the real challenge actually begins only after you successfully download the data. For every new dataset, we now need to hunt through the documentation to determine how the data is formatted and organized, so we can begin reading and processing appropriately. Then, we need to split the data into training, validation, and test sets (preferably converting to TFRecords). And when the data is so large as to not fit in memory, we will need to do some manual jiu-jitsu to read it and feed it efficiently to the training pipeline. We never said it was easy.\n\nAlternately, we could skip all the pain by consuming the high-performance, ready-to-use TensorFlow Datasets package. With several famous datasets available, it downloads, splits, and feeds our training pipeline using best practices in a few lines.\n\nLet's look at which datasets are available.\n\n    import tensorflow_datasets as tfds\n\n    # See available datasets\n    print(tfds.list_builders())\n\n    ===== Output =====\n    ['abstract_reasoning', 'bair_robot_pushing_small', 'caltech101', 'cats_vs_dogs',\n    'celeb_a', 'celeb_a_hq', 'chexpert', 'cifar10', 'cifar100', 'cifar10_corrupted',\n    'cnn_dailymail', 'coco2014', 'colorectal_histology',\n    'colorectal_histology_large', 'cycle_gan' ...\n\nThere are more than 100 datasets as of this writing, and that number is steadily increasing. Now, let's download, extract, and make an efficient pipeline using the training set of CIFAR-10:\n\n    train_dataset = tfds.load(name=\"cifar100\", split=tfds.Split.TRAIN)\n    train_dataset = train_dataset.shuffle(2048).batch(64)\n\nThat's it! The first time we execute the code, it will download and cache the dataset on our machine. For every future run, it will skip the network download and directly read from the cache.\n\n# Data Reading\n\nNow that the data is prepared, let's look for opportunities to maximize the throughput of the data reading pipeline.\n\n## Use tf.data\n\nWe could choose to manually read every file from our dataset with Python's built-in I\/O library. We could simply call `open` for each file and we'd be good to go, right? The main downside in this approach is that our GPU would be bottlenecked by our file reads. Every time we read a file, the GPU needs to wait. Every time the GPU starts processing its input, we wait before we read the next file from disk. Seems rather wasteful, doesn't it?\n\nIf there's only one thing you can take away from this chapter, let it be this: `tf.data` is the way to go for building a high-performance training pipeline. In the next few sections, we explore several aspects of `tf.data` that you can exploit to improve training speed.\n\nLet's set up a base pipeline for reading data:\n\n    files = tf.data.Dataset.list_files(\".\/training_data\/*.tfrecord\")\n    dataset = tf.data.TFRecordDataset(files)\n\n    dataset = dataset.shuffle(2048)\n                     .repeat()\n                     .map(lambda item: tf.io.parse_single_example(item, features))\n                     .map(_resize_image)\n                     .batch(64)\n\n## Prefetch Data\n\nIn the pipeline we discussed earlier, the GPU waits for the CPU to generate data, and then the CPU waits for the GPU to finish computation before generating data for the next cycle. This circular dependency causes idle time for both CPU and GPU, which is inefficient.\n\nThe `prefetch` function helps us here by delinking the production of the data (by the CPU) from the consumption of the data (by the GPU). Using a background thread, it allows data to be passed _asynchronously_ into an intermediate buffer, where it is readily available for a GPU to consume. The CPU now carries on with the next computation instead of waiting for the GPU. Similarly, as soon as the GPU is finished with its previous computation, and there's data readily available in the buffer, it starts processing.\n\nTo use it, we can simply call `prefetch` on our dataset at the very end of our pipeline along with a `buffer_size` parameter (which is the maximum amount of data that can be stored). Usually `buffer_size` is a small number; `1` is good enough in many cases:\n\n    dataset = dataset.prefetch(buffer_size=16)\n\nIn just a few pages, we show you how to find an optimal value for this parameter.\n\nIn summary, if there's an opportunity to overlap CPU and GPU computations, `prefetch` will automatically exploit it.\n\n## Parallelize CPU Processing\n\nIt would be a waste to have a CPU with multiple cores but doing all of our processing on only one of them. Why not take advantage of the rest? This is exactly where the `num_parallel_calls` argument in the `map` function comes in handy:\n\n    dataset = dataset.map(lambda item: tf.io.parse_single_example(item, features), \n                          **num_parallel_calls=4**)\n\nThis starts multiple threads to parallelize processing of the `map()` function. Assuming that there is no heavy application running in the background, we will want to set `num_parallel_calls` to the number of CPU cores on our system. Anything more will potentially degrade the performance due to the overhead of context switching.\n\n## Parallelize I\/O and Processing\n\nReading files from disk or worse, over a network, is a huge cause of bottlenecks. We might possess the best CPU and GPU in the world, but if we don't optimize our file reads, it would all be for naught. One solution that addresses this problem is to parallelize both I\/O and subsequent processing (also known as _interleaving_ ).\n\n    dataset = files.interleave(map_func, num_parallel_calls=4)\n\nIn this command, two things are happening:\n\n  * The input data is acquired in parallel (by default equal to the number of cores on the system).\n\n  * On the acquired data, setting the `num_parallel_calls` parameter allows the `map_func` function to execute on multiple parallel threads and read from the incoming data asynchronously.\n\nIf `num_parallel_calls` was not specified, even if the data were read in parallel, `map_func` would run synchronously on a single thread. As long as `map_func` runs faster than the rate at which the input data is coming in, there will not be a problem. We definitely want to set `num_parallel_calls` higher if `map_func` becomes a bottleneck.\n\n## Enable Nondeterministic Ordering\n\nFor many datasets, the reading order is not important. After all, we might be randomizing their ordering anyway. By default, when reading files in parallel, `tf.data` still attempts to produce their outputs in a _fixed round-robin order_. The disadvantage is that we might encounter a \"straggler\" along the way (i.e., an operation that takes a lot longer than others, such as a slow file read, and holds up all other operations). It's like a grocery store line where the person in front of us insists on using cash with the exact change, whereas everyone else uses a credit card. So instead of blocking all the subsequent operations that are ready to give output, we skip over the stragglers until they are done with their processing. This breaks the ordering while reducing wasted cycles waiting for the handful of slower operations:\n\n    options = tf.data.Options()\n    options.experimental_deterministic = False\n\n    dataset = tf.data.Dataset.list_files(\".\/training_data\/\")\n    dataset = dataset.with_options(options)\n    dataset = dataset.interleave(tf.data.TFRecordDataset, num_parallel_calls=4)\n\n## Cache Data\n\nThe `Dataset.cache()` function allows us to make a copy of data either in memory or as a file on disk. There are two reasons why you might want to cache a dataset:\n\n  * To avoid repeatedly reading from disk after the first epoch. This is obviously effective only when the cache is in memory and can fit in the available RAM.\n\n  * To avoid having to repeatedly perform expensive CPU operations on data (e.g., resizing large images to a smaller size).\n\n###### Tip\n\nCache is best used for data that is not going to change. It is recommended to place `cache()` before any random augmentations and shuffling; otherwise, caching at the end will result in exactly the same data and order in every run.\n\nDepending on our scenario, we can use one of the two following lines:\n\n    dataset = dataset.cache()                     _# in-memory_\n    dataset = dataset.cache(filename='tmp.cache') _# on-disk_\n\nIt's worth noting that in-memory cache is volatile and hence only shows performance improvements in the second epoch of every run. On the other hand, file-based cache will make every run faster (beyond the very first epoch of the first run).\n\n###### Tip\n\nIn the \"Reduce Size of Input Data\", we mentioned preprocessing the data and saving it as TFRecord files as input to future data pipelines. Using the **`cache()`** function directly after the preprocessing step in your pipeline would give a similar performance with a single word change in code.\n\n## Turn on Experimental Optimizations\n\nTensorFlow has many built-in optimizations, often initially experimental and turned off by default. Depending on your use case, you might want to turn on some of them to squeeze out just a little more performance from your pipeline. Many of these optimizations are detailed in the documentation for `tf.data.experimental.OptimizationOptions`.\n\n###### Note\n\nHere's a quick refresher on filter and map operations:\n\nFilter\n\nA filter operation goes through a list element by element and grabs those that match a given condition. The condition is supplied as a lambda operation that returns a boolean value.\n\nMap\n\nA map operation simply takes in an element, performs a computation, and returns an output. For example, resizing an image.\n\nLet's look at a few experimental optimizations that are available to us, including examples of two consecutive operations that could benefit from being fused together as one single operation.\n\n### Filter fusion\n\nSometimes, we might want to filter based on multiple attributes. Maybe we want to use only images that have both a dog and a cat. Or, in a census dataset, only look at families above a certain income threshold who also live within a certain distance to the city center. `filter_fusion` can help speed up such scenarios. Consider the following example:\n\n    dataset = dataset.filter(lambda x: x < 1000).filter(lambda x: x % 3 == 0)\n\nThe first filter performs a full pass over the entire dataset and returns elements that are less than 1,000. On this output, the second filter does another pass to further remove elements not divisible by three. Instead of doing two passes over many of the same elements, we could instead combine both the filter operations into one pass using an `AND` operation. That is precisely what the `filter_fusion` option enables\u2014combining multiple filter operations into one pass. By default, it is turned off. You can enable it by using the following statement:\n\n    options = tf.data.Options()\n    options.experimental_optimization.filter_fusion = True\n    dataset = dataset.with_options(options)\n\n### Map and filter fusion\n\nConsider the following example:\n\n    dataset = dataset.map(lambda x: x * x).filter(lambda x: x % 2 == 0)\n\nIn this example, the `map` function does a full pass on the entire dataset to calculate the square of every element. Then, the `filter` function discards the odd elements. Rather than doing two passes (more so in this particularly wasteful example), we could simply fuse the map and filter operations together by turning on the `map_and_filter_fusion` option so that they operate as a single unit:\n\n    options.experimental_optimization.map_and_filter_fusion = True\n\n### Map fusion\n\nSimilar to the aforementioned two examples, fusing two or more map operations prevents multiple passes from being performed on the same data and instead combines them in a single pass:\n\n    options.experimental_optimization.map_fusion = True\n\n## Autotune Parameter Values\n\nYou might have noticed that many of the code examples in this section have hardcoded values for some of the parameters. For the combination of the problem and hardware at hand, you can tune them for maximum efficiency. How to tune them? One obvious way is to manually tweak the parameters one at a time and isolate and observe the impact of each of them on the overall performance until we get the precise parameter set. But the number of knobs to tune quickly gets out of hand due to the combinatorial explosion. If this wasn't enough, our finely tuned script wouldn't necessarily be as efficient on another machine due to differences in hardware such as the number of CPU cores, GPU availability, and so on. And even on the same system, depending on resource usage by other programs, these knobs might need to be adjusted over different runs.\n\nHow do we solve this? We do the opposite of manual tuning: autotuning. Using hill-climbing optimization algorithms (which are a type of heuristic-driven search algorithms), this option automatically finds the ideal parameter combination for many of the `tf.data` function parameters. Simply use `tf.data.experimental.AUTOTUNE` instead of manually assigning numbers. It's the one parameter to rule them all. Consider the following example:\n\n    dataset = dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)\n\nIsn't that an elegant solution? We can do that for several other function calls in the `tf.data` pipeline. The following is an example of combining together several optimizations from the section \"Data Reading\" to make a high-performance data pipeline:\n\n    options = tf.data.Options()\n    options.experimental_deterministic = False\n\n    dataset = tf.data.Dataset.list_files(\"\/path\/*.tfrecord\")\n    dataset = dataset.with_options(options)\n    dataset = files.interleave(tf.data.TFRecordDataset,\n                                num_parallel_calls=tf.data.experimental.AUTOTUNE)\n    dataset = dataset.map(preprocess,\n                          num_parallel_calls=tf.data.experimental.AUTOTUNE)\n    dataset = dataset.cache() \n    dataset = dataset.repeat() \n    dataset = dataset.shuffle(2048)\n    dataset = dataset.batch(batch_size=64) \n    dataset = dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)\n\n# Data Augmentation\n\nSometimes, we might not have sufficient data to run our training pipeline. Even if we did, we might still want to manipulate the images to improve the robustness of our model\u2014with the help of data augmentation. Let's see whether we can make this step any faster.\n\n## Use GPU for Augmentation\n\nData preprocessing pipelines can be elaborate enough that you could write an entire book about them. Image transformation operations such as resizing, cropping, color transformations, blurring, and so on are commonly performed on the data immediately after it's read from disk into memory. Given that these are all matrix transformation operations, they might do well on a GPU.\n\nOpenCV, Pillow, and the built-in Keras augmentation functionality are the most commonly used libraries in computer vision for working on images. There's one major limitation here, though. Their image processing is primarily CPU based (although you can compile OpenCV to work with CUDA), which means that the pipeline might not be fully utilizing the underlying hardware to its true potential.\n\n###### Note\n\nAs of August 2019, there are efforts underway to convert Keras image augmentation to be GPU accelerated, as well.\n\nThere are a few different GPU-bound options that we can explore.\n\n### tf.image built-in augmentations\n\n`tf.image` provides some handy augmentation functions that we can seamlessly plug into a `tf.data` pipeline. Some of the methods include image flipping, color augmentations (hue, saturation, brightness, contrast), zooming, and rotation. Consider the following example, which changes the hue of an image:\n\n    updated_image = tf.image.adjust_hue(image, delta = 0.2)\n\nThe downside to relying on `tf.image` is that the functionality is much more limited compared to OpenCV, Pillow, and even Keras. For example, the built-in function for image rotation in `tf.image` only supports rotating images by 90 degrees counter-clockwise. If we need to be able to rotate by an arbitrary amount, such as 10 degrees, we'd need to manually build that functionality. Keras, on the other hand, provides that functionality out of the box.\n\nAs another alternative to the `tf.data` pipeline, the NVIDIA Data Loading Library (DALI) offers a fast data loading and preprocessing pipeline accelerated by GPU processing. As shown in Figure 6-4, DALI implements several common steps including resizing an image and augmenting an image in the GPU, immediately before the training. DALI works with multiple deep learning frameworks including TensorFlow, PyTorch, MXNet, and others, offering portability of the preprocessing pipelines.\n\n### NVIDIA DALI\n\n###### Figure 6-4. The NVIDIA DALI pipeline\n\nAdditionally, even JPEG decoding (a relatively heavy task) can partially make use of the GPU, giving it an additional boost. This is done using nvJPEG, a GPU-accelerated library for JPEG decoding. For multi-GPU tasks, this scales near linearly as the number of GPUs increases.\n\nNVIDIAs efforts culminated in a record-breaking MLPerf entry (which benchmarks machine learning hardware, software, and services), training a ResNet-50 model in 80 seconds.\n\n# Training\n\nFor those beginning their performance optimization journey, the quickest wins come from improving the data pipelines, which is relatively easy. For a training pipeline that is already being fed data fast, let's investigate optimizations for our actual training step.\n\n## Use Automatic Mixed Precision\n\n\" _One line to make your training two to three times faster!_ \"\n\nWeights in deep learning models are typically stored in single-precision; that is, 32-bit floating point, or as it's more commonly referenced: FP32. Putting these models in memory-constrained devices such as mobile phones can be challenging to accommodate. A simple trick to make models smaller is to convert them from single-precision (FP32) to half-precision (FP16). Sure, the representative power of these weights goes down, but as we demonstrate later in this chapter (\"Quantize the Model\"), neural networks are resilient to small changes, much like they are resilient to noise in images. Hence, we get the benefits of a more efficient model without sacrificing much accuracy. In fact, we can even reduce the representation to 8-bit integers (INT8) without a significant loss in accuracy, as we will see in some upcoming chapters.\n\nSo, if we can use reduced-precision representation during inference, could we do the same during training, as well? Going from 32-bit to 16-bit representation would effectively mean double the memory bandwidth available, double the model size, or double the batch size can be accommodated. Unfortunately, it turns out that using FP16 na\u00efvely _during training_ can potentially lead to a significant loss in model accuracy and might not even converge to an optimal solution. This happens because of FP16's limited range for representing numbers. Due to a lack of adequate precision, any updates to the model during training, if sufficiently small, will cause an update to not even register. Imagine adding 0.00006 to a weight value of 1.1. With FP32, the weight would be correctly updated to 1.10006. With FP16, however, the weight would remain 1.1. Conversely, any activations from layers such as Rectified Linear Unit (ReLU) could be high enough for FP16 to overflow and hit infinity (`NaN` in Python).\n\nThe easy answer to these challenges is to use automatic mixed-precision training. In this method, we store the model in FP32 as a master copy and perform the forward\/backward passes of training in FP16. After each training step is performed, the final update from that step is then scaled back up to FP32 before it is applied to the master copy. This helps avoid the pitfalls of FP16 arithmetic and results in a lower memory footprint, and faster training (experiments have shown increases in speed by two to three times), while achieving similar accuracy levels as training solely in FP32. It is noteworthy that newer GPU architectures like the NVIDIA Volta and Turing especially optimize FP16 operations.\n\nTo enable mixed precision during training, we simply need to add the following line to the beginning of our Python script:\n\n    os.environ['TF_ENABLE_AUTO_MIXED_PRECISION'] = '1'\n\n## Use Larger Batch Size\n\nInstead of using the entire dataset for training in one batch, we train with several minibatches of data. This is done for two reasons:\n\n  * Our full data (single batch) might not fit in the GPU RAM.\n\n  * We can achieve similar training accuracy by feeding many smaller batches, just as you would by feeding fewer larger batches.\n\nHaving smaller minibatches might not fully utilize the available GPU memory, so it's vital to experiment with this parameter, see its effect on the GPU utilization (using the `nvidia-smi` command), and choose the batch size that maximizes the utilization. Consumer GPUs like the NVIDIA 2080 Ti ship with 11 GB of GPU memory, which is plenty for efficient models like MobileNet family.\n\nFor example on hardware with the 2080 Ti graphics card, using 224 x 224 resolution images and MobileNetV2 model, the GPU can accommodate a batch size up to 864. Figure 6-5 shows the effect of varying batch sizes from 4 to 864, on both the GPU utilization (solid line) as well as the time per epoch (dashed line). As we can see in the figure, the higher the batch size, the higher the GPU utilization, leading to a shorter training time per epoch.\n\nEven at our max batch size of 864 (before running out of memory allocation), the GPU utilization does not cross 85%. This means that the GPU was fast enough to handle the computations of our otherwise very efficient data pipeline. Replacing MobileNetV2 with a heavier ResNet-50 model immediately increased GPU to 95%.\n\n###### Figure 6-5. Effect of varying batch size on time per epoch (seconds) as well as on percentage GPU utilization (Log scales have been used for both X- and Y-axes.)\n\n###### Tip\n\nEven though we showcased batch sizes up to a few hundreds, large industrial training loads distributed across multiple nodes often use much larger batch sizes with the help of a technique called Layer-wise Adaptive Rate Scaling (LARS). For example, Fujitsu Research trained a ResNet-50 network to 75% Top-1 accuracy on ImageNet in a mere 75 seconds. Their ammunition? 2048 Tesla V100 GPUs and a whopping batch size of 81,920!\n\n## Use Multiples of Eight\n\nMost of the computations in deep learning are in the form of \"matrix multiply and add.\" Although it's an expensive operation, specialized hardware has increasingly been built in the past few years to optimize for its performance. Examples include Google's TPUs and NVIDIA's Tensor Cores (which can be found in the Turing and Volta architectures). Turing GPUs provide both Tensor Cores (for FP16 and INT8 operations) as well as CUDA cores (for FP32 operations), with the Tensor Cores delivering significantly higher throughput. Due to their specialized nature, Tensor Cores require that certain parameters within the data supplied to them be divisible by eight. Here are just three such parameters:\n\n  * The number of channels in a convolutional filter\n\n  * The number of neurons in a fully connected layer and the inputs to this layer\n\n  * The size of minibatches\n\nIf these parameters are not divisible by eight, the GPU CUDA cores will be used as the fallback accelerator instead. In an experiment reported by NVIDIA, simply changing the batch size from 4,095 to 4,096 resulted in an increase in throughput of five times. Keep in mind that using multiples of eight (or 16 in the case of INT8 operations), in addition to using automatic mixed precision, is the bare minimum requirement to activate the Tensor Cores. For higher efficiency, the recommended values are in fact multiples of 64 or 256. Similarly, Google recommends multiples of 128 when using TPUs for maximum efficiency.\n\n## Find the Optimal Learning Rate\n\nOne hyperparameter that greatly affects our speed of convergence (and accuracy) is the learning rate. The ideal result of training is the global minimum; that is, the point of least loss. Too high a learning rate can cause our model to overshoot the global minimum (like a wildly swinging pendulum) and potentially never converge. Too low a learning rate can cause convergence to take too long because the learning algorithm will take very small steps toward the minimum. Finding the right initial learning rate can make a world of difference.\n\nThe naive way to find the ideal initial learning rate is to try a few different learning rates (such as 0.00001, 0.0001, 0.001, 0.01, 0.1) and find one that starts converging quicker than others. Or, even better, perform grid search over a range of values. This approach has two problems: 1) depending on the granularity, it might find a decent value, but it might not be the most optimal value; and 2) we need to train multiple times, which can be time consuming.\n\nIn Leslie N. Smith's 2015 paper, \"Cyclical Learning Rates for Training Neural Networks,\" he describes a much better strategy to find this optimal learning rate. In summary:\n\n  1. Start with a really low learning rate and gradually increase it until reaching a prespecified maximum value.\n\n  2. At each learning rate, observe the loss\u2014first it will be stagnant, then it will begin going down and then eventually go back up.\n\n  3. Calculate the rate of decrease of loss (first derivative) at each learning rate.\n\n  4. Select the point with the highest rate of decrease of loss.\n\nIt sounds like a lot of steps, but thankfully we don't need to write code for it. The keras_lr_finder library by Pavel Surmenok gives us a handy function to find it:\n\n    lr_finder = LRFinder(model)\n    lr_finder.find(x_train, y_train, start_lr=0.0001, end_lr=10, batch_size=512,\n                   epochs=5)\n    lr_finder.plot_loss(n_skip_beginning=20, n_skip_end=5)\n\nFigure 6-6 shows the plot of loss versus learning rate. It becomes evident that a learning rate of 10\u20134 or 10\u20133 might be too low (owing to barely any drop in loss), and similarly, above 1 might be too high (because of the rapid increase in loss).\n\n###### Figure 6-6. A graph showing the change in loss as the learning rate is increased\n\nWhat we are most interested in is the point of the greatest decrease in loss. After all, we want to minimize the time we spend in getting to the least loss during training. In Figure 6-7, we plot the _rate of change_ of loss\u2014the derivative of the loss with regard to the learning rate:\n\n    # Show Simple Moving Average over 20 points to smoothen the graph\n    lr_finder.plot_loss_change(sma=20, n_skip_beginning=20, n_skip_end=5,\n                               y_lim=(-0.01, 0.01))\n\n###### Figure 6-7. A graph showing the rate of change in loss as the learning rate is increased\n\nThese figures show that values around 0.1 would lead to the fastest decrease in loss, and hence we would choose it as our optimal learning rate.\n\n## Use tf.function\n\nEager execution mode, which is turned on by default in TensorFlow 2.0, allows users to execute code line by line and immediately see the results. This is immensely helpful in development and debugging. This is in contrast to TensorFlow 1.x, for which the user had to build all operations as a graph and then execute them in one go to see the results. This made debugging a nightmare!\n\nDoes the added flexibility from eager execution come at a cost? Yes, a tiny one, typically in the order of microseconds, which can essentially be ignored for large compute-intensive operations, like training ResNet-50. But where there are many small operations, eager execution can have a sizable impact.\n\nWe can overcome this by two approaches:\n\nDisabling eager execution\n\nFor TensorFlow 1.x, not enabling eager execution will let the system optimize the program flow as a graph and run it faster.\n\nUse `tf.function`\n\nIn TensorFlow 2.x, you cannot disable eager execution (there is a compatibility API, but we shouldn't be using that for anything other than migration from TensorFlow 1.x). Instead, any function that could benefit from a speedup by executing in graph mode can simply be annotated with `@tf.function`. It's worth noting that any function that is called within an annotated function will also run in graph mode. This gives us the advantage of speedup from graph-based execution without sacrificing the debugging capabilities of eager execution. Typically, the best speedup is observed on short computationally intensive tasks:\n\n    conv_layer = tf.keras.layers.Conv2D(224, 3)\n\n    def non_tf_func(image):\n      for _ in range(1,3):\n            conv_layer(image)\n      return\n\n    @tf.function\n    def tf_func(image):\n      for _ in range(1,3):\n            conv_layer(image)\n      return\n\n    mat = tf.zeros([1, 100, 100, 100])\n\n    # Warm up\n    non_tf_func(mat)\n    tf_func(mat)\n\n    print(\"Without @tf.function:\", timeit.timeit(lambda: non_tf_func(mat),\n    \t  number=10000), \" seconds\")\n    print(\"With @tf.function:\", timeit.timeit(lambda: tf_func(mat), number=10000),\n    \t  \"seconds\")\n\n    =====Output=====\n    Without @tf.function: 7.234016112051904 seconds\n    With @tf.function:    0.7510978290811181 seconds\n\nAs we can see in our contrived example, simply attributing a function with `@tf.function` has given us a speedup of 10 times, from 7.2 seconds to 0.7 seconds.\n\n## Overtrain, and Then Generalize\n\nIn machine learning, overtraining on a dataset is considered to be harmful. However, we will demonstrate that we can use overtraining in a controlled fashion to our advantage to make training faster.\n\nAs the saying goes, \"The perfect is the enemy of the good.\" We don't want our network to be perfect right off the bat. In fact, we wouldn't even want it to be any good initially. What we really want instead is for it to be learning _something_ quickly, even if imperfectly. Because then we have a good baseline that we can fine tune to its highest potential. And experiments have shown that we can get to the end of the journey faster than training conventionally.\n\n###### Note\n\nTo further clarify the idea of overtraining and then generalizing, let's look at an imperfect analogy of language learning. Suppose that you want to learn French. One way is to throw a book of vocabulary and grammar at you and expect you to memorize everything. Sure, you might go through the book every day and maybe in a few years, you might be able to speak some French. But this would not be the optimal way to learn.\n\nAlternatively, we could look at how language learning programs approach this process. These programs introduce you to only a small set of words and grammatical rules initially. After you have learned them, you will be able to speak some broken French. Maybe you could ask for a cup of coffee at a restaurant or ask for directions at a bus stop. At this point, you will be introduced constantly to a larger set of words and rules, and this will help you to improve over time.\n\nThis process is similar to how our model would learn gradually with more and more data.\n\nHow do we force a network to learn quickly and imperfectly? Make it overtrain on our data. The following three strategies can help.\n\n### Use progressive sampling\n\nOne approach to overtrain and then generalize is to progressively show more and more of the original training set to the model. Here's a simple implementation:\n\n  1. Take a sample of the dataset (say, roughly 10%).\n\n  2. Train the network until it converges; in other words, until it begins to perform well on the training set.\n\n  3. Train on a larger sample (or even the entire training set).\n\nBy repeatedly showing a smaller sample of the dataset, the network will learn features much more quickly, but only related to the sample shown. Hence, it would tend to overtrain, usually performing better on the training set compared to the test set. When that happens, exposing the training process to the entire dataset will tend to generalize its learning, and eventually the test set performance would increase.\n\n### Use progressive augmentation\n\nAnother approach is to train on the entire dataset with little to no data augmentation at first, and then progressively increase the degree of augmentation.\n\nBy showing the unaugmented images repeatedly, the network would learn patterns faster, and by progressively increasing the degree of augmentation, it would become more robust.\n\n### Use progressive resizing\n\nAnother approach, made famous by Jeremy Howard from fast.ai (which offers free courses on AI), is progressive resizing. The key idea behind this approach is to train first on images scaled down to smaller pixel size, and then progressively fine tune on larger and larger sizes until the original image size is reached.\n\nImages resized by half along both the width and height have a 75% reduction in pixels, and theoretically could lead to an increase in training speed of four times over the original images. Similarly, resizing to a quarter of the original height and width can in the best case lead to 16-times reduction (at a lower accuracy). Smaller images have fewer details visible, forcing the network to instead learn higher-level features including broad shapes and colors. Then, training with larger images will help the network learn the finer details, progressively increasing the test accuracy, as well. Just like a child is taught the high-level concepts first and then progressively exposed to more details in later years, the same concept is applied here to CNNs.\n\n###### Tip\n\nYou can experiment with a combination of any of these methods or even build your own creative methods such as training on a subset of classes and then generalizing to all the classes later.\n\n## Install an Optimized Stack for the Hardware\n\nHosted binaries for open source packages are usually built to run on a variety of hardware and software configurations. These packages try to appeal to the least common denominator. When we do `pip install` on a package, we end up downloading and installing this general-purpose, works-for-everyone binary. This convenience comes at the expense of not being able to take advantage of the specific features offered by a particular hardware stack. This issue is one of the big reasons to avoid installing prebuilt binaries and instead opt for building packages from source.\n\nAs an example, Google has a single TensorFlow package on `pip` that can run on an old Sandy Bridge (second-generation Core i3) laptop as well as a powerful 16-core Intel Xeon server. Although convenient, the downside of this is that this package does not take advantage of the highly powerful hardware of the Xeon server. Hence, for CPU-based training and inference, Google recommends compiling TensorFlow from source to best optimize for the hardware at hand.\n\nOne way to do this manually is by setting the configuration flags for the hardware before building the source code. For example, to enable support for AVX2 and SSE 4.2 instruction sets, we can simply execute the following build command (note the extra `m` character ahead of each instruction set in the command):\n\n    $ bazel build -c opt --copt= **-mavx2** --copt= **-msse4.2**\n    \/\/tensorflow\/tools\/pip_package:build_pip_package\n\nHow do you check which CPU features are available? Use the following command (Linux only):\n\n    $ lscpu | grep Flags\n\n    Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36\n    clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm\n    constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid\n    aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16\n    xtpr pdcm pcid dca sse4_1 **sse4_2** x2apic movbe popcnt\n    tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch\n    cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd ibrs ibpb stibp\n    tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle **avx2** smep bmi2\n    erms invpcid rtm cqm rdt_a rdseed adx smap intel_pt xsaveopt cqm_llc\n    cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts md_clear\n    flush_l1d\n\nBuilding TensorFlow from source with the appropriate instruction set specified as build flags should result in a substantial increase in speed. The downside here is that building from source can take quite some time, at least a couple of hours. Alternatively, we can use Anaconda to download and install a highly optimized variant of TensorFlow, built by Intel on top of their Math Kernel Library for Deep Neural Networks (MKL-DNN). The installation process is pretty straightforward. First, we install the Anaconda package manager. Then, we run the following command:\n\n    _# For Linux and Mac_\n    $ conda install tensorflow\n\n    _# For Windows_\n    $ conda install tensorflow-mkl\n\nOn Xeon CPUs, MKL-DNN often provides upward of two-times speedup in inference.\n\nHow about optimization for GPUs? Because NVIDIA abstracts away the differences between the various GPU internals with the CUDA library, there is usually no need to build from source. Instead, we could simply install a GPU variant of TensorFlow from `pip` (`tensorflow-gpu` package). We recommend the Lambda Stack one-liner installer for convenience (along with NVIDIA drivers, CUDA, and cuDNN).\n\nFor training and inference on the cloud, AWS, Microsoft Azure, and GCP all provide GPU machine images of TensorFlow optimized for their hardware. It's quick to spin up multiple instances and get started. Additionally, NVIDIA offers GPU-accelerated containers for on-premises and cloud setups.\n\n## Optimize the Number of Parallel CPU Threads\n\nCompare the following two examples:\n\n    _# Example 1_\n    X = tf.multiply(A, B)\n    Y = tf.multiply(C, D)\n\n    _# Example 2_\n    X = tf.multiply(A, B)\n    Y = tf.multiply( **X**, C)\n\nThere are a couple of areas in these examples where we can exploit inherent parallelism:\n\nBetween operations\n\nIn example 1, the calculation of Y does not depend on the calculation of X. This is because there is no shared data between those two operations, and thus both of them can execute in parallel on two separate threads.\n\nIn contrast, in example 2, the calculation of Y depends on the outcome of the first operation (X), and so the second statement cannot execute until the first statement completes execution.\n\nThe configuration for the maximum number of threads that can be used for interoperation parallelism is set using the following statement:\n\n    tf.config.threading.set_inter_op_parallelism_threads(num_threads)\n\nThe recommended number of threads is equal to the number of CPUs (sockets) on the machine. This value can be obtained by using the `lscpu` command (Linux only).\n\nPer-operation level\n\nWe can also exploit the parallelism within a single operation. Operations such as matrix multiplications are inherently parallelizable.\n\nFigure 6-8 demonstrates a simple matrix multiplication operation. It's clear that the overall product can be split into four independent calculations. After all, the product between one row of a matrix and one column of another matrix does not depend on the calculations for the other rows and columns. Each of those splits could potentially get its own thread and all four of them could execute at the same time.\n\n###### Figure 6-8. A matrix multiplication for A x B operation with one of the multiplications highlighted\n\nThe configuration for the number of threads that can be used for intraoperation parallelism is set using the following statement:\n\n    tf.config.threading.set_intra_op_parallelism_threads(num_threads)\n\nThe recommended number of threads is equal to the number of cores per CPU. You can obtain this value by using the `lscpu` command on Linux.\n\n## Use Better Hardware\n\nIf you have already maximized performance optimizations and still need faster training, you might be ready for some new hardware. Replacing spinning hard drives with SSDs can go a long way, as can adding one or more better GPUs. And let's not forget, sometimes the CPU can be the culprit.\n\nIn fact, you might not need to spend much money: public clouds like AWS, Azure, and GCP all provide the ability to rent powerful configurations for a few dollars per hour. Best of all, they come with optimized TensorFlow stacks preinstalled.\n\nOf course, if you have the cash to spend or have a rather generous expense account, you could just skip this entire chapter and buy the 2-petaFLOPS NVIDIA DGX-2. Weighing in at 163 kgs (360 pounds), its 16 V100 GPUs (with a total of 81,920 CUDA cores) consume 10 kW of power\u2014the equivalent of seven large window air conditioners. And all it costs is $400,000!\n\n###### Figure 6-9. The $400,000 NVIDIA DGX-2 deep learning system\n\n## Distribute Training\n\n\" _Two lines to scale training horizontally!_ \"\n\nOn a single machine with a single GPU, there's only so far that we can go. Even the beefiest GPUs have an upper limit in compute power. Vertical scaling can take us only so far. Instead, we look to scale horizontally\u2014distribute computation across processors. We can do this across multiple GPUs, TPUs, or even multiple machines. In fact, that is exactly what researchers at Google Brain did back in 2012, using 16,000 processors to run a neural network built to look at cats on YouTube.\n\nIn the dark days of the early 2010s, training on ImageNet used to take anywhere from several weeks to months. Multiple GPUs would speed things up, but few people had the technical know-how to configure such a setup. It was practically out of reach for beginners. Luckily, we live in the day of TensorFlow 2.0, in which setting up distributed training is a matter of introducing two lines of code:\n\n    mirrored_strategy = tf.distribute.MirroredStrategy()\n    with mirrored_strategy.scope():\n      model = tf.keras.applications.ResNet50()\n      model.compile(loss=\"mse\", optimizer=\"sgd\")\n\nTraining speed increases nearly proportionally (90\u201395%) in relation to the number of GPUs added. As an example, if we added four GPUs (of similar compute power), we would notice an increase of >3.6 times speedup ideally.\n\nStill, a single system can only support a limited number of GPUs. How about multiple nodes, each with multiple GPUs? Similar to `MirroredStrategy`, we can use `Multi``Worker``MirroredStrategy`. This is quite useful when building a cluster on the cloud. Table 6-1 presents a couple of distribution strategies for different use cases.\n\nTable 6-1. Recommended distribution strategies **Strategy** | **Use case**  \n---|---\n\n`MirroredStrategy`\n\n| Single node with two or more GPUs\n\n`MultiWorkerMirroredStrategy`\n\n| Multiple nodes with one or more GPUs each\n\nTo get the cluster nodes to communicate with one another for `MultiWorkerMirroredStrategy`, we need to configure the `TF_CONFIG` environment variable on every single host. This requires setting up a JSON object that contains the IP addresses and ports of all other hosts in the cluster. Manually managing this can be error prone, and this is where orchestration frameworks like Kubernetes really shine.\n\n###### Note\n\nThe open source Horovod library from Uber is another high-performance and easy-to-use distribution framework. Many of the record benchmark performances seen in the next section require distributed training on several nodes, and Horovod's performance helped them get the edge. It is worth noting that the majority of the industry uses Horovod particularly because distributed training on earlier versions of TensorFlow was a much more involved process. Additionally, Horovod works with all major deep learning libraries with minimal amount of code change or expertise. Often configured through the command line, running a distributed program on four nodes, each with four GPUs, can be done in a single command line:\n\n    $ horovodrun -np 16 -H\n    server1:4,server2:4,server3:4,server4:4 python\n    train.py\n\n## Examine Industry Benchmarks\n\nThree things were universally popular in the 1980s\u2014long hair, the Walkman, and database benchmarks. Much like the current hype of deep learning, database software was similarly going through a phase of making bold promises, some of which were marketing hype. To put these companies to the test, a few benchmarks were introduced, more famously among them was the Transaction Processing Council (TPC) benchmark. When someone needed to buy database software, they could rely on this public benchmark to decide where to spend their company's budget. This competition fueled rapid innovation, increasing speed and performance per dollar, moving the industry ahead faster than anticipated.\n\nInspired by TPC and other benchmarks, a few system benchmarks were created to standardize performance reporting in machine learning.\n\nDAWNBench\n\nStanford's DAWNBench benchmarks time and cost to get a model to 93% Top-5 accuracy on ImageNet. Additionally, it also does a time and cost leaderboard on inference time. It's worth appreciating the rapid pace of performance improvement for training such a massive network. When DAWNBench originally started in September 2017, the reference entry trained in 13 days at a cost of $2,323.39. In just one and a half years since then, although the cheapest training costs as low as $12, the fastest training time is 2 minutes 43 seconds. Best of all, most entries contain the training source code and optimizations that can be studied and replicated by us. This gives further guidance on the effects of hyperparameters and how we can use the cloud for cheap and fast training without breaking the bank.\n\nTable 6-2. Entries on DAWNBench as of August 2019, sorted by the least cost for training a model to 93% Top-5 accuracy **Cost (USD)** | **Training time** | **Model** | **Hardware** | **Framework**  \n---|---|---|---|---  \n$12.60 | 2:44:31 |\n\nResNet-50\n\nGoogle Cloud TPU\n\n| GCP n1-standard-2, Cloud TPU | TensorFlow 1.11  \n$20.89 | 1:42:23 |\n\nResNet-50\n\nSetu Chokshi\n\n(MS AI MVP)\n\n| Azure ND40s_v2 | PyTorch 1.0  \n$42.66 | 1:44:34 |\n\nResNet-50 v1\n\nGE Healthcare\n\n(Min Zhang)\n\n| 8*V100 (single p3.16x large) | TensorFlow 1.11 + Horovod  \n$48.48 | 0:29:43 |\n\nResNet-50\n\nAndrew Shaw, Yaroslav Bulatov, Jeremy Howard | 32 * V100\n\n(4x - AWS p3.16x large)\n\n| Ncluster + PyTorch 0.5\n\nMLPerf\n\nSimilar to DAWNBench, MLPerf is aimed at repeatable and fair testing of AI system performance. Although newer than DAWNBench, this is an industry consortium with much wider support, especially on the hardware side. It runs challenges for both training and inference in two divisions: open and closed. The closed division trains the same model with the same optimizers, so the raw hardware performance can be compared apples-to-apples. The open division, on the other hand, allows using faster models and optimizers to allow for more rapid progress. Compared to the more cost-effective entries in DAWNBench in Table 6-2, the top performers on MLPerf as shown in Table 6-3 might be a bit out of reach for most of us. The top-performing NVIDIA DGX SuperPod, composed of 96 DGX-2H with a total of 1,536 V100 GPUs, costs in the $35 to $40 million range. Even though 1,024 Google TPUs might themselves cost in the several millions, they are each available to rent on the cloud at $8\/hour on-demand pricing (as of August 2019), resulting in a net cost of under $275 for the less-than two minutes of training time.\n\nTable 6-3. Key closed-division entries on DAWNBench as of August 2019, showing training time for a ResNet-50 model to get to 75.9% Top-1 accuracy **Time (minutes)** | **Submitter** | **Hardware** | **Accelerator** | **# of accelerators**  \n---|---|---|---|---  \n1.28 | Google | TPUv3 | TPUv3 | 1,024  \n1.33 | NVIDIA | 96x DGX-2H | Tesla V100 | 1,536  \n8,831.3 | Reference | Pascal P100 | Pascal P100 | 1\n\nAlthough both the aforementioned benchmarks highlight training as well as inference (usually on more powerful devices), there are other inference-specific competitions on low-power devices, with the aim to maximize accuracy and speed while reducing power consumption. Held at annual conferences, here are some of these competitions:\n\n  * LPIRC: Low-Power Image Recognition Challenge\n\n  * EDLDC: Embedded Deep Learning Design Contest\n\n  * System Design Contest at Design Automation Conference (DAC)\n\n# Inference\n\nTraining our model is only half the game. We eventually need to serve the predictions to our users. The following points guide you to making your serving side more performant.\n\n## Use an Efficient Model\n\nDeep learning competitions have traditionally been a race to come up with the highest accuracy model, get on top of the leaderboard, and get the bragging rights. But practitioners live in a different world\u2014the world of serving their users quickly and efficiently. With devices like smartphones, edge devices, and servers with thousands of calls per second, being efficient on all fronts (model size and computation) is critically needed. After all, many machines would not be capable of serving a half gigabyte VGG-16 model, which happens to need 30 billion operations to execute, for not even that high of accuracy. Among the wide variety of pretrained architectures available, some are on the higher end of accuracy but large and resource intensive, whereas others provide modest accuracy but are much lighter. Our goal is to pick the architecture that can deliver the highest accuracy for the available computational power and memory budget of our inference device. In Figure 6-10, we want to pick models in the upper-left zone.\n\n###### Figure 6-10. Comparing different models for size, accuracy, and operations per second (adapted from \"An Analysis of Deep Neural Network Models for Practical Applications\" by Alfredo Canziani, Adam Paszke, and Eugenio Culurciello)\n\nUsually, the approximately 15 MB MobileNet family is the go-to model for efficient smartphone runtimes, with more recent versions like MobileNetV2 and MobileNetV3 being better than their predecessors. Additionally, by varying the hyperparameters of the MobileNet models like depth multiplier, the number of computations can be further reduced, making it ideal for real-time applications. Since 2017, the task of generating the most optimal architecture to maximize accuracy has also been automated with NAS. It has helped discover new (rather obfuscated looking) architectures that have broken the ImageNet accuracy metric multiple times. For example, FixResNeXt (based on PNASNet architecture at 829 MB) reaches a whopping 86.4% Top-1 accuracy on ImageNet. So, it was natural for the research community to ask whether NAS helps find architecture that's tuned for mobile, maximizing accuracy while minimizing computations. The answer is a resounding yes\u2014resulting in faster and better models, optimized for the hardware at hand. As an example, MixNet (July 2019) outperforms many state-of-the-art models. Note how we went from billions of floating-point operations to millions (Figure 6-10 and Figure 6-11).\n\n###### Figure 6-11. Comparison of several mobile-friendly models in the paper \"MixNet: Mixed Depthwise Convolution Kernels\" by Mingxing Tan and Quoc V. Le\n\nAs practitioners, where can we find current state-of-the-art models? _PapersWithCode.com\/SOTA_ showcases leaderboards on several AI problems, comparing paper results over time, along with the model code. Of particular interest would be the models with a low number of parameters that achieve high accuracies. For example, EfficientNet gets an amazing Top-1 84.4% accuracy with 66 million parameters, so it could be an ideal candidate for running on servers. Additionally, the ImageNet test metrics are on 1,000 classes, whereas our case might just require classification on a few classes. For those cases, a much smaller model would suffice. Models listed in Keras Application ( _tf.keras.applications_ ), TensorFlow Hub, and TensorFlow Models usually carry many variations (input image sizes, depth multipliers, quantizations, etc.).\n\n###### Tip\n\nShortly after Google AI researchers publish a paper, they release the model used in the paper on the TensorFlow Models repository.\n\n## Quantize the Model\n\n\" _Represent 32-bit weights to 8-bit integer, get 2x faster, 4x smaller models_ \"\n\nNeural networks are driven primarily by matrix\u2013matrix multiplications. The arithmetic involved tends to be rather forgiving in that small deviations in values do not cause a significant swing in output. This makes neural networks fairly robust to noise. After all, we want to be able to recognize an apple in a picture, even in less-than-perfect lighting. When we quantize, we essentially take advantage of this \"forgiving\" nature of neural networks.\n\nBefore we look at the different quantization techniques, let's first try to build an intuition for it. To illustrate quantized representations with a simple example, we'll convert 32-bit floating-point weights to INT8 (8-bit integer) using _linear quantization_. Obviously, FP32 represents 232 values (hence, 4 bytes to store), whereas INT8 represents 28 = 256 values (1 byte). To quantize:\n\n  1. Find the minimum and maximum values represented by FP32 weights in the neural network.\n\n  2. Divide this range into 256 intervals, each corresponding to an INT8 value.\n\n  3. Calculate a scaling factor that converts an INT8 (integer) back to a FP32. For example, if our original range is from 0 to 1, and INT8 numbers are 0 to 255, the scaling factor will be 1\/256.\n\n  4. Replace FP32 numbers in each interval with the INT8 value. Additionally, store the scaling factor for the inference stage where we convert INT8 values back to FP32 values. This scaling factor only needs to be stored once for the entire group of quantized values.\n\n  5. During inference calculations, multiply the INT8 values by the scaling factor to convert it back to a floating-point representation. Figure 6-12 illustrates an example of linear quantization for the interval [0, 1].\n\n###### Figure 6-12. Quantizing from a 0 to 1 32-bit floating-point range down to an 8-bit integer range for reduced storage space\n\nThere are a few different ways to quantize our models, the simplest one being reducing the bit representation of the weights from 32-bit to 16-bit or lower. As might be evident, converting 32-bit to 16-bit means half the memory size is needed to store a model. Similarly, converting to 8-bit would require a quarter of the size. So why not convert it to 1-bit and save 32x the size? Well, although the models are forgiving up to a certain extent, with each reduction, we will notice a drop in accuracy. This reduction in accuracy grows exponentially beyond a certain threshold (especially below 8 bits). To go below and still have a useful working model (like a 1-bit representation), we'd need to follow a special conversion process to convert them to binarized neural networks. XNOR.ai, a deep learning startup, has famously been able to bring this technique to production. The Microsoft Embedded Learning Library (ELL) similarly provides such tools, which have a lot of value for edge devices like the Raspberry Pi.\n\nThere are numerous benefits to quantization:\n\nImproved memory usage\n\nBy quantizing to 8-bit integer representation (INT8), we typically get a 75% reduction in model size. This makes it more convenient to store and load the model in memory.\n\nImproved performance\n\nInteger operations are faster than floating-point operations. Additionally, the savings in memory usage reduces the likelihood of having to unload the model from RAM during execution, which also has the added benefit of decreased power consumption.\n\nPortability\n\nEdge devices such as Internet of Things devices might not support floating-point arithmetic, so it would be untenable to keep the model as a floating-point in such situations.\n\nMost inference frameworks provide a way to quantize, including Core ML Tools from Apple, TensorRT from NVIDIA (for servers), and TensorFlow Lite, as well as the TensorFlow Model Optimization Toolkit from Google. With TensorFlow Lite, models can be quantized after training during conversion (called post-training quantization). To minimize accuracy losses even further, we can use the TensorFlow Model Optimization Toolkit during training. This process is called _quantization-aware training_.\n\nIt would be useful to measure the benefit provided by quantization. Metrics from the TensorFlow Lite Model optimization benchmarks (shown in Table 6-4) give us a hint, comparing 1) unquantized, 2) post-training quantized, and 3) quantization-aware trained models. The performance was measured on a Google Pixel 2 device.\n\nTable 6-4. Effects of different quantization strategies (8-bit) on models (source: TensorFlow Lite model optimization documentation) **Model** | **MobileNet** | **MobileNetV2** | **InceptionV3**  \n---|---|---|---\n\n**Top-1 accuracy** | **Original**\n\n| 0.709 | 0.719 | 0.78\n\n**Post-training quantized**\n\n| 0.657 | 0.637 | 0.772\n\n**Quantization-aware training**\n\n| 0.7 | 0.709 | 0.775\n\n**Latency (ms)** | **Original**\n\n| 124 | 89 | 1130\n\n**Post-training quantized**\n\n| 112 | 98 | 845\n\n**Quantization-aware training**\n\n| 64 | 54 | 543\n\n**Size (MB)** | **Original**\n\n| 16.9 | 14 | 95.7\n\n**Optimized**\n\n| 4.3 | 3.6 | 23.9\n\nSo, what do these numbers indicate? After quantization using TensorFlow Lite to INT8, we see roughly a four-times reduction in size, approximately two-times speedup in run time, and less than 1% change in accuracy. Not bad!\n\nMore extreme form of quantization, like 1-bit binarized neural networks (like XNOR-Net), claim a whopping 58-times speedup with roughly 32-times smaller size when tested on AlexNet, with a 22% loss in accuracy.\n\n## Prune the Model\n\nPick a number. Multiply it by 0. What do we get? Zero. Multiply your pick again by a small value neighboring 0, like 10\u20136, and we'll still get an insignificant value. If we replace such tiny weights (\u2192 0) in a model with 0 itself, it should have little effect on the model's predictions. This is called _magnitude-based weight pruning,_ or simply pruning, and is a form of _model compression_. Logically, putting a weight of 0 between two nodes in a fully connected layer is equivalent to deleting the edge between them. This makes a model with dense connections sparser.\n\nAs it happens, a large chunk of the weights in a model are close to 0. Pruning the model will result in many of those weights being set to 0. This happens with little impact to accuracy. Although this does not save any space by itself, it introduces a ton of redundancy that can be exploited when it comes time to save the model to disk in a compressed format such as ZIP. (It is worth noting that compression algorithms thrive on repeating patterns. The more the repetition, the higher the compressibility.) The end result is that our model can often be compressed by four times. Of course, when we finally need to use the model, it would need to be uncompressed before loading in memory for inference.\n\nThe TensorFlow team observed the accuracy loss shown in Table 6-5 while pruning the models. As expected, more efficient models like MobileNet observe higher (though still small) accuracy loss when compared with comparatively bigger models like InceptionV3.\n\nTable 6-5. Model accuracy loss versus pruning percentage **Model** | **Sparsity** | **Accuracy loss against original accuracy**  \n---|---|---  \nInceptionV3 | 50% | 0.1%  \nInceptionV3 | 75% | 2.5%  \nInceptionV3 | 87.5% | 4.5%  \nMobileNet | 50% | 2%\n\nKeras provides APIs to prune our model. This process can be done iteratively during training. Train a model normally or pick a pretrained model. Then, periodically prune the model and continue training. Having enough epochs between the periodic prunes allows the model to recover from any damage due to introducing so much sparsity. The amount of sparsity and number of epochs between prunes can be treated as hyperparameters to be tuned.\n\nAnother way of implementing this is by using Tencent's PocketFlow tool, a one-line command that provides several other pruning strategies implemented in recent research papers.\n\n## Use Fused Operations\n\nIn any serious CNN, the convolutional layer and batch normalization layer frequently appear together. They are kind of the Laurel and Hardy of CNN layers. Fundamentally, they are both linear operations. Basic linear algebra tells us that combining two or more linear operations will also result in a linear operation. By combining convolutional and batch normalization layers, we not only reduce the number of computations, but also decrease the amount of time spent in data transfer, both between main memory and GPU, and main memory and CPU registers\/cache. Making them one operation prevents an extra roundtrip. Luckily, for inference purposes, most inference frameworks either automatically do this fusing step or provide model converters (like TensorFlow Lite) to make this optimization while converting the model to the inference format.\n\n## Enable GPU Persistence\n\nLoading and initializing the GPU drivers take time. You might have noticed a delay every time a training or inference job was initiated. For frequent, short jobs, the overhead can become relatively expensive quickly. Imagine an image classification program for which the classification takes 10 seconds, 9.9 of which were spent in loading the driver. What we need is for the GPU driver to stay preinitialized in the background, and be ready for whenever our training jobs start. And that's where the NVIDIA GPU Persistence Daemon comes to the rescue:\n\n    $ nvidia-persistenced --user _{YOUR_USERNAME}_\n\nOur GPUs will use a bit more wattage during idle time, but they will be ready and available the next time a program is launched.\n\n# Summary\n\nIn this chapter, we explored different avenues for improving the speed and performance of our deep learning pipeline, from storing and reading the data to inference. A slow data pipeline often leads to a GPU starving for data, resulting in idle cycles. With several of the simple optimizations we discussed, our hardware can be put to its maximum efficiency. The handy checklist can serve as a ready reference. Feel free to make a copy for your desk (or your refrigerator). With these learnings, we hope to see your entry among the top performers of the MLPerf benchmark list.\n\n# Chapter 7. Practical Tools, Tips, and Tricks\n\nThis chapter contains material that we, your authors, have encountered during our professional work as well as while working on this book, primarily during experimentation. The material covered here doesn't necessarily fit in any single chapter; rather, it's material that deep learning practitioners could find useful on a day-to-day basis across a variety of tasks. In line with the \"practical\" theme, these questions cover a range of helpful pragmatic guidelines across topics including setting up an environment, training, model interoperability, data collection and labeling, code quality, managing experiments, team collaboration practices, privacy, and further exploration topics.\n\nDue to the fast-changing pace of the AI field, this chapter is a small subset of the \"living\" document hosted on the book's Github repository (see _http:\/\/PracticalDeepLearning.ai_) at _code\/chapter-9_ , where it is constantly evolving. If you have more questions or, even better, answers that might help other readers, feel free to tweet them @PracticalDLBook or submit a pull request.\n\n# Installation\n\n**Q:** _I came across an interesting and useful Jupyter Notebook on GitHub. Making the code run will require cloning the repository, installing packages, setting up the environment, and more steps. Is there an instant way to run it interactively?_\n\nSimply enter the Git repository URL into Binder ( _mybinder.org_ ), which will turn it into a collection of interactive notebooks. Under the hood, it will search for a dependency file, like _requirements.txt_ or _environment.yml_ in the repository's root directory. This will be used to build a Docker image, to help run the notebook interactively in your browser.\n\n**Q:** _What is the quickest way to get my deep learning setup running on a fresh Ubuntu machine with NVIDIA GPUs?_\n\nLife would be great if `pip install tensorflow-gpu` would solve everything. However, that's far from reality. On a freshly installed Ubuntu machine, listing all the installation steps would take at least three pages and more than an hour to follow, including installing NVIDIA GPU drivers, CUDA, cuDNN, Python, TensorFlow, and other packages. And then it requires carefully checking the version interoperability between CUDA, cuDNN and TensorFlow. More often than not, this ends in a broken system. A world of pain to say the least!\n\nWouldn't it be great if two lines could solve all of this effortlessly? Ask, and ye shall receive:\n\n    $ sudo apt update && sudo ubuntu-drivers autoinstall && sudo reboot \n    $ export LAMBDA_REPO=$(mktemp) \\\n    && wget -O${LAMBDA_REPO} \\\n    https:\/\/lambdalabs.com\/static\/misc\/lambda-stack-repo.deb \\\n    && sudo dpkg -i ${LAMBDA_REPO} && rm -f ${LAMBDA_REPO} \\\n    && sudo apt-get update && sudo apt-get install -y lambda-stack-cuda \\\n    && sudo reboot\n\nThe first line ensures that all the drivers are updated. The second line is brought to us by the Lambda Labs, a San Francisco\u2013based deep learning hardware and cloud provider. The command sets up the Lambda Stack, which installs TensorFlow, Keras, PyTorch, Caffe, Caffe2, Theano, CUDA, cuDNN, and NVIDIA GPU drivers. Because the company needs to install the same deep learning packages on thousands of machines, it automated the process with a one-line command and then open sourced it so that others can also make use of it.\n\n**Q:** _What is the fastest way to install TensorFlow on a Windows PC?_\n\n  1. Install Anaconda Python 3.7.\n\n  2. On the command line, run `conda install tensorflow-gpu`.\n\n  3. If you do not have GPUs, run `conda install tensorflow`.\n\nOne additional benefit of a CPU-based Conda installation is that it installs Intel MKL optimized TensorFlow, running faster than the version we get by using `pip install tensorflow`.\n\n**Q:** _I have an AMD GPU. Could I benefit from GPU speedups in TensorFlow on my existing system?_\n\nAlthough the majority of the deep learning world uses NVIDIA GPUs, there is a growing community of people running on AMD hardware with the help of the ROCm stack. Installation using the command line is simple:\n\n  1. `sudo apt install rocm-libs miopen-hip cxlactivitylogger`\n\n  2. `sudo apt install wget python3-pip`\n\n  3. `pip3 install --user tensorflow-rocm`\n\n**Q:** _Forget installation, where can I get preinstalled deep learning containers?_\n\nDocker is synonymous with setting up environments. Docker helps run isolated containers that are bundled with tools, libraries, and configuration files. There are several deep learning Docker containers available while selecting your virtual machine (VM) from major cloud providers AWS, Microsoft Azure, GCP, Alibaba, etc.) that are ready to start working. NVIDIA also freely provides NVIDIA GPU Cloud containers, which are the same high-performance containers used to break training speed records on the MLPerf benchmarks. You can even run these containers on your desktop machine.\n\n# Training\n\n**Q:** _I don't like having to stare at my screen constantly to check whether my training finished. Can I get a notification alert on my phone, instead?_\n\nUse Knock Knock, a Python library that, as the name suggests, notifies you when your training ends (or your program crashes) by sending alerts on email, Slack, or even Telegram! Best of all, it requires adding only two lines of code to your training script. No more opening your program a thousand times to check whether the training has finished.\n\n**Q:** _I prefer graphics and visualizations over plain text. Can I get real-time visualizations for my training process?_\n\nFastProgress progress bar (originally developed for fast.ai by Sylvain Gugger) comes to the rescue.\n\n**Q:** _I conduct a lot of experiments iteratively and often lose track of what changed between each experiment as well as the effect of the change. How do I manage my experiments in a more organized manner?_\n\nSoftware development has had the ability to keep a historical log of changes through version control. Machine learning, unfortunately, did not have the same luxury. That's changing now with tools like Weights and Biases, and Comet.ml. They allow you to keep track of multiple runs and to log training curves, hyperparameters, outputs, models, notes, and more with just two lines of code added to your Python script. Best of all, through the power of the cloud, you can conveniently track experiments even if you are away from the machine, and share the results with others.\n\n**Q:** _How do I check whether TensorFlow is using the GPU(s) on my machine?_\n\nUse the following handy command:\n\n    tf.test.is_gpu_available()\n\n**Q:** _I have multiple GPUs on my machine. I don't want my training script to consume all of them. How do I restrict my script to run on only a specific GPU?_\n\n_Use_ `CUDA_VISIBLE_DEVICES=GPU_ID`. Simply prefix the training script command as follows:\n\n    $ CUDA_VISIBLE_DEVICES=GPU_ID python train.py\n\nAlternatively, write the following lines early on in your training script:\n\n    import os\n    os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"GPU_ID\"\n\n`GPU_ID` can have values such as 0, 1, 2, and so on. You can see these IDs (along with GPU usage) using the `nvidia-smi` command. For assigning to multiple GPUs, use a comma-separated list of IDs.\n\n**Q:** _Sometimes it feels like there are too many knobs to adjust when training. Can it be done automatically, instead, to get the best accuracy?_\n\nThere are many options for automated hyperparameter tuning, including Keras-specific Hyperas and Keras Tuner, and more generic frameworks such as Hyperopt and Bayesian optimization that perform extensive experimentation to maximize our objective (i.e., maximizing accuracy in our case) more intelligently than simple grid searches.\n\n**Q:** _ResNet and MobileNet work well enough for my use case. Is it possible to build a model architecture that can achieve even higher accuracy for my scenario?_\n\nThree words: Neural Architecture Search (NAS). Let the algorithm find the best architecture for you. NAS can be accomplished through packages like Auto-Keras and AdaNet.\n\n**Q:** _How do I go about debugging my TensorFlow script?_\n\nThe answer is in the question: TensorFlow Debugger (`tfdbg)`.\n\n# Model\n\n**Q:** _I want to quickly know the input and output layers of my model without writing code. How can I accomplish that?_\n\nUse Netron. It graphically shows your model, and on clicking any layer, provides details on the architecture.\n\n**Q:** _I need to publish a research paper. Which tool should I use to draw my organic, free-range, gluten-free model architecture?_\n\nMS Paint, obviously! No, we're just kidding. We are fans of NN-SVG as well as PlotNeuralNet for creating high-quality CNN diagrams.\n\n**Q:** _Is there a one-stop shop for all models?_\n\nIndeed! Explore _PapersWithCode.com_ , _ModelZoo.co_ , and _ModelDepot.io_ for some inspiration.\n\n**Q:** _I've finished training my model. How can I make it available for others to use?_\n\nYou can begin by making the model available for download from GitHub. And then list it on the model zoos mentioned in the previous answer. For even wider adoption, upload it to TensorFlow Hub ( _tfhub.dev_ ).\n\nIn addition to the model, you should publish a \"model card,\" which is essentially like a r\u00e9sum\u00e9 of the model. It's a short report that details author information, accuracy metrics, and the dataset it was benchmarked on. Additionally, it provides guidance on potential biases and out-of-scope uses.\n\n**Q:** _I have a model previously trained in framework X, but I need to use it in framework Y. Do I need to waste time retraining it in framework Y?_\n\nNope. All you need is the power of the ONNX. For models not in the TensorFlow ecosystem, most major deep learning libraries support saving them in ONNX format, which can then be converted to the TensorFlow format. Microsoft's MMdnn can help in this conversion.\n\n# Data\n\n**Q:** _Could I collect hundreds of images on a topic in a few minutes?_\n\nYes, you can collect hundreds of images in three minutes or less with a Chrome extension called Fatkun Batch Download Image. Simply search for a keyword in your favorite image search engine, filter images by the correct usage rights (e.g., Public Domain), and press the Fatkun extension to download all images. See Chapter 12, where we use it to build a Not Hotdog app.\n\nBonus tip: to download from a single website, search for a keyword followed by site:website_address. For example, \"horse site:flickr.com.\"\n\n**Q:** _Forget the browser. How do I scrape Google for images using the command line?_\n\n    $ pip install google_images_download\n    $ googleimagesdownload -k=horse -l=50 -r=labeled-for-reuse\n\n`-k`, `-l`, and `-r` are shorthand for `keyword`, `limit` (number of images), and `usage_rights`, respectively. This is a powerful tool with many options for controlling and filtering what images to download from Google searches. Plus, instead of just loading the thumbnails shown by Google Images, it saves the original images linked by the search engine. For saving more than 100 images, install the `selenium` library along with `chromedriver`.\n\n**Q:** _Those were not enough for collecting images. I need more control. What other tools can help me download data in more custom ways beyond the search engine?_\n\nWith a GUI (no programming needed):\n\nScrapeStorm.com\n\nEasy GUI to identify rules for elements to extract\n\nWebScraper.io\n\nChrome-based scraping extension, especially for extracting structured output from single websites\n\n80legs.com\n\nCloud-based scalable scraper, for parallel, large tasks\n\nPython-based programmatic tools:\n\nScrapy.org\n\nFor more programmable controls on scraping, this is one of the most famous scrapers. Compared to building your own naive scraper to explore websites, it offers throttling rate by domain, proxy, and IP; can handle _robots.txt_ ; offers flexibility in browser headers to show to web servers; and takes care of several possible edge cases.\n\nInstaLooter\n\nA Python-based tool for scraping Instagram.\n\n**Q:** _I have the images for the target classes, but now need images for the negative (not item\/background) class. Any quick ways to build a big dataset of negative classes?_\n\nImageN offers 1,000 images\u20145 random images for 200 ImageNet categories\u2014which you can use as the negative class. If you need more, download a random sample programmatically from ImageNet.\n\n**Q:** _How can I search for a prebuilt dataset that suits my needs?_\n\nTry Google Dataset Search, _VisualData.io_ , and _DatasetList.com_.\n\n**Q:** _For datasets like ImageNet, downloading, figuring out the format, and then loading them for training takes far too much time. Is there an easy way to read popular datasets?_\n\nTensorFlow Datasets is a growing collection of datasets ready to use with TensorFlow. It includes ImageNet, COCO (37 GB), and Open Images (565 GB) among others. These datasets are exposed as `tf.data.Datasets`, along with performant code to feed them in your training pipeline.\n\n**Q:** _Training on the millions of ImageNet images will take a long, long time. Is there a smaller representative dataset I could try training on, to quickly experiment and iterate with?_\n\nTry Imagenette. Built by Jeremy Howard from fast.ai, this 1.4 GB dataset contains only 10 classes instead of 1,000.\n\n**Q:** _What are the largest readily available datasets that I could use for training?_\n\n  * Tencent ML Images: 17.7 million images with 11,000 category labels\n\n  * Open Images V4 (from Google): 9 million images in 19.7 K categories\n\n  * BDD100K (from UC Berkeley): Images from 100,000 driving videos, over 1,100 hours\n\n  * YFCC100M (from Yahoo): 99.2 million images\n\n**Q:** _What are some of the readily available large video datasets I could use?_\n\n**Name** | **Details**  \n---|---  \nYouTube-8M |\n\n6.1 million videos, 3,862 classes, 2.6 billion audio-visual features\n\n3.0 labels\/video\n\n1.53 terabytes of randomly sampled videos\n\nSomething Something\n\n(from Twenty Billion Neurons) | 221,000 videos in 174 action classes\n\nFor example, \"Pouring water into wine glass but missing so it spills next to it\"\n\nHumans performing predefined actions with everyday objects\n\nJester\n\n(from Twenty Billion Neurons) | 148,000 videos in 27 classes\n\nFor example, \"Zooming in with two fingers\"\n\nPredefined hand gestures in front of a webcam\n\n**Q:** _Are those the largest labeled datasets ever assembled in the history of time?_\n\nNope! Companies like Facebook and Google curate their own private datasets that are much larger than the public ones we can play with:\n\n  * Facebook: 3.5 billion Instagram images with noisy labels (first reported in 2018)\n\n  * Google \u2013 JFT-300M: 300 million images with noisy labels (first reported in 2017)\n\nSadly, unless you're an employee at one of these companies, you can't really access these datasets. Nice recruiting tactic, we must say.\n\n**Q:** _How can I get help annotating data?_\n\nThere are several companies out there that can assist with labeling different kinds of annotations. A few worth mentioning include SamaSource, Digital Data Divide, and iMerit, which employ people who otherwise have limited opportunities, eventually creating positive socioeconomic change through employment in underprivileged communities.\n\n**Q:** _Is there a versioning tool for datasets, like Git is for code?_\n\nQri and Quilt can help version control our datasets, aiding in reproducibility of experiments.\n\n**Q:** _What if I don't have access to a large dataset for my unique problem?_\n\nTry to develop a synthetic dataset for training! For example, find a realistic 3D model of the object of interest and place it in realistic environments using a 3D framework such as Unity. Adjust the lighting and camera position, zoom, and rotation to take snapshots of this object from many angles, generating an endless supply of training data. Alternatively, companies like AI.Reverie, CVEDIA, Neuromation, Cognata, Mostly.ai, and DataGen Tech provide realistic simulations for training needs. One big benefit of synthesized training data is that the labeling process is built into the synthesization process. After all, you would know what you are creating. This automatic labeling can save a lot of money and effort, compared to manual labeling.\n\n# Privacy\n\n**Q:** _How do I develop a more privacy-preserving model without going down the cryptography rabbit hole?_\n\nTensorFlow Encrypted might be the solution you're looking for. It enables development using encrypted data, which is relevant, especially if you are on the cloud. Internally, lots of secure multiparty computation and homomorphic encryptions result in privacy-preserving machine learning.\n\n**Q:** _Can I keep my model under wraps from prying eyes?_\n\nWell, unless you are on the cloud, weights are visible and can be reverse engineered. Use the Fritz library for protecting your model's IP when deployed on smartphones.\n\n# Education and Exploration\n\n**Q:** _I want to become an AI expert. Beyond this book, where should I invest my time to learn more?_\n\nThere are several resources on the internet to learn deep learning in depth. We highly recommend these video lectures from some of the best teachers, covering a variety of application areas from computer vision to natural language processing.\n\n  * Fast.ai (by Jeremy Howard and Rachel Thomas) features a free 14-video lecture series, taking a more learn-by-doing approach in PyTorch. Along with the course comes an ecosystem of tools and an active community that has led to many breakthroughs in the form of research papers and ready-to-use code (like three lines of code to train a state-of-the-art network using the fast.ai library).\n\n  * Deeplearning.ai (by Andrew Ng) features a five-course \"Deep Learning Specialization.\" It's free of cost (although you could pay a small fee to get a certificate) and will solidify your theoretical foundation further. Dr. Ng's first Coursera course on machine learning has taught more than two million students, and this series continues the tradition of highly approachable content loved by beginners and experts alike.\n\n  * We would be remiss if we didn't encourage you to note O'Reilly's Online Learning platform in this list. Helping more than two million users advance their careers, it contains hundreds of books, videos, live online trainings, and keynotes given by leading thinkers and practitioners at O'Reilly's AI and data conferences.\n\n**Q:** _Where can I find interesting notebooks to learn from?_\n\nGoogle Seedbank is a collection of interactive machine learning examples. Built on top of Google Colaboratory, these Jupyter notebooks can be run instantly without any installations. Some interesting examples include:\n\n  * Generating audio with GANs\n\n  * Action recognition on video\n\n  * Generating Shakespeare-esque text\n\n  * Audio-style transfer\n\n**Q:** _Where can I learn about the state of the art for a specific topic?_\n\nConsidering how fast the state of the art moves in AI, SOTAWHAT is a handy command-line tool to search research papers for the latest models, datasets, tasks, and more. For example, to look up the latest results on ImageNet, use `sotawhat imagenet` on the command line. Additionally, _paperswithcode.com\/sota_ also features repositories for papers, their source code, and released models, along with an interactive visual timeline of benchmarks.\n\n**Q:** _I am reading a paper on Arxiv and I really like it. Do I need to write code from scratch?_\n\nNot at all! The ResearchCode Chrome extension makes it easy to find code when browsing _arxiv.org_ or Google Scholar. All it takes is a press of the extension button. You can also look up code without installing the extension on the _ResearchCode.com_ website.\n\n**Q:** _I don't want to write any code, but I still want to interactively experiment with a model using my camera. How can I do that?_\n\nRunway ML is an easy-to-use yet powerful GUI tool that allows you to download models (from the internet or your own) and use the webcam or other input, such as video files, to see the output interactively. This allows further combining and remixing outputs of models to make new creations. And all of this happens with just a few mouse clicks; hence, it's attracting a large artist community!\n\n**Q:** _8-1If I can test without code, can I train without code, too?_\n\nWe discuss this in detail in Chapter 8 (web-based) and Chapter 12 (desktop-based). To keep it short, tools such as Microsoft's CustomVision.ai, Google's Cloud AutoML Vision, Clarifai, Baidu EZDL, and Apple's Create ML provide drag-and-drop training capabilities. Some of these tools take as little as a few seconds to do the training.\n\n# One Last Question\n\n**Q:** _Tell me a great deep learning prank?_\n\nPrint and hang poster shown in Figure 7-1 from _keras4kindergartners.com_ near the watercooler, and watch people's reactions.\n\n###### Figure 7-1. Satirical poster on the state of AI from keras4kindergartners.com\n\n# Chapter 8. Cloud APIs for Computer Vision: Up and Running in 15 Minutes\n\nDue to repeated incidents of near meltdown at the nearby nuclear power plant, the library of the city of Springfield (we are not allowed to mention the state) decided that it was too risky to store all their valuable archives in physical form. After hearing that the library from their rival city of Shelbyville started digitizing their records, they wanted to get in on the game as well. After all, their collection of articles such as \"Old man yells at cloud\" and \"Local man thinks wrestling is real\" and the hundred-year-old iconic photographs of the Gorge and the statue of the city's founder Jebediah Springfield are irreplaceable. In addition to making their archives resilient to catastrophes, they would make their archives easily searchable and retrievable. And, of course, the residents of Springfield now would be able to access all of this material from the comfort of their living room couches.\n\nThe first step in digitizing documents is, of course, scanning. That's the easy part. Then starts the real challenge\u2014processing and understanding all of this visual imagery. The team in Springfield had a few different options in front of them.\n\n  * Perform manual data entry for every single page and every single photograph. Given that the city has more than 200 years of rich history, it would take a really long time, and would be error prone and expensive. It would be quite an ordeal to transcribe all of that material.\n\n  * Hire a team of data scientists to build an image understanding system. That would be a much better approach, but there's just one tiny hitch in the plan. For a library that runs on charitable donations, hiring a team of data scientists would quickly exhaust its budget. A single data scientist might not only be the highest-paid employee at the library, they might also be the highest-earning worker in the entire city of Springfield (barring the wealthy industrialist Montgomery Burns).\n\n  * Get someone who knows enough coding to use the intelligence of ready-to-use vision APIs.\n\nLogically they went with the quick and inexpensive third option. They had a stroke of luck, too. Martin Prince, an industrious fourth grader from Springfield Elementary who happened to know some coding, volunteered to build out the system for them. Although Martin did not know much deep learning (he's just 10 years old, after all), he did know how to do some general coding, including making REST API calls using Python. And that was all he really needed to know. In fact, it took him just under 15 minutes to figure out how to make his first API call.\n\nMartin's _modus operandi_ was simple: send a scanned image to the cloud API, get a prediction back, and store it in a database for future retrieval. And obviously, repeat this process for every single record the library owned. He just needed to select the correct tool for the job.\n\nAll the big names\u2014Amazon, Google, IBM, Microsoft\u2014provide a similar set of computer-vision APIs that label images, detect and recognize faces and celebrities, identify similar images, read text, and sometimes even discern handwriting. Some of them even provide the ability to train our own classifier without having to write a single line of code. Sounds really convenient!\n\nIn the background, these companies are constantly working to improve the state of the art in computer vision. They have spent millions in acquiring and labeling datasets with a granular taxonomy much beyond the ImageNet dataset. We might as well make good use of their researchers' blood, sweat, and tears (and electricity bills).\n\nThe ease of use, speed of onboarding and development, the variety of functionality, richness of tags, and competitive pricing make cloud-based APIs difficult to ignore. And all of this without the need to hire an expensive data science team. Chapters Chapter 5 and Chapter 6 optimized for accuracy and performance, respectively; this chapter essentially optimizes for human resources.\n\nIn this chapter, we explore several cloud-based visual recognition APIs. We compare them all both quantitatively as well as qualitatively. This should hopefully make it easier to choose the one that best suits your target application. And if they still don't match your needs, we'll investigate how to train a custom classifier with just a few clicks.\n\n(In the interest of full disclosure, some of the authors of this book were previously employed at Microsoft, whose offerings are discussed here. We have attempted not to let that bias our results by building reproducible experiments and justifying our methodology.)\n\n# The Landscape of Visual Recognition APIs\n\nLet's explore some of the different visual recognition APIs out there.\n\n## Clarifai\n\nClarifai (Figure 8-1) was the winner of the 2013 ILSVRC classification task. Started by Matthew Zeiler, a graduate student from New York University, this was one of the first visual recognition API companies out there.\n\n###### Note\n\nFun fact: While investigating a classifier to detect NSFW (Not Safe For Work) images, it became important to understand and debug what was being learned by the CNN in order to reduce false positives. This led Clarifai to invent a visualization technique to expose which images stimulate feature maps at any layer in the CNN. As they say, necessity is the mother of invention.\n\n### What's unique about this API?\n\nIt offers multilingual tagging in more than 23 languages, visual similarity search among previously uploaded photographs, face-based multicultural appearance classifier, photograph aesthetic scorer, focus scorer, and embedding vector generation to help us build our own reverse-image search. It also offers recognition in specialized domains including clothing and fashion, travel and hospitality, and weddings. Through its public API, the image tagger supports 11,000 concepts.\n\n###### Figure 8-1. Sample of Clarifai's results\n\n## Microsoft Cognitive Services\n\nWith the creation of ResNet-152 in 2015, Microsoft was able to win seven tasks at the ILSVRC, the COCO Image Captioning Challenge as well as the Emotion Recognition in the Wild challenge, ranging from classification and detection (localization) to image descriptions. And most of this research was translated to cloud APIs. Originally starting out as Project Oxford from Microsoft Research in 2015, it was eventually renamed Cognitive Services in 2016. It's a comprehensive set of more than 50 APIs ranging from vision, natural language processing, speech, search, knowledge graph linkage, and more. Historically, many of the same libraries were being run at divisions at Xbox and Bing, but they are now being exposed to developers externally. Some viral applications showcasing creative ways developers use these APIs include _how-old.net_ (How Old Do I Look?), Mimicker Alarm (which requires making a particular facial expression in order to defuse the morning alarm), and _CaptionBot.ai_.\n\n### What's unique about this API?\n\nAs illustrated in Figure 8-2, the API offers image captioning, handwriting understanding, and headwear recognition. Due to many enterprise customers, Cognitive Services does not use customer image data for improving its services.\n\n###### Figure 8-2. Sample of Microsoft Cognitive Services results\n\n## Google Cloud Vision\n\nGoogle provided the winning entry at the 2014 ILSVRC with the help of the 22-layer GoogLeNet, which eventually paved the way for the now-staple Inception architectures. Supplementing the Inception models, in December 2015, Google released a suite of Vision APIs. In the world of deep learning, having large amounts of data is definitely an advantage to improve one's classifier, and Google has a lot of consumer data. For example, with learnings from Google Street View, you should expect relatively good performance in real-world text extraction tasks, like on billboards.\n\n### What's unique about this API?\n\nFor human faces, it provides the most detailed facial key points (Figure 8-3) including roll, tilt, and pan to accurately localize the facial features. The APIs also return similar images on the web to the given input. A simple way to try out the performance of Google's system without writing code is by uploading photographs to Google Photos and searching through the tags.\n\n###### Figure 8-3. Sample of Google Cloud Vision's results\n\n## Amazon Rekognition\n\nNo, that title is not a typo. Amazon Rekognition API (Figure 8-4) is largely based on Orbeus, a Sunnyvale, California-based startup that was acquired by Amazon in late 2015. Founded in 2012, its chief scientist also had winning entries in the ILSVRC 2014 detection challenge. The same APIs were used to power PhotoTime, a famous photo organization app. The API's services are available as part of the AWS offerings. Considering most companies already offer photo analysis APIs, Amazon is doubling down on video recognition offerings to offer differentiation.\n\n### What's unique about this API?\n\nLicense plate recognition, video recognition APIs, and better end-to-end integration examples of Rekognition APIs with AWS offerings like Kinesis Video Streams, Lambda, and others. Also, Amazon's API is the only one that can determine whether the subject's eyes are open or closed.\n\n###### Figure 8-4. Sample of Amazon Rekognition's results\n\n## IBM Watson Visual Recognition\n\nUnder the Watson brand, IBM's Visual Recognition offering started in early 2015. After purchasing AlchemyAPI, a Denver-based startup, AlchemyVision has been used for powering the Visual Recognition APIs (Figure 8-5). Like others, IBM also offers custom classifier training. Surprisingly, Watson does not offer optical character recognition yet.\n\n###### Figure 8-5. Sample of IBM Watson's Visual Recognition results\n\n## Algorithmia\n\nAlgorithmia is a marketplace for hosting algorithms as APIs on the cloud. Founded in 2013, this Seattle-based startup has both its own in-house algorithms as well as those created by others (in which case creators earn revenue based on the number of calls). In our experience, this API did tend to have the slowest response time.\n\n### What's unique about this API?\n\nColorization service for black and white photos (Figure 8-6), image stylization, image similarity, and the ability to run these services on-premises, or on any cloud provider.\n\n###### Figure 8-6. Sample of Algorithmia's style transfer results\n\nWith so many offerings, it can be overwhelming to choose a service. There are many reasons why we might choose one over another. Obviously, the biggest factors for most developers would be accuracy and price. Accuracy is the big promise that the deep learning revolution brings, and many applications require it on a consistent basis. Price of the service might be an additional factor to consider. We might also choose a service provider because our company already has a billing account with it, and it would take additional effort to integrate a different service provider. Speed of the API response might be another factor, especially if the user is waiting on the other end for a response. Because many of these API calls can be abstracted, it's easy to switch between different providers.\n\n# Comparing Visual Recognition APIs\n\nTo aid our decision making, let's compare these APIs head to head. In this section, we examine service offerings, cost, and accuracy of each.\n\n## Service Offerings\n\nTable 8-1 lists what services are being offered by each cloud provider.\n\nTable 8-1. Comparison shopping of vision API providers (as of Aug. 2019) **** | **Algorithmia** | **Amazon Rekognition** | **Clarifai** | **Microsoft Cognitive Services** | **Google Cloud Vision** | **IBM Watson Visual Recognition**  \n---|---|---|---|---|---|---  \nImage classification |\n\n\u2714 | \u2714 | \u2714 | \u2714 | \u2714 | \u2714\n\nImage detection |\n\n\u2714 | \u2714 | |\n\n\u2714 | \u2714\n\n|\n\nOCR |\n\n\u2714 | \u2714 | |\n\n\u2714 | \u2714\n\n|\n\nFace recognition |\n\n\u2714 | \u2714 | |\n\n\u2714 | |\n\nEmotion\n\nrecognition | \u2714 | |\n\n\u2714 | \u2714 | \u2714\n\n|\n\nLogo recognition |  | |\n\n\u2714 | \u2714 | \u2714\n\n|\n\nLandmark recognition |  | |\n\n\u2714 | \u2714 | \u2714 | \u2714\n\nCelebrity\n\nrecognition | \u2714 | \u2714 | \u2714 | \u2714 | \u2714 | \u2714\n\nMultilingual tagging |  | |\n\n\u2714 | |\n\n|\n\nImage description |  | |  | \u2714 | |\n\nHandwriting |  | |  | \u2714 | \u2714\n\n|\n\nThumbnail generation |\n\n\u2714 | |  | \u2714 | \u2714\n\n|\n\nContent moderation |\n\n\u2714 | \u2714 | \u2714 | \u2714 | \u2714\n\n|\n\nCustom classification training |  | |\n\n\u2714 | \u2714 | \u2714 | \u2714\n\nCustom detector training |  |  |  | \u2714 | \u2714\n\n|   \nMobile custom models |  |  |\n\n\u2714 | \u2714 | \u2714\n\n|   \nFree tier | 5,000 requests per month | 5,000 requests per month | 5,000 requests per month | 5,000 requests per month | 1,000 requests per month | 7,500\n\nThat's a mouthful of services already up and running, ready to be used in our application. Because numbers and hard data help make decisions easier, it's time to analyze these services on two factors: cost and accuracy.\n\n## Cost\n\nMoney doesn't grow on trees (yet), so it's important to analyze the economics of using off-the-shelf APIs. Taking a heavy-duty example of querying these APIs at about 1 query per second (QPS) service for one full month (roughly 2.6 million requests per month), Figure 8-7 presents a comparison of the different providers sorted by estimated costs (as of August 2019).\n\n###### Figure 8-7. A cost comparison of different cloud-based vision APIs\n\nAlthough for most developers, this is an extreme scenario, this would be a pretty realistic load for large corporations. We will eventually compare these prices against running our own service in the cloud to make sure we get the most bang for the buck fitting our scenario.\n\nThat said, many developers might find negligible charges, considering that all of the cloud providers we look at here have a free tier of 5,000 calls per month (except Google Vision, which gives only 1,000 calls per month for free), and then roughly $1 per 1,000 calls.\n\n## Accuracy\n\nIn a world ruled by marketing departments who claim their organizations to be the market leaders, how do we judge who is actually the best? What we need are common metrics to compare these service providers on some external datasets.\n\nTo showcase building a reproducible benchmark, we assess the text extraction quality using the COCO-Text dataset, which is a subset of the MS COCO dataset. This 63,686-image set contains text in daily life settings, like on a banner, street sign, number on a bus, price tag in a grocery store, designer shirt, and more. This real-world imagery makes it a relatively tough set to test against. We use the Word Error Rate (WER) as our benchmarking metric. To keep things simple, we ignore the position of the word and focus only on whether a word is present (i.e., bag of words). To be a match, the entire word must be correct.\n\nIn the COCO-Text validation dataset, we pick all images with one or more instances of legible text (full-text sequences without interruptions) and compare text instances of more than one-character length. We then send these images to various cloud vision APIs. Figure 8-8 presents the results.\n\n###### Figure 8-8. WER for different text extraction APIs as of August 2019\n\nConsidering how difficult the dataset is, these results are remarkable. Most state-of-the-art text extraction tools from earlier in the decade would not cross the 10% mark. This shows the power of deep learning. On a subset of manually tested images, we also noticed a year-on-year improvement in the performance of some of these APIs, which is another benefit enjoyed by cloud-based APIs.\n\nAs always, all of the code that we used for our experiment is hosted on GitHub (see _http:\/\/PracticalDeepLearning.ai_.\n\nThe results of our analysis depend significantly on the dataset we choose as well as our metrics. Depending on our dataset (which is in turn influenced by our use case) as well as our minimum quality metrics, our results can vary. Additionally, service providers are constantly improving their services in the background. As a consequence, these results are not set in stone and improve over time. These results can be replicated on any dataset with the scripts on GitHub.\n\n## Bias\n\nIn Chapter 1, we explored how bias can creep into datasets and how it can have real-life consequences for people. The APIs we explore in this chapter are no exception. Joy Buolamwini, a researcher at the MIT Media Lab, discovered that among Microsoft, IBM, and Megvii (also known as Face++), none were able to detect her face and gender accurately. Wondering if she had unique facial features that made her undetectable to these APIs, she (working along with Timnit Gebru) compiled faces of members of legislative branches from six countries with a high representation of women, building the Pilot Parliaments Benchmark (PPB; see Figure 8-9). She chose members from three African countries and three European countries to test for how the APIs performed on different skin tones. If you haven't been living under a rock, you can already see where this is going.\n\nShe observed that the APIs performed fairly well overall at accuracies between 85% and 95%. It was only when she started slicing the data across the different categories that she observed there was a massive amount of difference in the accuracies for each. She first observed that there was a significant difference between detection accuracies of men and women. She also observed that breaking down by skin tone, the difference in the detection accuracy was even larger. Then, finally, taking both gender and skin tone into consideration, the differences grew painfully starker between the worse detected group (darker females) and the best detected group (lighter males). For example, in the case of IBM, the detection accuracy of African women was a mere 65.3%, whereas the same API gave a 99.7% accuracy for European men. A whopping 34.4% difference! Considering many of these APIs are used by law enforcement, the consequences of bias seeping in might have life or death consequences.\n\n###### Figure 8-9. Averaged faces among different gender and skin tone, from Pilot Parliaments Benchmark (PPB)\n\nFollowing are a few insights we learned from this study:\n\n  * The algorithm is only as good as the data on which it's trained. And this shows the need for diversity in the training dataset.\n\n  * Often the aggregate numbers don't always reveal the true picture. The bias in the dataset is apparent only when slicing it across different subgroups.\n\n  * The bias does not belong to any specific company; rather, it's an industry-wide phenomenon.\n\n  * These numbers are not set in stone and reflect only the time at which the experiment was performed. As evident from the drastic change in numbers between 2017 (Figure 8-10) and a subsequent study in 2018 (Figure 8-11), these companies are taking bias removal from their datasets quite seriously.\n\n  * Researchers putting commercial companies to the test with public benchmarks results in industry-wide improvements (even if for the fear of bad PR, then so be it).\n\n###### Figure 8-10. Face detection comparison across APIs, tested in April and May 2017 on the PPB\n\n###### Figure 8-11. Face detection comparison across APIs in August 2018 on the PPB, conducted by Inioluwa Deborah Raji et al.\n\nHow about bias in image-tagging APIs? Facebook AI Research pondered over the question \"Does Object Recognition Work for Everyone?\" in a paper by the same title (Terrance DeVries et al.). The group tested multiple cloud APIs in February 2019 on Dollar Street, a diverse collection of images of household items from 264 different homes across 50 countries (Figure 8-12).\n\n###### Figure 8-12. Image-tagging API performance on geographically diverse images from the Dollar Street dataset\n\nHere are some of the key learnings from this test:\n\n  * Accuracy of object classification APIs was significantly lower in images from regions with lower income levels, as illustrated in Figure 8-13.\n\n  * Datasets such as ImageNet, COCO, and OpenImages severely undersample images from Africa, India, China, and Southeast Asia, hence leading to lower performance on images from the non-Western world.\n\n  * Most of the datasets were collected starting with keyword searches in English, omitting images that mentioned the same object with phrases in other languages.\n\n###### Figure 8-13. Average accuracy (and standard deviation) of six cloud APIs versus income of the household where the images were collected\n\nIn summary, depending on the scenario for which we want to use these cloud APIs, we should build our own benchmarks and test them periodically to evaluate whether these APIs are appropriate for the use case.\n\n# Getting Up and Running with Cloud APIs\n\nCalling these cloud services requires minimal code. At a high level, get an API key, load the image, specify the intent, make a POST request with the proper encoding (e.g., base64 for the image), and receive the results. Most of the cloud providers offer software development kits (SDKs) and sample code showcasing how to call their services. They additionally provide pip-installable Python packages to further simplify calling them. If you're using Amazon Rekognition, we highly recommend using its `pip` package.\n\nLet's reuse our thrilling image to test-run these services.\n\nFirst, let's try it on Microsoft Cognitive Services. Get an API key and replace it in the following code (the first 5,000 calls are free\u2014more than enough for our experiments):\n\n    cognitive_services_tagimage('DogAndBaby.jpg')\n\n    Results:\n\n    {\n        \"description\": {\n            \"tags\": [\"person\", \"indoor\", \"sitting\", \"food\", \"table\", \"little\",\n    \"small\", \"dog\", \"child\", \"looking\", \"eating\", \"baby\", \"young\", \"front\",\n    \"feeding\", \"holding\", \"playing\", \"plate\", \"boy\", \"girl\", \"cake\", \"bowl\",\n    \"woman\", \"kitchen\", \"standing\", \"birthday\", \"man\", \"pizza\"],\n            \"captions\": [{\n                \"text\": \"a little girl sitting at a table with a dog\",\n                \"confidence\": 0.84265453815486435\n            }]\n        },\n        \"requestId\": \"1a32c16f-fda2-4adf-99b3-9c4bf9e11a60\",\n        \"metadata\": {\n            \"height\": 427,\n            \"width\": 640,\n            \"format\": \"Jpeg\"\n        }\n    }\n\n\"A little girl sitting at a table with a dog\"\u2014pretty close! There are other options to generate more detailed results, including a probability along with each tag.\n\n###### Tip\n\nAlthough the ImageNet dataset is primarily tagged with nouns, many of these services go beyond and return verbs like \"eating,\" \"sitting,\" \"jumping.\" Additionally, they might contain adjectives like \"red.\" Chances are, these might not be appropriate for our application. We might want to filter out these adjectives and verbs. One option is to check their linguistic type against Princeton's WordNet. This is available in Python with the Natural Language Processing Toolkit (NLTK). Additionally, we might want to filter out words like \"indoor\" and \"outdoor\" (often shown by Clarifai and Cognitive Services).\n\nNow, let's test the same image using Google Vision APIs. Get an API key from their website and use it in the following code (and rejoice, because the first 1,000 calls are free):\n\n    google_cloud_tagimage('DogAndBaby.jpg')\n\n    Results:\n\n    {\n     \"responses\": [\n        {\n          \"labelAnnotations\": [\n            {\n              \"mid\": \"\/m\/0bt9lr\",\n              \"description\": \"dog\",\n              \"score\": 0.951077,\n              \"topicality\": 0.951077\n            },\n            {\n              \"mid\": \"\/m\/06z04\",\n              \"description\": \"skin\",\n              \"score\": 0.9230451,\n              \"topicality\": 0.9230451\n            },\n            {\n              \"mid\": \"\/m\/01z5f\",\n              \"description\": \"dog like mammal\",\n              \"score\": 0.88359463,\n              \"topicality\": 0.88359463\n            },\n            {\n              \"mid\": \"\/m\/01f5gx\",\n              \"description\": \"eating\",\n              \"score\": 0.7258142,\n              \"topicality\": 0.7258142\n            }\n    # other objects\n          ]\n        }\n     ]\n    }\n\nWasn't that a little too easy? These APIs help us get to state-of-the-art results without needing a Ph.D.\u2014in just 15 minutes!\n\n###### Tip\n\nEven though these services return tags and image captions with probabilities, it's up to the developer to determine a threshold. Usually, 60% and 40% are good thresholds for image tags and image captions, respectively.\n\nIt's also important to communicate the probability to the end-user from a UX standpoint. For example, if the result confidence is >80%, we might say prefix the tags with \"This image _contains...._ \" For <80%, we might want to change that prefix to \"This image _may contain..._ \" to reflect the lower confidence in the result.\n\n# Training Our Own Custom Classifier\n\nChances are these services were not quite sufficient to meet the requirements of our use case. Suppose that the photograph we sent to one of these services responded with the tag \"dog.\" We might be more interested in identifying the breed of the dog. Of course, we can follow Chapter 3 to train our own classifier in Keras. But wouldn't it be more awesome if we didn't need to write a single line of code? Help is on the way.\n\nA few of these cloud providers give us the ability to train our own custom classifier by merely using a drag-and-drop interface. The pretty user interfaces provide no indication that under the hood they are using transfer learning. As a result, Cognitive Services Custom Vision, Google AutoML, Clarifai, and IBM Watson all provide us the option for custom training. Additionally, some of them even allow building custom detectors, which can identify the location of objects with a bounding box. The key process in all of them being the following:\n\n  1. Upload images\n\n  2. Label them\n\n  3. Train a model\n\n  4. Evaluate the model\n\n  5. Publish the model as a REST API\n\n  6. Bonus: Download a mobile-friendly model for inference on smartphones and edge devices\n\nLet's see a step-by-step example of Microsoft's Custom Vision.\n\n  1. _Create a project_ (Figure 8-14): Choose a domain that best describes our use case. For most purposes, \"General\" would be optimal. For more specialized scenarios, we might want to choose a relevant domain.\n\n###### Figure 8-14. Creating a new project in Custom Vision\n\nAs an example, if we have an ecommerce website with photos of products against a pure white background, we might want to select the \"Retail\" domain. If we intend to run this model on a mobile phone eventually, we should choose the \"Compact\" version of the model, instead; it is smaller in size with only a slight loss in accuracy.\n\n  2. _Upload_ (Figure 8-15): For each category, upload images and tag them. It's important to upload at least 30 photographs per category. For our test, we uploaded more than 30 images of Maltese dogs and tagged them appropriately.\n\n###### Figure 8-15. Uploading images on CustomVision.ai\n\n  3. _Train_ (Figure 8-16): Click the Train button, and then in about three minutes, we have a spanking new classifier ready.\n\n###### Figure 8-16. The Train button in the upper-right corner of the CustomVision.ai page\n\n  4. _Analyze the model's performance_ : Check the precision and recall of the model. By default, the system sets the threshold at 90% confidence and gives the precision and recall metrics at that value. For higher precision, increase the confidence threshold. This would come at the expense of reduced recall. Figure 8-17 shows example output.\n\n  5. _Ready to go_ : We now have a production-ready API endpoint that we can call from any application.\n\nTo highlight the effect of the amount of data on model quality, let's train a dog breed classifier. We can use the Stanford Dogs dataset, a collection of more than 100 dog categories. For simplicity, we randomly chose 10 breeds, which have more than 200 images available. With 10 classes, a random classifier would have one-tenth, or 10%, the chance of correctly identifying an image. We should easily be able to beat this number. Table 8-2 shows the effect of training on datasets with different volumes.\n\nTable 8-2. Effect of number of training images on precision and recall **** | **30 training images\/class** | **200 training images\/class**  \n---|---|---  \nPrecision | 91.2% | 93.5%  \nRecall | 85.3% | 89.6%\n\n###### Figure 8-17. Relative precision and recall for our sample training set with 200 images per class\n\nBecause we haven't uploaded a test set, the performance figures reported here are on the full dataset using the common _k_ -fold cross-validation technique. This means the data was randomly divided into _k_ parts, then ( _k \u2013_ 1) parts were used for training, and the remaining part was used for testing. This was performed a few times, each time with a randomized subset of images, and the averaged results are reported here.\n\nIt is incredible that even with 30 images per class, the classifier's precision is greater than 90%, as depicted in Figure 8-18. And, surprisingly, this took slightly less than 30 seconds to train.\n\nNot only this, we can dig down and investigate the performance on each class. Classes with high precision might visibly be more distinct, whereas those with low precision might look similar to another class.\n\n###### Figure 8-18. Some of the possible tags returned by the API\n\nThis short and convenient approach is not without its downsides, as you will see in the following section. In that section, we also discuss mitigation strategies to help take advantage of this rather useful tool.\n\n## Top Reasons Why Our Classifier Does Not Work Satisfactorily\n\nThere are a number of reasons why a classifier would not perform well. The following are some of them:\n\nNot enough data\n\nIf we find that the accuracy is not quite sufficient for our needs, we might need to train the system with more data. Of course, 30 images per class just gets us started. But for a production-quality application, more images are better. 200 images per class are usually recommended.\n\nNonrepresentative training data\n\nOften, the images on the internet are far too clean, set up in studio lighting with clean backgrounds, and close to the center of the frame. Images that our application might see on a daily basis might not be represented quite so well. It's really important to train our classifier with real-world images for the best performance.\n\nUnrelated domain\n    Under the hood, Custom Vision is running transfer learning. This makes it really important to choose the correct domain when creating the project. As an example, if we are trying to classify X-ray images, transfer learning from an ImageNet-based model might not yield as accurate a result. For cases like that, training our own classifier manually in Keras would work best, as demonstrated in [Chapter 3](ch03.html#cats_versus_dogscolon_transfer_learning) (though this will probably take more than three minutes).\nUsing it for regression\n\nIn machine learning, there are two common categories of problems: classification and regression. Classification is predicting one or more classes for input. Regression, on the other hand, is predicting a numerical value given an input; for example, predicting house prices. Custom Vision is primarily a classification system. Using it to count objects by tagging the number of objects is the wrong approach, and will lead to unsatisfactory results.\n\nCounting objects is a type of regression problem. We can do it by localizing each instance of the object in an image (aka object detection) and counting their occurrences. Another example of a regression problem is predicting the age of a person based on their profile photo. We tackle both problems in later chapters.\n\nClasses are too similar\n\nIf our classes look too similar and rely heavily on smaller-level details for distinction, the model might not perform as well. For example, a five-dollar note and a 20-dollar note have very similar high-level features. It's at the lower-level details that show they are really distinct. As another example, it might be easy to distinguish between a Chihuahua and a Siberian Husky, but it's more difficult to distinguish between an Alaskan Malamute and a Siberian Husky. A fully retrained CNN, as demonstrated in Chapter 3, should perform better than this Custom Vision-based system.\n\n###### Tip\n\nA great feature of Custom Vision is that if the model is unsure of any image that it encounters via its API endpoint, the web UI will show those images for a manual review. We can review and manually tag new images on a periodic basis and continuously improve the quality of the model. These images tend to improve the classifier the most for two reasons: first, they represent real-world usage. Second, and more importantly, they have more impact on the model in comparison to images that the model can already easily classify. This is known as semisupervised learning.\n\nIn this section, we discussed a few different ways in which we can improve our model's accuracy. In the real world, that is not the end-all-be-all of a user's experience. How quickly we are able to respond to a request also matters a lot. In the following section, we cover a couple of different ways we can improve performance without sacrificing quality.\n\n# Comparing Custom Classification APIs\n\nAs you might have noticed throughout the book, we are pretty dogmatic about being data driven. If we are going to spend good money on a service, we better get the best bang for our buck. Time to put the hype to the test.\n\nFor a good number of classification problems, these custom cloud-based classifiers perform pretty well. To truly test their limits, we need something more challenging. We need to unleash the toughest doggone dataset, train this animal, and fetch some insightful results\u2014using the Stanford Dogs dataset.\n\nUsing the entire dataset might make it too easy for these classifiers (after all, ImageNet already has so many dog breeds), so we took it up a notch. Instead, we trained our own Keras classifier on the entire dataset and built a mini-dataset out of the top 34 worst-performing classes (each containing at least 140 images). The reason these classes performed poorly was because they often became confused with other similar-looking dog breeds. To perform better, they require a fine-grained understanding of the features. We divide the images into 100 randomly chosen images per class in the training dataset and 40 randomly chosen images per class in the test dataset. To avoid any class imbalances, which can have an impact on predictions, we chose the same number of train and test images for each class.\n\nLastly, we selected a minimum confidence threshold of 0.5 as it appeared to strike a good balance between precision and recall across all services. At a high confidence threshold such as 0.99, a classifier might be very accurate, but there might be only a handful of images with predictions; in other words, really low recall. On the other hand, a really low threshold of 0.01 would result in predictions for nearly all images. However, we should not rely on many of these results. After all, the classifier is not confident.\n\nInstead of reporting the precision and recall, we report the _F1 score_ (also known as _F-measure_ ), which is a hybrid score that combines both of those values:\n\nF1score= 2\u00d7precision\u00d7recall precision+recall\n\nAdditionally, we report the time it took to train, as shown in Figure 8-19. Beyond just the cloud, we also trained using Apple's Create ML tool on a MacBook Pro with and without data augmentations (rotate, crop, and flip).\n\nGoogle and Microsoft provide the ability to customize the duration of training. Google Auto ML allows us to customize between 1 and 24 hours. Microsoft provides a free \"Fast Training\" option and a paid \"Advanced Training\" option (similar to Google's offering) with which we can select the duration to be anywhere between 1 and 24 hours.\n\n###### Figure 8-19. A chart showing the F1 score for custom classifier services, as of August 2019 (higher is better)\n\nFollowing are some interesting takeaways from this experiment:\n\n  * Clarifai and Microsoft offered near-instant training for the 3,400 training images.\n\n  * Compared to \"Fast Training,\" Microsoft's \"Advanced Training\" performed slightly better (roughly a 1-point increase) for the extra one hour of training. Because \"Fast Training\" took less than 15 seconds to train, we can infer that its base featurizer was already good at extracting fine-grained features.\n\n  * Surprisingly, Apple's Create ML actually performed worse after adding in the augmentations, despite taking more than two extra hours to train, most of which was spent creating the augmentations. This was done on a top-of-the-line MacBook Pro and showed 100% GPU utilization in Activity Monitor.\n\nAdditionally, to test the featurizer's strength, we varied the amount of training data supplied to the service (Figure 8-20). Because Microsoft took less than 15 seconds to train, it was easy (and cheap!) for us to perform the experiment there. We varied between 30 and 100 images per class for training while keeping the same 40 images per class for testing.\n\n###### Figure 8-20. Effect of varying size of training data per class on test F1 score (higher is better)\n\nEven though Microsoft recommends using at least 50 images per class, going under that limit did not affect performance significantly. The fact that the F1 score did not vary as much as one would expect shows the value of transfer learning (enabling less data to build classifiers) and having a good featurizer capable of fine-grained classification.\n\nIt bears repeating that this experiment was intentionally made difficult to stress-test these classifiers. On average, they would have performed much better on the entire Stanford Dogs dataset.\n\n# Performance Tuning for Cloud APIs\n\nA photograph taken by a modern cell phone can have a resolution as high as 4000 x 3000 pixels and be upward of 4 MB in size. Depending on the network quality, it can take a few seconds to upload such an image to the service. With each additional second, it can become more and more frustrating for our users. Could we make this faster?\n\nThere are two ways to reduce the size of the image:\n\nResizing\n\nMost CNNs take an input image with a size of 224 x 224 or 448 x 448 pixels. Much of a cell phone photo's resolution would be unnecessary for a CNN. It would make sense to downsize the image prior to sending it over the network, instead of sending a large image over the network and then downsizing it on the server.\n\nCompression\n\nMost image libraries perform _lossy_ compression while saving a file. Even a little bit of compression can go a long way in reducing the size of the image while minimally affecting the quality of the image itself. Compression does introduce noise, but CNNs are usually robust enough to deal with some of it.\n\n## Effect of Resizing on Image Labeling APIs\n\nWe performed an experiment in which we took more than a hundred diverse unmodified images taken from an iPhone at the default resolution (4032 x 3024) and sent them to the Google Cloud Vision API to get labels for each of those images. We then downsized each of the original images in 5% increments (5%, 10%, 15%...95%) and collected the API results for those smaller images, too. We then calculated the agreement rate for each image using the following formula:\n\n% a g r e e m e n t r a t e = numberoflabelsinthebaselineimagealsopresentintestimage numberoflabelsinthebaselineimage \u00d7 100\n\nFigure 8-21 shows the results of this experiment. In the figure, the solid line shows the reduction in file size, and the dotted line represents the agreement rate. Our main conclusion from the experiment was that a 60% reduction in resolution led to a 95% reduction in file size, with little change in accuracy compared to the original images.\n\n###### Figure 8-21. Effect of resizing an image on agreement rate and file size reduction relative to the original image\n\n## Effect of Compression on Image Labeling APIs\n\nWe repeated the same experiment, but instead of changing the resolution, we changed the compression factor for each image incrementally. In Figure 8-22, the solid line shows the reduction in file size and the dotted line represents the agreement rate. The main takeaway here is that a 60% compression score (or 40% quality) leads to an 85% reduction in file size, with little change in accuracy compared to the original image.\n\n###### Figure 8-22. Effect of compressing an image on agreement rate and file size reduction relative to the original image\n\n## Effect of Compression on OCR APIs\n\nWe took a document containing 300-plus words at the default resolution of an iPhone (4032 x 3024), and sent it to the Microsoft Cognitive Services API to test text recognition. We then compressed it at 5% increments and then sent each image and compressed it. We sent these images to the same API and compared their results against the baseline to calculate the percentage WER. We observed that even setting the compression factor to 95% (i.e., 5% quality of the original image) had no effect on the quality of results.\n\n## Effect of Resizing on OCR APIs\n\nWe repeated the previous experiment, but this time by resizing each image instead of compressing. After a certain point, the WER jumped from none to almost 100%, with nearly all words being misclassified. Retesting this with another document having each word at a different font size showed that all words under a particular font size were getting misclassified. To effectively recognize text, OCR engines need the text to be bigger than a minimum height (a good rule of thumb is larger than 20 pixels). Hence the higher the resolution, the higher the accuracy.\n\nWhat have we learned?\n\n  * For text recognition, compress images heavily, but do not resize.\n\n  * For image labeling, a combination of moderate resizing (say, 50%) and moderate compression (say, 30%) should lead to heavy file size reductions (and quicker API calls) without any difference in quality of API results.\n\n  * Depending on your application, you might be working with already resized and compressed images. Every processing step can introduce a slight difference in the results of these APIs, so aim to minimize them.\n\n###### Tip\n\nAfter receiving an image, cloud APIs internally resize it to fit their own implementation. For us, this means two levels of resizing: we first resize an image to reduce the size, then send it to the cloud API, which further resizes the image. Downsizing images introduces distortion, which is more evident at lower resolutions. We can minimize the effect of distortion by resizing from a higher resolution, which is bigger by a few multiples. For example, resizing 3024x3024 (original) \u2192 302x302 (being sent to cloud) \u2192 224x224 (internally resized by APIs) would introduce much more distortion in the final image compared to 3024x3024 \u2192 896x896 \u2192 224x224. Hence, it's best to find a happy intermediate size before sending the images. Additionally, specifying advanced interpolation options like `BICUBIC` and `LANCZOS` will lead to more accurate representation of the original image in the smaller version.\n\n# Case Studies\n\nSome people say that the best things in life don't come easy. We believe this chapter proves otherwise. In the following section, we take a look at how some tech industry titans use cloud APIs for AI to drive some very compelling scenarios.\n\n## The New York Times\n\nIt might seem like the scenario painted at the beginning of the chapter was taken out of a cartoon, but it was, in fact, pretty close to the case of the _New York Times_ (NYT). With more than 160 years of illustrious history, NYT has a treasure trove of photographs in its archives. It stored many of these artifacts in the basement of its building three stories below the ground level, aptly called the \"morgue.\" The value of this collection is priceless. In 2015, due to a plumbing leak, parts of the basement were damaged including some of these archived records. Thankfully the damage was minimal. However, this prompted NYT to consider digitally archiving them to protect against another catastrophe.\n\nThe photographs were scanned and stored in high quality. However, the photographs themselves did not have any identifying information. What many of them did have were handwritten or printed notes on the backside giving context for the photographs. NYT used the Google Vision API to scan this text and tag the respective images with that information. Additionally, this pipeline provided opportunities to extract more metadata from the photographs, including landmark recognition, celebrity recognition, and so on. These newly added tags powered its search feature so that anyone within the company and outside could explore the gallery and search using keywords, dates, and so on without having to visit the morgue, three stories down.\n\n## Uber\n\nUber uses Microsoft Cognitive Services to identify each of its seven million-plus drivers in a couple of milliseconds. Imagine the sheer scale at which Uber must operate its new feature called \"Real-Time ID Check.\" This feature verifies that the current driver is indeed the registered driver by prompting them to take a selfie either randomly or every time they are assigned to a new rider. This selfie is compared to the driver's photo on file, and only if the face models are a match is the driver allowed to continue. This security feature is helpful for building accountability by ensuring the security of the passengers and by ensuring that the driver's account is not compromised. This safety feature is able to detect changes in the selfie, including a hat, beard, sunglasses, and more, and then prompts the driver to take a selfie without the hat or sunglasses.\n\n###### Figure 8-23. The Uber Drivers app prompts the driver to take a selfie to verify the identity of the driver (image source)\n\n## Giphy\n\nBack in 1976, when Dr. Richard Dawkins coined the term \"meme,\" little did he know it would take on a life of its own four decades later. Instead of giving a simple textual reply, we live in a generation where most chat applications suggest an appropriate animated GIF matching the context. Several applications provide a search specific to memes and GIFs, such as Tenor, Facebook messenger, Swype, and Swiftkey. Most of them search through Giphy (Figure 8-24), the world's largest search engine for animated memes commonly in the GIF format.\n\n###### Figure 8-24. Giphy extracts text from animations as metadata for searching\n\nGIFs often have text overlaid (like the dialogue being spoken) and sometimes we want to look for a GIF with a particular dialogue straight from a movie or TV show. For example, the image in Figure 8-24 from the 2010 _Futurama_ episode in which the \"eyePhone\" (sic) was released is often used to express excitement toward a product or an idea. Having an understanding of the contents makes the GIFs more searchable. To make this happen, Giphy uses Google's Vision API to extract the recognize text and objects\u2014aiding the search for the perfect GIF.\n\nIt's obvious that tagging GIFs is a difficult task because a person must sift through millions of these animations and manually annotate them frame by frame. In 2017, Giphy figured out two solutions to automate this process. The first approach was to detect text from within the image. The second approach was to generate tags based on the objects in the image to supplement the metadata for their search engine. This metadata is stored and searched using ElasticSearch to make a scalable search engine.\n\nFor text detection, the company used the OCR services from the Google Vision API on the first frame from the GIFs to confirm whether the GIF actually contained text. If the API replied in the affirmative, Giphy would send the next frames, receive their OCR-detected texts, and figure out the differences in the text; for instance, whether the text was static (remaining the same throughout the duration of the gif) or dynamic (different text in different frames). For generating the class labels corresponding to objects in the image, engineers had two options: label detection or web entities, both of which are available on Google Vision API. Label detection, as the name suggests, provides the actual class name of the object. Web entities provides an entity ID (which can be referenceable in the Google Knowledge Graph), which is the unique web URL for identical and similar images seen elsewhere on the net. Using these additional annotations gave the new system an increase in the click-through-rate (CTR) by 32%. Medium-to-long-tail searches (i.e., not-so-frequent searches) benefitted the most, becoming richer with relevant content as the extracted metadata surfaced previously unannotated GIFs that would have otherwise been hidden. Additionally, this metadata and click-through behavior of users provides data to make a similarity and deduplication feature.\n\n## OmniEarth\n\nOmniEarth is a Virginia-based company that specializes in collecting, analyzing, and combining satellite and aerial imagery with other datasets to track water usage across the country, scalably, and at high speeds. The company is able to scan the entire United States at a total of 144 million parcels of land within hours. Internally, it uses the IBM Watson Visual Recognition API to classify images of land parcels for valuable information like how green it is. Combining this classification with other data points such as temperature and rainfall, OmniEarth can predict how much water was used to irrigate the field.\n\nFor house properties, it infers data points from the image such as the presence of pools, trees, or irrigable landscaping to predict the amount of water usage. The company even predicted where water is being wasted due to malpractices like overwatering or leaks. OmniEarth helped the state of California understand water consumption by analyzing more than 150,000 parcels of land, and then devised an effective strategy to curb water waste.\n\n## Photobucket\n\nPhotobucket is a popular online image- and video-hosting community where more than two million images are uploaded every day. Using Clarifai's NSFW models, Photobucket automatically flags unwanted or offensive user-generated content and sends it for further review to its human moderation team. Previously, the company's human moderation team was able to monitor only about 1% of the incoming content. About 70% of the flagged images turned out to be unacceptable content. Compared to previous manual efforts, Photobucket identified 700 times more unwanted content, thus cleaning the website and creating a better UX. This automation also helped discover two child pornography accounts, which were reported to the FBI.\n\n## Staples\n\nEcommerce stores like Staples often rely on organic search engine traffic to drive sales. One of the methods to appear high in search engine rankings is to put descriptive image tags in the ALT text field for the image. Staples Europe, which serves 12 different languages, found tagging product images and translating keywords to be an expensive proposition, which is traditionally outsourced to human agencies. Fortunately, Clarifai provides tags in 20 languages at a much cheaper rate, saving Staples costs into five figures. Using these relevant keywords led to an increase in traffic and eventually increased sales through its ecommerce store due to a surge of visitors to the product pages.\n\n## InDro Robotics\n\nThis Canadian drone company uses Microsoft Cognitive Services to power search and rescue operations, not only during natural disasters but also to proactively detect emergencies. The company utilizes Custom Vision to train models specifically for identifying objects such as boats and life vests in water (Figure 8-25) and use this information to notify control stations. These drones are able to scan much larger ocean spans on their own, as compared to lifeguards. This automation alerts the lifeguard of emergencies, thus improving the speed of discovery and saving lives in the process.\n\nAustralia has begun using drones from other companies coupled with inflatable pods to be able to react until help reaches. Soon after deployment, these pods saved two teenagers stranded in the ocean, as demonstrated in Figure 8-26. Australia is also utilizing drones to detect sharks so that beaches can be vacated. It's easy to foresee the tremendous value these automated, custom training services can bring.\n\n###### Figure 8-25. Detections made by InDro Robotics\n\n###### Figure 8-26. Drone identifies two stranded swimmers and releases an inflatable pod that they cling onto (image source)\n\n# Summary\n\nIn this chapter, we explored various cloud APIs for computer vision, first qualitatively comparing the breadth of services offered and then quantitatively comparing their accuracy and price. We also looked at potential sources of bias that might appear in the results. We saw that with just a short code snippet, we can get started using these APIs in less than 15 minutes. Because one model doesn't fit all, we trained a custom classifier using a drag-and-drop interface, and tested multiple companies against one another. Finally, we discussed compression and resizing recommendations to speed up image transmission and how they affect different tasks. To top it all off, we examined how companies across industries use these cloud APIs for building real-world applications. Congratulations on making it this far! In the next chapter, we will see how to deploy our own inference server for custom scenarios.\n\n# Chapter 9. Scalable Inference Serving on Cloud with TensorFlow Serving and KubeFlow\n\nImagine this: you just built a top-notch classifier. Your goal, as the Silicon Valley motto goes, is to \" _make the world a better place_ ,\" which you're going to do... with a spectacular Dog\/Cat classifier. You have a solid business plan and you cannot wait to pitch your magical classifier to that venture capital firm next week. You know that the investors will question you about your cloud strategy, and you need to show a solid demo before they even consider giving you the money. How would you do this? Creating a model is half the battle, serving it is the next challenge, often the bigger one. In fact, for a long time it was common for training a model to only take a few weeks, but trying to serve it to a larger group of people was a months-long battle, often involving backend engineers and DevOps teams.\n\nIn this chapter, we answer a few questions that tend to come up in the context of hosting and serving custom-built models.\n\n  * How can I host my model on my personal server so that my coworkers can play with it?\n\n  * I am not a backend\/infrastructure engineer, but I want to make my model available so that it can serve thousands (or even millions) of users. How can I do this at a reasonable price without worrying about scalability and reliability issues?\n\n  * There are reasons (such as cost, regulations, privacy, etc.) why I cannot host my model on the cloud, but only on-premises (my work network). Can I serve predictions at scale and reliably in such a case?\n\n  * Can I do inference on GPUs?\n\n  * How much can I expect to pay for each of these options?\n\n  * Could I scale my training and serving across multiple cloud providers?\n\n  * How much time and technical know-how will it take to get these running?\n\nLet's begin our journey by looking at the high-level overview of the tools available to us.\n\n# Landscape of Serving AI Predictions\n\nThere is a multitude of tools, libraries, and cloud services available for getting trained AI models to serve prediction requests. Figure 9-1 simplifies them into four categories.\n\n###### Figure 9-1. A high-level overview and comparison of different inference serving options\n\nDepending on our inference scenarios, we can make an appropriate choice. Table 9-1 takes a deeper look.\n\nTable 9-1. Tools to serve deep learning models over the network **Category and examples** | **Expected time to first prediction** | **Pros and cons**  \n---|---|---\n\nHTTP servers\n\n  * Flask\n\n  * Django\n\n  * Apache OpenWhisk\n\n  * Python `http.server`\n\n| <5 minutes |\n\n\\+ Simple to run\n\n\\+ Often runs current Python code\n\n\u2013 Slow\n\n\u2013 Not optimized for AI\n\n**Hosted and managed cloud stacks**\n\n  * Google Cloud ML\n\n  * Azure ML\n\n  * Amazon Sage Maker\n\n  * Algorithmia\n\n| <15 minutes |\n\n\\+ Easier GUI\/command-line interfaces\n\n\\+ Highly scalable\n\n\\+ Fully managed, reduces the need for DevOps teams\n\n\u2013 Usually limited to CPU-based inference, can be slow for large models\n\n\u2013 Warm-up query time can be slow\n\n**Manually managed serving libraries**\n\n  * TensorFlow Serving\n\n  * NVIDIA TensorRT\n\n  * DeepDetect\n\n  * MXNet Model Serving\n\n  * Skymind Intelligence Layer with DeepLearning4J\n\n  * Seldon\n\n  * DeepStack AI Server\n\n| <15 minutes |\n\n\\+ High performance\n\n\\+ Allows manual controls on optimizations, batching, etc.\n\n\\+ Can run inference on GPU\n\n\u2013 More involved setup\n\n\u2013 Scaling over multiple nodes usually requires extra groundwork\n\n**Cloud AI orchestration frameworks**\n\n  * KubeFlow\n\n| ~1 hour |\n\n\\+ Makes scaling training and inference easy to manage\n\n\\+ Portability between cloud providers\n\n\\+ Consistent environments across development and production\n\n\\+ For data scientists, integration with familiar tools such as Jupyter Notebooks for sending models to production\n\n\\+ Enables composing conditional pipelines to automate testing, cascading models\n\n\\+ Uses existing manually managed serving libraries\n\n\u2013 Still evolving\n\n\u2013 For beginners, hosted and managed cloud stacks offer an easier learning curve\n\nIn this chapter, we explore a range of tools and scenarios. Some of these options are easy to use but limited in functionality. Others offer more granular controls and higher performance but are more involved to set up. We look at one example of each category and take a deeper dive to develop an intuition into when using one of those makes sense. We then present a cost analysis of the different solutions as well as case studies detailing how some of these solutions work in practice today.\n\n# Flask: Build Your Own Server\n\nWe begin with the most basic technique of _Build Your Own Server_ (BYOS). From the choices presented in the first column of Table 9-1, we've selected Flask.\n\n## Making a REST API with Flask\n\nFlask is a Python-based web application framework. Released in 2010 and with more than 46,000 stars on GitHub, it is under continuous development. It's also quick and easy to set up and is really useful for prototyping. It is often the framework of choice for data science practitioners when they want to serve their models to a limited set of users (e.g., sharing with coworkers on a corporate network) without a lot of fuss.\n\nInstalling Flask with `pip` is fairly straightforward:\n\n    $ pip install flask\n\nUpon installation, we should be able to run the following simple \"Hello World\" program:\n\n    from flask import Flask\n    app = Flask(__name__)\n\n    @app.route(\"\/hello\")\n    def hello():\n        return \"Hello World!\"\n\n    if __name__ == \"__main__\":\n        app.run()\n\nThe following is the command to run the \"Hello World\" program:\n\n    $ python hello.py\n     * Running on http:\/\/127.0.0.1:5000\/ (Press Ctrl+C to quit)\n\nBy default, Flask runs on port 5000. When we open the URL _http:\/\/localhost:5000\/hello_ in the browser, we should see the words \"Hello World!,\" as shown in Figure 9-2.\n\n###### Figure 9-2. Navigate to http:\/\/localhost:5000\/hello within a web browser to view the \"Hello World!\" web page\n\nAs you can see, it takes barely more than a few lines to get a simple web application up and running. One of the most important lines in that script is `@app.route(\"\/hello\")`. It specifies that the path `\/hello` after the hostname would be served by the method immediately beneath it. In our case, it merely returns the string \"Hello World!\" In the next step, we look at how to deploy a Keras model to a Flask server and create a route that will serve predictions by our model.\n\n## Deploying a Keras Model to Flask\n\nOur first step is to load our Keras model. The following lines load the model from the . _h5_ file. You'll find the scripts for this chapter on the book's GitHub (see _http:\/\/PracticalDeepLearning.ai_) in _code\/chapter-9_ :\n\n    from tf.keras.models import load_model\n    model = load_model(\"dogcat.h5\")\n\nNow, we create the route _\/infer_ that would support inference on our images. Naturally, we would support `POST` requests to accept images:\n\n    @app.route('\/infer', methods=[POST])\n    def infer():\n      file = request.files['file']\n      image = Image.open(file)\n      image = preprocess(image)\n\n      predictions = model.predict(image)\n      max_index = numpy.argmax(predictions)\n      _# We know the labels from the model we trained previously_\n      if max_index == 0:\n         return \"Cat\"\n      else:\n         return \"Dog\"\n\nTo test the inference, let's use the `curl` command, as follows, on a sample image containing a dog:\n\n    $ curl -X POST -F image=@dog.jpg 'http:\/\/localhost:5000\/infer'\n\n    {\"predictions\":[{\"label\":\"dog\",\"probability\":0.8525022864341736}]}\n\nAs expected, we get a prediction of \"dog.\" This has worked quite well so far. At this point, Flask runs only locally; that is, someone else on the network would not be able to make a request to this server. To make Flask available to others, we can simply change `app.run()` to the following:\n\n    app.run(host=\"0.0.0.0\")\n\nAt this point, we can give access to our model to anyone within our network. The next question would be\u2014can we do the same to make the model available to the general public? The answer to that question is an emphatic no! The Flask website has a prominent warning stating _\"WARNING: Do not use the development server in a production environment.\"_ Flask indeed does not support production work out of the box and would need custom code to enable that. In the upcoming sections, we look at how to host our models on systems that are meant for production use. With all of this in mind, let's recap some of the pros and cons of using Flask.\n\n## Pros of Using Flask\n\nFlask provides some advantages, namely:\n\n  * Quick to set up and to prototype\n\n  * Fast development cycle\n\n  * Lightweight on resources\n\n  * Broad appeal within the Python community\n\n## Cons of Using Flask\n\nAt the same time, Flask might not be your best choice, for the following reasons:\n\n  * Cannot scale; by default, it is not meant for production loads. Flask can serve only one request at one time\n\n  * Does not handle model versioning out of the box\n\n  * Does not support batching of requests out of the box\n\n# Desirable Qualities in a Production-Level Serving System\n\nFor any cloud service that is serving traffic from the public, there are certain attributes that we want to look for when deciding to use a solution. In the context of machine learning, there are additional qualities that we would look for while building inference services. We look at a few of them if this section.\n\n## High Availability\n\nFor our users to trust our service, it must be available almost always. For many serious players, they measure their availability metric in terms of \" _number of nines_.\" If a business claims that its service has four 9s availability, they mean the system is up and available 99.99% of the time. Even though 99% sounds impressive, Table 9-2 puts that downtime per year in perspective.\n\nTable 9-2. Downtime per year for different availability percentages **Availability %** | **Downtime per year**  \n---|---  \n99% (\"two nines\") | 3.65 days  \n99.9% (\"three nines\") | 8.77 hours  \n99.99% (\"four nines\") | 52.6 minutes  \n99.999% (\"five nines\") | 5.26 minutes\n\nImagine how ridiculous the situation would be if a major website like Amazon were only 99.9% available, losing millions in user revenue during the eight-plus hours of downtime. Five 9s is considered the holy grail. Anything less than three 9s is typically unsuitable for a high-quality production system.\n\n## Scalability\n\nTraffic handled by production services is almost never uniform across a larger time period. For example, the _New York Times_ experiences significantly more traffic during morning hours, whereas Netflix typically experiences a surge in traffic between the evening and late-night hours, when people chill. There are also seasonal factors in traffic. Amazon experiences orders of magnitude more traffic on Black Friday and during Christmas season.\n\nA higher demand requires a higher amount of resources being available and online to serve them. Otherwise, the availability of the system would be in jeopardy. A naive way to accomplish this would be to anticipate the highest volume of traffic the system would ever serve, determine the number of resources necessary to serve that level of traffic, and then allocate that amount all the time, in perpetuity. There are two problems with this approach: 1) if your planning was correct, the resources would be underutilized most of the time, essentially burning money; and 2) if your estimation was insufficient, you might end up affecting the availability of your service and end up with a far worse problem of losing the trust of your customers and ultimately their wallets.\n\nA smarter way to manage traffic loads is to monitor them as they are coming in and dynamically allocate and deallocate resources that are available for service. This ensures that the increased traffic is handled without loss of service while keeping operating costs to a minimum during low-traffic times.\n\nWhen scaling down resources, any resource that is about to be deallocated is quite likely to be processing traffic at that moment. It's essential to ensure that all of those requests be completed before shutting down that resource. Also, crucially, the resource must not process any new requests. This process is called _draining_. Draining is also crucial when machines are taken down for routine maintenance and\/or upgrades.\n\n## Low Latency\n\nConsider these facts. Amazon published a study in 2008 in which it found that every 100 ms increase in latency in its retail website resulted in a 1% loss of profit. A one-second delay in loading the website caused a whopping $1.6 billion in lost revenue! Google found that a 500 ms latency on mobile websites resulted in a traffic drop of 20%. In other words, a 20% decrease in the opportunity-to-serve advertisements. And this does not affect only industry giants. If a web page takes longer than three seconds to load on a mobile phone, 53% of users abandon it (according to a 2017 study by Google). It's clear that time is money.\n\nReporting average latency can be misleading because it might paint a cheerier picture than a ground reality. It's like saying if Bill Gates walks into a room, everyone is a billionaire on average. Instead, percentile latency is the typically reported metric. For example, a service might report 987 ms @ 99th percentile. This means that 99% of the requests were served in 987 ms or less. The very same system could have a 20 ms latency on average. Of course, as traffic to your service increases, the latency might increase if the service is not scaled up to give adequate resources. As such, latency, high availability, and scalability are intertwined.\n\n## Geographic Availability\n\nThe distance between New York and Sydney is nearly 10,000 miles (16,000 km). The speed of light in a vacuum is roughly 186,282 miles per second (300,000 km per second). Silica glass (used in fiber-optic cables) decreases the speed of light by a factor of roughly 30% down to 130,487 miles per second (210,000 km per second). On a piece of fiber-optic running in a straight line between these two cities, the roundtrip travel time alone for a single request is nearly 152 ms. Keep in mind that this does not account for the amount of time it takes for the request to be processed at the server, or the hops that the packets need to make across multiple routers along the way. This level of service would be unacceptable for many applications.\n\nServices that expect to be used throughout the world must be strategically located to minimize latency for the users in those regions. Additionally, resources can be dynamically scaled up or down depending on local traffic, thus giving more granular control. The major cloud providers have a presence on at least five continents (sorry penguins!).\n\n###### Tip\n\nWant to simulate how long the incoming requests would take from your computer to a particular datacenter around the world? Table 9-3 lists a few handy browser-based tools offered by cloud providers.\n\nTable 9-3. Latency measurement tools for different cloud providers **Service** | **Cloud provider**  \n---|---  \nAzureSpeed.com | Microsoft Azure  \nCloudPing.info | Amazon Web Services  \nGCPing.com | Google Cloud Platform\n\nAdditionally, to determine realistic combinations of latency from one location to another, _CloudPing.co_ measures AWS Inter-Region Latency, between more than 16 US-based AWS datacenters to one another.\n\n## Failure Handling\n\nThere's an old saying that there are only two things that are assured in life\u2014death and taxes. In the twenty-first century, this adage applies not just to humans but also computer hardware. Machines fail all the time. The question is never _if_ a machine will fail, it's _when_. One of the necessary qualities of production-quality service is its ability to gracefully handle failures. If a machine goes down, quickly bring up another machine to take its place and continue serving traffic. If an entire datacenter goes down, seamlessly route traffic to another datacenter so that users don't even realize that anything bad happened in the first place.\n\n## Monitoring\n\nIf you can't measure it, you can't improve it. Worse, does it even exist? Monitoring the number of requests, availability, latency, resource usage, number of nodes, distribution of traffic, and location of users is vital to understanding how a service performs; finding opportunities to improve it; and, more importantly, how much to pay. Most cloud providers already have built-in dashboards providing these metrics. Additionally, recording task-specific analytics like time for model inference, preprocessing, and so on can add another level of understanding.\n\n## Model Versioning\n\nWe have learned in this book (and will continue to learn all the way to the last page) that machine learning is always iterative. Particularly in the case of applications in the real world, data that the model can learn on is constantly being generated. Moreover, the incoming data distribution might shift over time compared to what it was trained on, leading to lower prediction power (a phenomenon called c _oncept drift_ ). To provide users with the best possible experience, we want to keep improving our models. Every time we train our model with newer data to further improve its accuracy and make the best version yet, we want to make it available for our users as quickly and seamlessly as possible. Any good production-quality inference system should provide the ability to provide different versions of a model, including the ability to swap a live version of the model with another version at a moment's notice.\n\n## A\/B Testing\n\nIn addition to supporting multiple versions of a model, there are reasons we'd want to serve different versions of the model at the same time depending on a variety of attributes such as the geographic location of the user, demographics, or simply by random assignment.\n\n_A\/B testing_ is a particularly useful tool when improving a model. After all, if our spanking new model was flawed in some way, we'd rather it be deployed to only a small subset of our users rather than 100% of them at the same time before we catch the flaw. Additionally, if a model meets criteria for success on that small subset, it provides validation for the experiment and justifies eventually being promoted to all users.\n\n## Support for Multiple Machine Learning Libraries\n\nLast but not least, we don't want to be locked into a single machine learning library. Some data scientists in an organization might train models in PyTorch, others in TensorFlow, or maybe scikit-Learn suffices for non\u2013deep learning tasks. The flexibility to support multiple libraries would be a welcome bonus.\n\n# Google Cloud ML Engine: A Managed Cloud AI Serving Stack\n\nConsidering all the desirable qualities we discussed in a production environment in the previous section, it's generally not a good idea to use Flask for serving users. If you do not have a dedicated infrastructure team and would like to spend more time making better models than deploying them, using a managed cloud solution is the right approach. There are several cloud-based Inference-as-a-Service solutions on the market today. We have chosen to explore the Google Cloud ML Engine partly because of the convenient TensorFlow integration and partly because it ties in nicely with the ML Kit material that we touch upon in Chapter 13.\n\n## Pros of Using Cloud ML Engine\n\n  * Easy-to-deploy models in production with web-based GUI\n\n  * Powerful and easily scalable to millions of users\n\n  * Provides deep insights into model usage\n\n  * Ability to version models\n\n## Cons of Using Cloud ML Engine\n\n  * High latency, offers only CPUs for inference (as of August 2019)\n\n  * Unsuitable for scenarios involving legal and data privacy issues where the data must not leave the network\n\n  * Imposes restrictions on architecture design of complex applications\n\n## Building a Classification API\n\nThe following step-by-step guide shows how to go about uploading and hosting our a Dog\/Cat classifier model on Google Cloud ML Engine:\n\n  1. Create a model on the Google Cloud ML Engine dashboard at _https:\/\/console.cloud.google.com\/mlengine\/models_. Because this is the first time we're using the dashboard, we need to click ENABLE API, as depicted in Figure 9-3.\n\n###### Figure 9-3. Listing page for machine learning models on the Google Cloud ML Engine dashboard\n\n  2. Give the model a name and a description (Figure 9-4).\n\n###### Figure 9-4. Model creation page on Google Cloud ML Engine\n\n  3. After the model is created, we can access the model on the listing page (Figure 9-5).\n\n###### Figure 9-5. Model listings page on Google Cloud ML Engine\n\n  4. Click the model to go to the model details page (Figure 9-6) and add a new version.\n\n###### Figure 9-6. Details page of the just-created Dog\/Cat classifier\n\n  5. Fill out the necessary information to create the new version. The last field at the bottom requires you to upload the model to Google Cloud Storage before we can use it. Click the Browse button to create a new bucket for storing the model (Figure 9-7).\n\n###### Figure 9-7. Creating a new version for a machine learning model\n\n  6. Create a new bucket with a unique name, a storage class, and region. After you create this bucket, go to _https:\/\/console.cloud.google.com\/storage\/browser_ (in a separate tab while keeping the current one open) to find this newly created bucket and upload the model there (Figure 9-8).\n\n###### Figure 9-8. Creating a new Google Cloud Storage bucket within the ML model version creation page\n\n  7. Our Dog\/Cat classifier model is an _.h5_ file. However, Google Cloud expects a SavedModel file. You can find the script to convert the _.h5_ file to SavedModel on the book's GitHub repository (see <http:\/\/PracticalDeepLearning.ai>) at _code\/chapter-9\/scripts\/h5_to_tf.ipynb_. Simply load the model and execute the rest of the notebook.\n\n  8. In the Google Cloud Storage browser, upload the newly converted model (Figure 9-9) to the bucket you created in step 6.\n\n###### Figure 9-9. Google Cloud Storage Browser page showing the uploaded Dog\/Cat classifier model in TensorFlow format\n\n  9. Specify the URI on the model version creation page for the model that you just uploaded (Figure 9-10).\n\n###### Figure 9-10. Add the URI for the model you uploaded to Google Cloud Storage\n\n  10. Click the Save button and wait for the model version to be created. As soon as the model version is created, you can begin making predictions against it.\n\n  11. If it's not already present on your machine, you can download and install the Google Cloud SDK from the installation website at _https:\/\/cloud.google.com\/sdk\/install_.\n\n  12. You can use the Cloud ML Engine REST API to make your requests. However, for brevity, use the command-line tools in the Cloud SDK. You first need to convert your image into a _request.json_ file using the _image-to-json.py_ script located at _code\/chapter-9_ :\n\n    $ python image-to-json.py --input dog.jpg --output request.json\n\n  13. Next, use the _request.json_ file created in the previous step to execute a request against our model:\n\n    $ time cloud ai-platform predict --model DogCat --version v1 \n                                     --json-instances\n    request.json\n\n    SCORES\n    [0.14749771356, 0.8525022864]\n\n    real     0m3.370s\n    user     0m0.811s\n    sys 0m0.182s\n\nAs you can see from the output, we get similar results as with our Flask server; that is, a prediction of \"dog\" with 85% confidence.\n\n###### Note\n\nIf this is your first time using **`gcloud`** , you need to run the following command to tie the command-line tool to your Google account:\n\n    $ gcloud auth login\n\nNext, select the project using the following command:\n\n    $ gcloud config set project {project_name}\n\nPiece of cake, wasn't it? In our example, we used the Google Cloud SDK to request a prediction for the sake of brevity. In a production scenario, you would want to execute the same prediction request using Google's API endpoints, instead; either by generating HTTP requests or by using their client libraries. We can follow the documentation on Google Cloud Docs for production scenarios.\n\nAt this point, the model is ready to be served to any user anywhere in the world using applications on the browser, mobile and edge devices, desktop, as well as cloud environments. Using a hosted stack is a pretty viable option for individuals and organizations who want the flexibility and reliability that the cloud provides while having to do minimal setup and maintenance work for the infrastructure.\n\nIn contrast, there are situations for which a hosted solution might not be the best approach. Reasons could include pricing models, data privacy issues, legal questions, technical issues, trust concerns, or contractual obligations. In such cases, a solution that is hosted and managed locally (or, \"on-premises\") would be preferable.\n\n###### Tip\n\nFor processing a large number of images at one time, you can modify _image-to-json.py_ to create a _request.json_ file that contains an array of multiple inputs.\n\n# TensorFlow Serving\n\nTensorFlow Serving is an open source library in the TensorFlow ecosystem for serving machine learning models fast. Unlike Flask, it's built for performance, with low overhead, and designed for use in production. TensorFlow Serving is widely used by large companies to serve their models for prediction services. It is one of the integral components of TensorFlow Extended (TFX)\u2014an end-to-end deep learning pipeline in the TensorFlow ecosystem.\n\nAs we saw when we looked at the desired qualities of a production system, TensorFlow serving offers low latency, failure handling, high throughput, and model versioning. Another benefit includes the ability to serve multiple models at the same time on the same service. It implements several techniques to speed up serving:\n\n  * During server startup, it starts a burst of threads for fast model loading.\n\n  * It uses separate thread pools for loading models and for serving inferences while giving higher priority to the threads in the inference pool. This is crucial to lowering request latency.\n\n  * It builds minibatches of incoming asynchronous requests for short periods of time. As we have seen, with the power of batching data on GPUs during training, it aims to bring similar efficiencies on asynchronous requests. As an example, waiting 500 ms to group together several requests for inference. While at worst case, this adds a 500 ms penalty for the first request in the batch, it reduces the average latency across requests and maximizes hardware utilization.\n\n###### Note\n\nTensorFlow serving gives you full control over the model rollout procedure. You can serve different models or different versions of the same kind of model in the same process. You just need to make sure that you know the name and location of the version that you want to remove or put into production.\n\n## Installation\n\nThere are a few different ways of setting up TensorFlow Serving:\n\n  * Building from source\n\n  * Downloading and installing using APT\n\n  * Deploying Docker images\n\nIf you're feeling adventurous, building from source might be the danger you seek. But if you just want to get up and running quickly, we recommend using Docker because it requires minimal steps to get the system up and running. What is Docker, you might ask? Docker provides virtualization of a Linux environment for applications that run within it. It provides isolation of resources that essentially operate as a clean slate for setting up an environment in which an application can run. Typically, an application and all of its dependencies are packaged into a single Docker container that can then be deployed repeatedly as necessary. Because the application is set up in a clean environment, it reduces the likelihood of configuration and deployment errors. This makes Docker very well suited for running applications in production.\n\nThe biggest benefit that Docker provides for us is alleviating \"dependency hell\" because all the necessary dependencies are packaged within the container. One additional advantage of using Docker is that the process of setting up your application remains more or less the same across different platforms, whether you use Windows, Linux, or Mac.\n\nThe Docker installation instructions, depending on the target platform, are available on the Docker home page. This should not take more than a few minutes because the setup is fairly straightforward. After you've installed Docker, you can run the following command to set up TensorFlow Serving for CPU:\n\n    $ docker run -p 8501:8501 \\\n    --mount type=bind,source=\/path\/to\/dogcat\/,target=\/models\/dogcat \\\n    -e MODEL_NAME=dogcat -t tensorflow\/serving\n\nFor GPU-enabled machines, run the following command, instead:\n\n    $ docker run -p 8501:8501 --runtime=nvidia \\\n    --mount type=bind,source=\/path\/to\/dogcat\/,target=\/models\/dogcat \\\n    -e MODEL_NAME=dogcat -t tensorflow\/serving\n\nIn either case, if everything went smoothly, you should have a REST API running on your local port 8501 serving our Dog\/Cat classifier.\n\n###### Note\n\nIn any inference request, the end-to-end latency is a summation of time taken by multiple steps along the process. This includes round-trip network time, request time to serialize\/deserialize the request and response objects, and, of course, time to perform the actual inference. One more component that adds overhead is the serving framework; that is, TensorFlow Serving. Google claims that the overhead contributed by TensorFlow Serving is minimal. In its experiments, it observed that TensorFlow Serving alone was able to handle approximately 100,000 QPS per core on a 16 vCPU Intel Xeon E5 2.6 GHz machine. Because it is measuring the overhead, this excludes the remote procedure call (RPC) time and the TensorFlow inference processing time.\n\nEven though TensorFlow Serving is a great choice for serving inferences from a single machine, it does not have built-in functionality for horizontal scaling. Instead, it is built to be used in conjunction with other systems that can supercharge TensorFlow Serving with dynamic scaling. We explore one such solution in the following section.\n\n# KubeFlow\n\nThroughout this book, we have explored the various steps of an end-to-end deep learning pipeline, from data ingestion, analysis, distributed training (including hyperparameter tuning) at scale, tracking experiments, deployment, and eventually to serving prediction requests at scale. Each of these steps is complex in its own right, with its set of tools, ecosystems, and areas of expertise. People dedicate their lifetimes developing expertise in just one of these fields. It's not exactly a walk in the park. The combinatorial explosion of the know-how required when factoring for the necessary backend engineering, hardware engineering, infrastructure engineering, dependency management, DevOps, fault tolerance, and other engineering challenges can result in a very expensive hiring process for most organizations.\n\nAs we saw in the previous section, Docker saves us the hassle of dependency management by making portable containers available. It helps us make TensorFlow Serving available across platforms easily without having to build it from source code or install dependencies manually. Great! But it still doesn't have an answer to many of the other challenges. How are we going to scale up containers to match rises in demand? How would we efficiently distribute traffic across containers? How do we ensure that the containers are visible to one another and can communicate?\n\nThese are questions answered by _Kubernetes_. Kubernetes is an orchestration framework for automatically deploying, scaling, and managing containers (like Docker). Because it takes advantage of the portability offered by Docker, we can use Kubernetes to deploy to developer laptops as well as thousand-machine clusters in an almost identical manner. This helps maintain consistency across different environments, with the added benefit of scalability in an accessible manner. It is worth noting that Kubernetes is not a dedicated solution for machine learning (neither is Docker); rather, it is a general-purpose solution to many of the problems faced in software development, which we use in the context of deep learning.\n\nBut let's not get ahead of ourselves just yet. After all, if Kubernetes were the be-all and end-all solution, it would have appeared in the chapter title! A machine learning practitioner using Kubernetes still needs to assemble all of the appropriate sets of containers (for training, deployment, monitoring, API management, etc.) that then need to be orchestrated together to make a fully functioning end-to-end pipeline. Unfortunately, many data scientists are trying to do exactly this in their own silos, reinventing the wheel building ad hoc machine learning-specific pipelines. Couldn't we save everyone the trouble and make one Kubernetes-based solution for machine learning scenarios?\n\nEnter _KubeFlow_ , which promises to automate a large chunk of these engineering challenges and hide the complexity of running a distributed, scalable, end-to-end deep learning system behind a web GUI-based tool and a powerful command-line tool. This is more than just an inference service. Think of it as a large ecosystem of tools that can interoperate seamlessly and, more importantly, scale up with demand. KubeFlow is built for the cloud. Though not just one cloud\u2014it's built to be compatible with all major cloud providers. This has significant implications on cost. Because we are not tied to a specific cloud provider, we have the freedom to move all of our operations at a moment's notice if a competing cloud provider drops its prices. After all, competition benefits consumers.\n\nKubeFlow supports a variety of hardware infrastructure, from developer laptops and on-premises datacenters, all the way to public cloud services. And because it's built on top of Docker and Kubernetes, we can rest assured that the environments will be identical whether deployed on a developer laptop or a large cluster in a datacenter. Every single way in which the developer setup is different from the production environment could result in an outage, so it's really valuable to have this consistency across environments.\n\nTable 9-4 shows a brief list of readily available tools within the KubeFlow ecosystem.\n\nTable 9-4. Tools available on KubeFlow **Tool** | **Functionality**  \n---|---  \nJupyter Hub | Notebook environment  \nTFJob | Training TensorFlow models  \nTensorFlow Serving | Serving TensorFlow models  \nSeldon | Serving models  \nNVIDIA TensorRT | Serving models  \nIntel OpenVINO | Serving models  \nKFServing | Abstraction for serving Tensorflow, XGBoost, scikit-learn, PyTorch, and ONNX models  \nKatib | Hyperparameter tuning and NAS  \nKubebench | Running benchmarking jobs  \nPyTorch | Training PyTorch models  \nIstio | API services, authentication, A\/B testing, rollouts, metrics  \nLocust | Load testing  \nPipelines | Managing experiments, jobs, and runs, scheduling machine learning workflows\n\nAs the joke goes in the community, with so many technologies prepackaged, KubeFlow finally makes our r\u00e9sum\u00e9s buzzword- (and recruiter-) compliant.\n\n###### Note\n\nMany people assume that KubeFlow is a combination of Kubernetes and TensorFlow, which, as you have seen, is not the case. It is that and much more.\n\nThere are two important parts to KubeFlow that make it unique: pipelines and fairing.\n\n## Pipelines\n\nPipelines give us the ability to compose steps across the machine learning to schedule complex workflows. Figure 9-11 shows us an example of a pipeline. Having visibility into the pipeline through a GUI tool helps stakeholders understand it (beyond just the engineers who built it).\n\n###### Figure 9-11. An end-to-end pipeline illustrated in KubeFlow\n\n## Fairing\n\nFairing allows us to manage the entire build, train, and deploy lifecycle directly through Jupyter Notebooks. Figure 9-12 shows how to start a new notebook server, where we can host all of our Jupyter Notebooks, run training on them, and deploy our models to Google Cloud using the following few lines of code, all the while being in the comfort of a very familiar Jupyter environment:\n\n    from fairing.deployers.gcp.gcpserving import GCPServingDeployer\n    GCPServingDeployer().deploy(model_dir, model_name, version_name)\n\n###### Figure 9-12. Creating a new Jupyter Notebook server on KubeFlow\n\n## Installation\n\nCreating a new KubeFlow deployment is a fairly straightforward process that is well documented on the KubeFlow website. You can set up KubeFlow using the browser for GCP. Alternatively, you can use the KubeFlow command-line tool to set up a deployment on GCP, AWS, and Microsoft Azure. Figure 9-13 shows a GCP deployment using the web browser.\n\n###### Figure 9-13. Creating a KubeFlow deployment on GCP using the browser\n\nAs of this writing, KubeFlow is in active development and shows no signs of stopping. Companies such as Red Hat, Cisco, Dell, Uber, and Alibaba are some of the active contributors on top of cloud giants like Microsoft, Google, and IBM. Ease and accessibility for solving tough challenges attract more people to any platform, and KubeFlow is doing exactly that.\n\n# Price Versus Performance Considerations\n\nIn Chapter 6, we looked at how to improve our model performance for inference (whether on smartphones or on a server). Now let's look from another side: the hardware performance and the price involved.\n\nOften while building a production system, we want the flexibility of choosing suitable hardware to strike the proper balance between performance, scale, and price for our scenario. Consider building an app that requires cloud-based inference. We can go set up our own stack manually (using Flask or TensorFlow Serving or KubeFlow), or we could use a managed Inference-as-a-Service stack (like the Google Cloud ML Engine). Assuming that our service went viral, let's see how much it would cost.\n\n## Cost Analysis of Inference-as-a-Service\n\nFor Google Cloud ML Engine, as of August 2019 in North America, it costs a rather inexpensive $0.0401 per hour of combined inference time on a single-core CPU machine. There's also an option for a quad-core CPU machine, but really, a single core should suffice for most applications. Running several queries to the server with a small image of 12 KB took roughly 3.5 seconds on average, as illustrated in Figure 9-14. This does sound slow, and is partly because of doing inference on a moderate-speed machine, and, more important on a CPU server. It's worth mentioning that this benchmark is on a warmed-up machine that has recently received an API request and hence has the model preloaded. For comparison, the first query takes between 30 and 60 seconds. This shows the importance of keeping the service running constantly or sending frequent warm-up queries. This happens because the Google Cloud ML engine takes down a model if it notices a prolonged period of nonuse.\n\n###### Figure 9-14. Google Cloud ML Engine showing incoming queries and latency of serving the calls, with end-to-end latency at user's end of about 3.5 seconds\n\nIf a request came in at every second for an entire month, there would be a total of 60 x 60 x 24 x 30 = 2,592,000 calls per month. Assuming that each inference takes 3.5 seconds, a single node would be insufficient. The cloud service would quickly realize that and, in response to the increased traffic, bring up three additional machines to handle the traffic. In total, with four machines running for a month at $0.0401 per hour per node, it would cost a grand total of $115.48. To put this into perspective, for two million calls, that's about the cost of a cup of Starbucks coffee a day for an entire month. And let's not forget this is without involving much of the DevOps team members, whose time is expensive. If we took the hypothetical scenario of a Yelp-like service for which users, on average, upload photos of food at 64 QPS, running inferences on them using a classification model would cost only $7,390.\n\n## Cost Analysis of Building Your Own Stack\n\nLess spending and high scalability, now that's a winning combination. But the one downside is the total roundtrip latency of each request. Taking matters into our own hands, getting a VM with a modest GPU on the cloud, and setting up our scaling pipeline (using KubeFlow or the native cloud load-balancing features with TensorFlow Serving), we would be able to respond either in milliseconds or batch a few incoming queries together (say every 500 ms) to serve them. As an example, looking at the inventory of VMs on Azure, for $2.07 per hour, we can rent out an ND6 machine that features an NVIDIA P40 GPU and 112 GiB RAM. By batching incoming requests every 500 ms to 1 second, this machine can serve 64 requests per second at a total cost of $1,490 per month, and faster than the Google Cloud ML Engine.\n\nIn summary, the cost savings and performance benefits of orchestrating our own cloud machine environment kicks in big time when working on large QPS scenarios, as demonstrated in Figure 9-15.\n\n###### Figure 9-15. Cost comparison of infrastructure as a service (Google Cloud ML Engine) versus building your own stack over virtual machines (Azure VM) (costs as of August 2019)\n\n###### Tip\n\nA common question that arises while benchmarking is what is my system's limit? JMeter can help answer this. JMeter is a load-testing tool that lets you perform stress testing of your system with an easy-to-use graphical interface. It lets you create reusable configurations to simulate a variety of usage scenarios.\n\n# Summary\n\nIn this chapter, we answered the question most engineers and developers ask: how do we serve model prediction requests at scale for applications in the real world? We explored four different methods of serving an image recognition model: using Flask, Google Cloud ML, TensorFlow Serving, and KubeFlow. Depending on the scale, latency requirements, and our skill level, some solutions might be more attractive than others. Finally, we developed an intuition into the cost effectiveness of different stacks. Now that we can show our fabulous classifier model off to the world, all that's left is to make our work go viral!\n\n# Chapter 10. AI in the Browser with TensorFlow.js and ml5.js\n\nWritten in collaboration with guest author: Zaid Alyafeai\n\nYou're a developer who dreams big. You have a kickass AI model that you would like a lot of people to try. How many is a lot? Ten thousand? A million? No, silly. You like to dream big. How about 100 million people? That's a nice round number. Now convincing 100 million people to download and install an app and make space for it on their phones is not an easy sell. But what if we told you that they all have an app already installed, just for you. No downloads. No installs. No app stores. What kind of black magic is this!? Of course, it's the web browser. And as a bonus, it also runs on your PC.\n\nThis is what Google did with its home page when it decided to launch its first-ever AI doodle to their billions of users (Figure 10-1). And what better theme to pick for it than the music of J.S. Bach. (Bach's parents wanted to call him J.S. Bach, 310 years before JavaScript was even created. They had quite the foresight!)\n\nTo explain briefly, the doodle allowed anyone to write one line (voice) of random notes for two measures using mouse clicks. When the user clicked a button labeled Harmonize, the input would then be processed against hundreds of musical pieces written by Bach that contain between two and four lines (voices) of music. The system would figure out which notes would sound best along with the user's input to create a much richer Bach-like sounding musical piece. The entire process ran in the browser, so Google would not need to scale up its machine learning prediction infrastructure at all.\n\n###### Figure 10-1. The Bach music harmonizer doodle from Google\n\nIn addition to the cost savings and the ability to run on any platform, with a browser we can provide users with a richer, more interactive experience because network latency is not a factor. And of course, because everything can be run locally after the model is downloaded, the end user can benefit from the privacy of their data.\n\nGiven that JavaScript is the language of the web browser, it's useful for us to delve into JavaScript-based deep learning libraries that can run our trained model within users' browsers. And that's exactly what we do in this chapter.\n\nHere, we focus on implementing deep learning models in the browser. First, we look at a brief history of different JavaScript-based deep learning frameworks before moving on to TensorFlow.js and eventually a higher-level abstraction for it called ml5.js. We also examine a few complex browser-based applications such as detecting the body pose of a person or converting a hand-drawn doodle to a photograph (using GANs). Finally, we talk about some practical considerations and showcase some real-world case studies.\n\n# JavaScript-Based Machine Learning Libraries: A Brief History\n\nSince the breakthrough of deep learning in recent years, many attempts have been made to make AI accessible to a wider range of people in the form of web-based libraries. Table 10-1 offers a brief overview of the different libraries in the order in which they were first released.\n\nTable 10-1. Historical overview of different JavaScript-based deep learning libraries (data captured as of August 2019) **** | **Active years** | **\u2605 on GitHub** | **Known for**  \n---|---|---|---  \nbrain.js | 2015\u2013present | 9,856 | Neural networks, RNNs, LSTMs, and GRUs  \nConvNetJS | 2014\u20132016 | 9,735 | Neural networks, CNNs  \nSynaptic | 2014\u2013present | 6,571 | Neural networks, LSTMs  \nMXNetJS | 2015\u20132017 | 420 | Running MXNet models  \nKeras.js | 2016\u20132017 | 4,562 | Running Keras models  \nCaffeJS | 2016\u20132017 | 115 | Running Caffe models  \nTensorFlow.js (formerly known as deeplearn.js) | 2017\u2013present | 11,282 | Running TensorFlow models on GPU  \nml5.js | 2017\u2013present | 2,818 | Easy to use on top of TF.js.  \nONNX.js | 2018\u2013present | 853 | Speed, running ONNX models\n\nLet's go through a few of these libraries in more detail and see how they evolved.\n\n## ConvNetJS\n\nConvNetJS is a JavaScript library that was designed in 2014 by Andrej Karpathy as part of a course during his Ph.D. at Stanford University. It trained CNNs in the browser, an exciting proposition, especially in 2014, considering the AI hype was starting to take off, and a developer wouldn't have had to go through an elaborate and painful setup process to get running. ConvNetJS helped introduce AI to so many people for the first time with interactive training demonstrations in the browser.\n\n###### Note\n\nIn fact, when MIT scientist Lex Fridman taught his popular self-driving course in 2017, he challenged students worldwide to train a simulated autonomous car using reinforcement learning\u2014in the browser using ConvNetJS\u2014as shown in Figure 10-2.\n\n###### Figure 10-2. Screenshot of DeepTraffic training a car with reinforcement learning using ConvNetJS\n\n## Keras.js\n\nKeras.js was introduced in 2016 by Leon Chen. It was a Keras port made to work in the browser by using JavaScript. Keras.js used WebGL to run computations on the GPU. It used shaders (special operations for pixel rendering) to run inferences, which made them run much faster than using just the CPU. Additionally, Keras.js could run on a Node.js server on a CPU to provide server-based inferences. Keras.js implemented a handful of convolutional, dense, pooling, activation, and RNN layers. It is no longer under active development.\n\n## ONNX.js\n\nCreated by Microsoft in 2018, ONNX.js is a JavaScript library for running ONNX models in browsers and on Node.js. ONNX is an open standard for representing machine learning models that is a collaboration between Microsoft, Facebook, Amazon, and others. ONNX.js is surprisingly fast. In fact, faster than even TensorFlow.js (discussed in the next section) in early benchmarks, as shown in Figure 10-3 and Figure 10-4. This could be attributed to the following reasons:\n\n  * ONNX.js utilizes WebAssembly (from Mozilla) for execution on the CPU and WebGL on the GPU.\n\n  * WebAssembly allows it to run C\/C++ and Rust programs in the web browser while providing near-native performance.\n\n  * WebGL provides GPU-accelerated computations like image processing within the browser.\n\n  * Although browsers tend to be single-threaded, ONNX.js uses Web Workers to provide a multithreaded environment in the background for parallelizing data operations.\n\n###### Figure 10-3. Benchmarking data for ResNet-50 on different JavaScript machine learning libraries on CPU (data source)\n\n###### Figure 10-4. Benchmarking data for ResNet-50 on different JavaScript machine learning libraries on GPU (data source)\n\n## TensorFlow.js\n\nSome libraries offered the ability to train within the browser (e.g., ConvNetJS), whereas other libraries offered blazing-fast performance (e.g., the now-defunct TensorFire). deeplearn.js from Google was the first library that supported fast GPU accelerated operations using WebGL while also providing the ability to define, train, and infer within the browser. It offered both an immediate execution model (for inference) as well as a delayed execution model for training (like in TensorFlow 1.x). Originally released in 2017, this project became the core of TensorFlow.js (released in 2018). It is considered an integral part of the TensorFlow ecosystem, and as a result, it is currently the most actively developed JavaScript deep learning library. Considering this fact, we focus on TensorFlow.js in this chapter. To make TensorFlow.js even simpler to use, we also look at ml5.js, which is built on top of TensorFlow.js and abstracts away its complexities, exposing a simple API with ready-to-use models from GANs to PoseNet.\n\n##### From the Creator's Desk\n\nBy Shanqing Cai, senior software engineer at Google and author of _Deep Learning with JavaScript_ (Manning)\n\nThe forebearer of TensorFlow.js, deeplearn.js, originated from an effort at Google to create an intuitive and interactive visualization to teach people how neural networks are trained. This visualization, today known as \"TensorFlow Playground\" and available at _https:\/\/playground.tensorflow.org_, used an early version of deeplearn.js to train a multilayered neural network entirely in the browser. In building TensorFlow Playground, the engineers became impressed by the potential of using WebGL to perform accelerated training and inference of deep learning models in the browser and at the client side. A team of engineers was assembled at Google to realize this vision, which gave birth to today's TensorFlow.js, a full-blown deep learning library that supports hundreds of operations and dozens of neural network layers and runs on environments ranging from the browser to Node.js, from native-mobile apps to cross-platform desktop apps.\n\n##### From the Creator's Desk\n\nBy Daniel Smilkov, software engineer at Google Brain and coauthor of TensorFlow.js\n\nBefore TensorFlow.js, Nikhil [Thorat] and I were building neural network interpretability tools in the browser. To enable a truly interactive experience, we wanted to run inference and compute gradients directly in the browser without sending data to a server. This led to deeplearn.js (you can still see the package with its API on npm), which we released in August 2017. The project got great momentum with creative coders being one of the earliest adopters. Riding on this momentum, the team grew quickly, and six months later we launched TensorFlow.js.\n\n# TensorFlow.js Architecture\n\nFirst, let's take a look at the high-level architecture of TensorFlow.js (see Figure 10-5). TensorFlow.js runs directly in the browser on desktop and mobile. It utilizes WebGL for GPU acceleration, but also can fall back to the browser's runtime for execution on the CPU.\n\nIt consists of two APIs: the Operations API and the Layers API. The Operations API provides access to lower-level operations such as tensor arithmetic and other mathematical operations. The Layers API builds on top of the Operations API to provide layers such as convolution, ReLU, and so on.\n\n###### Figure 10-5. A high-level overview of the TensorFlow.js and ml5.js ecosystem\n\nBeyond the browser, TensorFlow.js can also run on a Node.js server. Additionally, ml5.js uses TensorFlow.js to provide an even higher-level API along with several prebuilt models. Having access to all of these APIs at different levels of abstraction allows us to build web apps, not only to do simple inference, but also to train models within the browser itself.\n\nFollowing are some common questions that come up during the development life cycle for browser-based AI:\n\n  * How do I run pretrained models in the browser? Can I use my webcam feed for real-time interactivity?\n\n  * How can I create models for the browser from my TensorFlow trained models?\n\n  * Can I even train a model in the browser?\n\n  * How do different hardware and browsers affect performance?\n\nWe answer each of these questions in this chapter, starting with TensorFlow.js before moving on to ml5.js. We explore some rich built-in functionality contributed by the ml5.js community, which would otherwise take a lot of effort and expertise to implement directly on TensorFlow.js. We also look at approaches to benchmarking before looking at some motivating examples built by creative developers.\n\nNow let's take a look at how to take advantage of pretrained models to make inferences within the browser.\n\n# Running Pretrained Models Using TensorFlow.js\n\nTensorFlow.js offers lots of pretrained models that we can directly run in the browser. Some examples include MobileNet, SSD, and PoseNet. In the following example, we load a pretrained MobileNet model. The full code is located on the book's GitHub repository (see _http:\/\/PracticalDeepLearning.ai_) at _code\/chapter-10\/mobilenet-example\/_.\n\nFirst, we import the latest bundle of the library:\n\n    <script src=\"https:\/\/cdn.jsdelivr.net\/npm\/@tensorflow\/\n    tfjs@latest\/dist\/tf.min.js\"><\/script>\n\nWe then import the MobileNet model:\n\n    <script src=\"https:\/\/cdn.jsdelivr.net\/npm\/\n    @tensorflow-models\/mobilenet@1.0.0\"><\/script>\n\nNow we can make predictions using the following code:\n\n    <img id=\"image\" src=\"cat.jpg\" \/>\n    <p id=\"prediction_output\">Loading predictions...<\/p>\n    <script>\n     const image = document.getElementById(\"image\");\n     const predictionOutput = document.getElementById(\"prediction_output\");\n\n     _\/\/ Load the model._\n     mobilenet.load().then(model => {\n       _\/\/ Classify the image. And output the predictions_\n       model.classify(image).then(predictions => {\n         predictionOutput.innerText = predictions[0].className;\n       });\n     });\n    <\/script>\n\nFigure 10-6 shows a sample output.\n\n###### Figure 10-6. Class prediction output in the browser\n\nWe can alternatively load a model using a JSON file URL in the following manner:\n\n    const path = 'https:\/\/storage.googleapis.com\/tfjs-\n    models\/tfjs\/mobilenet_v1_1.0_224\/model.json';\n\n    model = tf.loadLayersModel(path).then(model => {\n        _\/\/ Load model and output predictions here_\n    });\n\nThe JSON file contains the architecture, the parameter names of the model, and the paths to the smaller sharded weight files. The sharding allows the files to be small enough to be cached by web browsers, which would make loading faster for any subsequent time the model would be needed.\n\n# Model Conversion for the Browser\n\nIn the previous section, we examined how to load a pretrained model that is already in the JSON format. In this section, we learn how to convert a pretrained Keras model (.`h5` format) to JSON that is compatible with TensorFlow.js. To do this, we need to install the conversion tool using `pip`.\n\n    $ pip install tensorflowjs\n\nAssuming that our trained Keras model is stored in a folder named _keras_model_ , we would be able to use the following command to convert it:\n\n    $ tensorflowjs_converter --input_format keras keras_model\/model.h5 web_model\/\n\nNow the `web_model` directory will contain the `.json` and `.shard` files that we can easily load using the `tf.loadLayersModel` method:\n\n    $ ls web_model\n    group1-shard1of4  group1-shard3of4  model.json group1-shard2of4  group1-shard4of4\n\nThat's it! Bringing our trained model to the browser is an easy task. For cases in which we don't already have an existing trained model, TensorFlow.js also allows us to train models directly in the browser. In the next section, we explore this by creating an end-to-end example of training a model using a webcam feed.\n\n###### Tip\n\nLoading a model locally requires running a web server. There are tons of options that we can use, ranging from the LAMP (Linux, Apache, MySQL, PHP) stack to installing `http-server` using npm, to even running Internet Information Services (IIS) on Windows to test model loading locally. Even Python 3 can run a simple web server:\n\n    $ python3 -m http.server 8080\n\n# Training in the Browser\n\nThe previous example used a pretrained model. Let's take it up a notch and train our own models directly in the browser using input from the webcam. Like in some of the previous chapters, we look at a simple example in which we exploit transfer learning to make the training process faster.\n\nAdapted from Google's Teachable Machine, we use transfer learning to construct a simple binary classification model, which will be trained using the webcam feed. To build this, we need a feature extractor (converts input images into features or embeddings), and then attach a network that converts these features into a prediction. Finally, we can train it with webcam inputs. The code is referenced in the book's GitHub repository (see _http:\/\/PracticalDeepLearning.ai_) at _code\/chapter-10\/teachable-machine_.\n\n###### Note\n\nGoogle Creative Lab built a fun interactive website called Teachable Machine where the user can train three classes on any type of classification problem by simply showing those objects in front of the webcam. The three classes are given simple labels of green, purple, and orange. During prediction, rather than showing bland-looking class probabilities in text on a web page (or even worse, in the console), Teachable Machine shows GIFs of cute animals or plays different sounds based on the class that is being predicted. As you can imagine, this would be a fun and engaging experience for kids in a classroom, and it would serve as a wonderful tool to introduce them to AI.\n\n###### Figure 10-7. Training live in the browser with Teachable Machine\n\n## Feature Extraction\n\nAs we explored in some of the early chapters in this book, training a large model from scratch is a slow process. It's a lot cheaper and quicker to use a pretrained model and use transfer learning to customize the model for our use case. We'll use that model to extract high-level features (embeddings) from the input image, and use those features to train our custom model.\n\nFor extracting the features, we load and use a pretrained MobileNet model:\n\n    const path = 'https:\/\/storage.googleapis.com\/tfjs-\n    models\/tfjs\/mobilenet_v1_1.0_224\/model.json';\n    const mobilenet = await tf.loadLayersModel(path);\n\nLet's inspect the inputs and outputs of the model. We know that the model was trained on ImageNet and the final layer predicts a probability of each one of the 1,000 classes:\n\n    const inputLayerShape = mobilenet.inputs[0].shape; \/\/ [null, 224, 224, 3]\n    const outputLayerShape = mobilenet.outputs[0].shape; \/\/ [null, 1000]\n    const numLayers = mobilenet.layers.length; \/\/ 88\n\nTo extract the features, we select a layer closer to the output. Here we select `conv_pw_13_relu` and make it the output of the model; that is, remove the dense layers at the end. The model we created is called a feature extraction model:\n\n    \/\/ Get a specific layer\n    const layer = mobilenet.getLayer('conv_pw_13_relu');\n\n    \/\/ Create a new feature extraction model\n    featureExtractionModel = tf.model({inputs: mobilenet.inputs, outputs:\n    layer.output});\n\n    featureExtractionModel.layers.length; _\/\/ 82_\n\nWe'll keep the feature extraction model unmodified during our training process. Instead, we add a trainable set of layers on top of it to build our classifier:\n\n    const trainableModel = tf.sequential({\n        layers: [\n            tf.layers.flatten({inputShape: [7, 7, 1024]}),\n            tf.layers.dense({\n            units: 64,\n            activation: 'relu',\n            kernelInitializer: 'varianceScaling',\n            useBias: true\n        }),\n        tf.layers.dense({\n            units: 2,\n            kernelInitializer: 'varianceScaling',\n            useBias: false,\n            activation: 'softmax'\n        })]\n    });\n\n## Data Collection\n\nHere, we collect images using the webcam feed and process them for feature extraction. The `capture()` function in the Teachable Machine is responsible for setting up the `webcamImage` for storing the captured images from the webcam in memory. Now, let's preprocess them to make them applicable to the feature extraction model:\n\n    function capture() {\n         return tf.tidy(() => {\n             \/\/ convert to a tensor\n             const webcamImage = tf.fromPixels(webcam);\n             \/\/ crop to 224x224\n             const croppedImage = cropImage(webcamImage);\n             \/\/ create batch and normalize\n             const batchedImage = croppedImage.expandDims(0);\n             return batchedImage.toFloat().div(tf.scalar(127)).sub(tf.scalar(1));\n         });\n    }\n\nAfter we capture the images, we can add the image and label to the training data:\n\n    function addTrainingExample(img, label) {\n         \/\/ Extract features.\n         const data = featureExtractionModel.predict(img);\n         \/\/ One-hot encode the label.\n         const oneHotLabel = tf.tidy(() =>\n         tf.oneHot(tf.tensor1d([label], 'int32'), 2));\n         \/\/ Add the label and data to the training set.\n    }\n\n## Training\n\nNext, we train the model, much as we trained in Chapter 3. Just like in Keras and TensorFlow, we add an optimizer and define the loss function:\n\n    const optimizer = tf.train.adam(learningRate);\n    model.compile({ optimizer: optimizer, loss: 'categoricalCrossentropy' });\n    model.fit(data, label, {\n        batchSize,\n        epochs: 5,\n        callbacks: {\n            onBatchEnd: async (batch, logs) => {\n                await tf.nextFrame();\n            }\n        }\n    }\n\n###### Note\n\nOne important thing to keep in mind is that on the GPU, memory allocated by TensorFlow.js is not released when a `tf.tensor` object goes out of scope. One solution is to call the `dispose()` method on every single object created. However, that would make the code more difficult to read, particularly when chaining is involved. Consider the following example:\n\n    const result = a.add(b).square().neg();\n    return result;\n\nTo cleanly dispose of all memory, we'd need to break it down into the following:\n\n    const sum = a.add(b);\n    const square = sum.square();\n    const result = square.neg();\n    sum.dispose();\n    square.dispose();\n    return result;\n\nInstead, we could simply use `tf.tidy()` to do the memory management for us, while keeping our code clean and readable. Our first line simply needs to be wrapped inside the `tf.tidy()` block, as shown here:\n\n    const result = tf.tidy(() => {\n        return a.add(b).square().neg();\n    });\n\nWith a CPU backend, the objects are automatically garbage collected by the browser. Calling `.dispose()` there does not have any effect.\n\nAs a simple use case, let's train the model to detect emotions. To do so, we need to simply add training examples belonging to either one of two classes: happy or sad. Using this data, we can start training. Figure 10-8 shows the final result after training the model on 30 images per each class.\n\n###### Figure 10-8. Predictions by our model on a webcam feed within a browser\n\n###### Tip\n\nUsually, when using a webcam for predictions, the UI tends to freeze. This is because the computation happens on the same thread as the UI rendering. Calling `await` `tf.nextFrame()` will release the UI thread, which will make the web page responsive and prevent the tab\/browser from freezing.\n\n## GPU Utilization\n\nWe can see the CPU\/GPU utilization during training and inference using the Chrome profiler. In the previous example, we recorded the utilization for 30 seconds and observed the GPU usage. In Figure 10-9, we see that the GPU is used a quarter of the time.\n\n###### Figure 10-9. GPU utilization shown in the Google Chrome profiler view\n\nSo far, we've discussed how to do everything from scratch, including loading the model, capturing video from the webcam, collecting the training data, training the model, and running an inference. Wouldn't it be great if all these steps could be taken care of under the hood and we could just focus on what to do with the results of the inference? In the next section, we discuss exactly that with ml5.js.\n\n# ml5.js\n\nml5.js is a higher abstraction of TensorFlow.js that makes it easy to use existing pretrained deep learning models in a unified way, with a minimal number of lines of code. The package comes with a wide range of built-in models, ranging from image segmentation to sound classification to text generation, as shown in Table 10-2. Further, ml5.js reduces the steps related to preprocessing, postprocessing, and so on, letting us concentrate on building the application that we want with these models. For each of these functionalities, ml5js comes with a demonstration and reference code.\n\nTable 10-2. Selected built-in models in ml5.js, showing the range of functionalities in text, image, and sound **Functionality** | **Description**  \n---|---  \nPoseNet | Detect the location of human joints  \nU-Net | Object segmentation; e.g., removing object background  \nStyle Transfer | Transfers style of one image to another  \nPix2Pix | Image-to-image translation; e.g., black-and-white to color  \nSketch RNN | Creates doodles based on an incomplete sketch  \nYOLO | Object detection; e.g., locates faces with bounding boxes  \nSound Classifier | Recognizes audio; e.g., whistle, clap, \"one,\" \"stop,\" etc.  \nPitch Detector | Estimates pitch of sound  \nChar RNN | Generates new text based on training on a large corpus of text  \nSentiment Classifier | Detects sentiment of a sentence  \nWord2Vec | Produces word embeddings to identify word relations  \nFeature Extractor | Generates features or embeddings from input  \nkNN Classifier | Creates a fast classifier using _k_ -Nearest Neighbor\n\nLet's see it in action. First, we import the latest bundle of ml5.js, which is similar to TensorFlow.js:\n\n    <script src=\"https:\/\/unpkg.com\/ml5@latest\/dist\/ml5.min.js\"\n    type=\"text\/javascript\"><\/script>\n\nNote that we no longer need to import anything related to TensorFlow.js, because it already comes included with ml5.js. We create a simple example where we use the same MobileNet scenario as earlier:\n\n    \/\/ Initialize the image classifier method with MobileNet\n    const classifier = ml5.imageClassifier('MobileNet', modelLoaded);\n\n    \/\/ Make a prediction with the selected image\n    classifier.predict(document.getElementById('image'), function(err, results) {\n      console.log(results);\n    });\n\nDone! In effectively three lines, a pretrained model is running in our browser. Now, let's open the browser's console to inspect the output presented in Figure 10-10.\n\n###### Figure 10-10. The top predicted classes with the probability of each class\n\n###### Tip\n\nIf you are unfamiliar with the browser console, you can simply access it by right-clicking anywhere within the browser window and selecting \"Inspect element.\" A separate window opens with the console inside it.\n\nWe can find the full source code for the previous example at _code\/chapter-10\/ml5js_.\n\nNote that ml5.js uses callbacks to manage asynchronous calls of the models. A callback is a function that is executed after the accompanying call is finished. For instance, in the last code snippet, after the model is loaded, the `modelLoaded` function is called, indicating that the model is loaded into memory.\n\n###### Note\n\np5.js is a library that works nicely in conjunction with ml5.js and makes it super easy to make model predictions in real time using a live video stream. You can find a code snippet demonstrating the power of p5.js at _code\/chapter-10\/p5js-webcam\/_.\n\nml5.js natively supports p5.js elements and objects. You can use p5.js elements for drawing objects, capturing webcam feeds, and more. Then, you can easily use such elements as input to the ml5.js callback functions.\n\n# PoseNet\n\nSo far in this book, we have primarily explored image classification problems. In later chapters, we take a look at object detection and segmentation problems. These types of problems form a majority of computer-vision literature. In this section, however, we choose to take a break from the usual and deal with a different kind of problem: keypoint detection. This has significant applications in a variety of areas including health care, fitness, security, gaming, augmented reality, and robotics. For instance, to encourage a healthy lifestyle through exercise, Mexico City installed kiosks that detect the squat pose and offer free subway tickets to passengers who can do at least 10 squats. In this section, we explore how to run something so powerful in our humble web browser.\n\nThe PoseNet model offers real-time pose estimation in the browser. A \"pose\" consists of the position of different keypoints (including joints) in the human body such as the top of the head, eyes, nose, neck, wrists, elbows, knees, ankles, shoulders, and hips. You can use PoseNet for single or multiple poses that might exist in the same frame.\n\nLet's build an example using PoseNet, which is readily available in ml5.js, to detect and draw keypoints (with the help of p5.js).\n\n###### Figure 10-11. Keypoints drawn using PoseNet on a picture of former President Obama in a snowball fight\n\nYou can find the code to detect keypoints for a still image at _code\/chapter-10\/posenet\/single.html_ :\n\n    <script src=\"http:\/\/p5js.org\/assets\/js\/p5.min.js\"><\/script>\n    <script src=\"http:\/\/p5js.org\/assets\/js\/p5.dom.min.js\"><\/script>\n    <script src=\"https:\/\/unpkg.com\/ml5@latest\/dist\/ml5.min.js\"><\/script>\n\n    <script>\n    function setup() {\n        \/\/ Set up camera here\n\n        \/\/ Call PoseNet model\n        const poseNet = ml5.poseNet(video, modelReady);\n\n        \/\/ PoseNet callback function\n        poseNet.on('pose', function (results) {\n            const poses = results;\n        });\n    }\n    <\/script>\n\nWe can also run a similar script (at _code\/chapter-10\/posenet\/webcam.html_ ) on our webcam.\n\nNow let us look at another example that is supported by ml5.js.\n\n##### From the Creator's Desk\n\nBy Cristobal Valenzuela, cofounder of Runway and contributor to ml5.js\n\nml5.js started as an experiment within NYU's Interactive Telecommunications Program (ITP) to understand creative applications of machine learning in the browser. Heavily inspired by Processing and p5.js, the original idea was to create a series of code experiments with deeplearn.js (now TensorFlow.js) and p5.js. In August of 2017, a few days after deeplearn.js was released, I created a repository called \"p5deeplearn: deeplearn.js meets p5\" as a tentative first approach. This repository was the genesis of ml5.js, with the \"5\" paying homage to p5.js and the Processing philosophy itself.\n\nWith the guidance of Dan Shiffman and the support of a Google Faculty Research Award to collaborate with the TensorFlow.js team, we organized a weekly working group to discuss developing the library itself and invited guest artists and researchers to participate in its development. The main idea behind ml5.js has always been to create an approachable and easy-to-use library space and community where we could encourage experimentation and creation with AI in the browser.\n\n##### From the Creator's Desk\n\nBy Daniel Shiffman, associate arts professor at the Interactive Telecommunications Program at NYU's Tisch School of the Arts and a director of The Processing Foundation\n\nDespite the ease of running a machine learning model in the browser with TensorFlow.js, beginners who want to code their own still face the challenge of mathematical notation and low-level technical code. Machine learning frameworks are commonly geared toward people with knowledge of calculus, linear algebra, statistics, data science, and several years of programming in a language like Python or C++. Although important for research and development of new machine learning models and architectures, starting from this point can turn away newcomers with other backgrounds. Rather than thinking creatively about how to use machine learning as an artistic platform, beginners can be overwhelmed by fine distinctions between scalars, vectors, matrices, operations, inputs layers, outputs, and more.\n\nThis is where ml5.js comes in. Its goal is to provide a neighborly approach to creating and exploring AI in the browser. _The \"5\" in ml5 is an homage to p5.js, a JavaScript library that serves as the primary inspiration and model for ml5._ p5.js is maintained by the Processing Foundation whose mission is \"to empower people of all interests and backgrounds to learn how to program and make creative work with code, especially those who might not otherwise have access to these tools and resources.\"\n\nWith Processing, an artist can invent their own tools rather than relying on those created and maintained by others. Today, being \"software literate\" is arguably more important than ever. Algorithms affect our lives in ways we couldn't have anticipated, and the explosion of machine learning research has further extended their reach. Our interactions are now negotiated not just with each other, but with autonomous vehicles, ever-listening assistants, and cameras that identify our faces and poses. Without access to and understanding of the machine learning models, underlying data, and outputs driving the software, how can we meaningfully engage, question, and propose alternatives? Machine learning faces a similar challenge of approachability as simply learning to code did more than 15 years ago.\n\nThe development of ml5.js is funded by a Google Faculty Research Award at ITP. ITP is a two-year graduate program in the NYU Tisch School of the Arts, whose mission is to explore the imaginative use of communications technologies\u2014how they might augment, improve, and bring delight and art into people's lives. Each week, a group of 10 to 15 ITP students, along with outside contributors and guest artists, meets at ITP to discuss API decisions, share creative projects, and teach each other about contributing to open source. To date, there have been more than 50 contributors to 17 releases of ml5.js.\n\n# pix2pix\n\n\"Hasta la vista, baby!\"\n\nThis is one of the most memorable lines in the history of film. Coincidentally, it was uttered by an AI cyborg in the 1991 classic _Terminator 2: Judgment Day_. By the way, its translation is \"Goodbye, baby!\" Language translation technology has come a long way since then. It used to be that language translation was built on phrase substitution rules. And now, it's replaced by much better performing deep learning systems that understand the context of the sentence to convert it into a similar meaning sentence in the target language.\n\nHere's a thought: if we can translate from sentence one to sentence two, could we translate a picture from one setting to another? Could we do the following:\n\n  * Convert an image from low resolution to higher resolution?\n\n  * Convert an image from black and white to color?\n\n  * Convert an image from daytime to nighttime view?\n\n  * Convert a satellite image of the earth into a map view?\n\n  * Convert an image from a hand-drawn sketch into a photograph?\n\nWell, image translation is not science fiction anymore. In 2017, Philip Isola et al. developed a way to convert a picture into another picture, conveniently naming it pix2pix. By learning from several pairs of before and after pictures, the pix2pix model is able to generate highly realistic renderings based on the input image. For example, as demonstrated in Figure 10-12, given a pencil sketch of a bag, it can recreate the photo of a bag. Additional applications include image segmentation, synthesizing artistic imagery, and more.\n\n###### Figure 10-12. Example of input and output pairs on pix2pix\n\n###### Note\n\nImagine a scenario with a bank teller and a currency counterfeiter. The job of the bank teller is to spot fake currency bills, whereas the counterfeiter's goal is to make it as difficult as possible for the bank teller to identify the fakes. They are clearly in an adversarial situation. Each time the cop spots the fake bills, the counterfeiter learns his mistake, takes it as an opportunity to improve (growth mindset after all), and tries to make it even more difficult for the bank teller to thwart him next time. This forces the bank teller to get better at recognizing fakes over time. This feedback cycle forces both of them to get better at what they do. This is the underlying principle driving GANs.\n\nAs Figure 10-13 illustrates, GANs consist of two networks, a Generator and a Discriminator, which have the same adversarial relationship as the counterfeiter and the bank teller. The Generator's job is to generate realistic-looking output, very similar to the training data. The Discriminator's responsibility is to identify whether the data passed to it by the Generator was real or fake. The output of the Discriminator is fed back into the Generator to begin the next cycle. Each time the Discriminator correctly identifies a generated output as a fake, it forces the Generator to get better in the next cycle.\n\nIt is worthwhile to note that GANs typically do not have control over the data to be generated. However, there are variants of GANs, such as _conditional GANs_ , that allow for labels to be part of the input, providing more control over the output generation; that is, conditioning the output. pix2pix is an example of a conditional GAN.\n\n###### Figure 10-13. A flowchart for a GAN\n\nWe used pix2pix to create a simple sketching app that works in the browser. The output images are really interesting to look at. Consider the examples shown in Figure 10-14 and Figure 10-15.\n\n###### Figure 10-14. Sketch-to-image example\n\n###### Figure 10-15. We can create colored blueprints (left) and pix2pix will convert them to realistic-looking human faces (right)\n\n###### Note\n\nFun fact: Ian Goodfellow came up with the idea for GANs while at a bar. This adds yet another item to the list of inventions, organizations, and companies whose ideas originated over drinks, including the creation of the RSA Algorithm, Southwest Airlines, and the game of Quidditch.\n\npix2pix works by training on pairs of images. In Figure 10-16, the image on the left is the input image or the conditional input. The image on the right is the target image, the realistic output that we want to generate (if you are reading the print version, you will not see the color image on the right).\n\n###### Figure 10-16. Training pairs for pix2pix: a B&W image and its original color image\n\nOne of the easier ports for training pix2pix is the TensorFlow-based implementation by Christopher Hesse. We can use a very simple script to train our own model:\n\n    python pix2pix.py \\\n      --mode train \\\n      --output_dir facades_train \\\n      --max_epochs 200 \\\n      --input_dir facades\/train \\\n      --which_direction BtoA\n\nAfter training has finished, we can save the model using the following command:\n\n    python tools\/export-checkpoint.py --checkpoint ..\/export --output_file\n    models\/MY_MODEL_BtoA.pict\n\nAfter that, we can use this simple code to load the saved weights to ml5.js. Note the transfer function that is used to retrieve the output in a canvas:\n\n    \/\/ Create a pix2pix model using a pre-trained network\n    const pix2pix = ml5.pix2pix('models\/customModel.pict', modelLoaded);\n\n    \/\/ Transfer using a canvas\n    pix2pix.transfer(canvas, function(err, result) {\n      console.log(result);\n    });\n\nWe can also draw strokes and allow real-time sketching. For instance, Figure 10-17 shows an example that draws Pikachu.\n\n###### Figure 10-17. Pix2Pix: Edges to Pikachu by Yining Shi, built on ml5.js\n\n##### From the Creator's Desk\n\nBy Cristobal Valenzuela, cofounder of Runway and contributor to ml5.js\n\nThere have been a lot of interesting projects, experiments, and creations built with ml5.js: artists have been using it to create art as well as live stage performances. Nikita Huggins and Ayodamola Okunseinde, two New York-based artists, have been exploring ways of increasing the diversity of the models the library currently supports, so students and developers have built applications for Hackathons and for all kinds of competitions. But what's super exciting is to see how it has been used in curricula around the world to teach about machine learning in an approachable manner, opening the possibilities for a broad audience of artists, creative coders, and students.\n\n# Benchmarking and Practical Considerations\n\nAs people who care deeply about how our end users perceive our product, it's important for us to treat them right. Two factors play a large role in how users experience our product: the model size, and the inference time based on the hardware. Let's take a closer look at each factor.\n\n## Model Size\n\nA typical MobileNet model is 16 MB. Loading this on a standard home or office network might just take a few seconds. Loading the same model on a mobile network would take even longer. The clock is ticking, and the user is becoming impatient. And this is before the model even gets a chance to start inference. Waiting for big models to load is more detrimental to the UX than their runtime, especially where internet speeds are not as fast as a broadband paradise like Singapore. There are a few strategies that can help:\n\nPick the smallest model for the job\n\nAmong pretrained networks, EfficientNet, MobileNet, or SqueezeNet tend to be the smallest (in order of decreasing accuracy).\n\nQuantize the model\n\nReduce the model size using the TensorFlow Model Optimization Toolkit before exporting to TensorFlow.js.\n\nBuild our own tiny model architecture\n\nIf the final product does not need heavy ImageNet-level classification, we could build our own smaller model. When Google made the J.S. Bach doodle on the Google home page, its model was only 400 KB, loading almost instantly.\n\n## Inference Time\n\nConsidering our model is accessible in a browser running on a PC or a mobile phone, we would want to pay careful attention to the UX, especially on the slowest hardware. During our benchmarking process, we ran _chapter10\/code\/benchmark.html_ within various browsers on different devices. Figure 10-18 presents the results of these experiments.\n\n###### Figure 10-18. Inference time for MobileNetV1 in Chrome on different devices\n\nFigure 10-18 implies the faster the hardware, the faster the model inference. Apple appears to be outdoing Android in terms of GPU performance. Though, clearly, it is not an \"apples-to-apples comparison.\"\n\nOut of curiosity, do different browsers run inference at the same speed? Let's find that out on an iPhone X; Figure 10-19 shows the results.\n\n###### Figure 10-19. Inference time in different browsers on iPhone X\n\nFigure 10-19 shows us the same speed in all browsers on an iPhone. This shouldn't be surprising, because all of these browsers use iPhone's WebKit-based built-in browser control called `WKWebView`. How about on a MacBook Pro? Take a look at Figure 10-20 to see.\n\n###### Figure 10-20. Inference time in different browsers on an i7 @ 2.6 GHz macOS 10.14 machine\n\nThe results might be surprising. Chrome is almost double the speed of Firefox in this example. Why is that? Opening a GPU monitor showed that Chrome had much higher GPU utilization compared to Firefox and slightly higher than Safari. The higher the utilization, the faster the inference. What this means is that depending on the operating system, browsers might have different optimizations to speed up the inference on the GPU, leading to different running times.\n\nOne key point to note is that these tests were performed on top-of-the-line devices. They do not necessarily reflect the kind of device an average user might have. This also has implications for battery usage, if run for prolonged periods of time. Accordingly, we need to set appropriate expectations regarding performance, particularly for real-time user experiences.\n\n# Case Studies\n\nNow that we know all the ingredients for deep learning on the browser, let's see what the industry is cooking.\n\n## Semi-Conductor\n\nHave you ever dreamed of conducting the New York Philharmonic Orchestra? With Semi-Conductor, your dream is half-way fulfilled. Open the website, stand in front of the webcam, wave your arms, and watch the entire orchestra perform Mozart's Eine Kleine Nachtmusik at your whim! As you might have guessed, it's using PoseNet to track the arm movements and using those movements to set the tempo, volume, and the section of instruments (Figure 10-21) playing the music (including violins, violas, cellos, and double bass). Built by the Google Creative Lab in Sydney, Australia, it uses a prerecorded musical piece broken up into tiny fragments, and each fragment is played at the scored speed and volume depending on the arm movements. Moving hands up increases volume, moving faster increases tempo. This interactive experience is possible only because PoseNet is able to run inferences at several frames a second (on a regular laptop).\n\n###### Figure 10-21. Control an orchestra by waving your arms on the Semi-Conductor demonstrator\n\n## TensorSpace\n\nCNNs can often feel...well, convoluted. Often treated as a black box, they can be difficult to understand. What do the filters look like? What activates them? Why did they make a certain prediction? They are shrouded in mystery. As with anything complex, visualizations can help open this black box and make it easier to understand. And that's where TensorSpace, the library that \"presents tensors in space,\" comes in.\n\nIt allows us to load models in 3D space, explore their structures in the browser, zoom and rotate through them, feed inputs, and understand how the image is processed and passed layer by layer all the way to the final prediction layer. The filters can finally be opened up to manual inspection without the need for any installation. And, as Figure 10-22 teases, if you're feeling savvy, you could even load this in virtual reality against any background!\n\n###### Figure 10-22. LeNet model visualized inside TensorSpace\n\n## Metacar\n\nSelf-driving cars are a complex beast. And using reinforcement learning to train them can take a lot of time, money, and monkeywrenching (not counting the crashes, initially). What if we could train them in the browser itself? Metacar solves this by providing a simulated 2D environment to train toy cars with reinforcement learning, all in the browser, as depicted in Figure 10-23. Just as with video games you progress to ever-more difficult levels, Metacar allows building multiple levels to improve the performance of your car. Utilizing TensorFlow.js, this is aimed at making reinforcement learning more accessible (in which we dive into greater detail in Chapter 17 while building a small-scale autonomous car).\n\n###### Figure 10-23. Metacar environment for training with reinforcement learning\n\n## Airbnb's Photo Classification\n\nAirbnb, the online property rental company, requires homeowners and renters to upload pictures of themselves for their profiles. Unfortunately, a few people try to find the most readily available picture they have\u2014their driver's license or passport. Given the confidential nature of the information, Airbnb uses a neural network running on TensorFlow.js to detect sensitive images and prevent their upload to the server.\n\n## GAN Lab\n\nSimilar to TensorFlow Playground (an in-browser neural network visualization tool), GAN Lab (Figure 10-24) is an elegant visualization tool for understanding GANs using TensorFlow.js. Visualizing GANs is a difficult process, so to simplify it, GAN Lab attempts to learn simple distributions and visualizes the generator and discriminator network output. For instance, the real distribution could be points representing a circle in 2D space. The generator starts from a random Gaussian distribution and gradually tries to generate the original distribution. This project is a collaboration between Georgia Tech and Google Brain\/PAIR (People + AI Research).\n\n###### Figure 10-24. Screenshot of GAN Lab\n\n# Summary\n\nWe first examined the evolution of JavaScript-based deep learning libraries and chose TensorFlow.js as the candidate to focus on. We ran pretrained models in real time on the webcam feed, and then even trained models in the browser. Tools like Chrome's profiler gave us insight into GPU usage. Then, to simplify development further, we used ml5.js, which enabled us to build demos like PoseNet and pix2pix in just a few lines of code. Finally, we benchmarked the performance of these models and libraries in the real world, ending with some interesting case studies.\n\nOne huge benefit of running neural networks in the browser is the vast reach that browsers have compared to any smartphone platform. Add to that the advantage of not having to overcome the user's reluctance to install yet another app. This also makes for a quick prototyping platform that enables inexpensive validation of hypotheses before investing significant amounts of time and money to build native experiences. TensorFlow.js, in conjunction with ml5.js, has accelerated the process of bringing the power of AI to the browser and broaden its reach to the masses.\n\n# Chapter 11. Real-Time Object Classification on iOS with Core ML\n\nSo far, we have seen our deep learning models running on the desktop, the cloud, and the browser. Although there are definite upsides to such a setup, it might not be ideal for all scenarios. In this chapter, we explore making predictions using deep learning models on mobile devices.\n\nBringing the computation closer to the user's device, rather than a distant remote server, can be advantageous for many reasons:\n\nLatency and interactivity\n\nSending an image, processing it in the cloud, and returning the result can take several seconds depending on the network quality and quantity of data being transferred. This can make for a poor UX. Decades of UX research, including Jakob Nielsen's findings in 1993, published in his book _Usability Engineering_ (Elsevier), showed the following:\n\n  * 0.1 second is about the limit for having the user feel that the system is reacting instantaneously.\n\n  * 1 second is about the limit for the user's flow of thought to stay uninterrupted.\n\n  * 10 seconds is about the limit for keeping the user's attention focused.\n\nAbout two decades later, Google published findings that half of all mobile browser users abandon a web page if it takes longer than three seconds to load. Forget three seconds, even a 100 ms increase in latency would result in a 1% decrease in sales for Amazon. That is a lot of lost revenue. Mitigating this by processing on the device instantaneously can make for rich and interactive UXs. Running deep learning models in real time, as is done with Snapchat Lenses, can increase engagement with users.\n\n24\/7 availability and reduced cloud costs\n\nObviously, sending less data to the cloud equates to less computing costs for the developer, leading to monetary savings. This reduces scaling costs, as well, when an app gains traction and grows to a large user base. For the users, computation on the edge is helpful, too, because they don't need to worry about data plan costs. Additionally, processing locally means 24\/7 availability, without the fear of losing connectivity.\n\nPrivacy\n\nFor the user, local computation preserves privacy by not sharing the data externally, which could potentially be mined for user information. For the developer, this ensures less headache dealing with Personally Identifiable Information (PII). With the European Union's General Data Protection Regulation (GDPR) and other user data protection laws coming up around the world, this becomes even more important.\n\nHopefully, these arguments were convincing for why AI on mobile is important. For someone building any kind of serious application, the following are a few common questions to consider during the course of development:\n\n  * How do I convert my model to run on a smartphone?\n\n  * Will my model run on other platforms?\n\n  * How do I run my model fast?\n\n  * How do I minimize the size of the app?\n\n  * How do I ensure that my app doesn't drain the battery?\n\n  * How do I update my model without going through the roughly two-day app review process?\n\n  * How do I A\/B test my models?\n\n  * Can I train a model on a device?\n\n  * How do I protect my intellectual property (i.e., model) from being stolen?\n\nIn the next three chapters, we look at how to run deep learning algorithms on smartphones using different frameworks. In the process, we answer these questions as they come up.\n\nIn this chapter, we delve into the world of mobile AI for iOS devices. We first look at the general end-to-end software life cycle (shown in Figure 11-1) and see how the different parts fit in with one another. We explore the Core ML ecosystem, its history, and the features offered. Next, we deploy a real-time object classification app on an iOS device, and learn performance optimizations and benchmarks. And finally, we analyze some real-world apps built on Core ML.\n\nTime to look at the big picture.\n\n# The Development Life Cycle for Artificial Intelligence on Mobile\n\nFigure 11-1 depicts the typical life cycle for AI on mobile devices.\n\n###### Figure 11-1. The mobile AI development life cycle\n\nLet's look at the phases in Figure 11-1 a bit more closely:\n\n  1. _Collect data_ : The data we collect should reflect the context in which the app would be used. Pictures taken by real users using smartphone cameras tend to be better training examples than pictures taken by professional photographers. We might not have this data on day one, but we can progressively collect more and more data as usage grows. A good starting point in many cases is to download images from search engines.\n\n  2. _Label data_ : We need associated labels for the data samples that we want our model to predict. High-quality (i.e., correct) labels are vital to a good model.\n\n  3. _Train model_ : We build the highest-accuracy neural network possible with the data and associated labels we have so far.\n\n  4. _Convert model_ : Export the model from the training framework into the mobile-compatible framework.\n\n  5. _Optimize performance_ : Due to the resource-constrained nature of a mobile device, it is crucial to make the model efficient for memory, energy, and processor usage.\n\n  6. _Deploy_ : Add the model to the app and ship it to users.\n\n  7. _Monitor_ : Track the app usage in the real world to find opportunities to improve further. Additionally, gather samples of real-word data from consenting users to feed to this life cycle, and then back to Step 1.\n\nIn the first portion of the book, we primarily explored phases 1, 2, and 3, and additionally general performance improvements. In this chapter, we focus on phases 4, 5, and 6. And in the next few chapters, we explore all of these phases in the context of mobile development.\n\n###### Tip\n\nBefore getting the app in the hands of real users (who might hate it if the app doesn't perform as well as they expected), it's common practice to gather feedback through a process called dogfooding. As the saying goes: Eat Your Own Dog Food. The process involves having an inner circle of loyal users who get to test early releases, thereby identifying bugs before they go out to the general public. For AI development, this inner circle might also contribute data and assess the success of the AI models in the real world. And as the models improve, they can be deployed to a gradually increasing number of test users before finally deploying to the public.\n\nLet's jump right in!\n\n# A Brief History of Core ML\n\nCore ML provides one of the simplest ways to run a deep neural network for inference on an Apple device such as the iPhone and iPad as well as MacBook, Apple TV, and Apple Watch. In addition to being easy to use, it is also optimized for the underlying hardware architecture. Alternative frameworks have become a lot better in the past few years, but it's difficult to beat the simplicity and performance offered by Core ML.\n\nTraditionally, to quickly run a CNN on an Apple device, developers _needed_ to write on Metal, a library that was offered to game developers to better utilize the GPU. Unfortunately, developing on Metal was akin to writing in assembly language or CUDA code for an NVIDIA GPU. It was tedious, error prone, and difficult to debug. Few developers dared to tread that path. DeepLearningKit (December 2015) by Amund Tveit was one effort to build an abstraction over Metal for deploying CNNs.\n\nAt the Apple Worldwide Developers Conference (WWDC) 2016, the company announced Metal Performance Shaders (MPS), a framework built on top of Metal, as a high-performance library for optimizing graphics and certain compute operations. It abstracted away a lot of low-level details, giving basic building blocks such as Convolution, Pooling, and ReLU. It allowed developers to write deep neural networks by combining these operations in code. For anyone who comes from the Keras world, this is a familiar and not-so-daunting task. Unfortunately, there is a lot of bookkeeping involved when writing MPS code, because you need to manually keep track of input and output dimensions at each step of the process. As an example, the code sample released by Apple for running the InceptionV3 model for recognizing 1,000 object categories was well over 2,000 lines long, with most of them defining the network. Now imagine changing the model slightly during training, and then having to dig through all 2,000 lines of code to reflect the same update in iOS code. Forge (April 2017), a library by Matthijs Hollemans, was an effort to simplify development over MPS by reducing the boilerplate code necessary to get a model running.\n\nAll of these hardships went away when Apple announced Core ML at WWDC 2017. This included an inference engine on iOS and an open source Python package called Core ML Tools to serialize CNN models from other frameworks like Keras and Caffe. The general workflow for building an app was: train a model in other packages, convert it to a _.mlmodel_ file, and deploy it in an iOS app running on the Core ML platform.\n\nCore ML supports importing a broad range of machine learning models built with first- and third-party frameworks and file formats. Figure 11-2 showcases a few of them (clockwise, starting in the upper left) such as TensorFlow, Keras, ONNX, scikit-learn, Caffe2, Apple's Create ML, LIBSVM, and TuriCreate (also from Apple). ONNX itself supports a large variety of frameworks, including PyTorch (Facebook), MXNet (Amazon), Cognitive Toolkit (Microsoft), PaddlePaddle (Baidu), and more, thereby ensuring compatibility with any major framework under the sky.\n\n###### Figure 11-2. Frameworks compatible with Core ML for model interchange as of 2019\n\n# Alternatives to Core ML\n\nDepending on the platform, there are a handful of options to achieve real-time predictions. These include general-purpose inference frameworks such as Core ML (from Apple), TensorFlow Lite (from Google), ML Kit (also from Google), and Fritz, as well as chip-specific accelerator frameworks including Snapdragon Neural Processing Engine (from Qualcomm) and Huawei AI Mobile Computing Platform (for Huawei's Neural Processing Unit). Table 11-1 presents a high-level comparison of all these frameworks.\n\nTable 11-1. A comparison of mobile device AI frameworks **Framework** | **Available for iOS** | **Available for Android** | **Dynamic updates** | **A\/B testing** | **On-device training** | **Model encryption**  \n---|---|---|---|---|---|---  \nCore ML | \u2713 | \u2014 | \u2713 | \u2014 | \u2713 | \u2014  \nTensorFlow Lite | \u2713 | \u2713 | \u2014 | \u2014 | Releases late 2019 | \u2014  \nML Kit | \u2713 | \u2713 | \u2713 | \u2713 | \u2014 | \u2014  \nFritz | \u2713 | \u2713 | \u2713 | \u2713 | \u2014 | \u2713\n\n## TensorFlow Lite\n\nIn November 2017, Google announced an on-device inference engine called TensorFlow Lite with the intention of extending the TensorFlow ecosystem beyond just servers and PCs. Prior to this, the options within the TensorFlow ecosystem were porting the entire TensorFlow library, itself ported to iOS (which was heavy and slow), and later on, its slightly stripped-down version called TensorFlow Mobile (which was still pretty bulky).\n\nTensorFlow Lite was rebuilt from the ground up targeting mobile and edge devices, optimizing for speed, model and interpreter size, and power consumption. It added support for GPU backend delegates, meaning that as long as there was GPU support implemented for a hardware platform, TensorFlow Lite could take advantage of the power of the GPU. On iOS, the GPU delegate uses Metal for acceleration. We discuss TensorFlow Lite at greater length in Chapter 13.\n\n## ML Kit\n\nML Kit is a high-level library from Google that provides many computer vision, NLP, and AI functionalities out of the box, including the ability to run TensorFlow Lite models. Some of the features include face detection, barcode scanning, smart reply, on-device translation, and language identification. However, the main selling point of ML Kit is its integration with Google Firebase. Features offered by Firebase include dynamic model updates, A\/B testing, and remote configuration-driven dynamic model selection (fancy words for choosing which model to use based on the customer). We explore ML Kit in greater detail in Chapter 13.\n\n## Fritz\n\nFritz is a startup founded with the goal of making the end-to-end process of mobile inference easier. It bridges the gap between machine learning practitioners and mobile engineers by providing easy-to-use command-line tools. On one side, it integrates training in Keras directly into the deployment pipeline, so a machine learning engineer could add a single line of Keras callback to deploy the model to users immediately after it finishes training. On the other side, a mobile engineer can benchmark the model without even needing to deploy to a physical device, simulating a model's performance virtually, assess a Keras model's compatibility with Core ML, and get analytics for each model. One unique selling point of Fritz is the model protection feature that prevents a model from deep inspection through obfuscation in the event that a phone is jailbroken.\n\n# Apple's Machine Learning Architecture\n\nTo gain a better understanding of the Core ML ecosystem, it's useful to see the high-level picture of all the different APIs that Apple offers as well as how they fit with one another. Figure 11-3 gives us a look at the different components that make up Apple's machine learning architecture.\n\n###### Figure 11-3. Different levels of APIs provided by Apple for app developers\n\n## Domain-Based Frameworks\n\nTo simplify common tasks in machine learning without requiring domain expertise, Apple provides many APIs out of the box, in domains including Vision, Natural Language, Speech, and Sound Analysis. Table 11-2 gives a detailed outline of the functionalities available on Apple operating systems.\n\nTable 11-2. Out-of-the-box machine learning functionality in Apple operating systems **Vision** | **Natural language** | **Other**  \n---|---|---\n\n  * Facial landmark detection\n\n  * Image similarity\n\n  * Saliency detection\n\n  * Optical character recognition\n\n  * Rectangle detection\n\n  * Face detection\n\n  * Object classification\n\n  * Barcode detection\n\n  * Horizon detection\n\n  * Human and animal detection\n\n  * Object tracking (for video)\n |   * Tokenization\n\n  * Language identification\n\n  * Parts of speech identification\n\n  * Text embedding\n |   * Speech recognition (on-device and on-cloud)\n\n  * Sound classification\n\n## ML Framework\n\nCore ML gives the ability to run inference on deep learning and machine learning models.\n\n## ML Performance Primitives\n\nThe following are some of the machine learning primitives in the Apple stack:\n\nMPS\n\nProvides low-level and high-performance primitives that utilize the GPU to aid running most CNN-based networks fast. And if Core ML does not support a model, MPS provides all of the building blocks to enable us to build them. Additionally, we might consider using MPS to hand roll a model for performance reasons (like guaranteeing that the model runs on the GPU).\n\nAccelerate and Basic Neural Network Subroutine\n\nAccelerate is Apple's implementation of the Basic Linear Algebra Subprogram (BLAS) library. It provides functionality for high-performance large-scale mathematical computations and image calculations, like Basic Neural Network Subroutine (BNNS), which helps to implement and run neural networks.\n\nNow that we have seen how Core ML and the domain-specific APIs fit into the overall architecture, let's see how little work is needed to run a machine learning model using Core ML and the Vision framework on an iOS app.\n\n###### Tip\n\nApple provides several downloadable models (Figure 11-4) for various computer-vision tasks, from classification to detecting objects (with bounding boxes), segmenting (identifying the pixels), depth estimation, and more. You can find them at _https:\/\/developer.apple.com\/machine-learning\/models\/_.\n\nFor classification, you can find many pretrained Core ML models on Apple's Machine Learning website, including MobileNet, SqueezeNet, ResNet-50, and VGG16.\n\n###### Figure 11-4. Ready-to-use models from the Apple Machine Learning website\n\n# Building a Real-Time Object Recognition App\n\nAlthough we don't intend to teach iOS development, we want to demonstrate running an object recognition model that classifies among 1,000 ImageNet categories in real time on an iOS device.\n\nWe will examine the minimal amount of code necessary to get this app running. The general outline is as follows:\n\n  1. Drag and drop an _.mlmodel_ file to the Xcode project.\n\n  2. Load the model into a Vision container (`VNCoreMLModel`).\n\n  3. Create a request (`VNCoreMLRequest`) based on that container and provide a function that would be called when the request is completed.\n\n  4. Create a request handler that can process the request based on an image provided.\n\n  5. Execute the request and print out the results.\n\nLet's take a look at the code to do this:\n\n    import CoreML\n    import Vision\n\n    \/\/ load the model\n    let model = try? VNCoreMLModel(for: Resnet50().model)!\n    \/\/ create a request with a callback\n    let classificationRequest = VNCoreMLRequest(model: model) { \n        (request, error) in\n        \/\/ print the results once the request is complete\n        if let observations = request.results as? [VNClassificationObservation] {\n            let results = observations\n                          .map{\"\\($0.identifier) - \\($0.confidence)\"}\n                          .joined(separator: \"\\n\")\n            print(results)\n        }\n    }\n    \/\/ create a request handler taking an image as an argument\n    let requestHandler = VNImageRequestHandler(cgImage: cgImage)\n    \/\/ execute the request\n    try? requestHandler.perform([classificationRequest])\n\nIn this code, `cgImage` can be an image from a variety of sources. It can be a photo from the photo library, or from the web.\n\nWe can also power real-time scenarios using the camera and pass the individual camera frames into this function. A regular iPhone camera can shoot up to 60 frames per second (FPS).\n\nBy default, Core ML crops the image along the longer edge. In other words, if a model requires square dimensions, Core ML will extract the biggest square in the center of the image. This can be a source of confusion for developers who find the top and bottom strips of an image are being ignored for making predictions. Depending on the scenario, we might want to use `.centerCrop`, `.scaleFit`, or the `.scaleFill` option, as illustrated in Figure 11-5, such as the following:\n\n    classificationRequest.imageCropAndScaleOption = .scaleFill\n\n###### Figure 11-5. How different scaling options modify the input image to Core ML models\n\nNow that we have gone through the meat of it, what could be more fun? How about actually running it on a phone! We've made a complete app available on the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_) at _code\/chapter-11_. With an iPhone or an iPad, we can deploy this pretty quickly and experiment with it, even without knowledge of iOS development. Here are the steps (note: this requires a Mac):\n\n  1. Download Xcode from the Apple developer website or the Mac App Store.\n\n  2. Plug in an iOS device. The phone needs to remain unlocked during deployment.\n\n  3. Change the current working directory to `CameraApp`:\n\n    $ cd code\/chapter-11\/CameraApp\n\n  4. Download the Core ML models from the Apple website using the available Bash script:\n\n    $ .\/download-coreml-models.sh\n\n  5. Open the Xcode project:\n\n    $ open CameraApp.xcodeproj\n\n  6. In the Project Hierarchy Navigator, in the upper-left corner, click the CameraApp project, as shown in Figure 11-6, to open the Project Information view.\n\n###### Figure 11-6. Project information view within Xcode\n\n  7. Because Xcode reserves a unique bundle identifier, use a unique name to identify the project.\n\n  8. Log in to an Apple account to let Xcode sign the app and deploy it to the device. Select a team to do the signing, as shown in Figure 11-7.\n\n###### Figure 11-7. Select a team and let Xcode automatically manage code signing\n\n  9. Click the \"Build and Run\" button (the right-facing triangle) to deploy the app on the device, as shown in Figure 11-8. This should typically take around 30 to 60 seconds.\n\n###### Figure 11-8. Select the device and click the \"Build and Run\" button to deploy the app\n\n  10. The device will not run the app right away, because it is not trusted. Go to Settings > General > Profiles and Device Management and select the row with your information on it, and then tap \"Trust {your_email_id},\" as shown in Figure 11-9.\n\n###### Figure 11-9. Profiles and Device Management screen\n\n  11. On the home screen, find CameraApp and run the app.\n\n###### Figure 11-10. Screenshot of the app\n\nThe output shows predictions of \"notebook, notebook computer\" with 84% confidence and \"laptop, laptop computer\" with 11% confidence.\n\nThat was all fun and good. Let's now get down to more serious business: converting models from different frameworks to Core ML models.\n\n###### Tip\n\nA great thing about Xcode is that the model input and output parameters are shown when we load the _.mlmodel_ file into Xcode, as demonstrated in Figure 11-11. This is especially helpful when we have not trained the model ourselves and don't necessarily want to write code (like `model.summary()` in Keras) to explore the model architecture.\n\n###### Figure 11-11. Xcode model inspector showing the inputs and outputs of the MobileNetV2 model\n\n# Conversion to Core ML\n\nIn the code sample we built, you saw the _Inceptionv3.mlmodel_ file. Ever wonder how that file came to be? Inception was, after all, trained by Google using TensorFlow. That file was converted from a _.pb_ file into a Core ML model. We might similarly have models that need to be converted from Keras, Caffe or any other framework to Core ML. The following are a few tools that enable model conversion to Core ML.\n\n  * Core ML Tools (Apple): from Keras ( _.h5_ ), Caffe ( _.caffemodel_ ), as well as machine learning libraries such as LIBSVM, scikit-learn, and XGBoost.\n\n  * tf-coreml (Google): from TensorFlow ( _.pb_ ).\n\n  * onnx-coreml (ONNX): from ONNX ( _.onnx_ ).\n\nWe take a look at the first two converters in detail.\n\n## Conversion from Keras\n\nCore ML Tools aids the translation of Keras, ONNX, and other model formats to the Core ML format ( _.mlmodel_ ). Install the `coremltools` framework using `pip`:\n\n    $ pip install --upgrade coremltools\n\nNow, let's see how we can use an existing Keras model and convert it to Core ML. The conversion is just a one-line command and then we can save the converted model, as shown here:\n\n    from tensorflow.keras.applications.resnet50 import ResNet50\n    model = ResNet50()\n\n    import coremltools\n    coreml_model = coremltools.converters.keras.convert(model)\n    coreml_model.save(\"resnet50.mlmodel\")\n\nThat's it! It can't get any simpler. We discuss how to convert models with unsupported layers (such as MobileNet) in Chapter 12.\n\n## Conversion from TensorFlow\n\nApple recommends using `tf-coreml` (from Google) for converting from TensorFlow-based models. In the following steps, we convert a pretrained TensorFlow model to Core ML. This process is a bit more involved than the single line of code we saw earlier.\n\nFirst, we perform the installation of `tfcoreml` using `pip`:\n\n    $ pip install tfcoreml --user --upgrade\n\nTo do the conversion, we need to know the first and last layers of the model. We can ascertain this by inspecting the model architecture using a model visualization tool like Netron. After loading the MobileNet model ( _.pb_ file) in Netron, we can visualize the entire model in the graphical interface. Figure 11-12 shows a small portion of the MobileNet model; in particular, the output layer.\n\n###### Figure 11-12. Output layer of MobileNet as seen in Netron\n\nWe simply need to plug that into the following Python code as an argument and run it:\n\n    import tfcoreml as tf_converter\n    tf_converter.convert(tf_model_path = \"input_model.pb\",\n         mlmodel_path = \"output_model.mlmodel\",\n         output_feature_names = [\"MobilenetV1\/Predictions\/Reshape_1:0\"])\n\nAs the script runs, we see each of the operations from each layer being converted to their corresponding Core ML equivalents. When finished, we should find the _.mlmodel_ exported in the directory.\n\nOur Core ML model is now ready. Drag it into Xcode to test it.\n\n# Dynamic Model Deployment\n\nAs developers, we will likely want to keep improving our models over time. After all, we'd want our users to have access to the latest and greatest model as quickly as we can hand it to them. One way to handle that is to send an update to the App Store every time we want to deploy a new model. That approach does not work very well, because we'd need to wait about two days for Apple's approval each time, potentially causing significant delays.\n\nAn alternate recommended approach would be to have the app dynamically download the _.mlmodel_ file and compile it within the user's device. There are several reasons why we might want to try this approach:\n\n  * We want to regularly update the model, independent of the cadence of our App Store releases.\n\n  * We want to keep the app download size small and have the user eventually download only the models relevant for their usage scenarios. This will reduce storage needs and bandwidth costs. Apple places a limit of 200 MB on App Store downloads via cellular networks, so it's vital to remain under that limit to not lose out on potential downloads.\n\n  * We want to A\/B test different models on different sets of users to further improve the quality of the model.\n\n  * We want to use different models for different user segments, regions, and locales.\n\nThe process to achieve this is rather simple: host the _.mlmodel_ on a server and design our app to dynamically download the file. After the model is on the user's device, we can run `MLModel.compileModel` on that file to generate a compiled version of the model:\n\n    let compiledModelUrl = try MLModel.compileModel(at: downloadedModelUrl)\n    let model = try MLModel(contentsOf: compiledModelUrl)\n\nKeep in mind that `compiledModelUrl` is an address to a temporary location. If you want to keep the model on the device longer beyond a session, you must move it to permanent storage.\n\n###### Note\n\nAlthough we can manually do model management directly with Core ML, it would still involve writing a lot of boilerplate code, on the backend and on the device. We would need to manually manage the versions of each model, the file storage, and configuration infrastructure, as well as any errors in the process. This is where ML Kit and Fritz really shine by providing these features out of the box. We discuss this in more detail in Chapter 13.\n\n# On-Device Training\n\nSo far, we have looked at scenarios that would be served well by a \"one size fits all\" neural network. However, some use cases would simply not work without personalizing the models. An example would be an app that organizes a user's photo gallery based on recognizing the faces of people in each picture. Given that an average user's phone is mostly filled with pictures of friends and family, a generic model trained on Danny DeVito's and Tom Hanks's faces would not be very useful for users (unless, of course, they belong to the DeVito or Hanks families).\n\nA real-life example would be on the iOS system keyboard, which learns a user's language patterns over time and begins to give more and more relevant suggestions for that user. And this becomes even more evident when that person uses slang words, nicknames, domain-specific terms, and so on, which might not be part of the common language dictionary. Personalized suggestions based on such data would be useless for any other user.\n\nIn these cases, we want to collect the data for that user and train a personalized model for that individual only. One way to accomplish this would be to collect and send the data to the cloud, train a new model there, and send back the updated model to the user. This approach reeks of scalability, cost, and privacy issues.\n\nInstead, Core ML provides functionality for on-device training, so that user's data never needs to leave the device. A Core ML model is updatable when its `isUpdatable` property is set to `true`. Additionally, the specific set of layers that need to be retrained (usually toward the tail end of the network) also need to have the same property set to `true`. Additional training parameters such as learning rate and optimizers can also be set on the model.\n\nEven though training can consume GPU and Neural Processor Unit (NPU; more on these in Chapter 13) cycles, the training can be scheduled in the background (using the `BackgroundTasks` framework), even when the device is idle and charging, typically at night. This will have the least impact on UX.\n\nTo do on-device training, we can call the `MLUpdateTask` function, along with the new data with which we will be retraining the model. We would also need to pass in the path for the newly updated model in the function call. After training is complete, the model at that path is ready to go:\n\n    let modelUrl = bundle.url(forResource: \"MyClassifier\",\n                              withExtension: \"mlmodelc\")!\n    let updatedModelUrl = bundle.url(forResource: \"MyClassifierUpdated\",\n                                     withExtension: \"mlmodelc\")!\n\n    let task = try MLUpdateTask(\n        forModelAt: modelUrl,\n        trainingData: trainData,\n        configuration: nil,\n        completionHandler: { [weak self] (updateContext) in\n              self.model = updateContext.model\n              updateContext.model.write(to: updatedModelUrl)\n        })\n\n    task.resume()\n\n## Federated Learning\n\nOn-device training is great\u2014except for one downside: the generic global model does not get a chance to improve. Wouldn't it be great for a developer to somehow use that data being generated on each user's device to improve the global model without needing to transfer their data from their device? This is where _federated learning_ comes in.\n\nFederated learning is a collaborative distributed training process. It essentially takes on-device training one step further by sending the incremental updates (on the personalized models from users' devices) to the cloud and aggregating many of these over the entire user base, subsequently enriching the global model for everyone. Keep in mind that no user data is being transmitted here and it is not possible to reverse engineer the user data from the aggregated feature set. With this approach, we are able to respect our users' privacy while also benefiting everyone through collective participation.\n\nOn-device training is a vital stepping stone for federated learning. Even though we are not there yet, this is the direction in which the industry is moving. We can expect to see more support for federated learning over time.\n\nTensorFlow Federated is one such implementation of federated learning, powering training for Google's GBoard, a keyboard app. Training on user devices occurs in the background when they are being charged.\n\n# Performance Analysis\n\nWriting prototypes is one thing. Making production-ready apps is another ball game entirely. There are several factors that influence the user's experience, and it is important to understand the trade-offs. Some of these include the supported device models, minimum operating system (OS) version, processing frame rate, and choice of the deep learning model. In this section, we explore the impact that some of these factors can have on the quality and performance of the product.\n\n## Benchmarking Models on iPhones\n\nAs is the case with benchmarking, it is best to perform experiments on publicly available models that we can download easily and get our hands dirty with!\n\nFirst, we ran our real-time object classification app on multiple iPhones produced from 2013 to 2018. Table 11-3 catalogs the results of those experiments.\n\nTable 11-3. Benchmarking inference time on different models on different iPhone versions Device model | iPhone 5s | iPhone 6 | iPhone 6s | iPhone 7+ | iPhone X | iPhone XS | iPhone 11 Pro  \n---|---|---|---|---|---|---|---  \n**Year released** | 2013 | 2014 | 2015 | 2016 | 2017 | 2018 | 2019  \n**RAM** | 1 GB | 1 GB | 2 GB | 2 GB | 2 GB | 4 GB | 4GB  \n**Processor chip** | A7 | A8 | A9 | A10 | A11 | A12 | A13  \n**Model** | **Accuracy (%)** | **Size (MB)** | **FPS** | **FPS** | **FPS** | **FPS** | **FPS** | **FPS** | **FPS**  \n**VGG-16** | 71 | 553 | 0.1 | 0.2 | 4.2 | 5.5 | 6.9 | 27.8 | 34.5  \n**InceptionV3** | 78 | 95 | 1.4 | 1.5 | 9 | 11.1 | 12.8 | 35.7 | 41.7  \n**ResNet-50** | 75 | 103 | 1.9 | 1.7 | 11.6 | 13.5 | 14.1 | 38.5 | 50  \n**MobileNet** | 71 | 17 | 7.8 | 9 | 19.6 | 28.6 | 28.6 | 55.6 | 71.4  \n**SqueezeNet** | 57 | 5 | 13.3 | 12.4 | 29.4 | 33.3 | 34.5 | 66.7 | 76.9\n\nThe difference between inference times in 2014 and 2015 is particularly striking. What happened in 2015? If you guessed GPUs, you'd be right. The iPhone 6S introduced a dedicated GPU for the first time, powering features such as \"Hey Siri.\"\n\n##### Benchmarking Methodology\n\nTo do reproducible benchmarking, we took the code of the example app from earlier in the chapter, and simply swapped the _.mlmodel_ file within Xcode with other pretrained models available from Apple. For each kind of model, we ran the app for a fixed amount of time and collected measurements on execution time from the debug console. After running it on different iPhones, we averaged the results from the 6th prediction to the 20th prediction for each run and tabulated those results.\n\nOn the other hand, let's look at the device market share among iPhones in the release month of the iPhone XS (Sept. 2018) in the United States, as shown in Table 11-4. Note that releases typically happen in September of each year.\n\nTable 11-4. US device market share among iPhones as of September 2018 (the release month of the iPhone XS; data from Flurry Analytics adjusted to exclude iPhone XS, XS Plus, and XS Max) Year released | iPhone model | Percentage  \n---|---|---  \n2017 | 8 Plus | 10.8%  \n2017 | X | 10.3%  \n2017 | 8 | 8.1%  \n2016 | 7 | 15.6%  \n2016 | 7 Plus | 12.9%  \n2016 | SE | 4.2%  \n2015 | 6S | 12.5%  \n2015 | 6S Plus | 6.1%  \n2014 | 6 | 10.7%  \n2014 | 6 Plus | 3.3%  \n2013 | 5S | 3.4%  \n2013 | 5C | 0.8%  \n2012 | 5 | 0.8%  \n2011 | 4S | 0.4%  \n2010 | 4 | 0.2%\n\nDeriving the cumulative percentages from Table 11-4, we get Table 11-5. This table expresses our potential market share based on the oldest device we choose to support. For example, iPhones released after September 2016 (two years before September 2018) have a total market share of 61.9%.\n\nTable 11-5. Cumulative market share of iPhones, year-on-year, going back in time Years under | Cumulative percentage  \n---|---  \n1 | 29.2%  \n2 | 61.9%  \n3 | 80.5%  \n4 | 94.5%  \n5 | 98.7%  \n6 | 99.5%\n\nCombining the benchmark and market share, there are some design choices and optimizations available to us:\n\nUsing a faster model\n    On an iPhone 6, VGG-16 runs about 40 times slower than MobileNet. On an iPhone XS, it's still about two times slower. Just choosing a more efficient model can have a dramatic impact on performance, often without the need to compromise an equivalent amount of accuracy. It's worth noting that MobileNetV2 and EfficientNet offer an even better combination of speed and accuracy.\nDetermining the minimum FPS to support\n    A feature to apply filters on a real-time camera feed will need to be able to process a high number of the frames delivered by the camera every second to ensure a smooth experience. In contrast, an app that processes a single image due to a user-initiated action does not need to worry about performance nearly as much. A lot of applications are going to be somewhere in between. It's important to determine the minimum necessary FPS and perform benchmarking similar to Table 11-5 to arrive at the best model(s) for the scenario across iPhone generations.\nBatch processing\n    GPUs are really good at parallel processing. So, it comes as no surprise that processing a batch of data is more efficient than performing inferences on each item individually. Core ML takes advantage of this fact by exposing batch processing APIs. Some user experiences, particularly those that are asynchronous and\/or memory intensive, can vastly benefit from these batching APIs. For example, any bulk processing of photos in the photo gallery. Instead of processing one image at a time, we can send a list of images together to the batch API, so that Core ML can optimize the performance on the GPU.\nDynamic model selection\n    Depending on the use case, we might not care nearly as much about high levels of accuracy as we do about a smooth experience (at some minimum FPS threshold). In these scenarios, we might choose a lighter, less-accurate model on a slower, older device, and a larger, more-accurate model on faster, modern devices. This can be used in conjunction with model deployment via the cloud so that we don't needlessly increase the app size.\nUsing _Sherlocking_ to our advantage\n\n\"Sherlocking\" is a term used for when first-party vendors (Apple in particular) make third-party software obsolete by baking in a feature themselves. A good example of that is when Apple released the Flashlight feature within the iPhone rendering all third-party flashlight apps (whether paid or free) obsolete. For illustration here, take the example of a hypothetical app that released a face tracking feature in 2017. A year later, Apple added a more accurate face tracking API within Core ML for iOS 12. Because the neural networks for the face tracking come baked into the OS, the app could update its code to use the built-in API. However, because the API would still not be available for any users in iOS 11, the app could use a hybrid approach in which it would remain backward compatible by using the old code path for iOS 11. The app developer could then stop bundling the neural network into the app, thereby reducing its size, potentially by several megabytes, and dynamically download the older model for users of iOS 11.\n\nGraceful degradation\n    Sometimes an older phone just cannot handle newer deep learning models due to performance needs. That's a perfectly reasonable situation to be in, particularly if the feature is cutting edge. In those scenarios, it's acceptable to use one of two approaches. The first is to offload the computation to the cloud. This obviously comes at the cost of interactivity and privacy. The other alternative is to disable the feature for those users and display a message to them stating why it is not available.\n\n# Measuring Energy Impact\n\nIn the previous chapters, we focused on hosting classifiers on servers. Although there are several factors that affect design decisions, energy consumption is often not among them. However, on the client side, battery power tends to be limited, and minimizing its consumption becomes a priority. How users perceive a product depends a lot on how much energy it consumes. Remember the good old days when GPS was a battery hog. Many apps that required location access used to get tons of one-star ratings for eating up too much battery power. And we definitely don't want that.\n\n###### Note\n\nPeople are pretty generous with bad reviews when they realize an app is eating their battery faster than anticipated, as the reviews shown in Figure 11-13 demonstrate.\n\n###### Figure 11-13. App Store reviews for YouTube and Snapchat that complain about heavy battery consumption\n\nHere we utilize the Energy Impact tab (Figure 11-14) in Xcode's Debug Navigator and generate Figure 11-15.\n\n###### Figure 11-14. Xcode Debug Navigator tab\n\n###### Figure 11-15. Xcode Energy Impact chart on an iPad Pro 2017 (note: this screenshot was taken at a different time than Figure 11-14, which is why the numbers are slightly different)\n\nFigure 11-15 demonstrates that the energy impact of running the process on the iPad Pro 2017 is high. A big reason for this is that in our sample code, we were processing every single frame that we were receiving from the camera. This meant that each frame that the camera captured was being sent to the GPU for processing, resulting in higher energy consumption. In many real-world applications, it is not necessary to classify every single frame. Even processing every other frame can result in significant energy savings without generally affecting the UX. In the next section, we explore the relationship between the frame processing rate and the energy impact.\n\n###### Note\n\nThe ratio of CPU to GPU usage is affected by a model's architecture, specifically the number of convolutional operations. Here we profile the average CPU and GPU utilization for different models (Figure 11-16). These numbers can be extracted from the Energy Impact chart in Xcode. Note that we'd ideally prefer models with more GPU utilization for performance efficiency.\n\n###### Figure 11-16. Comparing CPU and GPU utilization for different models on iOS 11\n\n## Benchmarking Load\n\nAs you would expect, running CNN models in real time on each frame causes high GPU\/CPU utilization, which in turn rapidly drains the battery. Users might also notice the phone heating up.\n\nInstead of running an analysis on every frame, how about we skip a few frames? MobileNet takes 20 ms to analyze one frame on an iPad Pro 2017. So, it classifies about 50 FPS. Instead, if we run it at 1 FPS, the GPU utilization is reduced from 42% to a mere 7%\u2014a reduction of more than 83%! For many applications, processing at 1 FPS might be sufficient while still keeping the device cool. For example, a security camera might perform reasonably well even with processing frames once every couple of seconds.\n\nBy varying the number of frames analyzed per second and measuring the percentage of GPU utilization, we observe a fascinating trend. It is evident from the graph in Figure 11-17 that the higher the FPS, the more the GPU is utilized. That has a significant impact on energy consumption. For apps that need to be run for longer periods of time, it might be beneficial to reduce the number of inferences per second.\n\n###### Figure 11-17. Varying the FPS and analyzing the load on an iPad Pro 2017\n\nThe values shown in the graph were derived from the Core Animation Instrument of Xcode Instruments. Following is the process we used to generate our results:\n\n  1. From Xcode, click Product and then select Profile.\n\n  2. After the Instruments window appears, select the Core Animation Instrument, as shown in Figure 11-18.\n\n###### Figure 11-18. The Instruments window in Xcode Instruments\n\n  3. Press the Record button to start running the app in the Profiling mode.\n\n  4. Wait a few seconds to begin collecting instrumentation data.\n\n  5. Measure the values in the GPU Hardware Utilization column (Figure 11-19).\n\n###### Figure 11-19. An app being profiled live in the Core Animation instrument\n\nSo far in this chapter, we've explored techniques for making production-worthy apps using Core ML. In the next section, we discuss examples of such apps in the real world.\n\n# Reducing App Size\n\nApp size can be crucially important to developers in certain markets. Apple does not allow apps larger than 200 MB to be downloaded via the cellular network. This becomes a crucial point of contention for apps that are used frequently on the go, such as ride-hailing apps like Uber and Lyft. Interesting tidbit here: Uber had to do some really stringent optimizations such as heavily reducing the number of Swift Optionals, Structs, and Protocols (which otherwise help with code maintainability) in order to bring the app binary size below the App Store limit (at the time, 150 MB). Without doing that, the company would have lost a lot of new users.\n\nIt's a given that any new AI functionality we introduce in the app will result in additional usage of storage. There are a few strategies that we can employ to reduce the impact of that.\n\n###### Note\n\nSegment, a San Francisco-based data analytics company, wanted to determine how much of an impact app size has on the number of installs. To experiment, the company purchased a mortgage calculator app that had a steady number of app downloads (approximately 50 per day). It took the app that was originally 3 MB in size and kept bloating it repeatedly with...well, images of Taylor Swift albums (hey, it was in the name of science and research!). Engineers observed significant drops in the number of daily installs as the company increased the app size. When the app crossed 100 MB (the maximum limit for cellular network downloads from the App Store at the time), the number of daily installs dropped a whopping 44%! Additionally, the app attracted several negative reviews with users expressing incredulity at the app size.\n\nThe moral of the story here is that the app size is far more crucial than we think, and we ought to be mindful of the amount of space our app consumes before releasing it out to the public.\n\n## Avoid Bundling the Model\n\nIf possible, we should avoid bundling the model with the App Store binary. The same amount of data needs to be downloaded anyway, so as long as the UX is not affected, we should delay model download to when the feature is actually going to be used. Additionally, we should prefer to download the model when on WiFi in order to preserve cellular bandwidth. Microsoft Translator and Google Translate implement a form of this where they do only cloud-based translations at the outset. Knowing that travelers use these apps a lot (where they might not have good internet access), they also offer an offline mode where the required language models are downloaded at the user's prompt in the background.\n\n## Use Quantization\n\nAs we discussed in Chapter 6, quantization is a good strategy to reduce model size drastically while still preserving performance. Essentially, it reduces 32-bit floating-point weights down to 16-bit floating point, and 8-bit integers all the way down to 1-bit. We definitely won't recommend going below 8 bits, though, due to the resulting loss in accuracy. We can achieve quantization with Core ML Tools for Keras models in just a couple of lines:\n\n    import coremltools\n\n    model_spec = coremltools.utils.load_spec(\"MyModel.mlmodel\")\n\n    _# 16-bit conversion_\n    model_fp16_spec =\n    coremltools.utils.convert_neural_network_spec_weights_to_fp16(model_spec)\n    coremltools.utils.save_spec(model_fp16_spec, \"MyModel_FP16.mlmodel\")\n\n    _# 8-bit or lower quantization_\n    num_bits = 8\n    model_quant_spec =\n    coremltools.models.neural_network.quantization_utils.quantize_weights(model_spec,\n    num_bits, \"linear\")\n    coremltools.utils.save_spec(model_quant_spec, \"MyModel_Quant.mlmodel\")\n\nTo illustrate the impact of quantization on a challenging dataset where small modifications could lead to drastic changes, we chose to build a classifier for Oxford's 102 Category Flower Dataset with Keras (roughly 14 MB in size), and quantized it to different bit representations while measuring its accuracy and decreasing its size. To measure the change in predictions, we compare the percent match between the full-precision model and the quantized model. We tested three quantization modes.\n\n  * Simple `linear` quantization, which we described in Chapter 6. The intervals are distributed equally in this strategy.\n\n  * Linear quantization using a lookup table, or `linear_lut`. In this technique, the intervals are distributed unequally, with denser areas getting smaller and more intervals, whereas sparser areas would have fewer and larger intervals. Because these intervals are unequal, they need to be stored in the lookup table, rather than being directly computed with simple arithmetic.\n\n  * Lookup table generated by _k_ -means, or `kmeans_lut`, which is often used in nearest neighbor classifiers.\n\nTable 11-6 shows our observations.\n\nTable 11-6. Quantization results for different target bit sizes and different quantization modes **Quantized to** | **Percent size reduction (approx.)** | **Percent match with 32-bit results**  \n---|---|---  \n**`linear`** | **`linear_lut`** | **`kmeans_lut`**  \n16-bit | 50% | 100% | 100% | 100%  \n8-bit | 75% | 88.37% | 80.62% | 98.45%  \n4-bit | 88% | 0% | 0% | 81.4%  \n2-bit | 94% | 0% | 0% | 10.08%  \n1-bit | 97% | 0% | 0% | 7.75%\n\nThis gives us a few insights.\n\n  * Lowering representation to 16-bits has no effect at all on accuracy. We can essentially halve the model size with no perceivable difference in accuracy.\n\n  * _k_ -means when used for building the lookup table outperforms simple linear division methods. Even at 4-bit, only 20% accuracy is lost, which is remarkable.\n\n  * Going to 8-bit quantized models gives 4 times smaller modes with little loss in accuracy (especially for `kmeans_lut` mode).\n\n  * Quantizing to lower than 8-bits results in a drastic drop in accuracy, particularly with the linear modes.\n\n## Use Create ML\n\nCreate ML is a tool from Apple to train models by simply dragging and dropping data into a GUI application on the Mac. It provides several templates including object classification\/detection, sound classification, and text classification, among others, putting training AI in the hands of novices without requiring any domain expertise. It uses transfer learning to tune only the handful of layers that will be necessary for our task. Because of this, the training process can be completed in just a few minutes. The OS ships with the bulk of the layers (that are usable across several tasks), whereas the task-specific layers can be shipped as part of the app. The resulting exported model ends up being really small (as little as 17 KB, as we will see shortly). We get hands-on with Create ML in the next chapter.\n\n# Case Studies\n\nLet's take a look at some real-world examples that use Core ML for mobile inferences.\n\n## Magic Sudoku\n\nSoon after the launch of ARKit on iOS in 2017, the Magic Sudoku app from Hatchlings, a game and mobile app startup, came out as one of the break-out hits. Simply point the phone at a Sudoku puzzle, and the app will show a solved Sudoku right on the piece of paper. As we might have guessed at this point, the system is using Core ML to run a CNN-based digit recognizer. The system goes through the following steps:\n\n  1. Use ARKit to get camera frames.\n\n  2. Use iOS Vision Framework to find rectangles.\n\n  3. Determine whether it is a Sudoku grid.\n\n  4. Extract 81 squares from the Sudoku image.\n\n  5. Recognize digits in each square using Core ML.\n\n  6. Use a Sudoku solver to finish the empty squares.\n\n  7. Project the finished Sudoku on the surface of original paper using ARKit, as shown in Figure 11-20.\n\n###### Figure 11-20. Step-by-step solution to solving ARKit (image source)\n\nThe Hatchlings team first started with using the MNIST digit recognition model, which mostly consists of handwritten digits. But this was not robust for printed fonts. The team then photographed thousands of pages of Sudoku books, extracting the squares using its pipeline until it had a large number of images of individual digits in a variety of fonts. To label this data, the team requested its fans classify each item as 0 through 9 and empty classes. Within 24 hours, the team got 600,000 digits scanned. The next step was to train a custom CNN, which needed to work fast, because the system needed to classify 81 square images. Deploying this model using Core ML, the app launched and become an overnight hit.\n\nRunning out with new users in the wild brought new cases not previously anticipated. Because most users do not have a Sudoku puzzle in front of them, they often search for Sudoku on computer screens, and this was difficult for the model to always recognize precisely. Additionally, due to a fixed focal length limitation of ARKit, the input image could be slightly blurry. To remedy this, the Hatchlings team collected additional examples of photos of computer screens with puzzles, blurred them slightly, and trained a new CNN with the additional data. With an App Store update, the whole experience became much more robust. In summary, while launching an app or a service, the app builder needs to constantly learn from new scenarios not previously anticipated.\n\n## Seeing AI\n\nSeeing AI is a talking camera app from Microsoft Research designed for the blind and low-vision community. It uses computer vision to describe people, text, handwriting, objects, currency, and more through spoken audio. Much of the image processing happens locally on the device, using Core ML. One of the core user needs is to recognize products\u2014this can usually be determined by scanning close-ups of barcodes. But for a blind user, the location of a barcode is unknown, making most barcode apps difficult to use. To solve this, the team built a custom CNN trained with images containing barcodes at a variety of angles, sizes, lighting, and orientation. The user now attempts to rotate the object in front of the iPhone, and when the CNN classifies the presence of a barcode (running in real time, frame by frame), it signals with an audible beep. The rate of beeping directly correlates with the area of the barcode visible to the camera. As the blind users begin to bring the barcode closer, it beeps faster. When it is close enough for the barcode reading library to be able to clearly see the barcode, the app decodes the Universal Product Code and speaks the product name. In the past, blind users typically had to purchase and carry around bulky laser barcode scanners, which usually cost more than $1,300. In fact, a charity raised millions to donate these hardware barcode readers to those who needed them. Now, deep learning can solve this problem for free. This is a good example of mixing computer vision and UX to solve a real-world problem.\n\n## HomeCourt\n\nFor anything in life, regular practice is a must if we want to get better at something, be it writing, playing a musical instrument, or cooking. However, the quality of practice is far more important than quantity. Using data to support our practice and monitor our progress does wonders to the rate at which we acquire a skill. This is exactly what NEX Team, a San Jose, California-based startup, set out to do for basketball practice with the help of its HomeCourt app. Running the app is easy: set it up on the ground or a tripod, point the camera at the basketball court, and hit record.\n\nThe app runs an object detector in real time on top of Core ML to track the ball, the people, and the basketball hoop. Among these people, the big question remains: who shot the ball? When the ball goes near the hoop, the app rewinds the video to identify the player who shot the ball. Then, it performs human pose estimation on the player to track the player's movements. If this wasn't impressive enough, it uses geometric transformations to convert this 3D scene of the court into a 2D map (as shown in the upper right of Figure 11-21) to track the locations from which the player shot the ball. The important thing to note here is that multiple models are being run simultaneously. Based on the tracked objects and the player's body position, joints, and so on, the app provides the player with stats and visualizations for the release height of each throw, release angle, release position, release time, player speed, and more. These stats are important because they have a high correlation with a successful attempt. Players are able to record an entire game as they play and then go home and analyze each shot so that they can determine their weak areas in order to improve for the next time.\n\nIn just under two years of existence, this tiny startup attracted hundreds of thousands of users as word of mouth spread, from amateurs to professionals. Even the National Basketball Association (NBA) partnered with NEX Team to help improve their players' performance using HomeCourt. All of this happened because they saw an unmet need and came up with a creative solution using deep learning.\n\n###### Figure 11-21. The HomeCourt app tracking a player's shots live as they are playing\n\n## InstaSaber + YoPuppet\n\nDo you know the largest profit maker in the _Star Wars_ franchise? It's not the movie ticket sales, and it's definitely not from selling DVDs. Here's a hint: it rhymes with merchandise. Lucas Film (and now Disney) made a lot of money selling R2D2 toys and Chewbacca costumes. Though the all-time favorite still remains the dear lightsaber. However, instead of looking cool like in the movies, the merchandise lightsabers are mostly made of plastic and frankly don't look as sci-fi. Plus, you swing it around a few times and there's a good chance you'll accidentally whack someone in the face.\n\nHart Woolery, the founder of 2020CV, decided to change that by bringing Hollywood-level VFX to our phone with the InstaSaber app. Simply roll up a sheet of paper, grip it, and point your phone's camera toward it and watch it transform into a glowing lightsaber, as illustrated in Figure 11-22. Wave your hand and not just watch it track it realistically in real time, but also hear the same sound effects as when Luke fought his father, (spoiler alert!) Darth Vader.\n\nTaking the tracking magic to the next level, he built YoPuppet, which tracks the joints in hands in real time to build virtual puppets that mimic the motion of the hand. The suspension of disbelief works only if it is truly real time with no lag, is accurate, and realistic looking.\n\nIn addition to being realistic, these apps were a lot of fun to play with. No wonder InstaSaber became an insta-hit with millions of viral views online and in the news. The AI potential even got billionaire investor Mark Cuban to say, \"you've got a deal\" and invest in 2020CV.\n\n###### Figure 11-22. Screenshots of InstaSaber and YoPuppet\n\n##### From the Creator's Desk\n\nBy Hart Woolery, founder and CEO of 2020CV\n\nSince starting 2020CV a few years ago, I've developed a number of apps that attempt to push the boundaries of these special effects. I will talk about two: InstaSaber and YoPuppet. InstaSaber turns a rolled-up sheet of paper into a virtual lightsaber, and YoPuppet transforms the user's hand into a virtual hand puppet. In both cases, I am using real-time camera frames as inputs to my models, and returning normalized x-y coordinates of feature locations inside those images. I got the idea for InstaSaber after attempting to track a handheld marker, because I was struggling at the time to track hands by themselves. The initial challenge\u2014as is generally the case for most machine learning models\u2014is gathering the data, labeling the data, and making sure you have enough data to create a generalized solution.\n\nJust how many labeled images do you need for a machine learning model that can decode an object reliably with multiple variables like lighting, skin tone, and occlusion? I think the answer depends largely on the problem, but at a minimum, it will generally be at least 10,000. Assuming it could take 30 seconds or more per image to label it correctly, that is a lot of time being invested. The simplest workaround is to augment your data (i.e., affine transforms, cropping, and color\/lighting manipulation). While this turns a finite amount of data into a nearly infinite set of variants, your model will still be constrained by the quality of your initial set. While I manually labeled images for InstaSaber, I found it far more efficient to synthetically generate images of hands in various positions and environments for YoPuppet (Figure 11-23). Once you are able to train and iterate on your model, the next challenge is getting them to perform well within the constraints of a mobile device.\n\n###### Figure 11-23. Real-time hand pose estimation within YoPuppet\n\nIt's fairly easy to get a machine learning model running on a mobile device these days with the advent of such frameworks as Core ML. However, getting a model to run in real time and deliver a great user experience\u2014without melting the user's phone\u2014is challenging. In the case of YoPuppet and InstaSaber, I also had to find the initial location of the object, and track it in real time so I could crop a square out of the full camera image, scale it down, and pass it to my model. One of the tricks I use to maintain a smooth frame rate is buffering the images so that the CPU and GPU can respectively preprocess and run inference on the images in tandem.\n\nOnce your app is published and out in the wild, the real-world feedback will often make you realize there are a lot of issues with your model you may have never considered (the bias is real!). For example, one user said that painted fingernails seemed to be causing issues with hand detection. Fortunately, this kind of feedback only helps you further generalize your model and make it more robust to variations in your objects. To be certain, the process of delivering a great experience for all your users is a never-ending challenge, but the reward is that you will be a pioneer in an emerging area of technology.\n\n# Summary\n\nIn this chapter, we went on a whirlwind tour of the world of Core ML, which provides an inference engine for running machine learning and deep learning algorithms on an iOS device. Analyzing the minimal code needed for running a CNN, we built a real-time object recognition app that classifies among 1,000 ImageNet categories. Along the way, we discussed some useful facts about model conversion. Moving to the next level, we learned about practical techniques like dynamic model deployment and on-device training, while also benchmarking different deep learning models on various iOS devices, providing a deeper understanding of the battery and resource constraints. We also looked at how we can optimize the app size using model quantization. Finally, to take some inspiration from the industry, we explored real-life examples where Core ML is being used in production apps.\n\nIn the next chapter, we build an end-to-end application by training using Create ML (among other tools), deploying our custom trained classifier, and running it using Core ML.\n\n# Chapter 12. Not Hotdog on iOS with Core ML and Create ML\n\n\"I'm a rich,\" said Jian-Yang, a newly minted millionaire in an interview with Bloomberg (Figure 12-1). What did he do? He created the Not Hotdog app (Figure 12-2) and made the world \"a better place.\"\n\n###### Figure 12-1. Jian-Yang being interviewed by Bloomberg News after Periscope acquires his \"Not Hotdog\" technology (image source: From HBO's Silicon Valley)\n\nTo the few of us who may be confused (including a third of the authors of this book), we are making a reference to HBO's _Silicon Valley_ , a show in which one of the characters is tasked with making SeeFood\u2014the \"Shazam for food.\" It was meant to classify pictures of food and give recipes and nutritional information. Hilariously, the app ends up being good only for recognizing hot dogs. Anything else would be classified as \"Not Hotdog.\"\n\nThere are a few reasons we chose to reference this fictitious app. It's very much a part of popular culture and something many people can easily relate to. It's an exemplar: easy enough to build, yet powerful enough to see the magic of deep learning in a real-world application. It is also very trivially generalizable to recognize more than one class of items.\n\n###### Figure 12-2. The Not Hotdog app in action (image source: Apple App Store listing for the Not Hotdog app)\n\nIn this chapter, we work through a few different approaches to building a Not Hotdog clone. The general outline of the end-to-end process is as follows:\n\n  1. Collect relevant data.\n\n  2. Train the model.\n\n  3. Convert to Core ML.\n\n  4. Build the iOS app.\n\nTable 12-1 presents the different options available for steps 1 through 3. Further along in the chapter, we do a deep dive into each of them.\n\nTable 12-1. Various approaches to getting a model ready for mobile deployment, right from scratch **Data collection** | **Training mechanism** | **Model conversion**  \n---|---|---\n\n  * Find or collect a dataset\n\n  * Fatkun Chrome browser extension\n\n  * Web scraper using Bing Image Search API\n |   * Web-based GUI: CustomVision.ai, IBM Watson, Clarifai, Google AutoML\n\n  * Create ML\n\n  * Fine-tune using any framework of choice like Keras\n |   * Create ML, CustomVision.ai, and other GUI tools generate . _mlmodel_.\n\n  * For Keras, use Core ML Tools.\n\n  * For TensorFlow trained models, use `tf-coreml`.\n\nLet's dive right in!\n\n# Collecting Data\n\nTo begin solving any computer-vision task using deep learning, we first need to have a dataset of images to train on. In this section, we use three different approaches to collecting the images of the relevant categories in increasing order of time required, from minutes to days.\n\n## Approach 1: Find or Collect a Dataset\n\nThe fastest way to get our problem solved is to have an existing dataset in hand. There are tons of publicly available datasets for which a category or a subcategory might be relevant to our task. For example, Food-101 ( _https:\/\/www.vision.ee.ethz.ch\/datasets_extra\/food-101\/_) from ETH Zurich contains a class of hot dogs. Alternatively, ImageNet contains 1,257 images of hot dogs. We can use a random sample of images from the remaining classes as \"Not Hotdog.\"\n\nTo download images from a particular category, you can use the ImageNet-Utils tool:\n\n  1. Search for the relevant category on the ImageNet website; for example, \"Hot dog.\"\n\n  2. Note the `wnid` (WordNet ID)in the URL: _http:\/\/image-net.org\/synset?wnid=n07697537_.\n\n  3. Clone the ImageNet-Utils repository:\n\n    $ git clone --recursive\n    \thttps:\/\/github.com\/tzutalin\/ImageNet_Utils.git\n\n  4. Download the images for the particular category by specifying the `wnid`:\n\n    $ .\/downloadutils.py --downloadImages --wnid n07697537\n\nIn case we can't find a dataset, we can also build our own dataset by taking pictures ourselves with a smartphone. It's essential that we take pictures representative of how our application would be used in the real world. Alternatively, crowdsourcing this problem, like asking friends, family, and coworkers, can generate a diverse dataset. Another approach used by large companies is to hire contractors who are tasked with collecting images. For example, Google Allo released a feature to convert selfies into stickers. To build it, they hired a team of artists to take an image and create the corresponding sticker so that they could train a model on it.\n\n###### Note\n\nMake sure to check the license under which the images in the dataset are released. It's best to use images released under permissive licenses such as Creative Commons.\n\n## Approach 2: Fatkun Chrome Browser Extension\n\nThere are several browser extensions that allow us to batch download multiple images from a website. One such example is the Fatkun Batch Download Image, a browser extension available on the Chrome browser.\n\nWe can have the entire dataset ready in the following short and quick steps.\n\n  1. Add the extension to our browser.\n\n  2. Search for the keyword either on Google or Bing Image search.\n\n  3. Select the appropriate filter for image licenses in the search settings.\n\n  4. After the page reloads, scroll to the bottom of it a few times repeatedly to ensure more thumbnails are loaded on the page.\n\n  5. Open the extension and select \"This Tab\" option, as demonstrated in Figure 12-3.\n\n###### Figure 12-3. Bing Search results for \"hot dog\"\n\n  6. Notice that all the thumbnails are selected by default. At the top of the screen, click the Toggle button to deselect all the thumbnails and select only the ones we need. We can set the minimum width and height to be 224 (most pretrained models take 224x224 as input size).\n\n###### Figure 12-4. Selecting images through the Fatkun extension\n\n  7. In the upper-right corner, click Save Image to download all of the selected thumbnails to our computer.\n\n###### Note\n\nNote that the images shown in the screenshots are iconic images (i.e., the main object is in direct focus with a clean background). Chances are that using such images exclusively in our model will cause it to not generalize well to real-world images. For example, in images with a clean white background (like on an ecommerce website), the neural network might incorrectly learn that white background equals hot dog. Hence, while performing data collection, ensure that your training images are representative of the real world.\n\n###### Tip\n\nFor the negative class of \"Not Hotdog,\" we want to collect random images that are abundantly available. Additionally, collect items that look similar to a hot dog but are not; for example, a submarine sandwich, bread, plate, hamburgers, and so on.\n\nAn absence of commonly co-occurring items with hot dogs like plates with food, tissues, ketchup bottles or packets, can mistakenly lead the model to think that those are the real hot dogs. So be sure to add these to the negative class.\n\nWhen you install a browser extension like Fatkun, it will request permissions to read and modify data on all the websites we visit. It might be a good idea to disable the extension when you're not using it for downloading images.\n\n## Approach 3: Web Scraper Using Bing Image Search API\n\nFor building larger datasets, using Fatkun to collect images can be a tedious process. Additionally, images returned by the Fatkun browser extension are thumbnails and not the original size images. For large-scale image collections, we can use an API for searching images, like the Bing Image Search API, where we can establish certain constraints such as the keyword, image size, and license. Google used to have the Image Search API, but it was discontinued in 2011.\n\nBing's Search API is an amalgamation of its AI-based image understanding and traditional information retrieval methods (i.e., using tags from fields like \"alt-text,\" \"metadata,\" and \"caption\"). Many times, we can end up with some number of irrelevant images because of misleading tags from these fields. As a result, we want to manually parse the collected images to make sure that they are actually relevant to our task.\n\nWhen we have a very large image dataset, it can be a daunting task to have to go through it manually and filter out all of the poor training examples. It's easier to approach this in an iterative manner, slowly improving the quality of the training dataset with each iteration. Here are the high-level steps:\n\n  1. Create a subset of the training data by manually reviewing a small number of images. For example, if we have 50k images in our original dataset, we might want to manually select around 500 good training examples for the first iteration.\n\n  2. Train the model on those 500 images.\n\n  3. Test the model on the remaining images and get the confidence value for each image.\n\n  4. Among the images with the least confidence values (i.e., often mispredictions), review a subset (say, 500) and discard the irrelevant images. Add the remaining images from this subset to the training set.\n\n  5. Repeat steps 1 through 4 for a few iterations until we are happy with the quality of the model.\n\nThis is a form of semisupervised learning.\n\n###### Tip\n\nYou can improve the model accuracy further by reusing the discarded images as negative training examples.\n\n###### Note\n\nFor a large set of images that don't have labels, you might want to use co-occurrence of other defining text as labels; for example, hashtags, emojis, alt-text, etc.\n\nFacebook built a dataset of 3.5 billion images using the hashtags from the text of corresponding posts as weak labels, training them, and eventually fine tuning them on the ImageNet dataset. This model beat the state-of-the-art result by 2% (85% top 1% accuracy).\n\nNow that we have collected our image datasets, let's finally begin training them.\n\n# Training Our Model\n\nBroadly speaking there are three easy ways to train, all of which we have discussed previously. Here, we provide a brief overview of a few different approaches.\n\n## Approach 1: Use Web UI-based Tools\n\nAs discussed in Chapter 8, there are several tools to build custom models by supplying labeled images and performing training using the web UI. Microsoft's CustomVision.ai, Google AutoML, IBM Watson Visual Recognition, Clarifai, and Baidu EZDL are a few examples. These methods are code free and many provide simple drag-and-drop GUIs for training.\n\nLet's look at how we can have a mobile-friendly model ready in less than five minutes using CustomVision.ai:\n\n  1. Go to _http:\/\/customvision.ai_, and make a new project. Because we want to export the trained model to a mobile phone, select a compact model type. Since our domain is food related, select \"Food (Compact),\" as shown in Figure 12-5.\n\n###### Figure 12-5. Define a new project on CustomVision.ai\n\n  2. Upload the images and assign tags (labels), as depicted in Figure 12-6. Upload at least 30 images per tag.\n\n###### Figure 12-6. Uploading images on the CustomVision.ai dashboard. Note that the tags have been populated as Hotdog and Not Hotdog\n\n  3. Click the Train button. A dialog box opens, as shown in Figure 12-7. Fast Training essentially trains the last few layers, whereas Advanced Training can potentially tune the full network giving even higher accuracy (and obviously take more time and money). The Fast Training option should be sufficient for most cases.\n\n###### Figure 12-7. Options for training type\n\n  4. In under a minute, a screen should appear, showing the precision and recall for the newly trained model per category, as shown in Figure 12-8. (This should ring a bell because we had discussed precision and recall earlier in the book.)\n\n###### Figure 12-8. Precision, Recall, and Average Precision of the newly trained model\n\n  5. Play with the probability threshold to see how it changes the model's performance. The default 90% threshold achieves pretty good results. The higher the threshold, the more precise the model becomes, but at the expense of reduced recall.\n\n  6. Press the Export button and select the iOS platform (Figure 12-9). Internally, CustomVision.ai converts the model to Core ML (or TensorFlow Lite if you're exporting for Android).\n\n###### Figure 12-9. The model exporter options in CustomVision.ai\n\nAnd we're done, all without writing a single line of code! Now let's look at an even more convenient way of training without coding.\n\n## Approach 2: Use Create ML\n\nIn 2018, Apple launched Create ML as a way for developers within the Apple ecosystem to train computer-vision models natively. Developers could open a _playground_ and write a few lines of Swift code to train an image classifier. Alternatively, they could use `CreateMLUI` import to display a limited GUI training experience within the playground. It was a good way to get Swift developers to deploy Core ML models without requiring much experience in machine learning.\n\nA year later, at the Apple Worldwide Developers Conference (WWDC) 2019, Apple lowered the barrier even further by announcing the standalone Create ML app on macOS Catalina (10.15). It provides an easy-to-use GUI to train neural networks without needing to write any code at all. Training a neural network simply became a matter of dragging and dropping files into this UI. In addition to supporting the image classifier, they also announced support for object detectors, NLP, sound classification, activity classification (classify activities based on motion sensor data from Apple Watch and iPhone), as well as tabular data (including recommendation systems).\n\nAnd, it's fast! Models can be trained in under a minute. This is because it uses transfer learning, so it doesn't need to train all of the layers in the network. It also supports various data augmentations such as rotations, blur, noise, and so on, and all you need to do is click checkboxes.\n\nBefore Create ML came along, it was generally a given that anyone seeking to train a serious neural network in a reasonable amount of time had to own an NVIDIA GPU. Create ML took advantage of the onboard Intel and\/or Radeon graphics cards, which allowed faster training on MacBooks without the need to purchase additional hardware. Create ML allows us to train multiple models, from different data sources, all at the same time. It can benefit particularly from powerful hardware such as the Mac Pro or even an external GPU (eGPU).\n\nOne major motivation to use Create ML is the size of the models it outputs. A full model can be broken down into a base model (which emits features) and lighter task-specific classification layers. Apple ships the base models into each of its operating systems. So, Create ML just needs to output the task-specific classifier. How small are these models? As little as a few kilobytes (compared to more than 15 MB for a MobileNet model, which is already pretty small for a CNN). This is important in a day and age when more and more app developers are beginning to incorporate deep learning into their apps. The same neural networks do not need to be unnecessarily replicated across several apps consuming valuable storage space.\n\nIn short, Create ML is easy, speedy, and tiny. Sounds too good to be true. Turns out the flip-side of having full vertical integration is that the developers are tied into the Apple ecosystem. Create ML exports only _.mlmodel_ files, which can be used exclusively on Apple operating systems such as iOS, iPadOS, macOS, tvOS, and watchOS. Sadly, Android integration is not yet a reality for Create ML.\n\nIn this section, we build the Not Hotdog classifier using Create ML:\n\n  1. Open the Create ML app, click New Document, and select the Image Classifier template from among the several options available (including Sound, Activity, Text, Tabular), as shown in Figure 12-10. Note that this is only available on Xcode 11 (or greater), on macOS 10.15 (or greater).\n\n###### Figure 12-10. Choosing a template for a new project\n\n  2. In the next screen, enter a name for the project, and then select Done.\n\n  3. We need to sort the data into the correct directory structure. As Figure 12-11 illustrates, we place images in directories that have the names of their labels. It is useful to have separate train and test datasets with their corresponding directories.\n\n###### Figure 12-11. Train and test data in separate directories\n\n  4. Point the UI to the training and test data directories, as shown in Figure 12-12.\n\n###### Figure 12-12. Training interface in Create ML\n\n  5. Figure 12-12 shows the UI after you select the train and test data directories. Notice that the validation data was automatically selected by Create ML. Additionally, notice the augmentation options available. It is at this point that we can click the Play button (the right-facing triangle; see Figure 12-13) to start the training process.\n\n###### Figure 12-13. Create ML screen that opens after loading train and test data\n\n###### Note\n\nAs you experiment, you will quickly notice that each augmentation that we add will make the training slower. To set a quick baseline performance metric, we should avoid using augmentations in the first run. Subsequently, we can experiment with adding more and more augmentations to assess how they affect the quality of the model.\n\n  6. When the training completes, we can see how the model performed on the training data, (auto-selected) validation data, and the test data, as depicted in Figure 12-14. At the bottom of the screen, we can also observe how long the training process took and the size of the final model. 97% test accuracy in under two minutes. And all that with a 17 KB output. Not too shabby.\n\n###### Figure 12-14. The Create ML screen after training completes\n\n  7. We're so close now\u2014we just need to export the final model. Drag the Output button (highlighted in Figure 12-14) to the desktop to create the _.mlmodel_ file.\n\n  8. We can double-click on the newly exported _.mlmodel_ file to inspect the input and output layers, as well as test drive the model by dropping images into it, as shown in Figure 12-15.\n\n###### Figure 12-15. The model inspector UI within Xcode\n\nThe model is now ready to be plugged into apps on any Apple device.\n\n###### Note\n\nCreate ML uses transfer learning, training only the last few layers. Depending on your use case, the underlying model that Apple provides you might be insufficient to make high-quality predictions. This is because you are unable to train the earlier layers in the model, thereby restricting the potential to which the model can be tuned. For most day-to-day problems, this should not be an issue. However, for very domain-specific applications like X-rays, or very similar-looking objects for which the tiniest of details matter (like distinguishing currency notes), training a full CNN would be a better approach. We look at doing so in the following section.\n\n## Approach 3: Fine Tuning Using Keras\n\nBy now, we have become experts in using Keras. This option can get us even higher accuracy if we are up for experimentation and are willing to spend more time training the model. Let's reuse the code from Chapter 3 and modify parameters such as directory and filename, batch size, and number of images. You will find the code on the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_) at _code\/chapter-12\/1-keras-custom-classifier-with-transfer-learning.ipynb_.\n\nThe model training should take a few minutes to complete, depending on the hardware, and at the end of the training, we should have a _NotHotDog.h5_ file ready on the disk.\n\n# Model Conversion Using Core ML Tools\n\nAs discussed in Chapter 11, there are several ways of converting our models to the Core ML format.\n\nModels generated from CustomVision.ai are directly available in Core ML format, hence no conversion is necessary. For models trained in Keras, Core ML Tools can help convert as follows. Note that because we are using a MobileNet model, which uses a custom layer called `relu6`, we need to import `CustomObjectScope`:\n\n    from tensorflow.keras.models import load_model\n    from tensorflow.keras.utils.generic_utils import CustomObjectScope\n    import tensorflow.keras\n\n    with CustomObjectScope({'relu6':\n    tensorflow.keras.applications.mobilenet.relu6,'DepthwiseConv2D':\n    tensorflow.keras.applications.mobilenet.DepthwiseConv2D}):\n        model = load_model('NotHotDog-model.h5')\n\n    **import** coremltools\n    coreml_model **=** coremltools **.**converters **.**keras **.**convert(model)\n    coreml_model **.**save('NotHotDog.mlmodel')\n\nNow that we have a Core ML model ready, all we need to do is build the app.\n\n# Building the iOS App\n\nWe can use the code from Chapter 11 and simply replace the _.mlmodel_ with the newly generated model file, as demonstrated in Figure 12-16.\n\n###### Figure 12-16. Loading the `.mlmodel` into Xcode\n\nNow, compile and run the app and you're done! Figure 12-17 presents the awesome results.\n\n###### Figure 12-17. Our app identifying the hot dog\n\n# Further Exploration\n\nCan we make this application more interesting? We can build an actual \"Shazam for food\" by training for all the categories in the Food-101 dataset, which we cover in the next chapter. Additionally, we can enhance the UI compared to the barebones percentages our current app shows. And, to make it viral just like \"Not Hotdog,\" provide a way to share the classifications to social media platforms.\n\n# Summary\n\nIn this chapter, we worked through an end-to-end pipeline of collecting data, training and converting a model, and using it in the real world on an iOS device. For each step of the pipeline, we have explored a few different options in varying degrees of complexity. And, we have placed the concepts covered in the previous chapters in the context of a real-world application.\n\nAnd now, like Jian-Yang, go make your millions!\n\n# Chapter 13. Shazam for Food: Developing Android Apps with TensorFlow Lite and ML Kit\n\nAfter developing the viral Not Hotdog app (that we looked at in Chapter 12), Jian-Yang was originally supposed to build a classifier to recognize all food in existence. In fact, the app was originally supposed to be called SeeFood\u2014an app that can \"see\" food and know it right away (Figure 13-1). In other words, the \"Shazam for Food.\" However, the app was too successful for its own good and was acquired by Periscope. The original vision of his investor, Erlich Bachman, remains unfulfilled. In this chapter, our mission is to fulfill this dream.\n\n###### Figure 13-1. Not Hotdog app listing on the Apple App Store\n\nWhere would such a feature be useful? For health nuts, it could look at a dish and provide the nutritional information, including the number of calories. Or, it could scan a few ingredients, and recommend a recipe based on them. Or, it could even look at a product in the market, and check whether it contains any blacklisted ingredients such as specific allergens.\n\nThis is an interesting problem to solve for several reasons because it represents several challenges:\n\nData collection challenge\n\nThere are more than a hundred cuisines around the world, each with hundreds if not thousands of dishes.\n\nAccuracy challenge\n\nIt should be right most of the time.\n\nPerformance challenge\n\nIt should run near instantly.\n\nPlatform challenge\n\nAn iPhone app alone would be insufficient. A lot of users in developing countries use less powerful smartphones, particularly Android devices. Cross-platform development is a must.\n\nMaking a food classifier app for one cuisine is tricky enough. Imagine having to do that for every food in existence\u2014and doing it on two platforms! An individual or a small team will quickly run into scaling issues trying to tackle this problem. In this chapter, we use this example as a motivation to explore the different parts of the mobile AI development life cycle that we explored in Chapter 11.\n\nThe material we explore here does not need to be limited to smartphones, either. We can apply our learnings beyond mobile to edge devices such as Google Coral and Raspberry Pi, which we discuss later in the book.\n\n# The Life Cycle of a Food Classifier App\n\nSo, we want to build a global multicuisine, multiplatform food classifier. It sounds like a daunting task, but we can break it down into manageable steps. As in life, we first need to crawl, then walk, and then run. The following is one potential approach to consider:\n\n  1. Collect a small initial set images for a single cuisine (e.g., Italian).\n\n  2. Label these images with their corresponding dish identifiers (e.g., `margherita_pizza`).\n\n  3. Train a classifier model.\n\n  4. Convert the model to a mobile framework-compatible format (e.g., _.tflite_ ).\n\n  5. Build a mobile app by integrating the model with a great UX.\n\n  6. Recruit alpha users and share the app with them.\n\n  7. Collect detailed usage metrics along with feedback from active users, including camera frames (which tend to reflect real-world usage) and corresponding proxy labels (indicating whether the classification was right or wrong).\n\n  8. Improve the model using the newly collected images as additional training data. This process needs to be iterative.\n\n  9. When the quality of the model meets the minimum quality bar, ship the app\/feature to more\/all users. Continue to monitor and improve the quality of the model for that cuisine.\n\n  10. Repeat these steps for each cuisine.\n\n###### Tip\n\nFor step 7, we could alternatively integrate the feedback within the UX itself. For example, the app could show a ranked list of predictions (by probability) that are likely candidates for a given picture. If our model is doing well, the user should be selecting the first option most of the time. Selecting a lower ranked prediction would essentially be considered an incorrect prediction. In the worst case, if none of the options were correct, allow the user to manually add a new label. That photo, along with the label (in all three scenarios), can be incorporated as training data.\n\nWe don't need a whole lot of data to get underway. Although each of the aforementioned steps might sound involved, we can significantly automate the process. What's cool about this approach is that the more the app is used, the better it becomes, automatically. It's as if it has a life of its own. We explore this self-evolving approach for a model toward the end of the chapter.\n\n###### Tip\n\nA die-hard fan of your app\/company would make for a good alpha user. Alpha users ideally are people who have a vested interest in the success of your product. For a food recognition app, a potential group of users could be fitness buffs who watch every calorie and ingredient they eat. These users understand that the quality of the app will not be up to par early on, but they also see their role in shaping it through consistent, constructive feedback. They voluntarily agree to a liberal data-sharing agreement in order to provide data such as usage metrics and image frames from day-to-day use. We would recommend that your users know exactly what information you would be collecting about them and allow them to opt out or delete. Don't be creepy!\n\nIn this chapter, we explore the different parts of the aforementioned life cycle and the tools that help us with each of those steps. In the end, we will take a holistic, end-to-end look at the entire mobile development life cycle, not just from this chapter, but also from the previous chapters, and combine them to see how we could effectively use them in building a production-quality, real-world application.\n\nOur journey begins with understanding the following tools from the Google ecosystem.\n\nTensorFlow Lite\n\nModel conversion and mobile inference engine.\n\nML Kit\n\nHigh-level software development kit (SDK) with several built-in APIs, along with the ability to run custom TensorFlow Lite models as well as integration with Firebase on Google Cloud.\n\nFirebase\n\nA cloud-based framework that provides the necessary infrastructure for production-quality mobile applications, including analytics, crash reporting, A\/B testing, push notifications, and more.\n\nTensorFlow Model Optimization Toolkit\n\nA set of tools for optimizing the size and performance of models.\n\n# An Overview of TensorFlow Lite\n\nAs mentioned in Chapter 11, Google released an on-device inference engine called TensorFlow Lite that expanded the reach of the TensorFlow ecosystem beyond the cloud and desktops. Prior to this, the options within the TensorFlow ecosystem were porting the entire TensorFlow library itself to iOS (which was heavy and slow) and, later on, its slightly stripped-down version called TensorFlow Mobile (an improvement, but still fairly bulky).\n\nTensorFlow Lite is optimized from the ground up for mobile, with the following salient features:\n\nSmall\n\nTensorFlow Lite comes packaged with a much lighter interpreter. Even with all the operators included, the interpreter is less than 300 KB. In typical usage with common models like MobileNet, we can expect that footprint to be less than 200 KB. For reference, the previous generation TensorFlow Mobile used to occupy 1.5 MB. Additionally, TensorFlow Lite uses s _elective registration\u2014_ it packages only the operations that it knows will be used by the model, minimizing unnecessary overheads.\n\nFast\n\nTensorFlow Lite provides a significant speedup because it is able to take advantage of on-device _hardware acceleration such as_ GPUs and NPUs, where available. In the Android ecosystem, it uses the Android Neural Networks API for acceleration. Analogously on iPhones, it uses the Metal API. Google claims two to seven times speedup over a range of tasks when using GPUs (relative to CPUs).\n\nTensorFlow uses Protocol Buffers (Protobufs) for deserialization\/serialization. Protobufs are a powerful tool for representing data due to their flexibility and extensibility. However, that comes at a performance cost that can be felt on low-power devices such as mobiles.\n\n_FlatBuffers_ turned out to be the answer to this problem. Originally built for video game development, for which low overhead and high performance are a must, they proved to be a good solution for mobiles as well in significantly reducing the code footprint, memory usage, and CPU cycles spent for serialization and deserialization of models. This also improved start-up time by a fair amount.\n\nWithin a network, there are some layers that have fixed computations at inference time; for example, the batch normalization layers, which can be precomputed because they rely on values obtained during training, such as mean and standard deviation. So, the batch normalization layer computation can be fused (i.e., combined) with the previous layer's computation ahead of time (i.e., during model conversion), thereby reducing the inference time and making the entire model much faster. This is known as _prefused activation_ , which TensorFlow Lite supports.\n\nThe interpreter uses static memory and a static execution plan. This helps in decreasing the model load time.\n\nFewer dependencies\n\nThe TensorFlow Lite codebase is mostly standard C\/C++ with a minimal number of dependencies. It makes it easier to package and deploy and additionally reduces the size of the deployed package.\n\nSupports custom operators\n\nTensorFlow Lite contains quantized and floating-point core operators, many of which have been tuned for mobile platforms, and can be used to create and run custom models. If TensorFlow Lite does not support an operation in our model, we can also write custom operators to get our model running.\n\nBefore we build our initial Android app, it would be useful to examine TensorFlow Lite's architecture.\n\n##### From the Creator's Desk\n\nBy Pete Warden, technical lead for Mobile and Embedded TensorFlow, author of _TinyML_ (O'Reilly)\n\nWhen I wrote my first neural network inference engine back in 2013, it was because I couldn't find any other way to run image recognition models on the cheap AWS servers that were all we could afford at our startup. It was a happy accident that the same code was easy to build for iOS and Android, and I quickly realized that the possibilities for deep learning on mobile platforms were endless. That code became the Jetpac SDK, and I had tremendous fun seeing all the applications people built with it in the early days. When we were acquired by Google, I met Yangqing Jia (of Caffe fame), who was working on an internal 20% project to run inference on mobile devices, so we ended up collaborating and eventually a lot of what we learned ended up in the public release of TensorFlow. A lot of people were skeptical about the value of neural networks on phones, but Jeff Dean was always a firm supporter and he helped us ensure we had Android support right from the start, and iOS very soon after. We learned a lot from those initial TensorFlow releases, and the Mobile Vision team within Google had done some amazing work with their own specialized internal framework, so we joined forces to create TensorFlow Lite as a library offering TensorFlow compatibility, but with a much smaller footprint.\n\nSince launch, TensorFlow Lite has been shipped in production on over two billion devices, but none of that could have happened without the community that's grown up around it. It's been amazing to see all the contributions, examples, and peer-to-peer support that has enabled people to build applications that wouldn't have been possible without the ecosystem working together.\n\n## TensorFlow Lite Architecture\n\nFigure 13-2 provides a high-level view of the TensorFlow Lite architecture.\n\n###### Figure 13-2. High-level architecture of the TensorFlow Lite ecosystem\n\nAs app developers, we will be working in the topmost layer while interacting with the TensorFlow Lite API (or optionally with ML Kit, which in turn uses TensorFlow Lite). The TensorFlow Lite API abstracts away all of the complexities involved in using a lower-level API such as the Android's Neural Network API. Recall that this is similar to how Core ML works within the Apple ecosystem.\n\nLooking at the other extreme, computations can be run on various types of hardware modules. The most common among them is the CPU, simply because of its ubiquity and flexibility. Modern smartphones are increasingly equipped with specialized modules, including the GPU and the newer NPU (especially built for neural network computations like on the iPhone X). Additionally, Digital Signal Processors (DSP) specialize in singular tasks such as facial authentication, fingerprint authentication, and wake word detection (like \"Hey Siri\").\n\nIn the world of Internet of Things (IoT), microcontrollers (MCUs) reign supreme. With no OS, no processor, and very little memory (KBs), these are cheap to produce in mass quantities and easy to incorporate into various applications. With TensorFlow Lite for Microcontrollers, developers can run AI on these bare-metal devices without needing internet connectivity. The pared-down version (roughly 20 KB) of the TensorFlow Lite Interpreter for MCUs is called the TensorFlow Lite Micro Interpreter.\n\nSo how does TensorFlow Lite interact with hardware? By using delegates that are platform-aware objects that expose a consistent platform-agnostic API. In other words, delegates shield the interpreter from needing to know anything about the specific hardware it runs on. They take on all or part of the responsibility of graph execution that would have otherwise been run on the CPU and instead run on the much more efficient GPUs and NPUs. On Android, the GPU delegate accelerates performance using OpenGL, whereas on iOS, the Metal API is used.\n\nGiven that TensorFlow Lite by itself is platform agnostic, it needs to call into a platform-specific library that implements a known contract. This contract is the TensorFlow Lite Delegate API. Within Android, this contract is fulfilled by the Android Neural Network API (available on devices running Android 8.1 and above). The Neural Network API is designed to provide a base layer of functionality for higher-level machine learning frameworks. The equivalent of the Neural Network API within the Apple world is Metal Performance Shaders.\n\nWith the information we have looked at so far, let's get hands on.\n\n# Model Conversion to TensorFlow Lite\n\nAt this point in the book, we should already have a model at hand (either pretrained on ImageNet or custom trained in Keras). Before we can plug that model into an Android app, we need to convert it to the TensorFlow Lite format (a . _tflite_ file).\n\nLet's take a look at how to convert the model using the TensorFlow Lite Converter tool, the `tflite_convert` command that comes bundled with our TensorFlow installation:\n\n    _# Keras to TensorFlow Lite_\n    $ tflite_convert \\\n      --output_file=my_model.tflite \\\n      --keras_model_file=my_model.h5\n\n    _# TensorFlow to TensorFlow Lite_\n    $ tflite_convert \\\n      --output_file=my_model.tflite \\\n      --graph_def_file=my_model\/frozen_graph.pb\n\nThe output of this command is the new _my_model.tflite_ file, which we can then plug into the Android app in the following section. Later, we look at how to make that model more performant by using the `tflite_convert` tool again. Additionally, the TensorFlow Lite team has created many pretrained models that are available in TensorFlow Lite format, saving us this conversation step.\n\n# Building a Real-Time Object Recognition App\n\nRunning the sample app from the TensorFlow repository is an easy way to play with the TensorFlow Lite API. Note that we would need an Android phone or tablet to run the app. Following are steps to build and deploy the app:\n\n  1. Clone the TensorFlow repository:\n\n    git clone https:\/\/github.com\/tensorflow\/tensorflow.git\n\n  2. Download and install Android Studio from _https:\/\/developer.android.com\/studio_.\n\n  3. Open Android Studio and then select \"Open an existing Android Studio project\" (Figure 13-3).\n\n###### Figure 13-3. Start screen of Android Studio\n\n  4. Go to the location of the cloned TensorFlow repository and then navigate further to _tensorflow\/tensorflow\/contrib\/lite\/java\/demo\/_ (Figure 13-4). Select Open.\n\n###### Figure 13-4. Android Studio \"Open Existing Project\" screen in the TensorFlow repository\n\n  5. On the Android device, enable Developer Options. (Note that we used a Pixel device here, which uses the stock Android OS. For other manufacturers, the instructions might be a little different.)\n\n    1. Go to Settings.\n\n    2. Scroll down to the About Phone or About Tablet option (Figure 13-5) and select it.\n\n###### Figure 13-5. System information screen on an Android phone; select the About Phone option here\n\n    3. Look for the Build Number row and tap it seven times. (Yeah, you read that right\u2014seven!)\n\n    4. You should see a message (Figure 13-6) confirming that developer mode is enabled.\n\n###### Figure 13-6. The About Phone screen on an Android device\n\n    5. If you are using a phone, tap the back button to go back to the previous menu.\n\n    6. You should see a \"Developer options\" button, directly above the \"About phone\" or \"About tablet\" option (Figure 13-7). Tab this button to reveal the \"Developer options\" menu (Figure 13-8).\n\n###### Figure 13-7. The System information screen showing \"Developer options\" enabled\n\n###### Figure 13-8. \"Developer options\" screen on an Android device with USB debugging enabled\n\n  6. Plug the Android device into the computer via a USB cable.\n\n  7. The Android device might show a message asking to allow USB debugging. Enable \"Always allow this computer,\" and then select OK (Figure 13-9).\n\n###### Figure 13-9. Allow USB debugging on the displayed alert\n\n  8. In Android Studio, on the Debug toolbar bar (Figure 13-10), click the Run App button (the right-facing triangle).\n\n###### Figure 13-10. Debug toolbar in Android Studio\n\n  9. A window opens displaying all the available devices and emulators (Figure 13-11). Choose your device, and then select OK.\n\n###### Figure 13-11. Select the phone from the deployment target selection screen\n\n  10. The app should install and begin running on our phone.\n\n  11. The app will request permission for your camera; go ahead and grant it permission.\n\n  12. A live view of the camera should appear, along with real-time predictions of object classification, plus the number of seconds it took to make the prediction, as shown in Figure 13-12.\n\n###### Figure 13-12. The app up-and-running app, showing real-time predictions\n\nAnd there you have it! We have a basic app running on the phone that takes video frames and classifies them. It's simple and it works reasonably well.\n\nBeyond object classification, the TensorFlow Lite repository also has sample apps (iOS and Android) for many other AI problems, including the following:\n\n  * Object detection\n\n  * Pose estimation\n\n  * Gesture recognition\n\n  * Speech recognition\n\nThe great thing about having these sample apps is that with basic instructions, someone without a mobile development background can get them running on a phone. Even better, if we have a custom trained model, we can plug it into the app and see it run for our custom task.\n\nThis is great for starting out. However, things are a lot more complicated in the real world. Developers of serious real-world applications with thousands or even millions of users need to think beyond just inference\u2014like updating and distributing models, testing different versions among subsets of users, maintaining parity between iOS and Android, and ultimately reducing the engineering cost of each. Doing all of this in-house can be expensive, time consuming, and frankly unnecessary. Naturally, platforms that provide these features would be enticing. This is where ML Kit and Firebase come in.\n\n# ML Kit + Firebase\n\nML Kit is a mobile SDK launched during the 2018 Google I\/O conference. It provides convenient APIs for novices as well as advanced ML developers to do a lot of common ML tasks. By default, ML Kit comes with a generic feature set in vision and language intelligence. Table 4-1 lists some of the common ML tasks that we can do in just a few lines.\n\nTable 13-1. ML Kit built-in features **Vision** | **Language**  \n---|---\n\nObject classification\n\nObject detection and tracking\n\nPopular landmark detection\n\nText recognition\n\nFace detection\n\nBarcode detection | Language identification\n\nOn-device translation\n\nSmart replies\n\nML Kit also gives us the ability to use custom trained TensorFlow Lite models for inference. Let's take a moment to appreciate why this is so great for developers. Imagine that we were building a business card scanner. We could bring in a custom business card detection model, recognize when a business card is visible and its boundaries (to make a nice visual user interface), run the built-in text recognition, and filter out text outside of those boundaries to prevent extraneous characters. Or, consider a language learning game that can be built by pointing at objects, running an object classifier, and then using the on-device translation API to announce the labels in French. It is entirely possible to build these relatively quickly using ML Kit. Although many of these features are available in Core ML, too, ML Kit has the added advantage of being cross-platform.\n\nML Kit, though, is only one piece of the puzzle. It integrates into Google's Firebase, the mobile and web application development platform that is part of Google Cloud. Firebase offers an array of features that are necessary infrastructure for production-quality apps, such as the following:\n\n  * Push notifications\n\n  * Authentication\n\n  * Crash reporting\n\n  * Logging\n\n  * Performance monitoring\n\n  * Hosted device testing\n\n  * A\/B testing\n\n  * Model management\n\nThe last point is very pertinent to us. One of the biggest benefits that Firebase gives us is the ability to host our custom models on the cloud and download them within the app as needed. Simply copy the models over to Firebase on Google Cloud, reference the model on ML Kit inside the app, and we're good to go. The A\/B testing feature gives us the ability to show different users different versions of the same model and measure the performance across the different models.\n\n###### Note\n\nFor a lot of the built-in functionality, ML Kit optionally provides a feature to process images in the cloud, where the models are much larger compared to on-device models. These larger models obviously require more powerful hardware to run them, with the benefit of providing some improvement in accuracy and potentially a much larger taxonomy (like thousands of object classes instead of hundreds). In fact, some functionality such as the landmark recognition feature works only on the cloud.\n\nThe cloud processing option is particularly useful when we need a little bit of extra accuracy and\/or the user's phone has low processing power that prevents it from running the on-device model well.\n\n## Object Classification in ML Kit\n\nFor our previous task of object classification in real time, if we use ML Kit instead of vanilla TensorFlow Lite, we can simplify our code to just the following lines (in Kotlin):\n\n    val image = FirebaseVisionImage.fromBitmap(bitmap)\n    val detector = FirebaseVision.getInstance().visionLabelDetector\n    val result = detector.detectInImage(image).addOnSuccessListener { labels ->\n        _\/\/ Print labels_\n    }\n\n## Custom Models in ML Kit\n\nIn addition to the prebuilt models provided by ML Kit, we can also run our own custom models. These models must be in the TensorFlow Lite format. Following is a simple piece of code to load a custom model that's bundled into the app:\n\n    val customModel = FirebaseLocalModelSource.Builder(\"my_custom_model\")\n            .setAssetFilePath(\"my_custom_model.tflite\").build()\n    FirebaseModelManager.getInstance().registerLocalModelSource(customModel)\n\nNext, we specify the model's input and output configuration (for a model that takes in an RGB image of size 224x224 and gives predictions for 1,000 class names):\n\n    val IMAGE_WIDTH = 224\n    val IMAGE_HEIGHT = 224\n    val modelConfig = FirebaseModelInputOutputOptions.Builder()\n            .setInputFormat(0, FirebaseModelDataType.FLOAT32, intArrayOf(1,\n    IMAGE_WIDTH, IMAGE_HEIGHT, 3))\n            .setOutputFormat(0, FirebaseModelDataType.FLOAT32, intArrayOf(1, 1000))\n            .build()\n\nNext, we create an array of a single image and normalize each pixel to the range [\u20131,1]:\n\n    val bitmap = Bitmap.createScaledBitmap(image, IMAGE_WIDTH, IMAGE_HEIGHT, true)\n    val input = Array(1) {\n    \tArray(IMAGE_WIDTH) { Array(IMAGE_HEIGHT) { FloatArray(3) } }\n    \t}\n    for (x in 0..IMAGE_WIDTH) {\n        for (y in 0..IMAGE_HEIGHT) {\n            val pixel = bitmap.getPixel(x, y)\n            input[0][x][y][0] = (Color.red(pixel) - 127) \/ 128.0f\n            input[0][x][y][1] = (Color.green(pixel) - 127) \/ 128.0f\n            input[0][x][y][2] = (Color.blue(pixel) - 127) \/ 128.0f\n        }\n    }\n\nNow, we set up an interpreter based on our custom model:\n\n    val options = FirebaseModelOptions.Builder()\n         .setLocalModelName(\"my_custom_model\").build()\n    val interpreter = FirebaseModelInterpreter.getInstance(options)\n\nNext, we run our input batch on the interpreter:\n\n    val modelInputs = FirebaseModelInputs.Builder().add(input).build()\n    interpreter.run(modelInputs, modelConfig).addOnSuccessListener { result ->\n        \/\/ Print results\n    }\n\nYup, it's really that simple! Here, we've seen how we can bundle custom models along with the app. Sometimes, we might want the app to dynamically download the model from the cloud for reasons such as the following:\n\n  * We want to keep the default app size small on the Play Store so as to not prevent users with data usage constraints from downloading our app.\n\n  * We want to experiment with a different variety of models and pick the best one based on the available metrics.\n\n  * We want the user to have the latest and greatest model, without having to go through the whole app release process.\n\n  * The feature that needs the model might be optional, and we want to conserve space on the user's device.\n\nThis brings us to hosted models.\n\n## Hosted Models\n\nML Kit, along with Firebase, gives us the ability to upload and store our model on Google Cloud and download it from the app when needed. After the model is downloaded, it functions exactly like it would have if we had bundled the model into the app. Additionally, it provides us with the ability to push updates to the model without having to do an entire release of the app. Also, it lets us do experiments with our models to see which ones perform best in the real world. For hosted models, there are two aspects that we need to look at.\n\n### Accessing a hosted model\n\nThe following lines inform Firebase that we'd like to use the model named `my_remote_custom_model:`\n\n    val remoteModel = FirebaseCloudModelSource.Builder(\"my_remote_custom_model\")\n            .enableModelUpdates(true).build()\n    FirebaseModelManager.getInstance().registerCloudModelSource(remoteModel)\n\nNotice that we set `enableModelUpdates` to enable us to push updates to the model from the cloud to the device. We can also optionally configure the conditions under which the model would be downloaded for the first time versus every subsequent time\u2014whether the device is idle, whether it's currently charging, and whether the download is restricted to WiFi networks only.\n\nNext, we set up an interpreter much like we did with our local model:\n\n    val options = FirebaseModelOptions.Builder()\n        .setCloudModelName(\"my_remote_custom_model\").build()\n    val interpreter = FirebaseModelInterpreter.getInstance(options)\n\nAfter this point, the code to perform the prediction would look exactly the same as that for local models.\n\nNext, we discuss the other aspect to hosted models\u2014uploading the model.\n\n### Uploading a hosted model\n\nAs of this writing, Firebase supports only models hosted on GCP. In this section, we walk through the simple process of creating, uploading, and storing a hosted model. This subsection presumes that we already have an existing GCP account.\n\nThe following lists the steps that we need to take to get a model hosted on the cloud:\n\n  1. Go to _https:\/\/console.firebase.google.com_. Select an existing project or add a new one (Figure 13-13).\n\n###### Figure 13-13. Home page of Google Cloud Firebase\n\n  2. On the Project Overview screen, create an Android app (Figure 13-14).\n\n###### Figure 13-14. The Project Overview screen on Google Cloud Firebase\n\n  3. Use the app ID from the project in Android Studio (Figure 13-15).\n\n###### Figure 13-15. App creation screen on Firebase\n\n  4. After clicking \"Register app,\" download the configuration file. This configuration file gives the necessary credentials to the app for it to access our cloud account. Add the configuration file and the Firebase SDK to the Android app as shown on the app creation page.\n\n  5. In the ML Kit section, select Get Started, and then select \"Add custom model\" (Figure 13-16).\n\n###### Figure 13-16. The ML Kit custom models tab\n\n  6. In the name field, enter `my_remote_custom_model` to match the name in the code.\n\n  7. Upload the model file from your computer (Figure 13-17).\n\n###### Figure 13-17. Uploading a TensorFlow Lite model file to Firebase\n\n  8. Tap the \"Publish\" button after the file upload completes.\n\nThat's it! Our model is now ready to be accessed and used from the app dynamically. Next, we examine how we can do A\/B testing between models using Firebase.\n\n## A\/B Testing Hosted Models\n\nLet's take the scenario in which we had a version 1 model named `my_model_v1` to start off with and deployed it to our users. After some usage by them, we obtained more data that we were able to train on. The result of this training was `my_model_v2` (Figure 13-18). We want to assess whether this new version would give us better results. This is where A\/B testing comes in.\n\n###### Figure 13-18. Currently uploaded custom models to Firebase\n\nWidely used by industry, A\/B testing is a statistical hypothesis testing technique, which answers the question \"Is B better than A?\" Here A and B could be anything of the same kind: content on a website, design elements on a phone app, or even deep learning models. A\/B testing is a really useful feature for us when actively developing a model and discovering how our users respond to different iterations of the model.\n\nUsers have been using `my_model_v1` for some time now, and we'd like to see whether the v2 iteration has our users going gaga. We'd like to start slow; maybe just 10% of our users should get v2. For that, we can set up an A\/B testing experiment as follows:\n\n  1. In Firebase, click the A\/B Testing section, and then select \"Create experiment\" (Figure 13-19).\n\n###### Figure 13-19. A\/B testing screen in Firebase where we can create an experiment\n\n  2. Select the Remote Config option.\n\n  3. In the Basics section, in the \"Experiment name\" box, enter the experiment name and an optional description (Figure 13-20), and then click Next.\n\n###### Figure 13-20. The Basics section of the screen to create a remote configuration experiment\n\n  4. In the Targeting section that opens, from the \"Target users\" drop-down menu, select our app and enter the percentage of target users (Figure 13-21).\n\n###### Figure 13-21. The Targeting section of the Remote Config screen\n\n  5. Select a goal metric that makes sense. We discuss this in a little more detail in the next section.\n\n  6. In the variants section (Figure 13-22), create a new parameter called `model_name` that reflects the name of the model a particular user would use. The control group gets the default model, which is `my_model_v1`. We also create an additional variant with the name `my_model_v2,` which goes to 10% of the users.\n\n###### Figure 13-22. The Variants section of the Remote Config screen\n\n  7. Select Review and then select \"Start experiment.\" Over time, we can increase the distribution of users using the variant.\n\nTa-da! Now we have our experiment up and running.\n\n### Measuring an experiment\n\nWhere do we go from here? We want to give our experiment some time to see how it performs. Depending on the type of experiment, we might want to give it a few days to a few weeks. The success of the experiments can be determined by any number of criteria. Google provides a few metrics out of the box that we can use, as shown in Figure 13-23.\n\n###### Figure 13-23. Analytics available when setting up an A\/B testing experiment\n\nSuppose that we want to maximize our estimated total revenue\u2014after all, we all want to become rich like Jian-Yang. We would measure our revenue from the users who have the experiment turned on against the baseline; that is, the users who do not have the experiment turned on. If our revenue per user increased against the baseline, we would consider the experiment a success. Conversely, we would conclude the opposite if there were no increase\/decrease in revenue per user. For successful experiments, we want to slowly roll them out to all users. At that point, it ceases to be an experiment and it \"graduates\" to become a core offering.\n\n## Using the Experiment in Code\n\nNow that we have set up an experiment, let's see how to include it in our application using code. To use the appropriate model within our code, we simply need to access the remote configuration object (`remoteConfig`) and get the model name from it. The model name that we get from the remote configuration object will depend on whether the user is included in the experiment. The following lines of code accomplish that:\n\n    val remoteConfig = FirebaseRemoteConfig.getInstance()\n    remoteConfig.fetch()\n    val modelName = remoteConfig.getString(\"current_best_model\")\n    val remoteModel = FirebaseCloudModelSource.Builder(modelName)\n            .enableModelUpdates(true).build()\n    FirebaseModelManager.getInstance().registerCloudModelSource(remoteModel)\n\nThe rest of the code to perform the prediction remains exactly the same as in the previous sections. Our app is now ready to use the correct model as dictated by our experiment.\n\n# TensorFlow Lite on iOS\n\nThe previous chapters demonstrate how easy it is to use Apple's Core ML on iOS. We can simply drag and drop a model into Xcode and start making inferences with just a couple of lines of code. In contrast, even looking at basic iOS examples such as those on TensorFlow Lite's repository, it becomes evident that it needs a significant amount of boilerplate code to get even the most basic of apps running. In contrast, TensorFlow Lite used in conjunction with ML Kit is a rather pleasant experience. In addition to being able to use a clean and concise API, we get all the features we detailed earlier in this chapter\u2014remotely downloading and updating models, model A\/B testing, and cloud fallback for processing. All of this without having to do much extra work. A developer writing a deep learning app for both iOS and Android might consider using ML Kit as a way to \"build once, use everywhere.\"\n\n# Performance Optimizations\n\nIn Chapter 6, we explored quantization and pruning, mostly from a theoretical standpoint. Let's see them up close from TensorFlow Lite's perspective and the tools to achieve them.\n\n## Quantizing with TensorFlow Lite Converter\n\nFor iOS, Apple provides `quantization_utils` in its Core ML Tools package. And for TensorFlow Lite, the equivalent is the already built-in `tflite_convert` tool, which we used earlier in this chapter. On the command line, we can specify the input file, the model graph, the data type to convert to, and the names of the input and output (which can be inspected using Netron, as shown in Chapter 11). Going from 32-bit to 8-bit integer representation means a four times smaller model, with relatively little loss in accuracy.\n\n    $ tflite_convert \\\n      --output_file=quantized-model.tflite \\\n      --graph_def_file=\/tmp\/some-graph.pb \\\n      --inference_type=QUANTIZED_UINT8 \\\n      --input_arrays=input \\\n      --output_arrays=MobilenetV1\/Predictions\/Reshape_1 \\\n      --mean_values=128 \\\n      --std_dev_values=127\n\nWhen it's finished, this command should give us the `quantized-model.tflite` model.\n\n## TensorFlow Model Optimization Toolkit\n\nThe TensorFlow Lite converter is the simplest way to get our models quantized. It's worth noting that the converter quantizes the models post-training. Due to the decrease in representation power, there can be a tiny but noticeable loss in accuracy. Could we do better? _Quantization-aware training_ , as the name suggests, accounts for the effects of quantization during training time and attempts to compensate and minimize the losses that would have happened in post-training quantization.\n\nAlthough both forms of quantization offer 75% reduction in the size of the model, experiments have shown the following:\n\n  * In the case of MobileNetV2, compared to an eight-point loss in accuracy with post-training quantization, quantization-aware training yielded only a one-point loss.\n\n  * For InceptionV3, quantization-aware training yielded a whopping 52% reduction in latency, compared to 25% reduction with post-training quantization.\n\n###### Note\n\nIt is worth noting that these accuracy metrics are on the 1,000 class ImageNet test set. Most problems have less complexity with a smaller number of classes. Post-training quantization should result in a smaller loss on such simpler problems.\n\nQuantization-aware training can be implemented with the TensorFlow Model Optimization Toolkit. This toolkit additional offers an increasing array of tools for model compression, including pruning. Additionally, the TensorFlow Lite model repository already offers these prequantized models using this technique. Table 13-2 lists the effects of various quantization strategies.\n\nTable 13-2. Effects of different quantization strategies (8-bit) on models (source: TensorFlow Lite Model optimization documentation) Model | MobileNet | MobileNetV2 | InceptionV3  \n---|---|---|---  \n**Top-1 accuracy** | **Original** | 0.709 | 0.719 | 0.78  \n**Post-training quantized** | 0.657 | 0.637 | 0.772  \n**Quantization-aware training** | 0.7 | 0.709 | 0.775  \n**Latency (ms)** | **Original** | 124 | 89 | 1130  \n**Post-training quantized** | 112 | 98 | 845  \n**Quantization-aware training** | 64 | 54 | 543  \n**Size (MB)** | **Original** | 16.9 | 14 | 95.7  \n**Optimized** | 4.3 | 3.6 | 23.9\n\n# Fritz\n\nAs we have seen so far, the primary purpose of Core ML and TensorFlow Lite is to serve fast mobile inferences. Have a model, plug it into the app, and run inferences. Then came ML Kit, which apart from its built-in AI capabilities, made it easier to deploy models and monitor our custom models (with Firebase). Taking a step back to look at the end-to-end pipeline, we had training, conversion to mobile format, optimizing its speed, deploying to users, monitoring the performance, keeping track of model versions. These are several steps spread across many tools. And there is a good chance that there is no single owner of the entire pipeline. This is because oftentimes these tools require some level of familiarity (e.g., having to deploy a model to users might be outside a data scientist's comfort zone). Fritz, a Boston-based startup, is attempting to bring down these barriers and make the full cycle from model development to deployment even more straightforward for both data scientists and mobile developers.\n\nFritz offers an end-to-end solution for mobile AI development that includes the following noteworthy features:\n\n  * Ability to deploy models directly to user devices from Keras after training completes using callbacks.\n\n  * Ability to benchmark a model directly from the computer without having to deploy it on a phone. The following code demonstrates this:\n\n    $ fritz model benchmark <path to keras model.h5>\n     ...\n     ------------------------\n     Fritz Model Grade Report\n     ------------------------\n\n     Core ML Compatible:              True\n     Predicted Runtime (iPhone X):    31.4 ms (31.9 fps)\n     Total MFLOPS:                    686.90\n     Total Parameters:                1,258,580\n     Fritz Version ID:                <Version UID>\n\n  * Ability to encrypt models so that our intellectual property can be protected from theft from a device by nefarious actors.\n\n  * Ability to implement many advanced computer-vision algorithms with just a few lines of code. The SDK comes with ready-to-use, mobile-friendly implementations of these algorithms, which run at high frame rates. For example, image segmentation, style transfer, object detection, and pose estimation. Figure 13-24 shows benchmarks of object detection run on various iOS devices.\n\n    let poseModel = FritzVisionPoseModel()\n    guard let poseResult = try? poseModel.predict(image) else { return }\n    let imageWithPose = poseResult.drawPose() _\/\/ Overlays pose on input._\n\n###### Figure 13-24. Performance of Fritz SDK's object detection functionality on different mobile devices, relative to the iPhone X\n\n  * Ability to customize the prebuilt models to custom datasets by retraining using their Jupyter notebooks. It's worth noting that this can be a difficult problem (even for professional data scientists) that is simplified significantly because the developer has only to ensure that the data is in the right format.\n\n  * Ability to manage all model versions from the command line.\n\n  * Ability to test out mobile readiness of our models using Fritz' open source app called Heartbeat (also available on iOS\/Android app stores). Assuming that we have a model ready without much mobile know how, we can clone the app, swap the existing model with our own, and get to see it run on the phone.\n\n  * A vibrant community of contributors blogging about the latest in mobile AI on _heartbeat.fritz.ai_.\n\n##### From the Creator's Desk\n\nBy Dan Abdinoor, CEO and cofounder of Fritz\n\nOur motivation to build Fritz came from two somewhat contradictory feelings\u2014inspiration and frustration. The inspiration came from seeing how many previously impossible problems were being solved with machine learning. The combination of data, compute, and model architectures were all starting to click. At the same time, we found ourselves frustrated by the lack of resources for deploying and maintaining machine learning anywhere other than the cloud. For reasons of performance, privacy, and cost, it makes more sense to shift inference to devices where data is being collected, as opposed to moving all that data to a cloud environment, processing it, and then sending predictions back down to the devices. This led us to build Fritz as a platform for developers to collect, train, deploy, and manage machine learning models for edge devices\u2014like mobile phones.\n\nIn many ways, the process of building a solution with machine learning is like a race with hurdles. We want to lower or remove those hurdles so that engineers, even those without deep learning or mobile expertise, can use what we built to make great mobile applications. One can get started quickly with efficient pretrained models, or custom train a new one. Our tools should make it easier to deploy models, understand how they are performing, and keep them updated and protected. The results of all this, in part, are models that perform well on edge devices.\n\nMaking it easier to implement machine learning features in mobile apps will lead to more powerful, personalized experiences that retain current users, attract new ones, and shift our thinking about what's possible. Ultimately, we believe machine learning is the most fundamental change to computer science in decades, and we want to help build, educate, and evolve the community.\n\n# A Holistic Look at the Mobile AI App Development Cycle\n\nUntil this point in the book, we've looked at a bunch of techniques and technologies that enable us to perform individual tasks in the mobile AI development cycle. Here, we tie everything together by exploring the kind of questions that come up throughout this life cycle. Figure 13-25 provides a broad overview of the development cycle.\n\n###### Figure 13-25. Mobile AI app development life cycle\n\n## How Do I Collect Initial Data?\n\nWe can apply a few different strategies to accomplish this:\n\n  * Find the object of interest and manually take photos with different angles, lighting, environments, framing, and so on.\n\n  * Scrape from the internet using browser extensions like Fatkun (Chapter 12).\n\n  * Find existing datasets (Google Dataset Search). For example, Food-101 for dishes.\n\n  * Synthesize your own dataset.\n\n    1. Place the object (foreground) in front of a green screen (background) and take a photograph of it. Replace the background with random images to synthesize a large dataset, while zooming, cropping, and rotating to create potentially hundreds of images.\n\n    2. If a green screen is not possible, segment (i.e., cut) the object out of the background from existing real-world images and repeat the previous step in order to build a robust, diverse dataset. It's important to note that, in points (a) and (b), there needs to be sufficient diversity in the foreground; otherwise, the network might overlearn that one example instead of understanding the object.\n\n    3. Find a realistic 3D model of the object of interest and place it in realistic environments using a 3D framework such as Unity. Adjust the lighting and camera position, zoom, and rotation to take snapshots of this object from many angles. We explored companies such as AI.Reverie and CVEDIA in Chapter 7 that work in this space. We use photo-realistic simulators to train models for the self-driving chapters (Chapter 16 and Chapter 17).\n\n## How Do I Label My Data?\n\nMost of the steps in the previous answer should have already given you labels for the data. For unlabeled collections, use labeling tools such as Supervisely, Labelbox, and Diffgram. For really large datasets in which annotating data by yourself is not feasible, socially responsible labeling services such as Digital Data Divide, iMerit, and Samasource, which provide income opportunities to disadvantaged populations, might be a good option.\n\n## How Do I Train My Model?\n\nThe following are the two broad approaches to train a model:\n\n  * With code: Use Keras and TensorFlow (Chapter 3).\n\n  * Without code: Use custom classifier services such as Google's Auto ML, Microsoft's CustomVision.ai, and Clarifai (benchmarked in Chapter 8), or Apple's ecosystem-only Create ML (Chapter 12).\n\n## How Do I Convert the Model to a Mobile-Friendly Format?\n\nThe following are a few different ways to convert a model to a mobile-compatible format:\n\n  * Use Core ML Tools (Apple only).\n\n  * Use TensorFlow Lite Converter for iOS and Android.\n\n  * Alternatively, use Fritz for the end-to-end pipeline.\n\n## How Do I Make my Model Performant?\n\nHere are some techniques to make a model performant:\n\n  * Start with an efficient model such as the MobileNet family, or even better, EfficientNet.\n\n  * Make the model smaller in size by quantizing and pruning it to improve loading and inference times, while keeping the model accuracy relatively intact (Chapter 6, Chapter 11, and Chapter 13). Expect up to a 75% reduction in size with little loss in accuracy.\n\n  * Use Core ML Tools for the Apple ecosystem or TensorFlow Model Optimization Kit for both iOS and Android.\n\n## How Do I Build a Great UX for My Users?\n\nObviously, that's going to depend on the kind of problem you are trying to solve. But a general guideline is to strike the proper balance between interactivity, performance, and resource usage (memory, CPU, battery, etc.). And, of course, an intelligent feedback mechanism that enables effortless data collection and feedback helps. Gamifying this experience would take it to an entirely new level.\n\nIn the case of the food classifier app, after the user takes a picture, our UI would show a list of top five candidate predictions (in decreasing order of confidence) for a given photo. For example, if the user takes a picture of a pizza, the candidate predictions shown on screen could be \"pie - 75%,\" \"pizza - 15%,\" \"naan - 6%,\" \"bread - 3%,\" \"casserole - 1%.\" In this case, the user would have selected the second prediction. For a perfect model, the user would always select the first prediction. Because our model is not perfect, the rank of the prediction the user selects becomes a signal that will help improve the model in the future. In the absolute worst case, in which none of the predictions were correct, the user should have a way to manually label the data. Providing an autosuggest feature during manual entry will help keep clean labels in the dataset.\n\nAn even better experience could be one in which the user never needs to click a picture. Rather, the predictions are available in real time.\n\n## How Do I Make the Model Available to My Users?\n\nThe following are some ways to deploy models to users:\n\n  * Bundle the model into the app binary and release it on the app stores.\n\n  * Alternatively, host the model in the cloud and have the app download the model, as necessary.\n\n  * Use a model management service like Fritz or Firebase (integrated with ML Kit).\n\n## How Do I Measure the Success of My Model?\n\nThe very first step is to determine the success criteria. Consider the following examples:\n\n  * \"My model should run inferences in under 200 ms at the 90th percentile.\"\n\n  * \"Users who use this model open the app every day.\"\n\n  * In the food classifier app, a success metric could be something along the lines of \"80% of users selected the first prediction in the list of predictions.\"\n\nThese success criteria are not set in stone and should continue to evolve over time. It's very essential to be data driven. When you have a new model version, run A\/B tests on a subset of users and evaluate the success criteria against that version to determine whether it's an improvement over the previous version.\n\n## How Do I Improve My Model?\n\nHere are some ways to improve the quality of our models:\n\n  * Collect feedback on individual predictions from users: what was correct, and, more important, what was incorrect. Feed these images along with the corresponding labels as input for the next model training cycle. Figure 13-26 illustrates this.\n\n  * For users who have explicitly opted in, collect frames automatically whenever the prediction confidence is low. Manually label those frames and feed them into the next training cycle.\n\n###### Figure 13-26. The feedback cycle of an incorrect prediction, generating more training data, leading to an improved model\n\n## How Do I Update the Model on My Users' Phones?\n\nHere are some ways to update the model on your users' phones:\n\n  * Bundle the new model into the next app release. This is slow and inflexible.\n\n  * Host the new model on the cloud and force the apps out in the world to download the new model. Prefer to do this when the user is on WiFi.\n\n  * Use a model management system such as Firebase (in conjunction with ML Kit) or Fritz to automate a lot of the grunt work involved.\n\nWith all of these questions answered, let's appreciate the beauty of how this app improves on its own.\n\n# The Self-Evolving Model\n\nFor applications that need only mature pretrained models, our job is already done. Just plug the model into an app and we're ready to go. For custom-trained models that rely especially on scarce training data, we can get the user involved in building a self-improving, ever-evolving model.\n\nAt the most fundamental level, each time a user uses the app, they're providing the necessary feedback (image + label) to improve the model further, as illustrated in Figure 13-27.\n\n###### Figure 13-27. The self-evolving model cycle\n\nJust like a university student must graduate multiple years before getting ready to step into the real world, a model needs to go through phases of development before it reaches its final users. The following is a popular model of phasing releases in software, which is also a useful guideline for releasing AI models.\n\n1. Dev\n\nThis is the initial phase of development in which the developers of the app are the only users of it. Data accumulation is slow, the experience is very buggy, and model predictions can be quite unreliable.\n\n2. Alpha\n\nAfter the app is ready to be tested by a few users beyond the developers, it is now considered to be in alpha. User feedback is extremely crucial here because it would serve to improve the app experience as well provide data at a greater scale to the pipeline. The experience is not nearly as buggy, and the model predictions are a little more reliable here. In some organizations, this phase is also known as _dogfooding_ (i.e., internal application testing by employees).\n\n3. Beta\n\nAn app in beta has significantly more users than in alpha. User feedback is also quite important here. The data collection rate is at an even larger scale. Many different devices are out in the real world collecting data to help improve the model rapidly. The experience is a lot more stable, and the model predictions are a lot more reliable, thanks to the alpha users. Beta apps tend to be hosted on Apple's TestFlight for iOS and Google Play Console for Android. Third-party services such as HockeyApp and TestFairy are also popular for hosting beta programs.\n\n4. Prod\n\nAn app in production is stable and is widely available. Although the data is coming in at large rates, the model is fairly stable at this point, and it does not learn as much. However, as real-world usage keeps growing, it's able to get better at edge cases it might not have been seen before in the first three phases. When the model is mature enough, small version improvements can be alpha\/beta tested or A\/B tested on subsets of the production audience.\n\nAlthough many data scientists assume the data to be available before starting development, people in the mobile AI world might need to bootstrap on a small dataset and improve their system incrementally.\n\nWith this self-evolving system for the Shazam for Food app, the most difficult choice that the developers must make should be the restaurants they visit to collect their seed data.\n\n# Case Studies\n\nLet's look at a few interesting examples that show how what we have learned so far is applied in the industry.\n\n## Lose It!\n\nWe need to confess here. The app we were building all along in this chapter already exists. Erlich Bachman's dream was already built a year before by a Boston-based company named Fit Now. Lose It! (Figure 13-28) claims to have helped 30 million users lose 85 million pounds by tracking what they eat. Users can point their phone's camera at barcodes, nutritional labels, and the food that they're just about to eat to track calories and macros they're consuming for every single meal.\n\nThe company first implemented its food scanning system with a cloud-based algorithm. Due to the resource demands and network delay, the experience could not be real time. To improve the experience, the Fit Now team migrated its models to TensorFlow Lite to optimize them for mobile and used ML Kit to seamlessly deploy to its millions of users. With a constant feedback loop, model updates, and A\/B testing, the app keeps improving essentially on its own.\n\n###### Figure 13-28. Snap It feature from Lose It! showing multiple suggestions for a scanned food item\n\n## Portrait Mode on Pixel 3 Phones\n\nOne of the key visual concepts that separates professional photographers from the amateur wannabes is bokeh, or blurring of the background behind the subject of the image. (Not to be confused with the Python visualization tool of the same name.) The closer the camera is located to the subject, the more the background appears blurred. With the help of professional lenses with a low _f_ -stop number (which often run into the thousands of dollars), one can produce spectacular blurs.\n\nBut if you don't have that kind of money, AI can help. Google's Pixel 3 offers a \"Portrait\" mode that creates the bokeh effect. Essentially, it uses a CNN to estimate the depth of each pixel in a scene, and determine which pixels belong to the foreground and the background. The background pixels are then blurred to varying intensities to create the bokeh effect, as demonstrated in Figure 13-29.\n\nDepth estimation is a compute-intensive task so running it fast is essential. Tensorflow Lite with a GPU backend comes to the rescue, with a roughly 3.5 times speedup over a CPU backend.\n\nGoogle accomplished this by training a neural network that specializes in depth estimation. It used a rig with multiple phone cameras (which it termed \"Frankenphone\") to take pictures of the same scene from slightly different viewpoints belonging to each camera. It then used the _parallax effect_ experienced by each pixel to precisely determine the depth of each pixel. This depth map was then fed into the CNN along with the images.\n\n###### Figure 13-29. Portrait effect on Pixel 3, which achieves separation between foreground and background using blurring\n\n## Speaker Recognition by Alibaba\n\nImagine speaking \"Hey Siri\" or \"Hey Google\" into a friend's phone, followed by the magical words \"read the last text.\" From a privacy standpoint, it would be terrifying if this worked. Obviously, most mobile operating systems today ask us to unlock the phone first before proceeding. This creates a not-so-smooth experience for the owner of the phone.\n\nAlibaba Machine Intelligence Labs solved this problem using the following approach. First, by converting speech into images of spectrograms (extracting audio features with mel-frequency cepstrum algorithm), training a CNN (using transfer learning), and finally deploying it to devices using Tensorflow Lite. That's right: CNNs can do more than just computer vision!\n\nBy recognizing speakers, they are able to personalize the content on devices with many users in a family (e.g., in its \"Talk to your TV\" feature on its Netflix-like TV app). Additionally, by recognizing whether the sound is a human voice, it is able to isolate the spoken commands to transcribe from background noise. To make processing faster with Tensorflow Lite, engineers kept the `USE_NEON` flag to accelerate instruction sets for ARM-based CPUs. The team reported a speed increase of four times with this optimization.\n\n## Face Contours in ML Kit\n\nWant to quickly build a Snapchat face filter\u2013like functionality without getting a Ph.D.? ML Kit also provides the goods to do that\u2014an API to identify facial contours, a set of points (in x and y coordinates) that follow the shape of facial features in the input image. In total, 133 points map to the various facial contours, including 16 points representing each eye, while 36 points map to the oval shape around the face, as demonstrated in Figure 13-30.\n\nInternally, ML Kit is running a deep learning model using TensorFlow Lite. In experiments with the new TensorFlow Lite GPU backend, Google found an increase in speed of four times on Pixel 3 and Samsung Galaxy S9, and a six-times speedup on iPhone 7, compared to the previous CPU backend. The effect of this is that we can precisely place a hat or sunglasses on a face in real time.\n\n###### Figure 13-30. 133 Face contour points identified by ML Kit (image source)\n\n## Real-Time Video Segmentation in YouTube Stories\n\nGreen screens are staple equipment for anyone in the video production industry. By choosing a color not matching the subject, the background can be changed during postproduction using the _chroma keying_ technique. Obviously, this requires expensive software and powerful machines for postproduction. Many YouTubers have the same needs but might not have the budget. And now they have a solution within the YouTube Stories app\u2014a real-time video segmentation option implemented over TensorFlow Lite.\n\nThe key requirements here are first, to run semantic segmentation fast (more than 30 frames per second), and second, be temporally consistent; for instance, achieving smooth frame-to-frame temporal continuity on the edges. If we attempt to run semantic segmentation over multiple frames, we'll quickly notice the edges of many of the segmented masks bouncing around. The key trick employed here is to pass the segmented masks of the person's face to the next frame as a prior. Because we traditionally work on three channels (RGB), the trick is to add a fourth channel, which essentially is the output of the previous frame, as illustrated in Figure 13-31.\n\n###### Figure 13-31. An input image (left) is broken down into its three components layers (R, G, B). The output mask of the previous frame is then appended with these components (image source)\n\nWith other optimizations to the structure of the base CNN, the YouTube team was able to run the system at more than 100 frames per second on an iPhone 7, and faster than 40 frames per second on a Pixel 2. Additionally, comparing the CPU versus GPU backend of TensorFlow Lite, an increase in speed of 18 times is observed when choosing the GPU backend over the CPU backend (much more acceleration compared to the usual two to seven times for other semantic segmentation tasks for images only).\n\n# Summary\n\nIn this chapter, we discussed the architecture of TensorFlow Lite, as it relates to Android's Neural Network API. We then worked on getting a simple object recognition app running on an Android device. We covered Google's ML Kit and talked about the reasons why you might want to use it. We additionally went through how to convert our TensorFlow models to TensorFlow Lite format so that they could be used within an Android device. We finally discussed a couple of examples of how TensorFlow Lite is being used to solve real-world problems.\n\nIn the next chapter, we look at how we can use real-time deep learning to develop an interactive, real-world application.\n\n# Chapter 14. Building the Purrfect Cat Locator App with TensorFlow Object Detection API\n\nBob gets frequent visits from a neighborhood feral cat. Those visits result in less than pleasant outcomes. You see, Bob has a nice fairly large garden that he tends to with great care. However, this furry little fella visits his garden every night and starts chewing on a bunch of plants. Months of hard work gets destroyed overnight. Clearly unhappy with this situation, Bob is keen on doing something about it.\n\nChanneling his inner Ace Ventura (Pet Detective), he tries to stay awake at night to drive the cat away, but clearly, that's not sustainable in the long term. After all, Red Bull has its limits. Reaching a breaking point, he decides to use the nuclear option: use AI in conjunction with his sprinkler system to literally \"turn the hose\" on the cat.\n\nIn his sprawling garden, he sets up a camera that can track the motion of the cat and turn on the nearest set of sprinklers to scare the cat away. He has an old phone lying around, and all he needs is a way to determine the cat's location in real time, so he can control accordingly which sprinkler to trigger.\n\nHopefully, after a few unwelcome baths, as demonstrated in Figure 14-1, the cat would be disincentivized from raiding Bob's garden.\n\n###### Figure 14-1. Building an AI cat sprinkler system, like this Havahart Spray Away Motion Detector\n\n###### Tip\n\nKeen observers might have noticed that we're attempting to find the position of the cat in 2D space, whereas the cat itself exists in 3D space. We can make assumptions to simplify the problem of locating the cat in real-world coordinates by assuming that the camera will always be in a fixed position. To determine the distance of the cat, we can use the size of an average cat to take measurements of its apparent size on a picture at regular intervals of distance from the camera. For example, an average cat might appear to have a bounding box height of 300 pixels at a distance of two feet from the camera, whereas it might occupy 250 pixels when it's three feet away from the camera lens.\n\nIn this chapter, we help Bob make his cat detector. (Note to animal-rights advocates: no cats were harmed in the making of this chapter.) We answer the following questions along the way:\n\n  * What are the different kinds of computer-vision tasks?\n\n  * How do I reuse a model pretrained for an existing class of objects?\n\n  * Can I train an object detector model without writing any code?\n\n  * I want more granular control for greater accuracy and speed. How do I train a custom object detector?\n\n# Types of Computer-Vision Tasks\n\nIn most of the previous chapters, we've essentially looked at one kind of problem: object classification. In that, we found out whether an image contained objects of a certain class. In Bob's scenario, in addition to knowing whether the camera is seeing a cat, it is necessary to know exactly where the cat is in order to trigger the nearest sprinkler. Determining the location of an object within an image is a different type of problem: _object detection_. Before we jump into object detection, let's look at the variety of frequent computer-vision tasks and the questions they purport to answer.\n\n## Classification\n\nWe have looked at several examples of classification throughout this book. Simply put, the task of classification assigns an image with the class(es) of the object(s) present in the image. It answers the question: \"Is there an object of class X in the image?\" Classification is one of the most fundamental computer-vision tasks. In an image, there might be objects of more than one class\u2014called multiclass classification. The label in this case for one image would be a list of all object classes that the image contains.\n\n## Localization\n\nThe downside of classification is that it does not tell us where in the image a particular object is, how big or small it is, or how many objects there are. It only tells us that an object of a particular class exists somewhere within it. _Localization_ , otherwise known as _classification with localization_ , can inform us which classes are in an image and where they are located within the image. The keyword here is _classes_ , as opposed to individual objects, because localization gives us only one bounding box per class. Just like classification, it cannot answer the question \"How many objects of class X are in this image?\"\n\nAs you will see momentarily, localization will work correctly only when it is guaranteed that there is a single instance of each class. If there is more than one instance of a class, one bounding box might encompass some or all objects of that class (depending upon the probabilities). But just one bounding box per class seems rather restrictive, doesn't it?\n\n## Detection\n\nWhen we have multiple objects belonging to multiple classes in the same image, localization will not suffice. Object detection would give us a bounding rectangle for every instance of a class, for all classes. It also informs us as to the class of the object within each of the detected bounding boxes. For example, in the case of a self-driving car, object detection would return one bounding box per car, person, lamp post, and so on that the car sees. Because of this ability, we also can use it in applications that need to count. For example, counting the number of people in a crowd.\n\nUnlike localization, detection does not have a restriction on the number of instances of a class in an image. This is why it is used in real-world applications.\n\n###### Warning\n\nOften people use the term \"object localization\" when, really, they mean \"object detection.\" It's important to note the distinction between the two.\n\n## Segmentation\n\nObject segmentation is the task of assigning a class label to individual pixels throughout an image. A children's coloring book where we use crayons to color various objects is a real-world analogy. Considering each pixel in the image is being assigned a class, it is a highly compute-intensive task. Although object localization and detection give bounding boxes, segmentation produces groups of pixels\u2014also known as _masks_. Segmentation is obviously much more precise about the boundaries of an object compared to detection. For example, a cosmetic app that changes a user's hair color in real-time would use segmentation. Based on the types of masks, segmentation comes in different flavors.\n\n### Semantic segmentation\n\nGiven an image, _semantic segmentation_ assigns one mask per class. If there are multiple objects of the same class, all of them are assigned to the same mask. There is no differentiation between the instances of the same class. Just like localization, semantic segmentation cannot count.\n\n### Instance-level segmentation\n\nGiven an image, _instance-level segmentation_ identifies the area occupied by each instance of each class. Different instances of the same class will be segmented uniquely.\n\nTable 14-1 lists the different computer-vision tasks.\n\nTable 14-1. Types of computer-vision tasks illustrated using image ID  from the MS COCO dataset **Task** | **Image** | **In lay terms** | **Outputs**  \n---|---|---|---  \nObject classification |\n\n| Is there a sheep in this image? | Class probabilities  \nObject localization |\n\n| Is there \"a\" sheep in the image and where is it? | Bounding box and class probability\n\nObject\n\ndetection |\n\n| What and where are \"all\" the objects in this image? | Bounding boxes, class probabilities, class ID prediction  \nSemantic segmentation |\n\n| Which pixels in this image belong to different classes; e.g., the class \"sheep,\" \"dog,\" \"person\"? | One mask per class  \nInstance-level segmentation |\n\n| Which pixels in this image belong to each instance of each class; e.g., \"sheep,\" \"dog,\" \"person\"? | One mask per instance of each class\n\n# Approaches to Object Detection\n\nDepending on the scenario, your needs, and technical know-how, there are a few ways to go about getting the object detection functionality in your applications. Table 14-2 takes a look at some approaches.\n\nTable 14-2. Different approaches to object detection and their trade-offs **** | **Custom classes** | **Effort** | **Cloud or local** | **Pros and cons**  \n---|---|---|---|---  \nCloud-based object detection APIs | No | <5 minutes | Cloud only |\n\n\\+ Thousands of classes\n\n\\+ State of the art\n\n\\+ Fast setup\n\n\\+ Scalable API\n\n\u2013 No customization\n\n\u2013 Latency due to network\n\nPretrained model | No | <15 minutes | Local |\n\n\\+ ~100 classes\n\n\\+ Choice of model to fit speed and accuracy needs\n\n\\+ Can run on the edge\n\n\u2013 No customization\n\nCloud-based model training | Yes | <20 minutes | Cloud and local |\n\n\\+ GUI-based, no coding for training\n\n\\+ Customizability\n\n\\+ Choice between powerful model (for cloud) and efficient model (for edge)\n\n\\+ Scalable API\n\n\u2013 Uses transfer learning, might not allow full-model retraining\n\nCustom training a model | Yes | 4 hours to 2 days | Local |\n\n\\+ Highly customizable\n\n\\+ Largest variety of models available for different speed and accuracy requirements\n\n\u2013 Time consuming and highly involved\n\nIt's worth noting that the options marked \"Cloud\" are already engineered for scalability. At the same time, the options marked as being available for \"Local\" inferences can also be deployed on the cloud. To achieve high scale, we can host them on the cloud similar to how we did so for the classification models in Chapter 9.\n\n# Invoking Prebuilt Cloud-Based Object Detection APIs\n\nAs we already saw in Chapter 8, calling cloud-based APIs is relatively straightforward. The process involves setting up an account, getting an API key, reading through the documentation, and writing a REST API client. To make this process easier, many of these services provide Python-based (and other language) SDKs. The main cloud-based object detection API providers are Google's Vision AI (Figure 14-2) and Microsoft's Cognitive Services.\n\nThe main advantages of using cloud-based APIs are that they are scalable out of the box and they support recognition of thousands of classes of objects. These cloud giants built their models using large proprietary datasets that are constantly growing, resulting in a very rich taxonomy. Considering that the number of classes in the largest public datasets with object detection labels is usually only a few hundred, there exists no nonproprietary solution out there that is capable of detecting thousands of classes.\n\nThe main limitation of cloud-based APIs, of course, is the latency due to network requests. It's not possible to power real-time experiences with this kind of lag. In the next section, we look at how to power real-time experiences without any data or training.\n\n###### Figure 14-2. Running a familiar photo on Google's Vision AI API to obtain object detection results\n\n# Reusing a Pretrained Model\n\nIn this section, we explore how easy it is to get an object detector running on a mobile phone with a pretrained model. Hopefully, you are already comfortable with running a neural network on mobile from some of the previous chapters. The code to run object detection models on iOS and Android is referenced on the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_) at _code\/chapter-14_. Changing functionality is simply a matter of replacing the existing _.tflite_ model file with a new one.\n\nIn the previous chapters, we often used MobileNet to do classification tasks. Although the MobileNet family, including MobileNetV2 (2018) and MobileNetV3 (2019) by itself are classification networks only, they can act as a backbone for object detection architectures like SSD MobileNetV2, which we use in this chapter.\n\n## Obtaining the Model\n\nThe best way to uncover the magic behind the scenes is to take a deeper look at a model. We'd need to get our hands on the TensorFlow Models repository, which features models for more than 50 deep learning tasks, including audio classification, text summarization, and object detection. This repository contains models as well as utility scripts that we use throughout this chapter. Without further ado, let's clone the repository on our machines:\n\n    $ git clone https:\/\/github.com\/tensorflow\/models.git && cd models\/research\n\nThen, we update the `PYTHONPATH` to make all scripts visible to Python:\n\n    $ export PYTHONPATH=\"${PYTHONPATH}:`pwd`:`pwd`\/slim\"\n\nNext, we copy over the Protocol Buffer (protobuf) compiler command script from our GitHub repository (see _http:\/\/PracticalDeepLearning.ai_) into our current directory to make the .`proto` files available to Python:\n\n    $ cp {path_to_book_github_repo}\/code\/chapter-14\/add_protoc.sh . && \\\n    chmod +x add_protoc.sh && \\\n    .\/add_protoc.sh\n\nFinally, we run the _setup.py_ script as follows:\n\n    $ python setup.py build\n    $ python setup.py install\n\nNow is the time to download our prebuilt model. In our case, we're using the SSD MobileNetV2 model. You can find a large list of models at TensorFlow's object detection model zoo, featuring models by the datasets on which they were trained, speed of inference, and Mean Average Precision (mAP). Within that list, download the model titled `ssd_mobilenet_v2_coco`, which, as the name suggests, was trained on the MS COCO dataset. Benchmarked at 22 mAP, this model runs at 31 ms on an NVIDIA GeForce GTX TITAN X with a 600x600 resolution input. After downloading that model, we unzip it into the _models\/research\/object_detection_ directory inside the TensorFlow repository. We can inspect the contents of the directory as follows:\n\n    $ cd object_detection\/\n    $ ls **ssd_mobilenet_v2_coco_2018_03_29**\n    checkpoint                      model.ckpt.data-00000-of-00001  model.ckpt.meta\n    saved_model\n    frozen_inference_graph.pb       model.ckpt.index                pipeline.config\n\n## Test Driving Our Model\n\nBefore we plug our model into a mobile app, it's a good idea to verify whether the model is able to make predictions. The TensorFlow Models repository contains a Jupyter Notebook where we can simply plug in a photograph to make predictions on it. You can find the notebook at _models\/research\/object_detection\/object_detection_tutorial.ipynb_. Figure 14-3 shows a prediction from that notebook on a familiar photograph.\n\n###### Figure 14-3. Object detection prediction in the ready-to-use Jupyter Notebook from the TensorFlow Models repository\n\n## Deploying to a Device\n\nNow that we've verified that the model works, it's time to convert it into a mobile-friendly format. For conversion to the TensorFlow Lite format, we use our familiar `tflite_convert` tool from Chapter 13. It should be noted that this tool operates on _.pb_ files, whereas the TensorFlow Object Detection model zoo provides only model checkpoints. So, we first need to generate a _.pb_ model from the checkpoints and graphs. Let's do that using the handy script that comes with the TensorFlow Models repository. You can find the _export_tflite_ssd_graph.py_ file in _models\/research\/object_detection_ :\n\n    $ python export_tflite_ssd_graph.py \\\n    --pipeline_config_path= **ssd_mobilenet_v2_coco_2018_03_29** \/pipeline.config \\\n    --trained_checkpoint_prefix= **ssd_mobilenet_v2_coco_2018_03_29** \/\n    model.ckpt.data-00000-of-00001 \\\n    --output_directory=tflite_model \\\n    --add_postprocessing_op=true\n\nIf the preceding script execution was successful, we'll see the following files in the _tflite_model_ directory:\n\n    $ ls tflite_model\n    tflite_graph.pb\n    tflite_graph.pbtxt\n\nWe will convert this to the TensorFlow Lite format using the `tflite_convert` tool.\n\n    $ tflite_convert --graph_def_file=tflite_model\/tflite_graph.pb \\\n    --output_file=tflite_model\/model.tflite\n\nNow the only thing that remains is to plug the model into an app. The app that we'll be using is referenced in the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_) at _code\/chapter-14_. We have already looked at how to swap a model in an app in Chapters , , and , so we won't discuss that here again. Figure 14-4 shows the result of the object detector model when plugged into the Android app.\n\n###### Figure 14-4. Running a real-time object detector model on an Android device\n\nThe reason why we're able to see the \"Cat\" prediction right away is because 'Cat' is one of the 80 categories already present in the MS COCO dataset that was used to train the model. This model might be sufficient for Bob to deploy on his devices (keeping in mind that although Bob might not use mobile phones to track his garden, the process for deploying a TensorFlow Lite model to edge hardware is quite similar). However, Bob would do well for himself to improve the precision of his model by fine tuning it on data that has been generated within his garden. In the next section, we explore how to use transfer learning to build an object detector by using only a web-based tool. If you've read Chapter 8, what comes next should feel familiar.\n\n# Building a Custom Detector Without Any Code\n\n> My cat's breath smells like cat food!\n> \n> Ralph Wiggum\n\nLet's face it. Your authors could go around and snap a lot of cat pictures within their gardens to perform this experiment. It would be tedious, and we might be rewarded with only a few extra percentage points of precision and recall. Or we could look at a really fun example that also tests the limits of CNNs. We are, of course, referring to _The Simpsons_. Given that most models are not trained with features from cartoon images, it would be interesting to see how our model performs here.\n\nFirst, we need to get our hands on a dataset. Thankfully, we have one readily available on Kaggle. Let's download the dataset and use CustomVision.ai (similar to Chapter 8) to train our Simpsons Classifier using the following steps. _Excellent!_\n\n###### Note\n\nContrary to what some may believe, we, the authors, do not have millions of dollars lying around. This means that we cannot afford to buy the rights to _The Simpsons_ from Fox. Copyright laws, as a consequence, prevent us from publishing any images from that show. Instead, in this section, we are using the next best thing\u2014public domain images from the United States Congress. In place of the images from the cartoon, we use the photographs of congresswomen and congressmen with the same first names as the Simpsons family: _Homer_ Hoch, _Marge_ Roukema, _Bart_ Gordon, and _Lisa_ Blunt Rochester. Keep in mind that this is only for publishing purposes; we have still trained on the original dataset from the cartoon.\n\n  1. Go to CustomVision.ai and start a new Object Detection Project (Figure 14-5).\n\n###### Figure 14-5. Creating a new Object Detection project in CustomVision.ai\n\n  2. Create tags for Homer, Marge, Bart, and Lisa. Upload 15 images for each character and draw a bounding box for each of those images. The dashboard should resemble Figure 14-6 right about now.\n\n###### Figure 14-6. Dashboard with bounding box and class name\n\n  3. Click the Train button and let the training commence. There are sliders to control the probability and overlap thresholds on the dashboard. We can play around with them and adjust them to something we like and with which get good precision and recall.\n\n  4. Click the Quick Test button and use any random image of the Simpsons (not previously used for training) to test the performance of the model. The results might not be great yet, but that's okay: we can fix it by training with more data.\n\n  5. Let's experiment with how many images we need to increase precision, recall, and mAP. Let's start by adding five more images for each class and retrain the model.\n\n  6. Repeat this process until we have acceptable results. Figure 14-7 shows the results from our experiment.\n\n###### Figure 14-7. Measuring improvement in percent mean average precision with increasing number of images per class\n\nWith our final model, we are able to get reasonably good detections for a random image off the internet, as can be seen in Figure 14-8. Although not expected, it is surprising that a model pretrained on natural images is able to fine tune on cartoons with relatively little data.\n\n###### Figure 14-8. Detected Simpsons characters with the final model, represented by US congress members of the same first name (see note at the beginning of this section)\n\nAs we already know by now, CustomVision.ai allows this model to be exported in a variety of formats including Core ML, TensorFlow Lite, ONNX, and more. We can simply download into our desired format ( _.tflite_ in our case) and plug this file into the app similar to see how we did with the pretrained model in the previous section. With real-time Simpsons detection in our hands, we could show it off to our friends and family the next time Principal Skinner talks about steamed hams on our TV.\n\n###### Note\n\nCustomVision.ai is not the only platform that allows you to label, train, and deploy online without writing any code. Google's Cloud AutoML (in beta as of October 2019) and Apple's Create ML (only for the Apple ecosystem) offer a very similar feature set. Additionally, Matroid allows you to build custom object detectors from video feeds, which can be extremely useful to train a network without spending too much effort on building a dataset.\n\nSo far, we have looked at three quick ways to get object detection running with a very minimal amount of code. We estimate that about 90% of use cases can be handled with one of these options. If you can solve for your scenario using these, you can stop reading this chapter and jump straight to the \"Case Studies\" section.\n\nIn a very small set of scenarios, we might want further fine-grained control on factors such as accuracy, speech, model size, and resource usage. In the upcoming sections, we look at the world of object detection in a little more detail, from labeling the data all the way to deploying the model.\n\n# The Evolution of Object Detection\n\nOver the years, as the deep learning revolution happened, there was a renaissance not only for classification, but also for other computer-vision tasks, including object detection. During this time, several architectures were proposed for object detection, often building on top of each other. Figure 14-9 shows a timeline of some of them.\n\n###### Figure 14-9. A timeline of different object detection architectures (image source: Recent Advances in Deep Learning for Object Detection by Xiongwei Wu et al.)\n\nIn object classification, our CNN architecture extracts the features from an image and computes the probabilities for a specific number of classes. These CNN architectures (ResNet, MobileNet) work as the _backbone_ for object detection networks. In object detection networks, our end result is bounding boxes (defined by the center of rectangle, height, width). Although covering the inner details of many of these will probably need significantly more in-depth exploration (which you can find on this book's GitHub website, linked at _http:\/\/PracticalDeepLearning.ai_), we can broadly divide them into two categories, as shown in Table 14-3.\n\nTable 14-3. Categories of object detectors **** | **Description** | **Pros and cons**  \n---|---|---  \n**Two-stage detectors** |\n\nGenerate category-agnostic bounding box candidates (regions of interest) using a Region Proposal Network (RPN).\n\nRun a CNN on these region proposals to assign each a category.\n\nExamples: Faster R-CNN, Mask R-CNN. | \\+ High Accuracy\n\n\u2013 Slower\n\n**One-stage detectors** |\n\nDirectly make categorical predictions on objects on each location of the feature maps; can be trained end-to-end training.\n\nExamples: YOLO, Single-Shot Detector (SSD). | \\+ Speed, better suited for real-time applications\n\n\u2013 Lower accuracy\n\n###### Note\n\nIn the world of object detection, Ross Girshik is a name you will encounter often while reading papers. Working before the deep learning era as well as during it, his work includes Deformable Part Models (DPM), R-CNN, Fast R-CNN, Faster R-CNN, You Only Look Once (YOLO), Mask R-CNN, Feature Pyramid Network (FPN), RetinaNet, and ResNeXt to name just a few, often breaking his own records year after year on public benchmarks.\n\n> I participated in several first-place entries into the PASCAL VOC object detection challenge, and was awarded a \"lifetime achievement\" prize for my work on deformable part models. I think this refers to the lifetime of the PASCAL challenge\u2014and not mine!\n> \n> \u2014Ross Girschik\n\n## Performance Considerations\n\nFrom a practitioner's point of view, the primary concerns tend to be accuracy and speed. These two factors usually have an inverse relationship with each other. We want to find the detector that strikes the ideal balance for our scenario. Table 14-4 summarizes some of those readily available pretrained models on the TensorFlow object detection model zoo. The speed reported was on a 600x600 pixel resolution image input, processed on an NVIDIA GeForce GTX TITAN X card.\n\nTable 14-4. Speed versus percent mean average precision for some readily available architectures on the TensorFlow object detection model zoo website **** | **Inference speed (ms)** | **% mAP on MS COCO**  \n---|---|---  \n**ssd_mobilenet_v1_coco** | 30 | 21  \n**ssd_mobilenet_v2_coco** | 31 | 22  \n**ssdlite_mobilenet_v2_coco** | 27 | 22  \n**ssd_resnet_50_fpn_coco** | 76 | 35  \n**faster_rcnn_nas** | 1,833 | 43\n\nA 2017 study done by Google (see Figure 14-10) revealed the following key findings:\n\n  * One-stage detectors are faster than two-stage, though less accurate.\n\n  * The higher the accuracy of the backbone CNN architecture on classification tasks (such as ImageNet), the higher the precision of the object detector.\n\n  * Small-sized objects are challenging for object detectors, which give poor performance (usually under 10% mAP). The effect is more severe on one-stage detectors.\n\n  * On large-sized objects, most detectors give similar performance.\n\n  * Higher-resolution images give better results because small objects appear larger to the network.\n\n  * Two-stage object detectors internally generate a large number of proposals to be considered as the potential locations of objects. The number of proposals can be tuned. Decreasing the number of proposals can lead to speedups with little loss in accuracy (the threshold will depend on the use case).\n\n###### Figure 14-10. The effect of object detection architecture as well as the backbone architecture (feature extractor) on the percent mean average precision and prediction time (note that the colors might not be clear in grayscale printing; refer to the book's GitHub website see [_http:\/\/PracticalDeepLearning.ai_] for the color image) (image source: Speed\/accuracy trade-offs for modern convolutional object detectors by Jonathan Huang et al.)\n\n# Key Terms in Object Detection\n\nThe terms used heavily in object detection include Intersection over Union, Mean Average Precision, Non-Maximum Suppression, and Anchor. Let's examine what each of these means.\n\n## Intersection over Union\n\nIn an object detection training pipeline, metrics such as Intersection over Union (IoU) are typically used as a threshold value to filter detections of a certain quality. To calculate the IoU, we use both the predicted bounding box and the ground truth bounding box to calculate two different numbers. The first is the area where both the prediction and the ground truth overlap\u2014the intersection. The second is the span covered by both the prediction and the ground truth\u2014the union. As the name suggests, we simply divide the total area of the intersection by the area of the union to get the IoU. Figure 14-11 visually demonstrates the IoU concept for two 2x2 squares that intersect in a single 1x1 square in the middle.\n\n###### Figure 14-11. A visual representation of the IoU ratio\n\nIn the perfect scenario in which the predicted bounding box matches the ground truth exactly, the IoU value would be 1. In the worst-case scenario, the prediction would have no overlap with the ground truth and the IoU value would be 0. As we can see, the IoU value would range between 0 and 1, with higher numbers indicating higher-quality predictions, as illustrated in Figure 14-12.\n\nTo help us filter and report results, we would set a minimum IoU threshold value. Setting the threshold to a really aggressive value (such as 0.9) would result in a loss of a lot of predictions that might turn out to be important further down the pipeline. Conversely, setting the threshold value too low would result in too many spurious bounding boxes. A minimum IoU of 0.5 is typically used for reporting the precision of an object detection model.\n\n###### Figure 14-12. IoU illustrated; predictions from better models tend to have heavier overlap with the ground truth, resulting in a higher IoU\n\nIt's worth reiterating here that the IoU value is calculated per instance of a category, not per image. However, to calculate the quality of the detector over a larger set of images, we would use IoU in conjunction with other metrics such as Average Precision.\n\n## Mean Average Precision\n\nWhile reading research papers on object detection, we often come across numbers such as AP@0.5. That term conveys the average precision at IoU=0.5. Another more elaborate representation would be AP@[0.6:0.05:0.75], which is the average precision from IoU=0.6 to IoU=0.75 at increments of 0.05. The Mean Average Precision (or mAP) is simply the average precision across all categories. For the COCO challenge, the MAP metric used is AP@[0.5:0.05:0.95].\n\n## Non-Maximum Suppression\n\nInternally, object detection algorithms make a number of proposals for the potential locations of objects that might be in the image. For each object in the image, it is expected to have multiple of these bounding box proposals at varying confidence values. Our task is to find the one proposal that best represents the real location of the object. A na\u00efve way would be to consider only the proposal with the maximum confidence value. This approach might work if there were only one object in the entire image. But this won't work if there are multiple categories, each with multiple instances in a single image.\n\nNon-Maximum Suppression (NMS) comes to the rescue (Figure 14-13). The key idea behind NMS is that two instances of the same object will not have heavily overlapping bounding boxes; for instance, the IoU of their bounding boxes will be less than a certain IoU threshold (say 0.5). A greedy way to achieve this is by repeating the following steps for each category:\n\n  1. Filter out all proposals under a minimum confidence threshold.\n\n  2. Accept the proposal with the maximum confidence value.\n\n  3. For all remaining proposals sorted in descending order of their confidence values, check whether the IoU of the current box with respect to one of the previously accepted proposals \u2265 0.5. If so, filter it; otherwise, accept it as a proposal.\n\n###### Figure 14-13. Using NMS to find the bounding box that best represents the location of the object in an image\n\n# Using the TensorFlow Object Detection API to Build Custom Models\n\nIn this section, we run through a full end-to-end example of building an object detector. We look at several steps in the process, including collecting, labeling, and preprocessing data, training the model, and exporting it to the TensorFlow Lite format.\n\nFirst, let's look at some strategies for collecting data.\n\n## Data Collection\n\nWe know by now that, in the world of deep learning, data reigns supreme. There are a few ways in which we can acquire data for the object that we want to detect:\n\nUsing ready-to-use datasets\n\nThere are a few public datasets available to train object detection models such as MS COCO (80 categories), ImageNet (200 categories), Pascal VOC (20 categories), and the newer Open Images (600 categories). MS COCO and Pascal VOC are used in most object detection benchmarks, with COCO-based benchmarks being more realistic for real-world scenarios due to more complex imagery.\n\nDownloading from the web\n\nWe should already be familiar with this process from Chapter 12 in which we collected images for the \"Hot Dog\" and \"Not Hot Dog\" classes. Browser extensions such as Fatkun come in handy for quickly downloading a large number of images from search engine results.\n\nTaking pictures manually\n\nThis is the more time consuming (but potentially fun) option. For a robust model, it's important to train it with pictures taken in a variety of different settings. With the object to detect in hand, we should take at least 100 to 150 images of it against different backgrounds, with a variety of compositions (framing) and angles, and in many different lighting conditions. Figure 14-14 shows some examples of pictures taken to train a Coke and Pepsi object detection model. Considering that a model can potentially learn the spurious correlation that red means Coke and blue means Pepsi, it's a good idea to mix objects with backgrounds that could potentially confuse it so that eventually we realize a more robust model.\n\n###### Figure 14-14. Photographs of objects taken in a variety of different settings to train an object detector model\n\nIt's worth taking pictures of the object in environments that it's unlikely to be used in order to diversify the dataset and improve the robustness of the model. It also has the added bonus of keeping an otherwise boring and tedious process a little entertaining. We can make it interesting by challenging ourselves to come up with unique and innovative photos. Figure 14-15 shows some creative examples of photos taken during the process of building a currency detector.\n\n###### Figure 14-15. Some creative photographs taken during the process of building a diverse currency dataset\n\nBecause we want to make a cat detector, we will be reusing the cat images from the \"Cats and Dogs\" Kaggle dataset that we used in Chapter 3. We randomly chose images from that set and split them into a train and test set.\n\nTo maintain uniformity of input size to our network, we resize all images to a fixed size. We will use the same size as part of our network definition and while converting to a _.tflite_ model. We can use the ImageMagick tool to resize all images at once:\n\n    $ apt-get install imagemagick\n    $ mogrify -resize 800x600 *.jpg\n\n###### Tip\n\nIf you have a large number of images (i.e., several tens of thousands of images) in the current directory, the preceding command might fail to enlist all of the images. A solution would be to use the `find` command to list all of the images and pipe the output to the `mogrify` command:\n\n    $ find . -type f | awk -F. '!a[$NF]++{print $NF}' |\n    xargs -I{} mogrify -resize 800x600 *.jpg\n\n## Labeling the Data\n\nAfter we have our data, the next step would be to label it. Unlike classification where simply placing an image in the correct directory would be sufficient, we need to manually draw the bounding rectangles for the objects of interest. Our tool of choice for this task is LabelImg (available for Windows, Linux, and Mac) for the following reasons:\n\n  * You can save annotations either as XML files in PASCAL VOC format (also adopted by ImageNet) or the YOLO format.\n\n  * It supports single and multiclass labels.\n\n###### Note\n\nIf you are the curious cat like us, you'd want to know what the difference is between the YOLO format and the PASCAL VOC format. YOLO happens to use simple _.txt_ files, one per image with the same name, to store information about the classes and the bounding boxes within the image. The following is what a _.txt_ file in YOLO format typically looks like:\n\n    class_for_box1 box1_x box1_y box1_w box1_h\n    class_for_box2 box2_x box2_y box2_w box2_h\n\nNote that the x, y, width, and height in this example are normalized using the full width and height of the image.\n\nPASCAL VOC format, on the other hand, is XML based. Similar to the YOLO format, we use one XML file per image (ideally with the same name). You can view the sample format on the Git repository at _code\/chapter-14\/pascal-voc-sample.xml_.\n\nFirst, download the application from GitHub to a directory on your computer. This directory will have an executable and a data directory containing some sample data. Because we will be using our own custom data, we don't need the data within the provided data directory. You can start the application by double-clicking the executable file.\n\nAt this point, we should have a collection of images that is ready to be used during the training process. We first divide our data randomly into training and test sets and place them in separate directories. We can do this split by simply dragging and dropping the images randomly into either directory. After the train and test directories are created, we load the training directory by clicking Open Dir, as shown in Figure 14-16.\n\n###### Figure 14-16. Click the Open Dir button and then select the directory that contains the training data\n\nAfter LabelImg loads the training directory, we can start our labeling process. We must go through each image and manually draw bounding boxes for each object (only cats in our case), as shown in Figure 14-17. After we make the bounding box, we are prompted to provide a label to accompany the bounding box. For this exercise, enter the name of the object, \"cat\". After a label is entered, we need only to select the checkbox to specify that label again for subsequent images. For images with multiple objects, we would make multiple bounding boxes and add their corresponding labels. If there are different kinds of objects, we simply add a new label for that object class.\n\n###### Figure 14-17. Select the Create RectBox from the panel on the left to make a bounding box that covers the cat\n\nNow, repeat this step for all of the training and test images. We want to obtain tight bounding boxes for each object, such that all parts of the object are bounded by the box. Finally, in the train and test directories, we see _.xml_ files accompanying each image file, as depicted in Figure 14-18. We can open the .`xml` file in a text editor and inspect metadata such as image file name, bounding box coordinates, and label name.\n\n###### Figure 14-18. Each image is accompanied by an XML file that contains the label information and the bounding box information\n\nIn Chapter 1, we mentioned that during the development of the MS COCO dataset, it took roughly three seconds to label the name of each object in an image, approximately 30 seconds to place a bounding box around each object, and 79 seconds to draw the outlines for each object.\n\nLabeling for detection can take about 10x the time as labeling for classification. Doing this on a large scale while maintaining high quality can get really expensive, really quickly. What's the expensive part here? Drawing the bounding boxes manually. What is something cheaper than drawing them? Verifying the correctness of already drawn boxes. And that's what we want to maximize.\n\nRather than labeling large quantities of data manually, we'd go through an iterative semiautomated process of labeling. In this approach you'd do the following:\n\n  1. Pick a small fraction of images from the pool of data.\n\n  2. Manually label it with the object class and bounding box information.\n\n  3. Train a model with the labeled data.\n\n  4. Make predictions on the remaining unlabeled data using this model.\n\n  5. If a prediction confidence is higher than a set threshold, assign that as the label.\n\n  6. Pick a sample of the remaining predictions under this threshold.\n\n  7. In this sample, verify the predictions for correctness. If a prediction is incorrect, manually fix it. In practice, this usually means moving the bounding box slightly, which is much quicker than drawing the full bounding boxes.\n\n  8. Add the images from the manually reviewed sample to the training dataset.\n\n  9. Repeat the process until the model achieves acceptable performance on a separate validation set.\n\nIn the first few iterations of this process, you might have a barely functioning model. But that's okay. By using this human-in-the-loop approach, we are constantly teaching it what it's bad at. Pretty soon, our model will become confident enough to automatically label a majority of the data, cutting down on the large costs of manually labeling large amounts of data.\n\nThis approach is one way of using the technique known as _active learning_. Famous labeling companies like Figure Eight use this heavily to reduce labeling costs and increase efficiency so that a large majority of time is eventually spent on verification.\n\n## Preprocessing the Data\n\nAt this point, we have handy XML data that gives the bounding boxes for all of our objects. However, for us to use TensorFlow for training on the data, we must preprocess it into a format that TensorFlow understands, namely TFRecords. Before we can convert our data into TFRecords, we must go through the intermediate step of consolidating the data from all the XML files into a single comma-separated values (CSV) file. TensorFlow provides helper scripts that assist us with these operations.\n\nNow that we have our environment set up, it's time to begin doing some real work:\n\n  1. Convert our cats dataset directory to a single CSV file using the `xml_to_csv` tool from the racoon_dataset repository by Dat Tran. We have provided a slightly edited copy of this file in our repository at _code\/chapter-14\/xml_to_csv.py_ :\n\n    $ python xml_to_csv.py -i {path to cats training dataset} -o {path to output\n    train_labels.csv}\n\n  2. Do the same for the test data:\n\n    $ python xml_to_csv.py -i {path to cats test dataset} -o {path to\n    test_labels.csv}\n\n  3. Make the _label_map.pbtxt_ file. This file contains the label and identifier mappings for all of our classes. We use this file to convert our text label into an integer identifier, which is what TFRecord expects. Because we have only one class, the file is rather small and looks as follows:\n\n    item {\n        id: 1\n        name: 'cat'\n    }\n\n  4. Generate the TFRecord format files that will contain the data to be used for training and testing our model later on. This file is also from the racoon_dataset repository by Dat Tran. We have provided a slightly edited copy of this file in our repository at _code\/chapter-14\/generate_tfrecord.py_. (It's worth noting that the `image_dir` path in the argument should be the same as the path in the XML files; LabelImg uses absolute paths.)\n\n    $ python generate_tfrecord.py \\\n    --csv_input={path to train_labels.csv} \\\n    --output_path={path to train.tfrecord} \\\n    --image_dir={path to cat training dataset}\n\n    $ python generate_tfrecord.py \\\n    --csv_input={path to test_labels.csv} \\\n    --output_path={path to test.tfrecord} \\\n    --image_dir={path to cat test dataset}\n\nWith our _train.tfrecord_ and _test.tfrecord_ files ready, we can now begin our training process.\n\n# Inspecting the Model\n\n(This section is informational only and is not necessary for the training process. You can skip ahead to \"Training\" if you want to.)\n\nWe can inspect our model using the `saved_model_cli` tool:\n\n    $ saved_model_cli show --dir ssd_mobilenet_v2_coco_2018_03_29\/saved_model \\\n    --tag_set serve \\\n    --signature_def serving_default\n\nThe output of that script looks like the following:\n\n    The given SavedModel SignatureDef contains the following input(s):\n      inputs['inputs'] tensor_info:\n          dtype: DT_UINT8\n          shape: (-1, -1, -1, 3)\n          name: image_tensor:0\n    The given SavedModel SignatureDef contains the following output(s):\n      outputs['detection_boxes'] tensor_info:\n          dtype: DT_FLOAT\n          shape: (-1, 100, 4)\n          name: detection_boxes:0\n      outputs['detection_classes'] tensor_info:\n          dtype: DT_FLOAT\n          shape: (-1, 100)\n          name: detection_classes:0\n      outputs['detection_scores'] tensor_info:\n          dtype: DT_FLOAT\n          shape: (-1, 100)\n          name: detection_scores:0\n      outputs['num_detections'] tensor_info:\n          dtype: DT_FLOAT\n          shape: (-1)\n          name: num_detections:0\n      outputs['raw_detection_boxes'] tensor_info:\n          dtype: DT_FLOAT\n          shape: (-1, -1, 4)\n          name: raw_detection_boxes:0\n      outputs['raw_detection_scores'] tensor_info:\n          dtype: DT_FLOAT\n          shape: (-1, -1, 2)\n          name: raw_detection_scores:0\n    Method name is: tensorflow\/serving\/predict\n\nThe following is how we can interpret the output from the previous command:\n\n  1. The shape of `inputs` being `(-1, -1, -1, 3)` indicates that the trained model can accept any arbitrarily sized image input that contains three channels. Due to this inherent flexibility in input size, the converted model would be larger than if the model had a fixed-size input. When we train our own object detector later in this chapter, we will be fixing the input size to 800 x 600 pixels to make the resulting model more compact.\n\n  2. Moving on to the outputs, the very first output is `detection_boxes (-1, 100, 4)`. This tells us how many detection boxes are possible, and what will they look like. In particular, the very first number (i.e., -1) indicates that we can have any number of boxes based on the detections for all 100 classes (second number) with four coordinates (third number)\u2014x, y, width, and height\u2014defining each box. In other words, we are not limiting the number of detections for each of the 100 classes.\n\n  3. For `detection_classes`, we have a list of two elements: the first defining how many objects we detected, and the second, a one-hot encoded vector with the detected class enabled.\n\n  4. `num_detections` is the number of objects detected in the image. It is a single floating point.\n\n  5. `raw_detection_boxes` defines the coordinates of each box that was detected for each object. All of these detections are before the application of NMS on all the detections.\n\n  6. `raw_detection_scores` is a list of two floating-point numbers, the first describing the total number of detections, and the second giving the total number of categories including the background if it is considered a separate category.\n\n## Training\n\nBecause we are using SSD MobileNetV2, we will need to create a copy of the _pipeline.config_ file (from the TensorFlow repository) and modify it based on our configuration parameters. First, make a copy of the configuration file and search for the string `PATH_TO_BE_CONFIGURED` within the file. This parameter indicates all of the places that need to be updated with the correct paths (absolute paths preferably) within our filesystem. We will also want to edit some parameters in the configuration file, such as number of classes (`num_classes`), number of steps (`num_steps`), number of validation samples (`num_examples`), and path of label mappings for both the train and test\/eval sections. (You'll find our version of the _pipeline.config_ file on the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_).\n\n    $ cp object_detection\/samples\/configs\/ssd_mobilenet_v2_coco.config\n    .\/pipeline.config\n    $ vim pipeline.config\n\n###### Note\n\nIn our example, we're using the SSD MobileNetV2 model to train our object detector. Under different conditions, you might want to choose a different model on which to base your training. Because each model comes with its own pipeline configuration file, you'd be modifying that configuration file, instead. The configuration file comes bundled with each model. You'd use a similar process in which you'd identify the paths that need to be changed and update them using a text editor accordingly.\n\nIn addition to modifying paths, you might want to modify other parameters within the configuration file such as image size, number of classes, optimizer, learning rate, and number of epochs.\n\nSo far, we've been at the _models\/research_ directory within the TensorFlow repository. Let's now move into the _object_detection_ directory and run the _model_main.py_ script from TensorFlow, which trains our model based on the configuration provided by the _pipeline.config_ file that we just edited:\n\n    $ cd object_detection\/\n    $ python model_main.py \\\n    --pipeline_config_path=..\/pipeline.config \\\n    --logtostderr \\\n    --model_dir=training\/\n\nWe'll know the training is happening correctly when we see lines of output that look like the following:\n\n    Average Precision (AP) @[ IoU=0.50:0.95 | area=all | maxDets=100 ] = 0.760\n\nDepending on the number of iterations that we are training for, the training will take from a few minutes to a couple of hours, so plan to have a snack, do the laundry, clean the litter box, or better yet, read the chapter ahead. After the training is complete, we should see the latest checkpoint file in the training\/ directory. The following files are also created as a result of the training process:\n\n    $ ls training\/\n\n    checkpoint                                   model.ckpt-13960.meta\n    eval_0                                       model.ckpt-16747.data-00000-of-00001\n    events.out.tfevents.1554684330.computername  model.ckpt-16747.index\n    events.out.tfevents.1554684359.computername  model.ckpt-16747.meta\n    export                                       model.ckpt-19526.data-00000-of-00001\n    graph.pbtxt                                model.ckpt-19526.index\n    label_map.pbtxt                            model.ckpt-19526.meta\n    model.ckpt-11180.data-00000-of-00001       model.ckpt-20000.data-00000-of-00001\n    model.ckpt-11180.index                     model.ckpt-20000.index\n    model.ckpt-11180.meta                      model.ckpt-20000.meta\n    model.ckpt-13960.data-00000-of-00001       pipeline.config\n    model.ckpt-13960.index\n\nWe convert the latest checkpoint file to the _.TFLite_ format in the next section.\n\n###### Tip\n\nSometimes, you might want to do a deeper inspection of your model for debugging, optimization, and\/or informational purposes. Think of it like a DVD extra of a behind-the-scenes shot of your favorite TV show. You would run the following command along with arguments pointing to the checkpoint file, the configuration file, and the input type to view all the different parameters, the model analysis report, and other golden nuggets of information:\n\n    $ python export_inference_graph.py \\\n    --input_type=image_tensor \\\n    --pipeline_config_path=training\/pipeline.config \\\n    --output_directory=inference_graph \\\n    --trained_checkpoint_prefix=training\/model.ckpt-20000\n\nThe output from the this command looks like the following:\n\n    =========================Options=========================\n    -max_depth                  10000\n    -step                       -1\n    ...\n\n    ==================Model Analysis Report==================\n    ...\n    Doc:\n    scope: The nodes in the model graph are organized by\n    their names, which is hierarchical like filesystem.\n    param: Number of parameters (in the Variable).\n\n    Profile:\n    node name | # parameters\n    _TFProfRoot (--\/3.72m params)\n     BoxPredictor_0 (--\/93.33k params)\n        BoxPredictor_0\/BoxEncodingPredictor\n        (--\/62.22k params)\n          BoxPredictor_0\/BoxEncodingPredictor\/biases\n          (12, 12\/12 params)\n     ...     \n    FeatureExtractor (--\/2.84m params)\n    ...\n\nThis report helps you to learn how many parameters you have, which could in turn help you in finding opportunities for model optimization.\n\n## Model Conversion\n\nNow that we have the latest checkpoint file (the suffix should match the number of epochs in the _pipeline.config_ file), we will feed it into the `export_tflite_ssd_graph` script like we did earlier in the chapter with our pretrained model:\n\n    $ python export_tflite_ssd_graph.py \\\n    --pipeline_config_path=training\/pipeline.config \\\n    --trained_checkpoint_prefix=training\/model.ckpt-20000 \\\n    --output_directory=tflite_model \\\n    --add_postprocessing_op=true\n\nIf the preceding script execution was successful, we'll see the following files in the _tflite_model_ directory:\n\n    $ ls tflite_model\n    tflite_graph.pb\n    tflite_graph.pbtxt\n\nWe have one last step remaining: convert the frozen graph files to the _.TFLite_ format. We can do that using the `tflite_convert` tool:\n\n    $ tflite_convert --graph_def_file=tflite_model\/tflite_graph.pb \\\n    --output_file=tflite_model\/cats.tflite \\\n    --input_arrays=normalized_input_image_tensor \\\n    --output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLi\n    te_Detection_PostProcess:2','TFLite_Detection_PostProcess:3' \\\n    --input_shape=1,800,600,3 \\\n    --allow_custom_ops\n\nIn this command, notice that we've used a few different arguments that might not be intuitive at first sight:\n\n  * For `--input_arrays`, the argument simply indicates that the images that will be provided during predictions will be normalized `float32` tensors.\n\n  * The arguments supplied to `--output_arrays` indicate that there will be four different types of information in each prediction: the number of bounding boxes, detection scores, detection classes, and the coordinates of bounding boxes themselves. This is possible because we used the argument `--add_postprocessing_op=true` in the export graph script in the previous step.\n\n  * For `--input_shape`, we provide the same dimension values as we did in the _pipeline.config_ file.\n\nThe rest are fairly trivial. We should now have a spanking new _cats.tflite_ model file within our _tflite_model\/_ directory. It is now ready to be plugged into an Android, iOS, or edge device to do live detection of cats. This model is well on its way to saving Bob's garden!\n\n###### Tip\n\nIn the floating-point MobileNet model, '`normalized_image_tensor`' has values between `[-1,1)`. This typically means mapping each pixel (linearly) to a value between [\u20131,1]. Input image values between 0 and 255 are scaled by (1\/128) and then a value of \u20131 is added to them to ensure the range is `[-1,1)`.\n\nIn the quantized MobileNet model, '`normalized_image_tensor`' has values between `[0, 255]`.\n\nIn general, see the `preprocess` function defined in the feature extractor class in the TensorFlow Models repository at _models\/research\/object_detection\/models_ directory.\n\n# Image Segmentation\n\nTaking a step further beyond object detection, to get more precise locations of objects, we can perform object segmentation. As we saw earlier in this chapter, this involves making category predictions for each pixel in an image frame. Architectures such as U-Net, Mask R-CNN, and DeepLabV3+ are commonly used to perform segmentation tasks. Similar to object detection, there has been a growing trend of running segmentation networks in real time, including on resource-constrained devices such as smartphones. Being real time opens up many consumer app scenarios like face filters (Figure 14-19), and industrial scenarios such as detecting drivable roads for autonomous cars (Figure 14-20).\n\n###### Figure 14-19. Colorizing hair with ModiFace app by accurately mapping the pixels belonging to the hair\n\n###### Figure 14-20. Image segmentation performed on frames from a dashcam (CamVid dataset)\n\nAs you might have grown accustomed to hearing in this book by now, you can accomplish much of this without needing much code. Labeling tools like Supervisely, LabelBox, and Diffgram can help not just label the data, but also load previously annotated data and train further on a pretrained object segmentation model. Moreover, these tools provide AI-assisted labeling, which can dramatically speed up (~10 times) the otherwise laborious and expensive labeling process. If this sounds exciting, you are in luck! We have included a bonus resource guide on how to learn and build segmentation projects on the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_).\n\n# Case Studies\n\nLet's now look at how object detection is being used to power real-world applications in the industry.\n\n## Smart Refrigerator\n\nSmart fridges have been gaining popularity for the past few years because they have started to become more affordable. People seem to like the convenience of knowing what's in their fridge even when they are not home to look inside. Microsoft teamed up with Swiss manufacturing giant Liebherr to use deep learning in its new generation SmartDeviceBox refrigerator (Figure 14-21). The company used Fast R-CNN to perform object detection within its refrigerators to detect different food items, maintain an inventory, and keep the user abreast of the latest state.\n\n###### Figure 14-21. Detected objects along with their classes from the SmartDeviceBox refrigerator (image source)\n\n## Crowd Counting\n\nCounting people is not just something that only cats do while staring out of windows. It's useful in a lot of situations including managing security and logistics at large sporting events, political gatherings, and other high-traffic areas. Crowd counting, as its name suggests, can be used to count any kind of object including people and animals. Wildlife conservation is an exceptionally challenging problem because of the lack of labeled data and intense data collection strategies.\n\n### Wildlife conservation\n\nSeveral organizations, including the University of Glasgow, University of Cape Town, Field Museum of Natural History, and Tanzania Wildlife Research Institute, came together to count the largest terrestrial animal migration on earth: 1.3 million blue wildebeest and 250,000 zebra between the Serengeti and the Masaai Mara National Reserve, Kenya (Figure 14-22). They employed data labeling through 2,200 volunteer citizen scientists using a platform called Zooniverse and used automated object detection algorithms like YOLO for counting the wildebeest and zebra population. They observed that the count from the volunteers and from the object detection algorithms were within 1% of each other.\n\n###### Figure 14-22. An aerial photograph of wildebeest taken from a small survey airplane (image source)\n\n### Kumbh Mela\n\nAnother example of dense crowd counting is from Prayagraj, India where approximately 250 million people visit the city once every 12 years for the Kumbh Mela festival (Figure 14-23). Crowd control for such large crowds has historically been difficult. In fact, in the year 2013, due to poor crowd management, a stampede broke out tragically resulting in the loss of 42 lives.\n\nIn 2019, the local state government contracted with Larsen & Toubro to use AI for a variety of logistical tasks including traffic monitoring, garbage collection, security assessments, and also crowd monitoring. Using more than a thousand CCTV cameras, authorities analyzed crowd density and designed an alerting system based on the density of people in a fixed area. In places of higher density, the monitoring was more intense. It made the Guinness Book of World Records for being the largest crowd management system ever built in the history of humankind!\n\n###### Figure 14-23. The 2013 Kumbh Mela, as captured by an attendee (image source)\n\n## Face Detection in Seeing AI\n\nMicrosoft's Seeing AI app (for the blind and low-vision community) provides a real-time facial detection feature with which it informs the user of the people in front of the phone's camera, their relative locations, and their distance from the camera. A typical guidance might sound like \"One face in the top-left corner, four feet away.\" Additionally, the app uses the facial-detection guidance to perform facial recognition against a known list of faces. If the face is a match, it will announce the name of the person, as well. An example of vocal guidance would be \"Elizabeth near top edge, three feet away,\" as demonstrated in Figure 14-24.\n\nTo identify a person's location in the camera feed, the system is running on a fast mobile-optimized object detector. A crop of this face is then passed for further processing to age, emotion, hair style, and other recognition algorithms on the cloud from Microsoft's Cognitive Services. For identifying people in a privacy-friendly way, the app asks the user's friends and family to take three selfies of their faces and generates a feature representation (embedding) of the face that is stored on the device (rather than storing any images). When a face is detected in the camera live feed in the future, an embedding is calculated and compared with the database of embeddings and associated names. This is based on the concept of one-shot learning, similar to Siamese networks that we saw in Chapter 4. Another benefit of not storing images is that the size of the app doesn't increase drastically even with a large number of faces stored.\n\n###### Figure 14-24. Face detection feature on Seeing AI\n\n## Autonomous Cars\n\nSeveral self-driving car companies including Waymo, Uber, Tesla, Cruise, NVIDIA, and others use object detection, segmentation, and other deep learning techniques for building their self-driving cars. Advanced Driver Assistance Systems (ADAS) such as pedestrian detection, vehicle type identification, and speed limit sign recognition are important components of a self-driving car.\n\nFor the decision-making involved in self-driving cars, NVIDIA uses specialized networks for different tasks. For example, WaitNet is a neural network used for doing fast, high-level detection of traffic lights, intersections, and traffic signs. In addition to being fast, it is also meant to withstand noise and work reliably in tougher conditions such as rain and snow. The bounding boxes detected by WaitNet are then fed to specialized networks for more detailed classification. For instance, if a traffic light is detected in a frame, it is then fed into LightNet\u2014a network used for detecting traffic light shape (circle versus arrow) and state (red, yellow, green). Additionally, if a traffic sign is detected in a frame, it's fed into SignNet, a network to classify among a few hundred types of traffic signs from the US and Europe, including stop signs, yield signs, directional information, speed limit information, and more. Cascading output from faster networks to specialized networks helps improve performance and modularize development of different networks.\n\n###### Figure 14-25. Detecting traffic lights and signs on a self-driving car using the NVIDIA Drive Platform (image source)\n\n# Summary\n\nIn this chapter, we explored the types of computer-vision tasks, including how they relate to one another and how they are different from one another. We took a deeper look at object detection, how the pipeline works, and how it evolved over time. We then collected data, labeled it, trained a model (with and without code), and deployed the model to mobile. We next looked at how object detection is being used in the real world by industry, government, and non-governmental organizations (NGOs). And then we took a sneak peek into the world of object segmentation. Object detection is a very powerful tool whose vast potential is still being discovered every day in numerous areas of our lives. What innovative use of object detection can you think of? Share it with us on Twitter @PracticalDLBook.\n\n# Chapter 15. Becoming a Maker: Exploring Embedded AI at the Edge\n\nContributed by guest author: Sam Sterckval\n\nYou know how to build a great AI application, but you want more. You don't want to be limited to just running AI software on some computer, you want to bring it out in the real physical world. You want to build devices to make things more interactive, to make life easier, to serve humanity, or perhaps just for the fun of it. Maybe you want to build an interactive painting that smiles at you when you look at it. A camera on your door that makes a loud alarm when an unauthorized person attempts to steal delivered packages. Maybe a robotic arm that sorts recyclables and trash. A device in the woods to prevent wildlife poaching, perhaps? Or a drone that can autonomously survey large areas and identify people in distress during floods. Maybe even a wheelchair that could drive on its own. What you need is a smart electronic device, but how would you build it, what would it cost, how powerful would it be? In this chapter, we begin to address those questions.\n\nWe look at how to implement AI on an embedded device\u2014a device that you might use in a \"maker\" project. Makers are people with a DIY spirit who use their creativity to build something new. Often starting as amateur hobbyists, makers are fun-loving problem solvers, roboticists, innovators, and sometimes entrepreneurs.\n\nThe aim of this chapter is to spark your ability to select the appropriate device for the task (which means not trying to run a heavy GAN on a tiny CPU or get a quadrillion-core GPU to run a \"Not Hotdog\" classifier), and set it up for the tests as quickly and easily as possible. We do this by exploring a few of the better-known devices out there, and seeing how we can use them to perform inferencing of our model. And finally, we look at how makers around the world are using AI to build robotic projects.\n\nLet's take our first step and look at the current landscape of embedded AI devices.\n\n# Exploring the Landscape of Embedded AI Devices\n\nIn this section, we explore a few well-known embedded AI devices, listed in Table 15-1. We talk about their inner workings and the differences between them before we go into testing.\n\nTable 15-1. Device list **Raspberry Pi 4** | **The most famous single-board computer, as of this writing**  \n---|---  \n**Intel Movidius NCS2** | A USB accelerator using a 16-core Visual Processing Unit (VPU)  \n**Google Coral USB** | A USB accelerator using a custom Google Application-Specific Integrated Circuit (ASIC)  \n**NVIDIA Jetson Nano** | A single-board computer using a combination of CPU and 128-core CUDA GPU  \n**PYNQ-Z2** | A single-board computer using the combination of CPU and 50k CLB Field-Programmable Gate Array (FPGA)\n\nInstead of saying one is better than the other, we want to learn how to choose a setup for a particular project. We don't typically see a Ferrari during the morning commute. Smaller, more compact cars are more common, and they get the job done just as well, if not better. Similarly, using a powerful $1,000-plus NVIDIA 2080 Ti GPU might be overkill for a battery-powered drone. Following are some questions we should be asking ourselves to understand which of these edge devices would best suit our needs:\n\n  1. How big is the device (for instance, compared to a coin)?\n\n  2. How much does the device cost? Consider whether you're on a tight budget.\n\n  3. How fast is the device? Will it process a single FPS or 100 FPS?\n\n  4. How much power (in watts) does the device typically need? For battery-powered projects, this can be essential.\n\nWith these questions in mind, let's explore some of the devices in Figure 15-1 one by one.\n\n###### Figure 15-1. Family photo of Embedded AI devices; starting at the top, going clockwise: PYNQ-Z2, Arduino UNO R3, Intel Movidius NCS2, Raspberry Pi 4, Google Coral USB Accelerator, NVIDIA Jetson Nano, and a \u20ac1 coin for reference in the middle\n\n## Raspberry Pi\n\nBecause we will be talking about embedded devices for makers, let's begin with the one universally synonymous with electronic projects: the Raspberry Pi (Figure 15-2). It's cheap, it's easy to build on, and it has a huge community.\n\n###### Figure 15-2. Raspberry Pi 4\n\n**Size** | 85.6 mm x 56.5 mm  \n---|---  \n**Price** | Starting at $35  \n**Processing unit** | ARM Cortex-A72 CPU  \n**Power rating** | 15W\n\nFirst of all, what is a Raspberry Pi? It's a \"single board computer,\" which simply means that all calculations can be performed on a single printed circuit board (PCB). The fourth version of the single-board computer houses a Broadcom SoC (system-on-a-chip) containing (most important) an ARMv8-A72 quad-core CPU, a VideoCore VI 3D unit, and some video decoders and encoders. It comes with a choice of RAM size, up to 4 GB.\n\nThis version is a big step up from the Raspberry Pi 3 in terms of performance, but it does lose a bit of efficiency. This is due to the Raspberry Pi 3 having an ARMv8-A53 core, which is the high-efficiency version, whereas the ARMv8-A72 core (used in version 4) is the high-performance version. We can see this reflected in the power supply recommendation, which was 2.5 amps for the Raspberry Pi 3. For the Raspberry Pi 4, this became 3 amps. It is also important to note that the Raspberry Pi 4 has USB 3 ports, where the Raspberry Pi 3 only has USB 2 ports. This will prove to be important information in the future.\n\nOn to some machine learning stuff now. However powerful the Raspberry Pi 4 has become, it still mainly consists of four sequential CPU cores (which now support Out of Order OoO] execution). It has the VideoCore VI, but there is no TensorFlow version available for this architecture as of this writing. Koichi Nakamura from Idein Inc. has built [py-videocore, a Python library for accessing the Quad Processing Units (QPUs; the GPU-like cores of the Raspberry Pi's SoC). He has accelerated neural networks with it before, but simple acceleration of TensorFlow isn't possible yet. C++ libraries for accessing these cores are also available. But as you might suspect, these cores are not that powerful, so even when used to accelerate neural networks, they might not yield the desired results. For the sake of simplicity, we will not go into these libraries, because digging into the algorithms is beyond the scope of this book. And, and as we will see further on, this might not be necessary at all.\n\nIn the end, the Raspberry Pi has proven to be an immensely useful piece of hardware for numerous tasks. It is often used for educational purposes, and by makers. You would be surprised how many industries use the Raspberry Pi in industrial environments (there is a thing called the netPi, which is just a Raspberry Pi with a robust enclosure).\n\n## Intel Movidius Neural Compute Stick\n\nLet's now dive straight into one of the reasons the Raspberry Pi is the go-to base for a lot of projects: The Intel Movidius Neural Compute Stick 2 (Figure 15-3), the second generation of a USB accelerator created by Intel.\n\n###### Figure 15-3. Intel Neural Compute Stick 2\n\n**Size** | 72.5 mm x 27 mm  \n---|---  \n**Price** | $87.99  \n**Processing unit** | Myriad X VPU  \n**Power rating** | 1W\n\nWhat it does is pretty simple: you give it a neural network and some data, and it does all the calculations necessary for inference. And it's connected only through USB, so you can just hook it up to your Raspberry Pi, run your inferencing on the USB accelerator, and free up the CPU of your Raspberry Pi to do all the other cool stuff you want it to do.\n\nIt is based on the Myriad VPU. The first generation had a Myriad 2 VPU, this one has a Myriad X VPU. The VPU contains 16 SHAVE (Streaming Hybrid Architecture Vector Engine) cores, which are kind of like GPU cores but less tailored to graphics-related operations. It features on-chip RAM, which is especially useful when doing neural network inferencing because these networks tend to create a lot of data while they compute, which then can be stored right next to the core, which reduces access time drastically.\n\n## Google Coral USB Accelerator\n\nThe Google Coral (Figure 15-4), containing the Google Edge TPU, is the second USB accelerator we will discuss. First, a little explanation on why we are looking at two different USB accelerators. The Intel stick, as mentioned, has a number of SHAVE cores, which have multiple available instructions, like GPU cores, and thus act as a processing unit. The Google Edge TPU, on the other hand, is an ASIC (Application Specific Integrated Circuit), which also does some processing, but it serves a single purpose (hence the \"Specific\" keyword). An ASIC comes with a few properties inherent to the hardware, some of which are really nice, others less so:\n\nSpeed\n\nBecause all electronic circuits inside the TPU serve a single purpose, there is no overhead in terms of decoding operations. You pump in input data and weights, and it gives you a result, almost instantly.\n\nEfficiency\n\nAll ASICs serve a single purpose, so no extra energy is required. The performance\/watt figure for ASICs is usually the highest in the business.\n\nFlexibility\n\nAn ASIC can do only what it was designed for; in this case, that would be accelerating TensorFlow Lite neural networks. You will need to stick to the Google Edge TPU compiler and 8-bit _.tflite_ models.\n\nComplexity\n\nGoogle is, in essence, a software company and it knows how to make things easy to use. That is exactly what it has done here, as well. The Google Coral is incredibly easy to get started with.\n\nSo how does the Google Edge TPU work? Information on the Edge TPU itself has not been shared, but information about the Cloud TPU has, so the assumption is that the Edge TPU works in broadly the same way as the Cloud TPU. It has dedicated hardware to do the multiply-add, activation, and pooling operations. All the transistors on the chip have been connected in such a way that it can take weights and input data, and it will calculate the output in a highly parallel fashion. The single biggest part of the chip (apart from the on-chip memory, that is) is a part that does exactly what it sounds like: the \"Matrix Multiply Unit.\" It uses a rather clever, though not so new, principle called _systolic execution_. This execution principle helps lower memory bandwidth by storing intermediate results in the processing elements rather than in memory.\n\n###### Figure 15-4. Google Coral USB Accelerator\n\n**Size** | 65 mm x 30 mm  \n---|---  \n**Price** | $74.99  \n**Processing unit** | Google Edge TPU  \n**Power rating** | 2.5W\n\n###### Note\n\nWhat's the main difference between the two USB accelerators we've discussed? The Google Coral is more powerful, but a (little) less flexible than the Intel Movidius NCS2. That being said, the Coral is far easier to set up and work with, certainly after you have the trained and converted model.\n\n## NVIDIA Jetson Nano\n\nOn to a different kind of hardware: the NVIDIA Jetson Nano (Figure 15-5), a single board AI computer. Kind of like the Raspberry Pi, but with a 128-core CUDA-enabled Maxwell GPU. The addition of the GPU makes it kind of similar to a Raspberry Pi with an Intel NCS2, but where the NCS2 has 16 cores, this Jetson has 128. What more does it contain? A quad-core A57 ARMv8 CPU, which is the predecessor of the ARMv8 A72 in the Raspberry Pi 4 (it is also a few months older than the Raspberry Pi 4), 4 GB of Low-Power Double Data Rate 4 (LPDDR4) RAM memory, which is, quite conveniently, shared between the CPU and GPU, which allows you to process data on the GPU, without copying it.\n\n###### Figure 15-5. NVIDIA Jetson Nano\n\n**Size** | 100 mm x 79 mm  \n---|---  \n**Price** | $99.00  \n**Processing unit** | ARM A57 + 128 core Maxwell GPU  \n**Power rating** | 10W (Can spike higher under high load)\n\nThe thing about this single board computer is that, as said before, these GPU cores are CUDA enabled, so yes, they can run TensorFlow-GPU or any-other-framework, which will make a big difference compared to a Raspberry Pi, especially when you also plan to train networks. Also, it's dirt cheap for a GPU. At this point in time, the Jetson Nano is yours for $99, and this includes a rather high-performance ARM CPU and a 128-core GPU. In comparison, the Raspberry Pi 4 with 4 GB of memory is around $55, the Coral USB accelerator is around $75, and the Movidius NCS2 is about $90, as well. The latter two are not standalone, and will at least need an additional Raspberry Pi to actually do something, and the Pi has no GPU that can easily accelerate deep learning applications.\n\nOne more note about the Jetson Nano: it can accelerate default TensorFlow 32-bit floating-point operations, but it will get much more efficient when 16-bit floating-point operations are used, and even more efficient if you use its own TensorRT framework. Luckily, the company has a nice little open source library called TensorFlow-TensorRT (TF-TRT) that will accelerate the available operations with TensorRT automatically while allowing TensorFlow to do the rest. This library offers grand speedups of around four times compared to TensorFlow-GPU. With all this in mind, this makes the Jetson Nano easily the most flexible device.\n\n##### From the Creators' Desk\n\nBy John Welsh and Chitoku Yato, NVIDIA Jetson Nano Team\n\nIt's exciting to be working on the NVIDIA Jetson platform, which provides the software and hardware capabilities to drive AI-powered robots, drones, Intelligent Video Analytics (IVA) applications, and other autonomous machines that think for themselves. Jetson Nano brings the power of modern AI to a small, easy-to-use platform.\n\nA few years ago, we had a group of high school interns who were developing summer projects using deep learning at the edge. The projects they came up with were creative, technically impressive, and fun. We were very excited to come out with Jetson Nano, because we felt that these kinds of projects were exactly what Jetson Nano would enable for a larger audience.\n\nIt's really rewarding to be part of the talented team here at NVIDIA that's bringing the tools to inspire these communities to get up and running fast with AI. We're continually amazed by some of the cool projects that makers, developers, and learners are building on Jetson Nano\u2014from DIY robocars to home automation\u2014and it's just the beginning. Our goal is to make AI more accessible to all.\n\n## FPGA + PYNQ\n\nNow is the time to fasten your seatbelts, because we are about to take a deep dive into the dark world of electronics. The PYNQ platform (Figure 15-6), based on the Xilinx Zynq family of chips is, for the most part, a totally different side of the electronics world compared to the other devices discussed in this topic. If you do a bit of research, you'll find out it has a dual-core ARM-A9 CPU, at a whopping 667 MHz. The first thing you'll think is \"Are you serious? That is ridiculous compared to the 1.5 GHz quad-core A72 from the Raspberry Pi?!\" And you'd be right, the CPU in this thing is, for the most part, absolutely worthless. But it has something else on board\u2014an FPGA.\n\n###### Figure 15-6. Xilinx PYNQ-Z2\n\n**Size** | 140 mm x 87 mm  \n---|---  \n**Price** | $119.00  \n**Processing unit** | Xilinx Zynq-7020  \n**Power rating** | 13.8W\n\n### FPGAs\n\nTo understand what an FPGA is, we need to first look at some familiar concepts. Let's begin with a CPU: a device that knows a list of operations known as the Instruction Set Architecture (ISA). This is a set that defines everything the CPU can do, and usually contains operations such as \"Load Word\" (a word is typically a number of bits equal to the datapath size of the CPU, usually 32-bit or 64-bit), which will load some value into an internal register of the CPU, and \"Add,\" which can add up two of the internal registers, and store the result in a third register, for example.\n\nThe reason the CPU can do this is because it contains a whole bunch of transistors (look at these as electrical switches if you like) that are hardwired in such a way that the CPU automatically translates the operations and does whatever that operation was intended to do. A software program is just a really long list of these operations, in a precisely thought-out order. A single-core CPU will take in operation per operation and carry them out, at pretty impressive speeds.\n\nLet's look at parallelism. Neural networks, as you know, consist of mostly convolutions or linear nodes, which can all be translated to matrix multiplications. If we look at the mathematics behind these operations, we can spot that each output point of a layer can be calculated independently from the other output points, as demonstrated in Figure 15-7.\n\n###### Figure 15-7. Matrix multiplication\n\nWe can now see that all of these operations could actually be done in a parallel fashion. Our single-core CPU won't be able to do so, because it can do only one operation at a time, but that's where multicore CPUs come in. A quad-core CPU can do four operations at a time, which means a theoretical speedup of four times. The SHAVE cores in the Intel Movidius NCS2 and the CUDA cores in the Jetson Nano might not be as complex as a CPU core, but they are good enough for these multiplications and additions, and instead of having four of these, the NCS2 has 16, and the Jetson Nano has 128. A bigger GPU like a RTX 2080 Ti even has 4,352 CUDA cores. It's easy to see now why GPUs are better at performing deep learning tasks compared to CPUs.\n\nLet's get back to FPGAs. Whereas CPUs and GPUs are huge collections of transistors hardwired to carry out a set of instructions, you can think of an FPGA as that same huge collection of transistors, but not wired. You can choose how they are wired, and rewire them whenever you want, making them reconfigurable. You can wire them to be a CPU. You can also find schematics and projects for which people have wired them to be a GPU. But most interesting here is that they can even be wired to the exact same architecture as your deep learning neural network, which would actually make them a physical implementation of your network. The word \"wired\" is used intentionally here. Often, you'll find people talking about this configuration with the word \"program,\" but this can be confusing. What you're doing is reconfiguring the actual hardware, unlike a CPU or GPU for which you download a program that the hardware can run.\n\n### PYNQ platform\n\nWe usually call the \"wiring\" file for an FPGA a _bitstream_ or a _bitmap_ , because what you flash onto the chip is basically just a map of the connections that should be made. As you can imagine, making these bitstreams is quite a lot more complicated than running a Python script. That's where PYNQ comes in. Its tagline is \"Python productivity for Zynq.\" The company installs the PYNQ image that automatically runs a Jupyter Notebook on the ARM inside the chip, and it comes with a few basic bitstreams; however, more bitstreams will likely become available in the future. Within the PYNQ world, these bitstreams are called _overlays_.\n\nIf you go looking for examples, you'll quickly find an example called \"BNN-PYNQ,\" which has a simple VGG-style, six-layer CNN that can run at 3,000-ish FPS on the PYNQ-Z1 and Z2, and close to 10,000 FPS on the ZCU104, which has the Ultrascale+ version of the Zynq chip onboard. These numbers look pretty insane, but take into account that they run on 32x32 pixel images, rather than the usual 224x224 pixel images and that this network is \"binarized,\" which means it has weights and activations of one bit, instead of the 32-bits of TensorFlow. To have a better comparison of performance, I tried to recreate a similar network in Keras.\n\nI built the FP32 model and trained it on the CIFAR-10 dataset. The CIFAR-10 dataset is easily available with Keras using `keras.datasets.cifar10`. The FP32 model reached a roughly 17% error rate, which is, surprisingly, only 2% better than the binary model. Inference speeds are around 132 FPS on an Intel i9 eight-core CPU. There is, to my knowledge, no easy way of using a binary network efficiently and easily on a CPU\u2014you'll need to dig into some specific Python packages or some C code to get the most out of the hardware. You could potentially achieve a speedup of three to five times. This would, however, still be far less than a low-end FPGA, and a CPU will usually draw more power in doing so. Of course, everything has a flip side, and for FPGAs that has to be the complexity of the design. Open source frameworks exist, namely the FINN framework, backed by Xilinx. Other manufacturers offer additional frameworks, but none of them come close to how easy to use software packages and frameworks like TensorFlow are. Designing a neural network for an FPGA would also involve a lot of electronics knowledge, and thus is beyond the scope of this book.\n\n## Arduino\n\nArduino (Figure 15-8) can make your life a lot easier when trying to interact with the real world through sensors and actuators.\n\nAnMicroController Units (MCUs) Arduino is a microcontroller development board built around the 8-bit Advanced Virtual RISC (AVR) ATMega328p microcontroller running at 16 MHz (so yeah, let's not try to run a decent neural network on that). A microcontroller is something everyone has come into contact with\u2014you can find microcontrollers in almost everything that is slightly electronic, ranging from every screen\/TV, to your desktop keyboard, heck bicycle lights may even contain a microcontroller to enable the flashing mode.\n\nThese boards come in a variety of shapes, sizes, and performances, and although the MCU in the Arduino UNO is pretty low performance, more potent MCUs are available. The ARM Cortex-M series is probably the most well known. Recently TensorFlow and uTensor (an extremely lightweight machine learning inference framework aimed at MCUs) joined forces to enable the possibility to run TensorFlow models on these MCUs.\n\n##### From the Creator's Desk\n\nBy Pete Warden, technical lead for Mobile and Embedded TensorFlow, author of _TinyML_\n\nWhen I joined Google in 2014, I was amazed to find that the \"Ok Google\" team was using models that were only 13 KB in size, and ran on tiny DSPs (Digital Signal Processors), to spot wake words. They were very experienced and pragmatic engineers, so I knew they weren't doing this to be fashionable; they'd experimented and this was the best solution they could find. At that time, I was used to models being megabytes in size, so finding out that tiny models could be useful was a revelation, and I couldn't help wondering what other problems they could be useful for. Since then, I've been keen to open up the technology they pioneered to a wider audience so we can discover what some of these applications might be. This is why we released our first version of TensorFlow Lite for Microcontrollers in early 2019, and it has been amazing to hear from all of the creative engineers building innovative new features, and even entirely new products, using these new capabilities.\n\n###### Figure 15-8. Arduino UNO R3\n\n**Size** | 68.6 mm x 53.3 mm  \n---|---  \n**Price** | Less than $10.00  \n**Processing unit** | AVR ATMega328p  \n**Power rating** | 0.3W\n\nAlthough the Arduino UNO is not really useful for performing lots of calculations, it has many other benefits that make it a mandatory part of the workbench of a maker. It's cheap, it's simple, it's reliable, and it has a huge community surrounding it. Most sensors and actuators have some kind of library and shield for Arduino, which makes it a really easy platform to interact with. The AVR architecture is a kind of dated 8-bit modified Harvard architecture, but it's really easy to learn, especially with the Arduino framework on top of it. The huge community around Arduino obviously brings a lot of tutorials with it, so just look up any Arduino tutorial for your sensor of choice, and you'll surely find what you need for your project.\n\n###### Tip\n\nUsing Python's serial library, you can easily interface with an Arduino using a Universal Asynchronous Receiver\/Transmitter (UART) through a USB cable to send commands or data back and forth.\n\n# A Qualitative Comparison of Embedded AI Devices\n\nTable 15-2 summarizes the platforms we've talked about, and Figure 15-9 plots performance versus complexity for each device.\n\nTable 15-2. Pros and cons of tested platforms **Embedded device** | **Pros** | **Cons**  \n---|---|---  \n**Raspberry Pi 4** |\n\n  * Easy\n\n  * Large community\n\n  * Readily available\n\n  * Cheap\n |   * Lacks computing power\n\n**Intel Movidius NCS2** |\n\n  * Easy optimizations\n\n  * Supports most platforms\n |   * Rather expensive for the speedup\n\n  * Not so powerful\n\n**Google Coral USB** |\n\n  * Easy getting started\n\n  * Huge speedup\n\n  * Price versus speedup\n |   * Supports only _.tflite_\n\n  * Needs eight-bit quantization\n\n**NVIDIA Jetson Nano** |\n\n  * All-in-one\n\n  * Training is possible\n\n  * Cheap\n\n  * Really easy to achieve decent performance\n\n  * CUDA GPU\n |   * Performance might still be lacking\n\n  * Advanced optimizations can be complex\n\n**PYNQ-Z2** |\n\n  * Potentially huge power efficiency\n\n  * Actual hardware design\n\n  * Potentially huge performance\n\n  * Even more versatile than the Jetson Nano\n |   * Massively complex\n\n  * Long design flow\n\n###### Figure 15-9. Performance versus complexity-to-use (the size of the circle represents price)\n\nIn this section, we attempt to test some platforms by performing 250 classifications of the same image using MobileNetV2, with the top platform trained on the ImageNet dataset. We use the same image every time because this will allow us to eliminate part of the data bottlenecks that can occur in systems that don't have enough RAM to store all weights and lots of different images.\n\nOk, let's find an image first. Who doesn't like cats? So, let's have an image of a beautiful cat (Figure 15-10) that is exactly 224 x 224 pixels, that way we don't need to scale the image.\n\n###### Figure 15-10. The image of the cat we will be using for our experiments\n\nIt's always nice to have a benchmark, so let's run the model on our PC first. Chapter 2 explains how to do this, so let's go ahead and write a script that will invoke the prediction 250 times, and measure how long this takes. You can find the full script on the book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_) at _code\/chapter-15_ :\n\n    $ python3 benchmark.py\n    Using TensorFlow backend.\n    tf version : 1.14.0\n    keras version : 2.2.4\n    input tensor shape : (1, 224, 224, 3)\n    warmup prediction\n    [('n02124075', 'Egyptian_cat', 0.6629321)]\n    starting now...\n    Time[s] : 16.704\n    FPS     : 14.968\n    Model saved.\n\nNow that we've established our benchmark, it's time to begin looking at how to run this model on the embedded devices and see whether and how we can take the performance to a useful level.\n\n# Hands-On with the Raspberry Pi\n\nWe begin with the most well-known and most basic of the previously discussed devices: the Raspberry Pi (RPi for short).\n\nWe start by installing Raspbian, a Linux variant made especially for the Raspberry Pi. For the sake of brevity, let's assume that we have a Raspberry Pi with everything installed, updated, connected, and ready. We can go straight to installing TensorFlow on the Pi, which should work using the `pip` installer:\n\n    $ pip3 install tensorflow\n\n###### Note\n\nDue to the recent switch to Raspbian Buster, we encountered some problems, which were resolved by using the following `piwheel`:\n\n    $ pip3 install\n    https:\/\/www.piwheels.org\/simple\/tensorflow\/tensorflow\n    -1.13.1-cp37-none-linux_armv7l.whl\n\nInstalling Keras is straightforward with `pip`, but don't forget to install `libhdf5-dev`, which you will need if you want to load weights into a neural network:\n\n    $ pip3 install keras\n    $ apt install libhdf5-dev\n\nBecause installing OpenCV, called in code as cv2, on the RPi can be a burden (especially when a recent switch of OS has happened), we can load the image using the PIL instead of OpenCV. This means replacing the import of cv2 with an import for PIL, and changing the code to load the image to use this library:\n\n    #input_image = cv2.imread(input_image_path)\n    #input_image = cv2.cvtColor(input_image, cv2.COLOR_BGR2RGB)\n    input_image = PIL.Image.open(input_image_path)\n    input_image = np.asarray(input_image)\n\n###### Tip\n\nUnlike OpenCV, PIL loads images in the RGB format, so a conversion from BGR to RGB is no longer needed.\n\nAnd that's it! The exact same script should run on your Raspberry Pi now:\n\n    $ python3 benchmark.py\n    Using TensorFlow backend.\n    tf version : 1.14.0\n    keras version : 2.2.4\n    input tensor shape : (1, 224, 224, 3)\n    warmup prediction\n    [('n02124075', 'Egyptian_cat', 0.6629321)]\n    starting now...\n    Time[s] : 91.041\n    FPS     : 2.747\n\nAs you can see, it runs a lot slower, dropping to less than three FPS. Time to think about how we can speed this up.\n\nThe first thing we could do is have a look at TensorFlow Lite. TensorFlow has, as discussed in Chapter 13, a converter built in, which lets us easily convert any model to a TensorFlow Lite ( _.tflite_ ) model.\n\nWe proceed by writing a small script, very similar to the original benchmark script, which will just set up everything to use the TensorFlow Lite model and run the 250 predictions. This script is, of course, also available on the book's GitHub webiste (see _http:\/\/PracticalDeepLearning.ai_).\n\nLet's go ahead and run this so that we can evaluate our effort to speed things up.\n\n    $ python3 benchmark_tflite.py\n    Using TensorFlow backend.\n    input tensor shape : (1, 224, 224, 3)\n    conversion to tflite is done\n    INFO: Initialized TensorFlow Lite runtime.\n    [[('n02124075', 'Egyptian_cat', 0.68769807)]]\n    starting now (tflite)...\n    Time[ms] : 32.152\n    FPS      : 7.775\n\n###### Note\n\nThe Google Coral website has 8-bit quantized CPU models available, which you can run using TensorFlow Lite.\n\nFor the Raspberry Pi 4, we observed an increase in speed of three times (see Table 15-3). Not bad for a quick try, but we still only achieved 7.8 FPS, which is kind of nice and perfectly fine for a lot of applications. But what if our original plan was to do something on live video, for example? Then this is not enough. Quantizing the model can achieve even more performance, but this will be only a marginal increase given that the CPU in the Raspberry Pi has not been optimized for 8-bit integer operations.\n\nTable 15-3. Raspberry Pi benchmarks **Setup** | **FPS**  \n---|---  \nRaspberry Pi 4 (tflite, 8-bit) | 8.6  \nRaspberry Pi 4 (tflite) | 7.8  \nRaspberry Pi 3B+ (tflite, 8-bit) | 4.2  \nRaspberry Pi 4 | 2.7  \nRaspberry Pi 3B+ | 2.1\n\nSo how can we speed things up even more so that they become useful for applications like autonomous racecars or automatic drones? We touched on it before: the USB accelerators.\n\n# Speeding Up with the Google Coral USB Accelerator\n\nThat's right, these accelerators come in quite handy right now. All of our hardware setup can stay exactly the same, we just need to add a little bit here and there. So let's find out: how do we get the Google Coral USB Accelerator to run on the Raspberry Pi, and will it speed things up?\n\nFirst things first. Assuming that we have a USB accelerator on our hands, the first thing we should always do is look for a \"Getting Started\" guide from the vendor. Check out _https:\/\/coral.withgoogle.com\/_, head to _Docs >USB Accelerator>Get Started_, and you're rolling in mere minutes.\n\nAs of this writing, the Coral USB Accelerator is not yet fully compatible with the Raspberry Pi 4, but after fiddling around with the _install.sh_ script, it is pretty easy to get it running. You need to simply add a part to the install script so that it recognizes the Pi 4, as demonstrated in Figure 15-11.\n\n###### Figure 15-11. Google Coral install script changes\n\nNow the _install.sh_ script will run correctly. Then, we need to rename a _.so_ file, so that it will also work with Python 3.7. To do this, use the Unix copy command:\n\n    $ sudo cp \/usr\/local\/lib\/python3.7\/dist-\n    packages\/edgetpu\/swig\/_edgetpu_cpp_wrapper.cpython-\n    35m-arm-linux-gnueabihf.so\n    \/usr\/local\/lib\/python3.7\/dist-\n    packages\/edgetpu\/swig\/_edgetpu_cpp_wrapper.cpython-\n    37m-arm-linux-gnueabihf.so\n\nAfter you do this, everything should work correctly, and the demonstrations from Google should run.\n\nNow that we have the Coral running, let's try to benchmark the same MobileNetV2 model again. But this time, it must be the quantized version of it, because Google's Edge TPU supports only 8-bit integer operations. Google supplied us with the quantized and converted MobileNetV2 model, ready to use with the Edge TPU. We can download it from the Coral website under _Resources > See pre-compiled models > MobileNetV2(ImageNet) > Edge TPU Model_.\n\n###### Note\n\nIf you want to create, train, quantize, and convert your own model for the Edge TPU, this is a bit more work and cannot be done using Keras. You can find information on how to do this on the Google Coral website.\n\nWe have a working USB accelerator and a model to run on it. Now, let's make a new script to test its performance. The file we are about to make is again very similar to the previous one and takes a lot straight from the examples Google supplied with the Coral. And, as always, this file is available for download at GitHub website (see _http:\/\/PracticalDeepLearning.ai_):\n\n    $ python3 benchmark_edgetpu.py\n    INFO: Initialized TensorFlow Lite runtime.\n    warmup prediction\n    Egyptian cat\n    0.59375\n    starting now (Edge TPU)...\n    Time[s] : 1.042\n    FPS     : 240.380\n\nYes, that is correct. It did all 250 classifications already! For the Raspberry Pi 4 with USB3, it took only 1.04 seconds, which translates to 240.38 FPS! Now that's a speedup. As you can expect with these speeds, live video would be no problem at all.\n\nLots of precompiled models are available, for all kinds of different purposes, so check them out. You might be able to find a model that suits your needs so that you don't need to go through the flow (or struggle) to create, train, quantize, and convert your own.\n\nThe Raspberry Pi 3 has no USB3 ports, only USB2. We can clearly see in Table 15-4 that this creates a data bottleneck for the Coral.\n\nTable 15-4. Google Coral benchmarks **Setup** | **FPS**  \n---|---  \ni7-7700K + Coral (tflite, 8-bit) | 352.1  \nRaspberry Pi 4 + Coral (tflite, 8-bit) | 240.4  \nJetson Nano + Coral (tflite, 8-bit) | 223.2  \nRPi3 + Coral (tflite, 8-bit) | 75.5\n\n# Port to NVIDIA Jetson Nano\n\nThat USB accelerator is nice, but what if your project needs a real niche model, trained on some crazy, self-made, dataset? As we said earlier, creating a new model for the Coral Edge TPU is quite a bit more work than just creating a Keras model. So, is there an easy way to have a custom model running on the edge with some decent performance? NVIDIA to the rescue! With its Jetson Nano, the company has created a replacement for the Raspberry Pi that has a CUDA enabled, and rather efficient, GPU on board, which lets you accelerate not only TensorFlow, but anything you like.\n\nAgain, we must begin by installing all needed packages. First, you want to download a copy of NVIDIA's JetPack, which is its version of a Debian-like OS for the Jetson platform. Then, to install TensorFlow and the needed packages, NVIDIA walks us through how to do that here.\n\n###### Note\n\nNote there is a known issue with `pip3`: \"`cannot import name 'main'`\".\n\nHere's the solution:\n\n    $ sudo apt install nano\n    $ sudo nano \/usr\/bin\/pip3\n\n    Replace : main => __main__\n    Replace : main() => __main__._main()\n\n###### Tip\n\nUpdating all `pip` packages can take a long time; if you want to ensure that your Jetson did not freeze, you might want to first install `htop`:\n\n    $ sudo apt install htop\n    $ htop\n\nThis lets you monitor CPU utilization. As long as this is working, your Jetson is, too.\n\n###### Note\n\nThe Jetson Nano tends to run rather hot when compiling, so we recommend using some kind of fan. The development board has a connector, but keep in mind this will feed 5V to the fan, whereas most fans are rated for 12V.\n\nSome 12V fans might do the trick out of the box, some might need a little push to get started. You will need a 40 mm x 40 mm fan. 5V versions are also available.\n\nUnfortunately, the NVIDIA walkthrough is not as easy as the Google one. Chances are you will go through a bit of a struggle with the occasional table-flip moment.\n\nBut hang in there; you'll get there eventually.\n\n###### Tip\n\nInstalling Keras on the Jetson requires scipy, which requires `libatlas-base-dev` and `gfortran`, so start by installing the latter, and move to the front:\n\n    $ sudo apt install libatlas-base-dev gfortran\n    $ sudo pip3 install scipy\n    $ sudo pip3 install keras\n\nAfter everything is done, we can choose to run the _benchmark.py_ file directly, which will be, because of the GPU, a lot faster than it was on the Raspberry Pi:\n\n    $ python3 benchmark.py\n    Using TensorFlow backend.\n    tf version : 1.14.0\n    keras version : 2.2.4\n    input tensor shape : (1, 224, 224, 3)\n    warmup prediction\n    [('n02124075', 'Egyptian_cat', 0.6629321)]\n    starting now...\n    Time[s] : 20.520\n    FPS     : 12.177\n\nThis immediately shows the power of the Jetson Nano: it runs almost exactly like any other PC with a GPU. However, 12 FPS is still not huge, so let's look at how to optimize this.\n\n###### Tip\n\nYou also can attach the Google Coral to your Jetson Nano, which opens up the possibility to run one model on the Coral, and another one on the GPU simultaneously.\n\nThe GPU of the Jetson is actually built specifically for 16-bit-floating point operations, so inherently, this will be the best trade-off between precision and performance. As mentioned earlier, NVIDIA has a package called TF-TRT, which facilitates optimization and conversion. However, it is still a bit more complex than the Coral example (keep in mind that Google supplied us with the precompiled model file for the Coral). You'll need to freeze the Keras model and then create a TF-TRT inference graph. Doing this can take quite a while if you need to find everything laying around on GitHub, so it's bundled on this book's GitHub website (see _http:\/\/PracticalDeepLearning.ai_).\n\nYou can use the _tftrt_helper.py_ file to optimize your own models for inferencing on the Jetson Nano. All it really does is freeze the Keras model, removes the training nodes, and then use NVIDIA's TF-TRT Python package (included in TensorFlow Contrib) to optimize the model.\n\n    $ python3 benchmark_jetson.py\n    FrozenGraph build.\n    TF-TRT model ready to rumble!\n    input tensor shape : (1, 224, 224, 3)\n    [[('n02124075', 'Egyptian_cat', 0.66293204)]]\n    starting now (Jetson Nano)...\n    Time[s] : 5.124\n    FPS     : 48.834\n\n48 FPS, that's a speedup of four times with just a few lines of code. Exciting, isn't it? You can achieve similar speedups on your own custom models, tailored to the specific needs of your project; these performance benchmarks can be found in Table 15-5.\n\nTable 15-5. NVIDIA Jetson Nano benchmarks **Setup** | **FPS**  \n---|---  \nJetson Nano + Coral (tflite, 8-bit) | 223.2  \nJetson Nano (TF-TRT, 16-bit) | 48.8  \nJetson Nano 128CUDA | 12.2  \nJetson Nano (tflite, 8-bit) | 11.3\n\n# Comparing the Performance of Edge Devices\n\nTable 15-6 shows a quantitative comparison of several edge devices running the MobileNetV2 model.\n\nTable 15-6. Full benchmarking results **Setup** | **FPS**  \n---|---  \ni7-7700K + Coral (tflite, 8-bit) | 352.1  \ni7-7700K + GTX1080 2560CUDA | 304.9  \nRaspberry Pi 4 + Coral | 240.4  \nJetson Nano + Coral (tflite, 8-bit) | 223.2  \nRPi3 + Coral (tflite, 8-bit) | 75.5  \nJetson Nano (TF-TRT, 16-bit) | 48.8  \ni7-7700K (tflite, 8-bit) | 32.4  \ni9-9880HQ (2019 MacBook Pro) | 15.0  \nJetson Nano 128CUDA | 12.2  \nJetson Nano (tflite, 8-bit) | 11.3  \ni7-4870HQ (2014 MacBook Pro) | 11.1  \nJetson Nano (tflite, 8-bit) | 10.9  \nRaspberry Pi 4 (tflite, 8-bit) | 8.6  \nRaspberry Pi 4 (tflite) | 7.8  \nRPi3B+ (tflite, 8-bit) | 4.2  \nRaspberry Pi 4 | 2.7  \nRPi3B+ | 2.1\n\nWe can summarize the takeaways from these experiments as follows:\n\n  1. Good optimizations make a big difference. The world of computing optimization is still growing rapidly, and we will see big things coming up in the next few years.\n\n  2. More is always better when talking about computing units for AI.\n\n  3. ASIC > FPGA > GPU > CPU in terms of performance\/watt.\n\n  4. TensorFlow Lite is awesome, especially for small CPUs. Keep in mind that it is specifically aimed at small CPUs, and using it for x64 machines will not be that interesting.\n\n# Case Studies\n\nThis chapter is about makers. And it would be incomplete without showcasing what they make. Let's look at a few things that have been created by running AI on the edge.\n\n## JetBot\n\nNVIDIA has been at the forefront of the deep learning revolution, enabling researchers and developers with powerful hardware and software. And in 2019, they took the next step to enable makers, too, by releasing the Jetson Nano. We already know the power of this hardware, so now it's time to build something with it\u2014like a DIY miniature car. Luckily, NVIDIA has us covered with JetBot (Figure 15-12), the open source robot that can be controlled by an onboard Jetson Nano.\n\nThe JetBot wiki features beginner-friendly step-by-step instructions on building it. Here's a high-level look:\n\n  1. Buy the parts specified in the Bill of Materials, like the motor, caster, camera, and WiFi antenna. There are about 30 parts, costing around $150, on top of the $99 Jetson Nano.\n\n  2. It's time to channel our inner MacGyver, get a set of pliers and a cross-tip screwdriver, and assemble the parts. The end result should be something resembling Figure 15-12.\n\n  3. Next, we'd set up the software. This involves flashing the JetBot image onto an SD card, booting the Jetson Nano, connecting it to WiFi, and finally connecting to our JetBot from a web browser.\n\n  4. Finally, we run through the example notebooks provided. The provided notebooks enable us, through little code, to not only control the bot from a web browser, but also use its camera to collect training data from streaming video, train a deep learning model right on the device (the Jetson Nano is a mini-GPU after all), and use it for tasks such as avoiding obstacles and following objects such as a person or a ball.\n\n###### Figure 15-12. NVIDIA JetBot\n\nMakers are already using this base framework to extend the JetBot's capabilities, such as attaching lidar sensors to it for understanding the environment better, bringing JetBot to different physical forms like tanks, Roomba vacuum cleaners (JetRoomba), object-following spiders (JetSpider), and more. Because all roads eventually lead to racing cars, the Jetson Nano team eventually released a similar open source recipe book for a racing car called JetRacer. It's based on a faster chassis, higher camera frame rate, and optimizations for inference (with TensorRT) to handle the speed. The end result: JetRacer is already being raced at events like the DIY Robocars meetups.\n\nThe slowest step here is...waiting for the hardware to arrive.\n\n##### From the Creators' Desk\n\nBy John Welsh and Chitoku Yato, NVIDIA Jetson Nano Team\n\nWe wanted to develop a reference platform to show developers what's possible with this tiny piece of hardware\u2014the Jetson Nano. While dreaming up the project that would eventually turn into JetBot, we realized its platform requirements matched what we envisioned many makers would need for their next-level AI projects.\n\nSpecifically, JetBot required a platform capable of real-time image processing that worked with an affordable camera and would be easy to interface with maker hardware. As the engineering development of Jetson Nano matured, these boxes were checked. With Jetson Nano as the brain powering JetBot, we felt it would be an excellent platform for makers interested in getting started with AI for a wide range of projects.\n\nWe focused on making JetBot easy to use and educational to inspire developers to start creating with Jetson Nano. Although JetBot is shown in NVIDIA demonstrations across the globe, its vision is much more far-reaching. It's meant to be an example of what's possible when makers combine our technology with their creativity and ingenuity. Since JetBot's release, there have been many different spin-offs, and we hope it's a great guide to learn, explore, and, most important, have fun with AI.\n\n## Squatting for Metro Tickets\n\nWhen good health can't motivate people to exercise, what else would? The answer turns out to be...free train tickets! To raise public awareness for its growing obesity crisis, Mexico City installed ticket machines with cameras at subway stations that give free tickets, but only if you exercise. Ten squats get a free ticket (Figure 15-13). Moscow, too, implemented a similar system at its metro stations, but apparently, officials there had a much higher fitness standard at 30 squats per person.\n\nInspired by these, we could foresee how to build our own \"squat-tracker.\" We could achieve this in multiple ways, the simplest being to train our own classifier with the classes \"squatting\" and \"not squatting.\" This would clearly involve building a dataset with thousands of pictures of people either squatting or not squatting.\n\nA much better way to accomplish this would involve running PoseNet (which tracks body joints, as we saw in Chapter 10) on, say, Google Coral USB Accelerator. With its 10-plus FPS, we could track how many times the hip points drop low enough and count much more accurately. All we would need is a Raspberry Pi, a Pi Camera, a Coral USB Accelerator, and a public transport operator to provide a ticket printer and endless free tickets, and we could start making our cities fitter than ever.\n\n###### Figure 15-13. Squatting for tickets\n\n## Cucumber Sorter\n\nMakoto Koike, the son of cucumber farmers in Japan, observed that his parents spent a lot of time postharvest sorting through cucumbers. The sorting required separating cucumbers into nine categories based on fine-grained analysis of very small features such as minute texture differences, tiny scratches, and prickles, along with larger attributes such as size, thickness, and curvature. The system was complicated and hiring part-time workers was pretty much a no-go because of how long it would take to train them. He couldn't sit by idly watching his parents go through the labor-intensive task for more than eight hours every single day.\n\nOne detail we didn't mention yet is that Mr. Koike was formerly an embedded systems designer at Toyota. He realized that a lot of the visual inspection could be automated using the power of deep learning. He took more than 7,000 pictures of different cucumbers that his mother had manually sorted over a three-month period. He then trained a classifier that could look at a picture of a cucumber and predict with a high degree of accuracy to which category it belonged. But a classifier is not going to do much on its own now, will it?\n\nUsing his experience in embedded systems, he deployed a classifier to a Raspberry Pi that was connected to a camera, a conveyor-belt system, and a sorting arm that pushed each cucumber into one of the multiple bins based on the predictions from the Raspberry Pi (Figure 15-14). The Raspberry Pi ran a small cucumber\/not-cucumber classifier as a first pass. For images classified as cucumber, the Raspberry Pi would send the image to a server that ran the more sophisticated multiple category cucumber classifier. Keep in mind that this was in 2015, shortly after TensorFlow was released (long before the days of TensorFlow Lite, MobileNet, and the extra power of the current Raspberry Pi 4 or even 3). With a machine like this, his parents could spend more time in actual farming rather than sorting through already harvested cucumbers for days on end.\n\n###### Figure 15-14. Makoto Koike's cucumber sorting machine (image source)\n\nEven though the debate on AI taking over jobs appears daily in the news, Mr. Koike's story truly highlights the real value that AI provides: making humans more productive and augmenting our abilities, along with making his parents proud.\n\n# Further Exploration\n\nThere are two equally important aspects to focus on when becoming a maker: the hardware and the software. One is incomplete without the other. There are more DIY projects to explore than one can build in a lifetime. But the key is to find some that you're passionate about, build them end to end, and feel the excitement when they start to actually function. Gradually, you'll develop the know-how so that the next time you read a project, you can guess how to build it yourself.\n\nThe best way to get started is to get hands-on and build more and more projects. Some excellent sources for inspiration include _Hackster.io_ , _Hackaday.io_ , and _Instructables.com_ , which feature a large range of project themes, platforms, and skill levels (from beginner to advanced), often along with the step-by-step instructions.\n\nTo reduce the overhead of assembling hardware for beginners, there are several kits available in the market. For example, AI Kits for Raspberry Pi including AIY Vision Kit (Google), AIY Speech Kit (Google), GoPiGo (Dexter Industries), and DuckieTown platform (MIT), to name a few. They lower the barriers for getting started with electronic hardware projects, making the new area more approachable and hence usable for everyone, from high schoolers to rapid prototypers.\n\nOn the software side of things, running AI faster on low-power devices means making the models more performant. Model compression techniques like quantization and pruning (as mentioned in Chapter 6) can help get us there. To go even faster (at some loss in accuracy), BNNs, which represent models in one bit instead of the normal 32 bits, are a potential solution. Companies like XNOR.ai and Microsoft's Embedded Learning Library allow us to make such models. For running AI on even lower-power devices like microcontrollers with less than 100 KB of memory, Pete Warden's book _TinyML_ is an excellent learning resource.\n\n# Summary\n\nIn this chapter, we introduced some well-known embedded AI devices and looked at how to run a Keras model on some of them. We presented a high-level view of their inner workings, and how they achieve the computing capacity needed for running neural networks. Eventually benchmarking them to develop an intuition on which might be suitable for our projects depending on size, costs, latency, and power requirements. Going from platforms to robotic projects, we looked at some ways in which makers around the world are using them to build electronic projects. The devices discussed in this chapter are obviously just a handful of those available. New devices will keep popping up as the edge becomes a more desirable place to be for AI.\n\nWith the power of edge AI, makers can imagine and bring their dream projects to reality, which might have been considered science fiction just a few years ago. On your journey in this DIY AI world, remember, you are among a tight-knit community of tinkerers with a lot of openness and willingness to share. The forums are active, so if you ever get stuck on a problem, there are helpful souls to call upon. And hopefully, with your project, you can inspire others to become makers themselves.\n\n# Chapter 16. Simulating a Self-Driving Car Using End-to-End Deep Learning with Keras\n\nContributed by guest authors: Aditya Sharma and Mitchell Spryn\n\nFrom the Batmobile to Knightrider, from robotaxis to autonomous pizza delivery, self-driving cars have captured the imagination of both modern pop culture and the mainstream media. And why wouldn't they? How often does it happen that a science-fiction concept is brought to life? More important, autonomous cars promise to address some key challenges urban societies are facing today: road accidents, pollution levels, cities becoming increasingly more congested, hours of productivity lost in traffic, and the list goes on.\n\nIt should come as no surprise that a full self-driving system is quite complex to build, and cannot be tackled in one book chapter. Much like an onion, any complex problem contains layers that can be peeled. In our case, we intend to tackle one fundamental problem here: how to steer. Even better, we don't need a real car to do this. We will be training a neural network to drive our car autonomously within a simulated environment, using realistic physics and visuals.\n\nBut first, a brief bit of history.\n\n##### SAE Levels of Driving Automation\n\n\"Autonomous driving\" is a broad term that is often used to describe everything from Adaptive Cruise Control (ACC) technology on your Chrysler 200 to the steering wheel-free cars of the future. To bring some structure around the definition of autonomous driving (or the term more popular in Europe: automated driving), the Society of Automotive Engineers (SAE) International proposed to divide automation capabilities of driving vehicles into six levels, ranging from \"No Automation\" at Level 0 to \"Full Automation\" at Level 5. Figure 16-1 gives an overview of these. These levels are somewhat subject to interpretation, but they give the industry a reference point when talking about autonomous driving. ACC is an example of Level 1 automation, whereas Tesla's cars fall somewhere between Levels 2 and 3. Although there are many startups, technology companies, and automakers working on Level 4 cars, we haven't seen any in production yet. It will take us a while before we realize our dream of a truly fully autonomous Level 5 vehicle.\n\n###### Figure 16-1. The SAE Levels of Driving Automation (image source)\n\n# A Brief History of Autonomous Driving\n\nEven though it might seem like autonomous driving is a very recent technological development, research in this field originated more than 30 years ago. Scientists at Carnegie Mellon University first tried their hands at this seemingly impossible feat back in 1984 when they started working on what would be known as the Navlab 1. \"Navlab\" was a shorthand for Carnegie Mellon's Navigation Laboratory, a not-so-fancy name for what would turn out to be the first step into a very impressive future for humanity. The Navlab 1 was a Chevrolet panel van (Figure 16-2) with racks of hardware onboard that included, among other things, workstations for the research team and a Sun Microsystems supercomputer. It was considered cutting edge at the time, a self-contained computing system on wheels. Long-distance wireless communication was pretty much nonexistent back then, and cloud computing was not even a concept. Putting the entire research team in the back of a highly experimental self-driving van might seem like a dangerous idea, but the Navlab had a maximum speed of 20 MPH and navigated on only empty roads. This ensured the safety of the scientists working onboard. What is very impressive to note though, is that the technology used in this car laid the foundation for technology used in the self-driving cars of today. For example, Navlab 1 used video cameras and a rangefinder (an early version of lidar) to observe its surroundings. For navigation, they used a single-layer neural network to predict steering angles given sensor data. Pretty neat isn't it?\n\n###### Figure 16-2. The NavLab 1 in all its glory (image source)\n\n##### Talk About Your Robo-Achievers\n\nThere were many more versions of the Navlab 1. Navlab 5 was actually able to steer itself all the way from Pittsburgh to San Diego in 1995, driving all but 50 of the 2,850 mile journey by itself. For this achievement, Navlab 5 was inducted into the Robot Hall of Fame. In the world of robots, being in the Hall of Fame pretty much makes you elite royalty. If you are ever in Pittsburgh and want to meet some other members of this prestigious class, be sure to check out the Roboworld exhibit at the Carnegie Science Center.\n\n# Deep Learning, Autonomous Driving, and the Data Problem\n\nEven 35 years ago scientists knew that neural networks were going to play a key role in making self-driving cars a reality. Back then, of course, we did not have the technology needed (GPUs, cloud computing, FPGAs, etc.) to train and deploy deep neural networks at scale for autonomous driving to become a reality. Today's autonomous cars use deep learning to perform all kinds of tasks. Table 16-1 lists some examples.\n\nTable 16-1. Examples of deep learning applications in self-driving cars **Task** | **Examples**  \n---|---\n\n**Perception**\n\n| Detect drivable area, lane markings, intersections, crosswalks, etc.  \nDetect other vehicles, pedestrians, animals, objects on the road, etc.  \nDetect traffic signs and signals\n\n**Navigation**\n\n| Predict and output steering angle, throttle, etc., based on sensor input  \nPerform overtaking, lane changes, U-turns, etc.  \nPerform merges, exits, navigate roundabouts, etc.  \nDrive in tunnels, over bridges, etc.  \nRespond to traffic rule breakers, wrong-way drivers, unexpected environment changes, etc.  \nNavigate stop-and-go traffic\n\nWe know that deep learning requires data. In fact, a general rule of thumb is that more data ensures better model performance. In previous chapters, we saw that training image classifiers can require millions of images. Imagine how much data it would take to train a car to drive itself. Things are also very different with self-driving cars when it comes to the data needed for validation and testing. Not only must a car be able to drive, it must do it safely. Testing and validating model performance is therefore fundamentally critical and requires a lot more data than what is needed for training, unlike traditional deep learning problems. However, it is difficult to accurately predict the exact amount of data needed, and estimates vary. A 2016 study shows that it would take 11 billion miles of driving for a car to become as good as a human driver. For a fleet of 100 self-driving cars collecting data 24 hours per day, 365 days per year at an average speed of 25 miles per hour, this would take more than 500 years!\n\nNow, of course, it is highly impractical and costly to collect 11 billion miles of data by having fleets of cars drive around in the real world. This is why almost everyone working in this space\u2014be it a large automaker or a startup\u2014uses simulators to collect data and validate trained models. Waymo (Google's self-driving car team), for example, had driven nearly 10 million miles on the road as of late 2018. Even though this is more than any other company on the planet it represents less than one percent of our 11-billion-mile figure. On the other hand, Waymo has driven seven billion miles in simulation. Although everyone uses simulation for validation, some companies build their own simulation tools, whereas others license them from companies like Cognata, Applied Intuition, and AVSimulation. There are some great open source simulation tools available as well: AirSim from Microsoft, Carla from Intel, and Apollo simulation from Baidu. Thanks to these tools, we don't need to connect the CAN bus of our car to learn the science behind making it drive itself.\n\nIn this chapter, we use a custom-built version of AirSim, built specifically for Microsoft's Autonomous Driving Cookbook.\n\n##### Simulation and the Autonomous Driving Industry\n\nSimulation plays a very important role in autonomous driving. Before the autonomy stack can be safely deployed on real-world vehicles, it must go through hours and hours of testing and validation inside simulated worlds that accurately represent the environment these vehicles will be driving in. Simulators enable closed-loop validation of algorithms. The ego vehicle in simulation receives sensor data and performs actions based on it, which has a direct affect on the vehicle's environment and influences the next set of sensor readings; hence, closing the loop.\n\nDifferent types of simulators are utilized based on which part of the autonomy stack is being tested (Figure 16-3). For tasks that require a visually realistic representation of the environment, like testing the perception stack, we use photorealistic simulators. These simulators also try to accurately mimic sensor physics, material dynamics, and so on. AirSim and Carla are examples of open source photo-realistic simulators. On the other hand, for tasks that can do without realistic visuals, we can save a lot of GPU resources by employing post-perception simulators. Most of these tasks utilize the results of a perception system to do path planning and navigation. Baidu's Apollo is an example of an open source simulator in this category.\n\n###### Figure 16-3. AirSim's photo realistic simulation versus Apollo postperception simulation (image source)\n\nCognata, an autonomous driving simulation startup based in Israel, is a leader in the space with its cloud-scale, photo-realistic simulation platform. Cognata is addressing some of the key challenges in this industry using a very innovative approach that uses DNNs. \"I came to this field after working on computer vision\u2013based ADAS [Advanced Driver Assistance Systems] applications for close to a decade,\" said Cognata CEO and founder Danny Atsmon. \"I found out that using OpenGL-based rendering techniques does not yield the feature quality which is needed to train or validate a deep neural network and came to a decision that a new rendering technique is needed for that. Our rendering engine uses DNNs to learn distributions from real recorded data and renders the simulated data in a way that mimics the actual sensor.\"\n\nWe are indeed living in a fantastic time in which hardware and software innovations are coming together like never before to solve one of the most complex technological challenges we have ever seen: the self-driving car.\n\n# The \"Hello, World!\" of Autonomous Driving: Steering Through a Simulated Environment\n\nIn this section, we implement the \"Hello, World!\" problem of autonomous driving. Self-driving cars are complex, with dozens of sensors streaming gigabytes of data and multiple decisions being made while driving down a road. Much like programming, for the \"Hello, World!\" of self-driving cars, we strip the requirements down to basic fundamentals:\n\n  * The car always stays on the road.\n\n  * For external sensor input, the car uses a single image frame from a single camera mounted on the front of the hood. No other sensors (lidar, radar, etc.) are used.\n\n  * Based on this single-image input, going at a constant speed, the car predicts its output steering angle. No other possible outputs (brake, throttle, gear changes, etc.) are predicted.\n\n  * There are no other vehicles, pedestrians, animals, or anything else on the road or relevant surroundings.\n\n  * The road being driven on is single lane, with no markings or traffic signs and signals and no rules for staying on the left or right side of the road.\n\n  * The road mostly changes in terms of turns (which will require changing steering angle to navigate) and not in terms of elevation (which would require a change in throttle, applying brakes, etc.).\n\n## Setup and Requirements\n\nWe will be implementing the \"Hello, World!\" problem in AirSim's Landscape map (Figure 16-4). AirSim is an open source photo-realistic simulation platform and is a popular research tool for training data collection and validation of deep learning\u2013based autonomous system model development.\n\n###### Figure 16-4. The Landscape map in AirSim\n\nYou can find the code for this chapter in the end-to-end deep learning tutorial from the _Autonomous Driving Cookbook_ on GitHub in the form of Jupyter Notebooks. In the cookbook repository, go to _AirSimE2EDeepLearning\/_. We can do so by running the following:\n\n    $ git clone https:\/\/github.com\/Microsoft\/AutonomousDrivingCookbook.git\n    $ cd AutonomousDrivingCookbook\/AirSimE2EDeepLearning\/\n\nWe will use Keras, a library with which we are already familiar, to implement our neural network. It's not necessary for us to learn any new deep learning concepts to work on this problem other than what has already been introduced in prior chapters of this book.\n\nFor this chapter, we have created a dataset by driving around in AirSim. The uncompressed size of the dataset is 3.25 GB. This is a little larger than datasets we've been using, but remember, we are implementing the \"Hello, World!\" of a highly complex problem. For comparison, it is fairly normal practice for automakers to collect multiple petabytes of data on the road every day. You can download the dataset by going to _aka.ms\/AirSimTutorialDataset_.\n\nWe can run the provided code and train our model on a Windows or Linux machine. We created a standalone build of AirSim for this chapter, which you can download from _aka.ms\/ADCookbookAirSimPackage_. After you've downloaded it, extract the simulator package to the location of your choice. Take note of the path because you will need it later. Please note that this build runs only on Windows but is needed only to see our final trained model in action. If you're using Linux, you can take the model files and run them on a Windows machine with the provided simulator package.\n\nAirSim is a highly photo-realistic environment, which means it is capable of generating lifelike pictures of the environment for the model to train on. You may have encountered similar graphics while playing high-definition video games. Given the size of the data and the graphic quality of the provided simulator package, a GPU is certainly preferable for running the code as well as the simulator.\n\nFinally, there are some additional tools and Python dependencies needed to run the code provided. You can find the details for these at Environment Setup in the README file in the code repository. At a high level, we need Anaconda with Python 3.5 (or higher), TensorFlow (to run as a backend to Keras), and h5py (for storing and reading data and model files). After we have our Anaconda environment set up, we can install the needed Python dependencies by running the _InstallPackages.py_ file as root or administrator.\n\nTable 16-2 provides a summary of all the requirements that we just defined.\n\nTable 16-2. Setup and requirements summary **Item** | **Requirements\/Link** | **Notes\/Comments**  \n---|---|---  \nCode repository | _https:\/\/oreil.ly\/_IJl9_ | Can be run on a Windows or Linux machine  \nJupyter Notebooks used |\n\n_DataExplorationAndPreparation.ipynb_\n\n_TrainModel.ipynb_\n\n_TestModel.ipynb_\n\n|   \nDataset download | _aka.ms\/AirSimTutorialDataset_ | 3.25 GB in size; needed to train model  \nSimulator download | _aka.ms\/ADCookbookAirSimPackage_ |\n\nNot needed to train model, only to deploy and run it; runs only on Windows\n\nGPU | Recommended for training, required for running simulator |   \nOther tools + Python dependencies | Environment Setup section in README file in the repository |\n\nWith all that taken care of, let's get started!\n\n# Data Exploration and Preparation\n\nAs we journey deeper into the world of neural networks and deep learning, you will notice that much of the magic of machine learning doesn't come from how a neural network is built and trained, but from the data scientist's understanding of the problem, the data, and the domain. Autonomous driving is no different. As you will see shortly, having a deep understanding of the data and the problem we are trying to solve is key when teaching our car how to drive itself.\n\nAll the steps here are also detailed in the Jupyter Notebook _DataExplorationAndPreparation.ipynb_. We start by importing all the necessary modules for this part of the exercise:\n\n    import os\n    import random\n\n    import numpy as np\n    import pandas as pd\n    import h5py\n\n    import matplotlib.pyplot as plt\n    from PIL import Image, ImageDraw\n\n    import Cooking #This module contains helper code. Please do not edit this file.\n\nNext, we make sure our program can find the location of our unzipped dataset. We also provide it a location to store our processed (or \"cooked\") data:\n\n    # << Point this to the directory containing the raw data >>\n    RAW_DATA_DIR = 'data_raw\/'\n\n    # << Point this to the desired output directory for the cooked (.h5) data >>\n    COOKED_DATA_DIR = 'data_cooked\/'\n\n    # The folders to search for data under RAW_DATA_DIR\n    # For example, the first folder searched will be RAW_DATA_DIR\/normal_1\n    DATA_FOLDERS = ['normal_1', 'normal_2', 'normal_3', 'normal_4', 'normal_5',\n    \t'normal_6', 'swerve_1', 'swerve_2', 'swerve_3']\n\n    # The size of the figures and illustrations used\n    FIGURE_SIZE = (10,10)\n\nLooking at the provided dataset we will see that it contains nine folders: _normal_1_ through _normal_6_ and _swerve_1_ through _swerve_3_. We come back to the naming of these folders later. Within each of these folders are images as well as _.tsv_ (or _.txt_ ) files. Let's check out one of these images:\n\n    sample_image_path = os.path.join(RAW_DATA_DIR, 'normal_1\/images\/img_0.png')\n    sample_image = Image.open(sample_image_path)\n    plt.title('Sample Image')\n    plt.imshow(sample_image)\n    plt.show()\n\nWe should see the following output. We can see that this image was captured by a camera placed centrally on the hood of the car (which is slightly visible at the bottom in Figure 16-5). We can also see the Landscape environment used for this problem from the simulator. Notice that no other vehicles or objects are on the road, keeping in line with our requirements. Although the road appears to be snowy, for our experiment, the snow is only visual and not physical; that is, it will have no effect on the physics and movement of our car. Of course, in the real world, it is critical to accurately replicate the physics of snow, rain, etc., so that our simulators can produce highly accurate versions of reality. This is a very complicated task to undertake and many companies dedicate a lot of resources to it. Thankfully for our purpose, we can safely ignore all that and treat the snow as just white pixels in our image.\n\n###### Figure 16-5. Plot showing the contents of the file normal_1\/images\/img_0.png\n\nLet's also get a preview of what the rest of the data looks like by displaying the contents of one of the _.tsv\/.txt_ files:\n\n    sample_tsv_path = os.path.join(RAW_DATA_DIR, 'normal_1\/airsim_rec.txt')\n    sample_tsv = pd.read_csv(sample_tsv_path, sep='\\t')\n    sample_tsv.head()\n\nWe should see the following output:\n\nThere is a lot going on here, so let's break it down. First, we can see that each row in the table corresponds to an image frame (we are displaying only the first five frames here). Because this is at the beginning of our data collection drive, we can see that the car hasn't started moving yet: the speed and throttle are 0 and the gear is neutral. Because we are trying to predict only steering angles for our problem, we won't be predicting the speed, throttle, brake, or gear values. These values will be used, however, as part of the input data (more on this later).\n\n## Identifying the Region of Interest\n\nLet's get back to our sample image now. All the observations we made so far were from the eyes of a human being. For a moment, let's take a look at the image again, only this time we step into the shoes (or tires) of a car that is just beginning to learn how to drive itself and is trying to understand what it is seeing.\n\nIf I'm the car, looking at this image that my camera presented me, I can divide it into three distinct parts (Figure 16-6). First, there is the lower third, which more or less looks consistent. It is made up of mostly straight lines (the road, lane divider, road fences, etc.) and has a uniform coloring of white, black, gray, and brown. There is also a weird black arc at the very bottom (this is the hood of the car). The upper third of the image, like the lower third, is also consistent. It is mostly gray and white (the sky and the clouds). Lastly, there is the middle third and it has a lot going on. There are big brown and grey shapes (mountains), which are not uniform at all. There are also four tall green figures (trees), their shape and color are different from anything else I see, so they must be super important. As I am presented more images, the upper third and the lower third don't change all that much but the middle third undergoes a lot of change. Hence, I conclude that this is the part that I should be focusing on the most.\n\n###### Figure 16-6. The three parts of the image as seen by the self-driving car\n\nDo you see the challenge in front of our car? It has no way of knowing what part of the image being presented to it is important. It might try to associate the changing steering angles with the changing scenery instead of focusing on turns in the road, which are very subtle compared to everything else that is going on. Not only are the changes in the scenery confusing, but they are also not relevant to the problem we are trying to solve here at all. The sky above, although mostly unchanging, also does not provide us with any information relevant to the steering of our car. Finally, the part of the hood being captured in every picture is also nonrelevant.\n\nLet us fix this by focusing only on the relevant portion of the image. We do this by creating a region of interest (ROI) in our image. As we can imagine, there is no hard-and-fast rule to dictate what our ROI should look like. It really depends on our use case. For us, because the camera is in a fixed position and there are no elevation changes on the road, the ROI is a simple rectangle focusing on the road and road boundaries. We can see the ROI by running the following code snippet:\n\n    sample_image_roi = sample_image.copy()\n\n    fillcolor=(255,0,0)\n    draw = ImageDraw.Draw(sample_image_roi)\n    points = [(1,76), (1,135), (255,135), (255,76)]\n    for i in range(0, len(points), 1):\n        draw.line([points[i], points[(i+1)%len(points)]], fill=fillcolor, width=3)\n    del draw\n\n    plt.title('Image with sample ROI')\n    plt.imshow(sample_image_roi)\n    plt.show()\n\nWe should see the output shown in Figure 16-7.\n\n###### Figure 16-7. The ROI for our car to focus on during training\n\nOur model will focus only on the road now and not be confused by anything else going on in the environment. This also reduces the size of the image by half, which will make training our neural network easier.\n\n## Data Augmentation\n\nAs mentioned earlier, it is always better to have as much data as we can get our hands on while training deep learning models. We have seen in previous chapters the importance of data augmentation and how it helps to not only get more training data but to also avoid overfitting. Most data augmentation techniques discussed previously for image classification, however, will not be useful for our current autonomous driving problem. Let's take the example of rotation. Rotating images randomly by 20 degrees is very helpful when training a classifier for a smartphone camera, but the camera on the hood of our car is fixed in place and will never see a rotated image (unless our car is doing flips, at which point we have larger concerns). The same goes for random shifts; we will never encounter those while driving. There is, however, something else we can do to augment the data for our current problem.\n\nLooking closely at our image, we can observe that flipping it on the y-axis produces an image that could have easily come from a different test run (Figure 16-8). We can effectively double the size of our dataset if we used the flipped versions of images along with their regular ones. Flipping an image this way will require us to modify the steering angle associated with it as well. Because our new image is a reflection of our original, we can just change the sign of the corresponding steering angle (e.g., from 0.212 to \u20130.212).\n\n###### Figure 16-8. Flipping image on the y-axis\n\nThere is another thing we can do to augment our data. Autonomous cars need to be ready for not only changes on the roads, but also for changes in external conditions like weather, time of day, and available light. Most simulators available today allow us to create these conditions synthetically. All of our data was collected in bright lighting conditions. Instead of going back to the simulator to collect more data under a variety of lighting conditions, we could just introduce random lighting changes by adjusting image brightness while training on the data we have. Figure 16-9 shows an example.\n\n###### Figure 16-9. Reducing image brightness by 40%\n\n## Dataset Imbalance and Driving Strategies\n\nDeep learning models are only as good as the data on which they are trained. Unlike humans, the AI of today cannot think and reason for itself. Hence, when presented with a completely new situation it will fall back on what it has seen before and make predictions based on the dataset it was trained on. Additionally, the statistical nature of deep learning makes it ignore instances that happen infrequently as aberrations and outliers, even if they are of significance. This is called _dataset imbalance_ , and it is a common nuisance for data scientists.\n\nImagine that we are trying to train a classifier that looks at dermatoscopic images to detect whether a lesion is benign or malignant. Suppose that our dataset has one million images, 10,000 of which contain lesions that are malignant. It is very likely that the classifier we train on this data always predicts images to be benign and never malignant. This happens because deep learning algorithms are designed to minimize error (or maximize accuracy) while training. By classifying all images as benign, our model attains 99% accuracy and calls it a day, while severely failing at the job it was trained to perform: detecting cancerous lesions. This problem is usually solved by training predictors on a more balanced dataset.\n\nAutonomous driving is not immune to the problem of dataset imbalance. Let's take our steering-angle prediction model. What do we know about steering angles from our day-to-day driving experience? While driving down a road we mostly drive straight. If our model was trained on only normal driving data, it would never learn how to navigate turns due to their relative scarcity in the training dataset. To solve this, it is important that our dataset contains data not only for a normal driving strategy but also for a \"swerved\" driving strategy in a statistically significant amount. To illustrate what we mean by this, let's go back to our _.tsv_ \/ _.txt_ files. Earlier in the chapter, we pointed out the naming of our data folders. It should now be clear that our dataset has data from six collection runs using a normal driving strategy and from three collection runs using a swerve driving strategy.\n\nLet us aggregate the data from all _.tsv_ \/ _.txt_ files into a single DataFrame to make the analysis easier:\n\n    full_path_raw_folders = [os.path.join(RAW_DATA_DIR, f) for f in DATA_FOLDERS]\n\n    dataframes = []\n    for folder in full_path_raw_folders:\n        current_dataframe = pd.read_csv(os.path.join(folder, 'airsim_rec.txt'), \n                                        sep='\\t')\n        current_dataframe['Folder'] = folder\n        dataframes.append(current_dataframe) \n    dataset = pd.concat(dataframes, axis=0)\n\nLet's plot some steering angles from both driving strategies on a scatter plot:\n\n    min_index = 100\n    max_index = 1500\n    steering_angles_normal_1 = dataset[dataset['Folder'].apply(lambda v: 'normal_1'\n    in v)]['Steering'][min_index:max_index]\n    steering_angles_swerve_1 = dataset[dataset['Folder'].apply(lambda v: 'swerve_1' \n    in v)]['Steering'][min_index:max_index]\n\n    plot_index = [i for i in range(min_index, max_index, 1)]\n    fig = plt.figure(figsize=FIGURE_SIZE)\n    ax1 = fig.add_subplot(111)\n    ax1.scatter(plot_index, steering_angles_normal_1, c='b', marker='o',\n    label='normal_1')\n    ax1.scatter(plot_index, steering_angles_swerve_1, c='r', marker='_',\n    label='swerve_1')\n    plt.legend(loc='upper left');\n    plt.title('Steering Angles for normal_1 and swerve_1 runs')\n    plt.xlabel('Time')\n    plt.ylabel('Steering Angle')\n    plt.show()\n\nFigure 16-10 shows the results.\n\n###### Figure 16-10. Plot showing steering angles from the two driving strategies\n\nLet's also plot the number of data points collected with each strategy:\n\n    dataset['Is Swerve'] = dataset.apply(lambda r: 'swerve' in r['Folder'], axis=1)\n    grouped = dataset.groupby(by=['Is Swerve']).size().reset_index()\n    grouped.columns = ['Is Swerve', 'Count']\n\n    def make_autopct(values):\n        def my_autopct(percent):\n            total = sum(values)\n            val = int(round(percent*total\/100.0))\n            return '{0:.2f}% ({1:d})'.format(percent,val)\n        return my_autopct\n\n    pie_labels = ['Normal', 'Swerve']\n    fig, ax = plt.subplots(figsize=FIGURE_SIZE)\n    ax.pie(grouped['Count'], labels=pie_labels, autopct =\n    make_autopct(grouped['Count']), explode=[0.1, 1], textprops={'weight': 'bold'},\n    colors=['lightblue', 'salmon'])\n    plt.title('Number of data points per driving strategy')\n    plt.show()\n\nFigure 16-11 shows those results.\n\n###### Figure 16-11. Dataset split for the two driving strategies\n\nLooking at Figure 16-10, as we had expected, we see that the normal driving strategy produces steering angles that we would observe during everyday driving, mostly straight on the road with the occasional turn. By contrast, the swerve driving strategy mostly focuses on sharp turns and hence has higher values for steering angles. As shown in Figure 16-11, a combination of these two strategies gives us a nice, albeit unrealistic in real life, distribution to train on with a 75\/25 split. This also further solidifies the importance of simulation for autonomous driving because it is unlikely we'd be collecting data using the swerve strategy in real life with an actual car.\n\nBefore closing our discussion on data preprocessing and dataset imbalance, let's take a final look at the distribution of our steering angles for the two driving strategies by plotting them on a histogram (Figure 16-12):\n\n    bins = np.arange(-1, 1.05, 0.05)\n    normal_labels = dataset[dataset['Is Swerve'] == False]['Steering']\n    swerve_labels = dataset[dataset['Is Swerve'] == True]['Steering']\n\n    def steering_histogram(hist_labels, title, color):\n        plt.figure(figsize=FIGURE_SIZE)\n        n, b, p = plt.hist(hist_labels.as_matrix(), bins, normed=1, facecolor=color)\n        plt.xlabel('Steering Angle')\n        plt.ylabel('Normalized Frequency')\n        plt.title(title)\n        plt.show()\n\n    steering_histogram(normal_labels, 'Normal driving strategy label distribution',\n    'g')\n    steering_histogram(swerve_labels, 'Swerve driving strategy label distribution',\n    'r')\n\n###### Figure 16-12. Steering angle distribution for the two driving strategies\n\nAs noted earlier, the swerve driving strategy gives us a much broader range of angles compared to the normal driving strategy, as shown in Figure 16-12. These angles will help our neural network react appropriately in case it ever finds the car going off the road. The problem of imbalance in our dataset is somewhat solved, but not entirely. We still have a lot of zeros, across both driving strategies. To balance our dataset further, we could just ignore a portion of these during training. Although this gives us a very balanced dataset, it greatly reduces the overall number of data points available to us. We need to keep this in mind when we build and train our neural network.\n\nBefore moving on, let's make sure our data is suitable for training. Let's take the raw data from all folders, split it into train, test, and validation datasets, and compress them into HDF5 files. The HDF5 format allows us to access data in chunks without reading the entire dataset into memory at once. This makes it ideal for deep learning problems. It also works seamlessly with Keras. The following code will take some time to run. When it has finished running, we will have three dataset files: _train.h5_ , _eval.h5_ , and _test.h5_.\n\n    train_eval_test_split = [0.7, 0.2, 0.1]\n    full_path_raw_folders = [os.path.join(RAW_DATA_DIR, f) for f in DATA_FOLDERS]\n    Cooking.cook(full_path_raw_folders, COOKED_DATA_DIR, train_eval_test_split)\n\nEach dataset file has four parts:\n\nImage\n\nA NumPy array containing the image data.\n\nPrevious_state\n\nA NumPy array containing the last known state of the car. This is a (steering, throttle, brake, speed) tuple.\n\nLabel\n\nA NumPy array containing the steering angles that we wish to predict (normalized on the range -1..1).\n\nMetadata\n\nA NumPy array containing metadata about the files (which folder they came from, etc.).\n\nWe are now ready to take the observations and learnings from our dataset and start training our model.\n\n# Training Our Autonomous Driving Model\n\nAll the steps in this section are also detailed in the Jupyter Notebook _TrainModel.ipynb_. As before, we begin by importing some libraries and defining paths:\n\n    from keras.preprocessing.image import ImageDataGenerator\n    from keras.models import Sequential, Model\n    from keras.layers import Conv2D, MaxPooling2D, Dropout, Flatten, Dense, Lambda,\n    Input, concatenate\n    from tensorflow.keras.optimizers import Adam, SGD, Adamax, Nadam\n    from tensorflow.keras.callbacks import ReduceLROnPlateau, ModelCheckpoint,\n    \tCSVLogger\n    from tensorflow.keras.callbacks import EarlyStopping\n    import tensorflow.keras.backend as K\n    from tensorflow.keras.preprocessing import image\n    from tensorflow.keras.models import Sequential, Model, load_model\n    from tensorflow.keras.preprocessing.image import ImageDataGenerator\n    from tensorflow.keras.layers import Conv2D, MaxPooling2D, Dropout, Flatten, Dense, \n    from tensorflow.keras.layers import Lambda, Input, concatenate, BatchNormalization\n\n    from keras_tqdm import TQDMNotebookCallback\n    import json\n    import os\n    import numpy as np\n    import pandas as pd\n    from Generator import DriveDataGenerator\n    from Cooking import checkAndCreateDir\n    import h5py\n    from PIL import Image, ImageDraw\n    import math\n    import matplotlib.pyplot as plt\n\n    # << The directory containing the cooked data from the previous step >>\n    COOKED_DATA_DIR = 'data_cooked\/'\n    # << The directory in which the model output will be placed >>\n    MODEL_OUTPUT_DIR = 'model'\n\nLet's also set up our datasets:\n\n    train_dataset = h5py.File(os.path.join(COOKED_DATA_DIR, 'train.h5'), 'r')\n    eval_dataset = h5py.File(os.path.join(COOKED_DATA_DIR, 'eval.h5'), 'r')\n    test_dataset = h5py.File(os.path.join(COOKED_DATA_DIR, 'test.h5'), 'r')\n\n    num_train_examples = train_dataset['image'].shape[0]\n    num_eval_examples = eval_dataset['image'].shape[0]\n    num_test_examples = test_dataset['image'].shape[0]\n\n    batch_size=32\n\n## Drive Data Generator\n\nWe were introduced to the concept of Keras data generators in previous chapters. A data generator iterates through the dataset and reads data in chunks from the disk. This allows us to keep both our CPU and GPU busy, increasing throughput. To implement ideas discussed in the previous section we created our own Keras generator called `DriveDataGenerator`.\n\nLet's recap some of our observations from the previous section:\n\n  * Our model should focus only on the ROI within each image.\n\n  * We can augment our dataset by flipping images horizontally and reversing the sign of the steering angle.\n\n  * We can further augment the data by introducing random brightness changes in the images. This will simulate different lighting conditions and make our model more robust.\n\n  * We can randomly drop a percentage of data points where the steering angle is zero so that the model sees a balanced dataset when training.\n\n  * Our overall number of data points available will be reduced significantly post-balancing.\n\nLet's see how the `DriveDataGenerator` takes care of the first four of these items. We will return to last item when we begin designing our neural network.\n\nThe ROI is a simple image crop. [76,135,0,255] (shown in the code block that follows) is the [x1,x2,y1,y2] value of the rectangle representing the ROI. The generator extracts this rectangle from each image. We can use the parameter `roi` to modify the ROI.\n\nThe horizontal flip is fairly straightforward. When generating batches, random images are flipped along the y-axis and their steering angle values are reversed if the parameter `horizontal_flip` is set to `True`.\n\nFor random brightness changes, we introduce the parameter `brighten_range`. Setting this parameter to `0.4` modifies the brightness of images in any given batch randomly up to 40%. We do not recommend increasing this beyond `0.4`. To compute brightness, we transform the image from RGB to HSV space, scale the \"V\" coordinate up or down, and transform back to RGB.\n\nFor dataset balancing through dropping zeros, we introduce the parameter `zero_drop_percentage`. Setting this to `0.9` will randomly drop 90% of the 0-label data points in any given batch.\n\nLet's initialize our generator with these parameters:\n\n    data_generator = DriveDataGenerator(rescale=1.\/255., horizontal_flip=True,\n    brighten_range=0.4)\n\n    train_generator = data_generator.flow\\\n        (train_dataset['image'], train_dataset['previous_state'], \n        train_dataset['label'], batch_size=batch_size, zero_drop_percentage=0.95, \n        roi=[76,135,0,255])\n    eval_generator = data_generator.flow\\\n        (eval_dataset['image'], eval_dataset['previous_state'],\n        eval_dataset['label'],\n    batch_size=batch_size, zero_drop_percentage=0.95, roi=[76,135,0,255]\n\nLet's visualize some sample data points by drawing labels (steering angle) over their corresponding images:\n\n    def draw_image_with_label(img, label, prediction=None):\n        theta = label * 0.69 #Steering range for the car is +- 40 degrees -> 0.69\n    # radians\n        line_length = 50\n        line_thickness = 3\n        label_line_color = (255, 0, 0)\n        prediction_line_color = (0, 255, 255)\n        pil_image = image.array_to_img(img, K.image_data_format(), scale=True)\n        print('Actual Steering Angle = {0}'.format(label))\n        draw_image = pil_image.copy()\n        image_draw = ImageDraw.Draw(draw_image)\n        first_point = (int(img.shape[1]\/2), img.shape[0])\n        second_point = (int((img.shape[1]\/2) + (line_length * math.sin(theta))),\n    int(img.shape[0] - (line_length * math.cos(theta))))\n        image_draw.line([first_point, second_point], fill=label_line_color,\n    width=line_thickness)\n\n        if (prediction is not None):\n            print('Predicted Steering Angle = {0}'.format(prediction))\n            print('L1 Error: {0}'.format(abs(prediction-label)))\n            theta = prediction * 0.69\n            second_point = (int((img.shape[1]\/2) + ((line_length\/2) *\n    math.sin(theta))), int(img.shape[0] - ((line_length\/2) * math.cos(theta))))\n            image_draw.line([first_point, second_point], fill=prediction_line_color,\n    width=line_thickness * 3)\n\n        del image_draw\n        plt.imshow(draw_image)\n        plt.show()\n\n    [sample_batch_train_data, sample_batch_test_data] = next(train_generator)\n    for i in range(0, 3, 1):\n        draw_image_with_label(sample_batch_train_data[0][i],\n    \tsample_batch_test_data[i])\n\nWe should see an output similar to Figure 16-13. Observe that we are now looking at only the ROI when training the model, thereby ignoring all the nonrelevant information present in the original images. The line indicates the ground truth steering angle. This is the angle the car was driving at when that image was taken by the camera during data collection.\n\n###### Figure 16-13. Drawing steering angles on images\n\n## Model Definition\n\nWe are now ready to define the architecture of our neural network. It is here that we must take into account the problem of our dataset being very limited in size after removing zeros. Because of this limitation, we cannot build a network that is too deep. Because we are dealing with images, we will need a few convolutional\/max-pooling pairs to extract features.\n\nImages alone, however, might not be enough to lead our model to convergence. Using only images to train also does not align with how driving decisions are made in the real world. When driving down the road, we are not only perceiving our surrounding environment, we are also aware of how fast we are going, how much we are turning and, the state of our gas and brake pedals. Input from sensors like cameras, lidars, radars, and so on being fed into a neural network corresponds to only one portion of all information a driver has at hand while making a driving decision in the real world. An image presented to our neural network could have been taken from a stationary car or a car driving at 60 mph; the network would have no way of knowing which. Turning the steering wheel by two degrees to the right while driving at 5 mph will yield very different results compared to doing the same at 50 mph. In short, a model trying to predict steering angles should not rely on sensory input alone. It also needs information about the current state of the car. Thankfully for us, we do have this information available.\n\nAt the end of the previous section, we noted that our datasets have four parts. For each image, in addition to the steering angle label and metadata, we also recorded the last known state of the car corresponding to the image. This was stored in the form of a (steering, throttle, brake, speed) tuple, and we will use this information, along with our images, as input to the neural network. Note that this does not violate our \"Hello, World!\" requirements, because we are still using the single camera as our only external sensor.\n\nPutting all we discussed together, you can see the neural network that we will be using for this problem in Figure 16-14. We are using three convolutional layers with 16, 32, 32 filters, respectively, and a (3,3) convolutional window. We are merging the image features (output from convolutional layers) with an input layer supplying the previous state of the car. The combined feature set is then passed through two fully connected layers with 64 and 10 hidden neurons, respectively. The activation function used in our network is ReLU. Notice that unlike the classification problems we have been working on in previous chapters, the last layer of our network is a single neuron with no activation. This is because the problem we are trying to solve is a regression problem. The output of our network is the steering angle, a floating-point number, unlike the discrete classes that we were predicting earlier.\n\n###### Figure 16-14. Network architecture\n\nLet's now implement our network. We can use `model.summary()`:\n\n    image_input_shape = sample_batch_train_data[0].shape[1:]\n    state_input_shape = sample_batch_train_data[1].shape[1:]\n    activation = 'relu'\n\n    # Create the convolutional stacks\n    pic_input = Input(shape=image_input_shape)\n\n    img_stack = Conv2D(16, (3, 3), name=\"convolution0\", padding='same',\n    activation=activation)(pic_input)\n    img_stack = MaxPooling2D(pool_size=(2,2))(img_stack)\n    img_stack = Conv2D(32, (3, 3), activation=activation, padding='same',\n    name='convolution1')(img_stack)\n    img_stack = MaxPooling2D(pool_size=(2, 2))(img_stack)\n    img_stack = Conv2D(32, (3, 3), activation=activation, padding='same',\n    name='convolution2')(img_stack)\n    img_stack = MaxPooling2D(pool_size=(2, 2))(img_stack)\n    img_stack = Flatten()(img_stack)\n    img_stack = Dropout(0.2)(img_stack)\n\n    # Inject the state input\n    state_input = Input(shape=state_input_shape)\n    merged = concatenate([img_stack, state_input])\n\n    # Add a few dense layers to finish the model\n    merged = Dense(64, activation=activation, name='dense0')(merged)\n    merged = Dropout(0.2)(merged)\n    merged = Dense(10, activation=activation, name='dense2')(merged)\n    merged = Dropout(0.2)(merged)\n    merged = Dense(1, name='output')(merged)\n\n    adam = Nadam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)\n    model = Model(inputs=[pic_input, state_input], outputs=merged)\n    model.compile(optimizer=adam, loss='mse')\n\n### Callbacks\n\nOne of the nice features of Keras is the ability to declare _callbacks_. Callbacks are functions that are executed after each epoch of training and help us to gain an insight into the training process as well as control hyperparameters to an extent. They also let us define conditions to perform certain actions while training is underway; for example, stopping the training early if the loss stops decreasing. We will use a few callbacks for our experiment:\n\n`ReduceLROnPlateau`\n\nIf the model is near a minimum and the learning rate is too high, the model will circle around that minimum without ever reaching it. This callback will allow the model to reduce its learning rate when the validation loss hits a plateau and stops improving, allowing us to reach the optimal point.\n\n`CSVLogger`\n\nThis lets us log the output of the model after each epoch into a CSV file, which will allow us to track the progress without needing to use the console.\n\n`ModelCheckpoint`\n\nGenerally, we will want to use the model that has the lowest loss on the validation set. This callback will save the model each time the validation loss improves.\n\n`EarlyStopping`\n\nWe will want to stop training when the validation loss stops improving. Otherwise, we risk overfitting. This option will detect when the validation loss stops improving and will stop the training process when that occurs.\n\nLet's now go ahead and implement these callbacks:\n\n    plateau_callback = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=3,\n    min_lr=0.0001, verbose=1)\n\n    checkpoint_filepath = os.path.join(MODEL_OUTPUT_DIR, 'models', '{0}_model.{1}\n    -{2}.h5'.format('model', '{epoch:02d}', '{val_loss:.7f}'))\n    checkAndCreateDir(checkpoint_filepath)\n    checkpoint_callback = ModelCheckpoint(checkpoint_filepath, save_best_only=True,\n    verbose=1)\n\n    csv_callback = CSVLogger(os.path.join(MODEL_OUTPUT_DIR, 'training_log.csv'))\n\n    early_stopping_callback = EarlyStopping(monitor='val_loss', patience=10, \n                                                                verbose=1)\n\n    nbcallback = TQDMNotebookCallback()\n    setattr(nbcallback, 'on_train_batch_begin', **lambda** x,y: **None**)\n    setattr(nbcallback, 'on_train_batch_end', **lambda** x,y: **None**)\n    setattr(nbcallback, 'on_test_begin', **lambda** x: **None**)\n    setattr(nbcallback, 'on_test_end', **lambda** x: **None**)\n    setattr(nbcallback, 'on_test_batch_begin', **lambda** x,y: **None**)\n    setattr(nbcallback, 'on_test_batch_end', **lambda** x,y: **None**)\n\n    callbacks=[plateau_callback, csv_callback, checkpoint_callback,\n    early_stopping_callback, nbcallback]\n\nWe're now all set to kick off training. The model will take a while to train, so this will make for a nice Netflix break. The training process should terminate with a validation loss of approximately .0003:\n\n    history = model.fit_generator(train_generator,\n    steps_per_epoch=num_train_examples \/\/ batch_size, epochs=500,\n    \tcallbacks=callbacks, validation_data=eval_generator, \n    \tvalidation_steps=num_eval_examples \/\/ batch_size, verbose=2)\n\n    Epoch 1\/500\n    Epoch 00001: val_loss improved from inf to 0.02338, saving model to\n    model\\models\\model_model.01-0.0233783.h5\n    - 442s - loss: 0.0225 - val_loss: 0.0234\n    Epoch 2\/500\n    Epoch 00002: val_loss improved from 0.02338 to 0.00859, saving model to\n    model\\models\\model_model.02-0.0085879.h5\n    - 37s - loss: 0.0184 - val_loss: 0.0086\n    Epoch 3\/500\n    Epoch 00003: val_loss improved from 0.00859 to 0.00188, saving model to\n    model\\models\\model_model.03-0.0018831.h5\n    - 38s - loss: 0.0064 - val_loss: 0.0019\n   ..................................\n\nOur model is now trained and ready to go. Before we see it in action, let's do a quick sanity check and plot some predictions against images:\n\n    [sample_batch_train_data, sample_batch_test_data] = next(train_generator)\n    predictions = model.predict([sample_batch_train_data[0],\n    sample_batch_train_data[1]])\n    for i in range(0, 3, 1):\n        draw_image_with_label(sample_batch_train_data[0][i],\n                              sample_batch_test_data[i], predictions[i])\n\nWe should see an output similar to Figure 16-15. In this figure, the thick line is the predicted output, and the thin line is the label output. Looks like our predictions are fairly accurate (we can also see the actual and predicted values above the images). Time to deploy our model and see it in action.\n\n###### Figure 16-15. Drawing actual and predicted steering angles on images\n\n# Deploying Our Autonomous Driving Model\n\nAll of the steps in this section are also detailed in the Jupyter Notebook _TestModel.ipynb_. Now that we have our model trained, it is time to spin up our simulator and use our model to drive our car around.\n\nAs before, begin by importing some libraries and defining paths:\n\n    from tensorflow.keras.models import load_model\n    import sys\n    import numpy as np\n    import glob\n    import os\n\n    if ('..\/..\/PythonClient\/' not in sys.path):\n        sys.path.insert(0, '..\/..\/PythonClient\/')\n    from AirSimClient import *\n\n    # << Set this to the path of the model >>\n    # If None, then the model with the lowest validation loss from training will be\n    # used\n    MODEL_PATH = None\n\n    if (MODEL_PATH == None):\n        models = glob.glob('model\/models\/*.h5') \n        best_model = max(models, key=os.path.getctime)\n        MODEL_PATH = best_model\n\n    print('Using model {0} for testing.'.format(MODEL_PATH))\n\nNext, load the model and connect to AirSim in the Landscape environment. To start the simulator, on a Windows machine, open a PowerShell command window at the location where we unzipped the simulator package and run the following:\n\n    .\\AD_Cookbook_Start_AirSim.ps1 landscape\n\nBack in the Jupyter Notebook, run the following to connect the model to the AirSim client. Ensure that the simulator is running _before_ kicking this off:\n\n    model = load_model(MODEL_PATH)\n\n    client = CarClient()\n    client.confirmConnection()\n    client.enableApiControl(True)\n    car_controls = CarControls()\n    print('Connection established!')\n\nWith the connection established, let's now set the initial state of the car as well as some buffers used to store the output from the model:\n\n    car_controls.steering = 0\n    car_controls.throttle = 0\n    car_controls.brake = 0\n\n    image_buf = np.zeros((1, 59, 255, 3))\n    state_buf = np.zeros((1,4))\n\nThe next step is to set up the model to expect an RGB image from the simulator as the input. We need to define a helper function for this:\n\n    def get_image():\n        image_response = client.simGetImages([ImageRequest(0, AirSimImageType.Scene,\n    False, False)])[0]\n        image1d = np.fromstring(image_response.image_data_uint8, dtype=np.uint8)\n        image_rgba = image1d.reshape(image_response.height, image_response.width, 4)\n\n        return image_rgba[76:135,0:255,0:3].astype(float)\n\nFinally, let's set up an infinite loop for our model to read an image from the simulator along with the current state of the car, predict the steering angle, and send it back to the simulator. Because our model predicts only steering angles, we will need to supply a control signal to maintain speed ourselves. Let's set it up so that the car will attempt to run at a constant 5 m\/s:\n\n    while (True):\n        car_state = client.getCarState()\n\n        if (car_state.speed < 5):\n            car_controls.throttle = 1.0\n        else:\n            car_controls.throttle = 0.0\n\n        image_buf[0] = get_image()\n        state_buf[0] = np.array([car_controls.steering, car_controls.throttle,\n    car_controls.brake, car_state.speed])\n        model_output = model.predict([image_buf, state_buf])\n        car_controls.steering = round(0.5 * float(model_output[0][0]), 2)\n\n        print('Sending steering = {0}, throttle = {1}'.format(car_controls.steering,\n    car_controls.throttle))\n\n        client.setCarControls(car_controls)\n\nWe should see output similar to the following:\n\n    Sending steering = 0.03, throttle = 1.0\n    Sending steering = 0.03, throttle = 1.0\n    Sending steering = 0.03, throttle = 1.0\n    Sending steering = 0.03, throttle = 1.0\n    Sending steering = 0.03, throttle = 1.0\n    Sending steering = -0.1, throttle = 1.0\n    Sending steering = -0.12, throttle = 1.0\n    Sending steering = -0.13, throttle = 1.0\n    Sending steering = -0.13, throttle = 1.0\n    Sending steering = -0.13, throttle = 1.0\n    Sending steering = -0.14, throttle = 1.0\n    Sending steering = -0.15, throttle = 1.0\n\nWe did it! The car is driving around nicely on the road, keeping to the right side, for the most part, carefully navigating all the sharp turns and instances where it could potentially go off the road. Kudos on training our first-ever autonomous driving model!\n\n###### Figure 16-16. Trained model driving the car\n\nBefore we wrap up, there are a couple of things worth noting. First, notice that the motion of the car is not perfectly smooth. This is because we are working with a regression problem and making a steering angle prediction for every image frame seen by the car. One way to fix this would be to average out predictions over a buffer of consecutive images. Another idea could be to turn this into a classification problem. More specifically, we could define buckets for the steering angles (..., \u20130.1, \u20130.05, 0, 0.05, 0.1, ...), bucketize the labels, and predict the correct bucket for each image.\n\nIf we let the model run for a while (a little more than five minutes), we observe that the car eventually veers off the road randomly and crashes. This happens on a portion of the track with a steep incline. Remember our last requirement during the problem setup? Elevation changes require manipulating throttle and brakes. Because our model can control only the steering angle, it doesn't do too well on steep roads.\n\n# Further Exploration\n\nHaving been trained on the \"Hello, World!\" scenario, our model is obviously not a perfect driver but don't be disheartened. Keep in mind that we have barely scratched the surface of the possibilities at the intersection of deep learning and self-driving cars. The fact that we were able to have our car learn to drive around almost perfectly using a very small dataset is something to be proud of!\n\n##### End-to-End Deep Learning for Self-Driving Cars of the Real World!\n\nAt the beginning of 2016, end-to-end deep learning gained popularity among roboticists. The hypothesis was that given enough visual data, a DNN should be able to learn a task in an \"end-to-end\" manner, mapping inputs directly to actions. Hence, it should be possible to train robots to perform complicated tasks using only image-based input from cameras into DNNs, resulting in actions.\n\nOne of the most popular works on this in the context of autonomous driving came from NVIDIA when it introduced its DAVE-2 system to the world. Unlike traditional systems, DAVE-2 used a single DNN to output driving control signals from image pixels (Figure 16-17). NVIDIA demonstrated that a small amount of training data from less than a hundred hours of driving was sufficient to train the car to operate in diverse conditions, on highways and local, residential roads in sunny, cloudy, and rainy conditions. You can find more details on this work at _https:\/\/devblogs.nvidia.com\/deep-learning-self-driving-cars_.\n\n###### Figure 16-17. Architecture of NVIDIA's DAVE-2 system (image source)\n\nHere are some new ideas for us to build on top of what we learned in this chapter. You can implement all of these ideas using the setup we already have in place for this chapter.\n\n## Expanding Our Dataset\n\nAs a general rule, using more data helps improve model performance. Now that we have the simulator up and running, it will be a useful exercise to expand our dataset by doing more data collection runs. We can even try combining data from various different environments available in AirSim and see how our model trained on this data performs in different environments.\n\nWe used only RGB data from a single camera in this chapter. AirSim allows us to do a lot more. For example, we can collect images in depth view, segmentation view, surface normal view, and more for each of the cameras available. So, we can potentially have 20 different images (for five cameras operating in all four modes) for each instance. Can using all this extra data help us improve the model we just trained?\n\n## Training on Sequential Data\n\nOur model currently uses a single image and a single vehicle state for each prediction. This is not really how we drive in real life, though. Our actions always take into account the recent series of events leading up to that given moment. In our dataset, we have timestamp information available for all our images, which we can use to create sequences. We can modify our model to make predictions using the previous _N_ images and states. For example, given the past 10 images and past 10 states, predict the next steering angle. (Hint: This might require the use of RNNs.)\n\n## Reinforcement Learning\n\nAfter we learn about reinforcement learning in the next chapter, we can come back and try the Distributed Deep Reinforcement Learning for Autonomous Driving tutorial from the _Autonomous Driving Cookbook_. Using the Neighborhood environment in AirSim, which is included in the package we downloaded for this chapter, we will learn how to scale a deep reinforcement learning training job and reduce training time from under a week to less than an hour using transfer learning and the cloud.\n\n# Summary\n\nThis chapter gave us a sneak peek into how deep learning enables the autonomous driving industry. Taking advantage of the skills acquired in the previous chapters, we implemented the \"Hello, World!\" problem of autonomous driving using Keras. Exploring the raw data at hand, we learned how to preprocess it to make it suitable for training a high-performing model. And we were able to accomplish this with a very small dataset. We were also able to take our trained model and deploy it to drive a car in the simulated world. Wouldn't you agree that there's just something magical about that?\n\n# Chapter 17. Building an Autonomous Car in Under an Hour: Reinforcement Learning with AWS DeepRacer\n\nContributed by guest author: Sunil Mallya\n\nIf you follow technology news, you probably have seen a resurgence in debates about when computers are going to take over the world. Although that's a fun thought exercise, what's triggered the resurgence in these debates? A large part of the credit can be attributed to the news of computers beating humans at decision-making tasks\u2014winning in chess, achieving high scores in video games like Atari (2013), beating humans in a complex game of Go (2016), and, finally, beating human teams at Defence of the Ancients (Dota) 2 in 2017. The most astonishing thing about these successes is that the \"bots\" learned the games by playing against one another and reinforcing the strategies that they found to bring them success.\n\nIf we think more broadly on this concept, it's no different than how humans teach their pets. To train a dog, every good behavior is reinforced by rewarding the dog with a treat and lots of hugs, and every undesired behavior is discouraged by asserting \"bad doggie.\" This concept of reinforcing good behaviors and discouraging bad ones essentially forms the crux of _reinforcement learning_.\n\nComputer games, or games in general, require a sequence of decisions to be made, so traditional supervised methods aren't well suited, because they often focus on making a single decision (e.g., is this an image of a cat or a dog?). The inside joke in the reinforcement learning community is that we just play video games all day (spoiler alert: It's true!). At present, reinforcement learning is being applied across industries to optimize stock trading, manage heating and cooling in large buildings and datacenters, perform real-time bidding of advertisements, optimize video-streaming quality, and even optimize chemical reactions in a lab. Given these examples of production systems, we would strongly advocate using reinforcement learning for sequential decision making and optimization problems. In this chapter, we focus on learning this paradigm of machine learning and applying it to a real-world problem: building a one-eighteenth-scale, self-driving autonomous car in less than an hour.\n\n# A Brief Introduction to Reinforcement Learning\n\nSimilar to deep learning, reinforcement learning has seen a renaissance in the past few years ever since previously held human video game records were broken. Reinforcement learning theory had its heyday in the 1990s, but didn't break in to large-scale production systems due to computational requirements and difficulty in training these systems. Traditionally reinforcement learning has been considered compute heavy; in contrast, neural networks are data heavy. But advancements in deep neural networks have also benefited reinforcement learning. Neural networks are now being used to represent the reinforcement learning model, thus giving birth to deep reinforcement learning. In this chapter, we use the term reinforcement learning and deep reinforcement learning interchangeably, but in almost all cases if not stated otherwise, when we refer to reinforcement learning, we are talking about deep reinforcement learning.\n\nDespite recent advancements, the landscape for reinforcement learning has not been developer friendly. Interfaces for training deep learning models have progressively become simpler, but this hasn't quite caught up in the reinforcement learning community. Another challenging aspect for reinforcement learning has been the significant computational requirements and time to model convergence (it has completed learning)\u2014it literally takes days, if not weeks, to create a converged model. Now, assuming that we had the patience, knowledge of neural networks, and the monetary bandwidth, educational resources regarding reinforcement learning are few and far between. Most of the resources are targeted to advanced data scientists, and sometimes out of reach for developers. That one-eighteenth-scale autonomous car we referenced earlier? That's AWS' DeepRacer (Figure 17-1). One of the biggest motivations behind AWS DeepRacer was to make reinforcement learning accessible to developers. It is powered by Amazon SageMaker reinforcement learning, which is a general-purpose reinforcement learning platform. And, let's get real: who doesn't like self-driving cars?\n\n###### Figure 17-1. The AWS DeepRacer one-eighteenth-scale autonomous car\n\n# Why Learn Reinforcement Learning with an Autonomous Car?\n\nIn recent years, self-driving technology has seen significant investments and success. DIY self-driving and radio-controlled (RC) car-racing communities have since become popular. This unparalleled enthusiasm in developers building scaled autonomous cars with real hardware, testing in real-world scenarios has compelled us to use a \"vehicle,\" literally, to educate developers on reinforcement learning. Although there existed other algorithms to build self-driving cars, like traditional computer vision or supervised learning (behavioral cloning), we believe reinforcement learning has an edge over these.\n\nTable 17-1 summarizes some of the popular self-driving kits available for developers, and the technologies that enable them. One of the key benefits of reinforcement learning is that models can be trained exclusively in a simulator. But reinforcement learning systems come with their own set of challenges, the biggest of them being the simulation-to-real (sim2real) problem. There's always a challenge of deploying models trained completely in simulation in to real-world environments. DeepRacer tackles this with some simple, yet effective solutions that we discuss later in the chapter. Within the first six months following the launch of DeepRacer in November 2018, close to 9,000 developers trained their models in the simulator and successfully tested them on a real-world track.\n\nTable 17-1. Landscape of autonomous self-driving technology **** | **Hardware** | **Assembly** | **Technology** | **Cost** | ****  \n---|---|---|---|---|---  \nAWS DeepRacer | Intel Atom with 100 GFLOPS GPU | Preassembled | Reinforcement learning | $399 |\n\nOpenMV | OpenMV H7 | DIY (two hours) | Traditional computer vision | $90 |\n\nDuckietown | Raspberry Pi | Preassembled | Reinforcement learning, behavioral cloning | $279\u2013$350 |\n\nDonkeyCar | Raspberry Pi | DIY (two to three hours) | Behavioral closing | $250 |\n\nNVIDIA JetRacer | Jetson Nano | DIY (three to five hours) | Supervised learning | ~$400 |\n\n##### From the Creator's Desk\n\nBy Chris Anderson, CEO 3DR and founder of DIY Robocars\n\nThe future of robocars in DIY communities will be:\n\nFaster\n    Consistently beating humans under all conditions, starting with straight speed and getting to track\/racing offensive and defensive tactics. We want our cars to be the nonplayer characters of real-world Mario Kart races set to ludicrous difficulty. ;-)\nEasier\n    In the same way that a great game is \"easy to pick up, but hard to master,\" we want our cars to drive well out of the box but require clever software approaches to win.\nCheaper\n    The cost and machine learning barrier to entry in to DIY racing leagues will significantly decrease.\n\n# Practical Deep Reinforcement Learning with DeepRacer\n\nNow for the most exciting part of this chapter: building our first reinforcement learning-based autonomous racing model. Before we embark on the journey, let's build a quick cheat sheet of terms that will help you to become familiar with important reinforcement learning terminology:\n\nGoal\n\nFinishing a lap around the track without going off track.\n\nInput\n\nIn a human driven car, the human visualizes the environment and uses driving knowledge to make decisions and navigate the road. DeepRacer is also a vision-driven system and so we use a single camera image as our input into the system. Specifically, we use a grayscale 120x160 image as the input.\n\nOutput (Actions)\n\nIn the real world, we drive the car by using the throttle (gas), brake, and steering wheel. DeepRacer, which is built on top of a RC car, has two control signals: the throttle and the steering, both of which are controlled by traditional pulse-width modulation (PWM) signals. Mapping driving to PWM signals can be nonintuitive, so we discretize the driving actions that the car can take. Remember those old-school car-racing games we played on the computer? The simplest of those used the arrow keys\u2014left, right, and up\u2014to drive the car. Similarly, we can define a fixed set of actions that the car can take, but with more fine-grained control over throttle and steering. Our reinforcement learning model, after it's trained, will make decisions on which action to take such that it can navigate the track successfully. We'll have the flexibility to define these actions as we create the model.\n\n###### Note\n\nA servo in an RC hobby car is generally controlled by a PWM signal, which is a series of pulses of varying width. The position in which the servo needs to end up is achieved by sending a particular width of the pulse signal. The parameters for the pulses are min pulse width, max pulse width, and repetition rate.\n\nAgent\n\nThe system that learns and makes decisions. In our case, it's the car that's learning to navigate the environment (the track).\n\nEnvironment\n\nWhere the agent learns by interacting with actions. In DeepRacer, the environment contains a track that defines where the agent can go and be in. The agent explores the environment to collect data to train the underlying deep reinforcement learning neural network.\n\nState (s)\n\nThe representation of where the agent is in an environment. It's a point-in-time snapshot of the agent. For DeepRacer, we use an image as the state.\n\nActions (a)\n\nSet of decisions that the agent can make.\n\nStep\n\nDiscrete transition from one state to the next.\n\nEpisode\n\nThis refers to an attempt by the car to achieve its goal; that is, complete a lap on the track. Thus, an episode is a sequence of steps, or experience. Different episodes can have different lengths.\n\nReward (r)\n\nThe value for the action that the agent took given an input state.\n\nPolicy (\u03c0)\n\nDecision-making strategy or function; a mapping from state to actions.\n\nValue function (V)\n\nThe mapping of state to values, in which value represents the expected reward for an action given the state.\n\nReplay or experience buffer\n\nTemporary storage buffer that stores the experience, which is a tuple of (s,a,r,s`'`), where \"s\" stands for an observation (or state) captured by the camera, \"a\" for an action taken by the vehicle, \"r\" for the expected reward incurred by the said action, and \"s`'`\" for the new observation (or new state) after the action is taken.\n\nReward function\n\nAny reinforcement learning system needs a guide, something that tells the model as it learns what's a good or a bad action given the situation. The reward function acts as this guide, which evaluates the actions taken by the car and gives it a reward (scalar value) that indicates the desirability of that action for the situation. For example, on a left turn, taking a \"left turn\" action would be considered optimal (e.g., reward = 1; on a 0\u20131 scale), but taking a \"right turn\" action would be bad (reward = 0). The reinforcement learning system eventually collects this guidance based on the reward function and trains the model. This is the most critical piece in training the car and the part that we'll focus on.\n\nFinally, when we put together the system, the schematic flow looks as follows:\n\n    Input (120x160 grayscale Image) \u2192 (reinforcement learning Model) \u2192 \n    Output (left, right, straight)\n\nIn AWS DeepRacer, the reward function is an important part of the model building process. We must provide it when training our AWS DeepRacer model.\n\nIn an episode, the agent interacts with the track to learn the optimal set of actions it needs to take to maximize the expected cumulative reward. But a single episode doesn't produce enough data for us to train the agent. So, we end up collecting many episodes' worth of data. Periodically, at the end of every _n_ th episode, we kick off a training process to produce an iteration of the reinforcement learning model. We run many iterations to produce the best model we can. This process is explained in detail in the next section. After the training, the agent executes autonomous driving by running inference on the model to take an optimal action given an image input. The evaluation of the model can be done in either the simulated environment with a virtual agent or a real-world environment with a physical AWS DeepRacer car.\n\nIt's finally time to create our first model. Because the input into the car is fixed, which is a single image from the camera, we need to focus only on the output (actions) and the reward function. We can follow the steps below to begin training the model.\n\n##### Simulation to Real Transfer Problem (Sim2Real)\n\nData collection for robotic applications can be extremely difficult and time consuming. So, there are great advantages to training these applications in a simulated environment. However, creating certain scenarios in the real world can be impractical, like simulating a collision with another car or a human. Therefore, there can be reality gaps in the simulator\u2014perceptual differences and simulation fidelity (image quality, sampling rate, etc.), physical differences (friction, mass, density, etc.), and incorrect physical modeling (interactions and collisions between physical objects in the simulator). All of these can lead to the simulation environment being vastly different from the real world.\n\n## Building Our First Reinforcement Learning\n\nTo do this exercise, you will need an AWS account. Log into the AWS Console using your account credentials, as shown in Figure 17-2.\n\n###### Figure 17-2. The AWS login console\n\nFirst, let's make sure we are in the North Virginia region, given that the service is available only in that region, and then navigate to the DeepRacer console page: _https:\/\/console.aws.amazon.com\/deepracer\/home?region=us-east-1#getStarted_.\n\nAfter you select \"Reinforcement learning,\" the model page opens. This page shows a list of all the models that have been created and the status of each model. To create a model, start the process here.\n\n###### Figure 17-3. Workflow for training the AWS DeepRacer model\n\n## Step 1: Create Model\n\nWe are going to create a model that can be used by the AWS DeepRacer car to autonomously drive (take actions) around a race track. We need to select the specific race track, provide the actions that our model can choose from, provide a reward function that will be used to incentivize our desired driving behavior, and configure the hyperparameters used during training.\n\n###### Figure 17-4. Creating a model on the AWS DeepRacer console\n\n## Step 2: Configure Training\n\nIn this step, we select our training environment, configure action space, write our reward function, and adjust other training related settings before kicking off our training job.\n\n### Configure the simulation environment\n\nTraining our reinforcement learning model takes place on a simulated race track, and we can choose the track to train our model. We'll use AWS RoboMaker, a cloud service that makes building robotic applications easy, to spin up the simulation environment.\n\nWhen training a model, we pick the track most similar to the final track we intend to race on. As of July 2019, AWS DeepRacer provides seven tracks that we can train on. While configuring such a complementary environment isn't required and doesn't guarantee a good model, it will maximize the odds that our model will perform best on the race track. Furthermore, if we train on a straight track, it's unlikely that our model would have learned how to turn. Just like in the case of supervised learning in which it's unlikely that the model will learn something that's not part of the training data, in reinforcement learning, the agent is unlikely to learn anything out of scope from the training environment. For our first exercise, select the re:Invent 2018 track, as shown in Figure 17-5.\n\nTo train a reinforcement learning model, we must choose a learning algorithm. Currently, the AWS DeepRacer console supports only the proximal policy optimization (PPO) algorithm. The team eventually will support more algorithms, but PPO was chosen for faster training times and superior convergence properties. Training a reinforcement learning model is an iterative process. First, it's a challenge to define a reward function to cover all important behaviors of an agent in an environment at once. Second, hyperparameters are often tuned to ensure satisfactory training performance. Both require experimentation. A prudent approach is to start with a simple reward function, which will be our approach in this chapter, and then progressively enhance it. AWS DeepRacer facilitates this iterative process by enabling us to clone a trained model, in which we could enhance the reward function to handle previous ignored variables or we can systematically adjust hyperparameters until the result converges. The easiest way to detect this convergence is look at the logs and see whether the car is going past the finish line; in other words, whether progress is 100%. Alternatively, we could visually observe the car's behavior and confirm that it goes past the finish line.\n\n###### Figure 17-5. Track selection on the AWS DeepRacer console\n\n### Configure the action space\n\nNext, we configure the action space that our model will select both during and after training. An action (output) is a combination of speed and steering angle. Currently in AWS DeepRacer, we use a discrete action space (a fixed set of actions) as opposed to a continuous action space (turn _x_ degrees with _y_ speed, where _x_ and _y_ take real values). This is because it's easier to map to values on the physical car, which we dive into later in the \"Racing the AWS DeepRacer Car\" To build this discrete action space, we specify the maximum speed, the speed levels, the maximum steering angle, and the steering levels, as depicted in Figure 17-6.\n\n###### Figure 17-6. Defining the action space on the AWS DeepRacer console\n\nFollowing are the configuration parameters for the action space:\n\nMaximum steering angle\n\nThis is the maximum angle in degrees that the front wheels of the car can turn to the left and to the right. There is a limit as to how far the wheels can turn, and so the maximum turning angle is 30 degrees.\n\nSteering angle granularity\n\nRefers to the number of steering intervals between the maximum steering angle on either side. Thus, if our maximum steering angle is 30 degrees, +30 degrees is to the left and \u201330 degrees is to the right. With a steering granularity of 5, the following steering angles, as shown in Figure 17-6, from left to right, will be in the action space: 30 degrees, 15 degrees, 0 degrees, \u201315 degrees, and \u201330 degrees. Steering angles are always symmetrical around 0 degrees.\n\nMaximum speed\n\nRefers to the maximum speed the car will drive in the simulator as measured in meters per second (m\/s).\n\nSpeed levels\n\nRefers to the number of speed levels from the maximum speed (including) to zero (excluding). So, if our maximum speed is 3 m\/s and our speed granularity is 3, our action space will contain speed settings of 1 m\/s, 2 m\/s, and 3 m\/s. Simply put, 3 m\/s divided by 3 = 1 m\/s, so go from 0 m\/s to 3 m\/s in increments of 1 m\/s. 0 m\/s is not included in the action space.\n\nBased on the previous example the final action space will include 15 discrete actions (three speeds x five steering angles), which should be listed in the AWS DeepRacer service. Feel free to tinker with other options, just remember that larger action spaces might take a bit longer to train.\n\n###### Tip\n\nBased on our experience, here are some tips on how to configure the action space:\n\n  * Our experiments have shown that models with a faster maximum speed take longer to converge than those with a slower maximum speed. In some cases (reward function and track dependent), it can take longer than 12 hours for a 5 m\/s model to converge.\n\n  * Our model will not perform an action that is not in the action space. Similarly, if the model is trained on a track that never required the use of this action\u2014for example, turning won't be incentivized on a straight track\u2014the model won't know how to use this action, because it won't be incentivized to turn. As you begin thinking about building a robust model, make sure that you keep the action space and training track in mind.\n\n  * Specifying a fast speed or a wide steering angle is great, but we still need to think about our reward function and whether it makes sense to drive full speed into a turn, or exhibit zigzag behavior on a straight section of the track.\n\n  * We also need to keep physics in mind. If we try to train a model at faster than 5 m\/s, we might see our car spin out on corners, which will probably increase the time to convergence of our model.\n\n### Configure reward function\n\nAs we explained earlier, the reward function evaluates the quality of an action's outcome given the situation, and rewards the action accordingly. In practice the reward is calculated during training after each action is taken, and forms a key part of the experience used to train the model. We then store the tuple (state, action, next state, reward) in a memory buffer. We can build the reward function logic using a number of variables that are exposed by the simulator. These variables represent measurements of the car, such as steering angle and speed; the car in relation to the racetrack, such as (x, y) coordinates; and the racetrack, such as waypoints (milestone markers on the track). We can use these measurements to build our reward function logic in Python 3 syntax.\n\nAll of the parameters are available as a dictionary to the reward function. Their keys, data types, and descriptions are documented in Figure 17-7, and some of the more nuanced ones are further illustrated in Figure 17-8.\n\n###### Figure 17-7. Reward function parameters (a more in-depth review of these parameters is available in the documentation)\n\n###### Figure 17-8. Visual explanation of some of the reward function parameters\n\nTo build our first model, let's pick an example reward function and train our model. Let's use the default template, shown in Figure 17-9, in which the car tries to follow the center dashed lines. The intuition behind this reward function is to take the safest navigation path around the track because being in the center keeps the car farthest from being off track. The reward function does the following: creates three tiers around the track, using the three markers, and then proceeds to reward the car more for driving in the second tier as opposed to the center or the last tier. Also note the differences in the size of the reward. We provide a reward of 1 for staying in the narrow center tier, 0.5 for staying in the second (off-center) tier, and 0.1 for staying in the last tier. If we decrease the reward for the center tier, or increase the reward for the second tier, we are essentially incentivizing the car to use a larger portion of the track surface. Remember that time you did your driving test? The examiner probably did the same and cut off points as you got closer to the curb or the lane makers. This could come in handy, especially when there are sharp corners.\n\n###### Figure 17-9. An example reward function\n\nHere's the code that sets this up:\n\n    def reward_function(params):\n        '''\n        Example of rewarding the agent to follow center line\n        '''\n\n        # Read input parameters\n        track_width = params['track_width']\n        distance_from_center = params['distance_from_center']\n\n        # Calculate 3 markers that are at varying distances away from the center line\n        marker_1 = 0.1 * track_width\n        marker_2 = 0.25 * track_width\n        marker_3 = 0.5 * track_width\n\n        # Give higher reward if the car is closer to center line and vice versa\n        if distance_from_center <= marker_1:\n            reward = 1.0\n        elif distance_from_center <= marker_2:\n            reward = 0.5\n        elif distance_from_center <= marker_3:\n            reward = 0.1\n        else:\n            reward = 1e-3 # likely crashed\/ close to off track\n\n        return float(reward)\n\nBecause this is the first training run, let's focus on understanding the process of creating and evaluating a basic model, and then focus on optimizing it. In this case, we skip the algorithm settings and hyperparameter sections and use the defaults.\n\n###### Note\n\n_FAQ: Should the rewards be in a certain range, and also can I give negative rewards?_\n\nThere are no real constraints on what we can reward and not reward, but as a good practice it's easier to understand rewards when they are in a 0\u20131 or 0\u2013100 scale. What's more important is that our reward scale gives relative rewards for the actions appropriately. For example, on a right turn, we should reward the right action with high reward, a left action with close to 0 reward, perhaps a straight action with an in-between reward or higher than the left action because it might not be a completely bad action to take.\n\n### Configure stop conditions\n\nThis is the last section before we begin training. Here we specify the maximum time our model will train for. This is provided as a convenient mechanism for us to terminate our training given that we are billed for the amount of training time.\n\nSpecify 60 minutes and then select \"Start training.\" If there is an error, we will be taken to the error location. After we start training, it can take up to six minutes to spin up the services (such as Amazon SageMaker, AWS Robomaker, AWS Lambda, AWS Step Function) needed to start training. Remember, we can always stop training early if we determine that the model has converged (as explained in the next section) by clicking the \"stop\" button.\n\n## Step 3: Model Training\n\nAfter our model has begun training, we can select it from the listed models on the DeepRacer console. We can then see quantitatively how the training is progressing by looking at the total reward over time graph and also qualitatively from a first-person view from the car in the simulator (see Figure 17-10).\n\nAt first, our car will not be able to drive on a straight road, but as it learns better driving behavior, we should see its performance improving and the reward graph increasing. Furthermore, when our car drives off of the track it will be reset on the track. We might observe that the reward graph is spiky.\n\n###### Note\n\n_FAQ: Why is the reward graph spiky?_\n\nThe agent starts with high exploration, and gradually begins to exploit the trained model. Because the agent always takes random actions for a fraction of its decisions, there might be occasions for which it totally makes the wrong decision and ends up going off track. This is typically high at the beginning of training, but eventually the spikiness should reduce as the model begins to learn.\n\nLogs are always a good source of more granular information regarding our model's training. Later in the chapter, we examine how we can use the logs programmatically to gain a deeper understanding of our model training. In the meantime, we can look at the log files for both Amazon SageMaker and AWS RoboMaker. The logs are output to Amazon CloudWatch. To see the logs, hover your mouse over the reward graph and select the three dots that appear below the refresh button. Then, select \"View logs.\" Because the model training will take an hour, this might be a good time to skip to the next section and learn a bit more about reinforcement learning.\n\n###### Figure 17-10. Training graph and simulation video stream on the AWS DeepRacer console\n\n## Step 4: Evaluating the Performance of the Model\n\nIn reinforcement learning, the best way to gauge the ability of the model is to run it such that it only exploits\u2014that is, it doesn't take random actions. In our case, first test it on a track similar to the one it's trained on to see whether it can replicate the trained behavior. Next, try it on a different track to test generalizability. After our model training is complete, we can commence model evaluation. From our model details page, where we observed training, select \"Start evaluation.\" We can now select the track on which we want to evaluate the performance of our model and also the number of laps. Select the \"re:Invent 2018\" track and 5 laps and then select Start. When it's done, we should see a similar page as that shown in Figure 17-11, which summarizes the results of our model's attempts to go around the track and complete the lap.\n\n###### Figure 17-11. Model evaluation page on the AWS DeepRacer console\n\nGreat job! We have successfully built our first reinforcement learning\u2013enabled autonomous car.\n\n###### Note\n\n_FAQ: When I run evaluation, I see only 100% completion rates sometimes?_\n\nAs the model runs inference and navigates around the track in the simulator, due to the fidelity of the simulator, the car can end up in slightly different positions for the same action\u2014in other words, a 15 degree left turn action might result in only 14.9 degrees. Practically, we observe only very small deviations in the simulator, but these small deviations can add up over time. A well-trained model is able to recover from close-to-off-track positions, but an undertrained model is likely to not recover from close calls.\n\nNow that our model has been successfully evaluated, we move on to improving it and learn to achieve better lap times. But before that, it's necessary for you to understand more about the theory behind reinforcement learning and dig deeper into the learning process.\n\n###### Note\n\nWe can get started with the AWS DeepRacer console service to create our first model, train it (for up to six hours), evaluate it, and submit it to the AWS DeepRacer League for free.\n\n# Reinforcement Learning in Action\n\nLet's take a closer look at reinforcement learning in action. Here, we discuss some theory, the inner workings, and lots of practical insights related to our autonomous car project.\n\n## How Does a Reinforcement Learning System Learn?\n\nFirst, we must understand _explore_ versus _exploit_. Similar to how a child learns, a reinforcement learning system learns from _exploring_ and finding out what's good and bad. In the case of the child, the parent guides or informs the child of its mistakes, qualifies the decisions made as good or bad, or to what degree they were good or bad. The child memorizes the decisions it took for given situations and tries to replay, or _exploit_ , those decisions at opportune moments. In essence, the child is trying to maximize the parent's approval or cheer. Early in a child's life, it's more attentive to the suggestions from its parents, thus creating opportunities to learn. And later in life, as adults who seldom listen to their parents, exploit the concepts they have learned to make decisions.\n\nAs shown in Figure 17-12, at every timestep \"t,\" the DeepRacer car (agent) receives its current observation, the state (St), and on that basis the car selects an action (At). As a result of the action the agent took, it gets a reward Rt+1 and moves to state St+1, and this process continues throughout the episode.\n\n###### Figure 17-12. Reinforcement learning theory basics in a nutshell\n\nIn the context of DeepRacer, the agent explores by taking random actions, and the reward function, which essentially acts just like the parent, tells the car whether the actions it took for the given state were good or not. Usually, \"goodness\" of the action for a given state is expressed as a number, with higher numbers meaning it's close to optimal and lower meaning it's not. The system records all of this information for every step, specifically: _current state, action taken, reward for the action, and next state_ (s,a,r,s`'`), in what we call the experience replay buffer, which is essentially a temporary memory buffer. The entire idea here being that the car can learn from what decisions were good and other decisions that were bad. The key point is that we _start with a high degree of exploration, and slowly increase exploitation._\n\nIn the DeepRacer simulator, we sample the input image state at 15 frames per second. At each step or image frame captured, the car transitions from one state to another. Each image represents the state the car is in, and eventually after the reinforcement learning model is trained, it will look to exploit by inferring which action to take. To make decisions, we could either take random actions or use our model for a recommendation. As the model trains, the training process balances the exploration and exploitation. To begin, we explore more, given that our model is unlikely to be good, but as it learns through experience, we shift more toward exploitation by giving the model more control. Figure 17-13 depicts this flow. This transition could be a linear decay, an exponential decay, or any similar strategy that's usually tuned based on the degree of learning we hypothesize. Typically, an exponential decay is used in practical reinforcement learning training.\n\n###### Figure 17-13. The DeepRacer training flow\n\nAfter an action is taken by a random choice or the model prediction, the car transitions to a new state. Using the reward function, the reward is computed and assigned to the outcome. This process will continue, for each state, until we get to a terminal state; that is, the car goes off track or completes a lap, at which point the car will be reset and then repeated. A step is a transition from one state to another, and at each step a (state, action, new state, reward) tuple is recorded. The collection of steps from the reset point until the terminal state is called an _episode_.\n\nTo illustrate an episode, let's look at the example of a miniature race track in Figure 17-14, where the reward function incentivizes following the centerline because it's the shortest and quickest path from start to finish. This episode consists of four steps: at step 1 and 2 the car follows the center line, and then it turns left 45 degrees at step 3, and continues in that direction only to finally crash at step 4.\n\n###### Figure 17-14. Illustration of an agent exploring during an episode\n\nWe can think of these episodes as experience, or training data for our model. Periodically, the reinforcement learning system takes a random subset of these recordings from the memory buffer and trains a DNN (our deep reinforcement learning model) to produce a model that can best navigate the environment based on our reward function \"guide\"; in other words, get maximum cumulative rewards. As time progresses or with more episodes, we see a model that becomes better over time. Of course, the supercritical caveat being that the reward function is well defined and directing the agent toward the goal. If our reward function is bad, our model will not learn the correct behavior.\n\nLet's take a slight diversion to understand a bad reward function. As we were initially designing the system, the car had freedom to go in any direction (left, right, forward, back). Our simplest reward function didn't distinguish between directions, and in the end the model learned the best way to accumulate awards by just rocking backward and forward. This was then overcome by incentivizing the car to go forward. Luckily for developers now, the DeepRacer team made it simpler by allowing the car to move only in the forward direction, so we don't even need to think of such behavior.\n\nNow back to reinforcement learning. _How do we know if the model is getting better?_ Let's return to the concepts of explore and exploit. Remember that we begin with high exploration, low exploitation, and gradually increase the rate of exploitation. This means that at any point in the training process, for a certain percentage of time the reinforcement learning system will explore; that is, take random actions. As the experience buffer grows, for any given state, it stores all kinds of decisions (actions) including the ones that resulted in the highest reward to the lowest reward. The model essentially uses this information to learn and predict the best action to take for the given state. Suppose that the model initially took an action \"go straight\" for state S and received a reward of 0.5, but the next time it arrived in state S, it took action \"go right\" and received a reward of 1. If the memory buffer had several samples of the reward as \"1\" for the same state S for action \"go right,\" the model eventually learns \"go right\" is the optimal action for that state. Over time, the model will explore less and exploit more, and this percentage allocation is changed either linearly or exponentially to allow for the best learning.\n\nThe key here is that if the model is taking optimal actions, the system learns to keep selecting those actions; if it isn't selecting the best action, the system continues to try to learn the best action for a given state. Practically speaking, there can exist many paths to get to the goal, but for the mini racetrack example in Figure 17-13, the fastest path is going to be right down the middle of the track, as practically speaking any turns would slow down the car. During training, in early episodes as the model begins to learn, we observe that it might end up at the finish line (goal) but might not be taking the optimal path, as illustrated in Figure 17-15 (left) where the car zigzags, thus taking more time to get to the finish line and earning cumulative rewards of only 9. But because the system still continues to explore for a fraction of the time, the agent gives itself opportunities to find better paths. As more experience is built, the agent learns to find an optimal path and converge to an optimal policy to reach the goal with total cumulative rewards of 18, as illustrated in Figure 17-15 (right).\n\n###### Figure 17-15. Illustration of different paths to the goal\n\nQuantitatively speaking, for each episode, you should see a trend of increasing rewards. If the model is making those optimal decisions, the car must be on track and navigating the course, resulting in accumulating rewards. However, you might see even after high reward episodes the graph dipping, which can happen in early episodes because the car might still have a high degree of exploration, as mentioned earlier.\n\n## Reinforcement Learning Theory\n\nNow that we understand how a reinforcement learning system learns and works, especially in the context of AWS DeepRacer, let's look at some formal definitions and general reinforcement learning theory. This background will be handy when we solve other problems using reinforcement learning.\n\n### The Markov decision process\n\nMarkov decision process (MDP) is a discrete stochastic state-transition process framework that is used for modeling decision making in a control process. The markov property defines that each state is solely dependent on the previous state. This is a convenient property because it means that to make a state transition, all of the necessary information must be available in the current state. Theoretical results in reinforcement learning rely on the problem being formulated as an MDP, hence it's important to understand how to model a problem as an MDP to be solved using reinforcement learning.\n\n### Model free versus model based\n\nModel, in this context, refers to the learned representation of the environment. This is useful because we can potentially learn the dynamics in the environment and train our agent using the model rather than having to use the real environment every time. However, in reality, learning the environment isn't easy, and it's often easier to have a representation of the real world in simulation and then jointly learn both perception and dynamics as part of the agent navigation rather than one after the other in a sequence. In this chapter, we focus only on model-free reinforcement learning.\n\n### Value based\n\nFor every action taken by the agent, there's a corresponding reward assigned by the reward function. For any given state-action pair, it's helpful to know its value (reward). If such a function were to exist, we could compute the maximum reward that can be achieved in any state and simply select the corresponding action to navigate the environment. For example, in a game of 3x3 tic-tac-toe, there are a finite number of game situations, so we could build a lookup table to give us the best move given the situation. But in a game of chess, given the size of the board and complexity of the game, such a lookup table would be computationally expensive to build and memory storage would be large. Thus, in a complex environment it's difficult to tabulate state-action value pairs or define a function that can map state-action pairs to values. So, instead we try to use a neural network by parameterizing the value function and using a neural network to approximate the value for every action given a state observation. An example of a value-based algorithm is Deep Q-Learning.\n\n### Policy based\n\nA policy is a set of rules that the agent learns to navigate the environment. Simply put the policy function tells the agent which action is the best action to take from its current state. Policy-based reinforcement learning algorithms like REINFORCE and Policy Gradients find the optimal policy without the need to map values to states. In reinforcement learning, we parameterize the policy. In other words, we allow a neural network to learn what the optimal policy function is.\n\n### Policy based or value based\u2014why not both?\n\nThere's always been a debate about whether to use policy-based or value-based reinforcement learning. Newer architectures try to learn both the value function and policy function together rather than keeping one fixed. This approach in reinforcement learning is called _actor critic_.\n\nYou can associate the actor with the policy and the critic with the value function. The actor is responsible for taking actions, and the critic is responsible for estimating the \"goodness,\" or the value of those actions. The actor maps states to actions, and the critic maps state-action pairs to values. In the actor-critic paradigm, the two networks (actor and critic) are trained separately using gradient ascent to update the weights of our deep neural network. (Remember that our goal is to maximize cumulative rewards; thus, we need to find global maxima.) As the episodes roll by, the actor becomes better at taking actions that lead to higher reward states, and the critic also becomes better at estimating the values of those actions. The signal for both the actor and critic to learn comes purely from the reward function.\n\nThe value for a given state-action pair is called _Q value_ , denoted by Q(s,a). We can decompose the Q value into two parts: the estimated value and a quantitative measure of the factor by which the action is better than others. This measure is called the _Advantage_ function. We can view Advantage as the difference between the actual reward for a given state-action pair and the expected reward at that state. The higher the difference, the farther we are from selecting an optimal action.\n\nGiven that estimating the value of the state could end up being a difficult problem, we can focus on learning the advantage function. This allows us to evaluate the action not only based on how good it is, but also based on how much better it could be. This allows us to converge to an optimal policy much easier than other simpler policy gradient-based methods because, generally, policy networks have high variance.\n\n### Delayed rewards and discount factor (\u03b3)\n\nRewards are given for every state transition based on the action taken. But the impact of these rewards might be nonlinear. For certain problems the immediate rewards might be more important, and in some cases the future rewards more important. As an example, if we were to build a stock-trading algorithm, the future rewards might have higher uncertainty; thus, we would need to discount appropriately. The discount factor (\u03b3) is a multiplication factor between [0,1], where closer to zero indicates rewards in the immediate future are more important. With a high value that's closer to 1, the agent will focus on taking actions that maximize future rewards.\n\n## Reinforcement Learning Algorithm in AWS DeepRacer\n\nFirst, let's look at one of the simplest examples of a policy optimization reinforcement learning algorithm: vanilla policy gradient.\n\n###### Figure 17-16. Training process for the vanilla policy gradient algorithm\n\nWe can think of a deep reinforcement learning model as consisting of two parts: the input embedder and policy network. The input embedder will extract features from the image input and pass it to the policy network, which makes decisions; for instance, predict which action is the best for the given input state. Given that our input is an image, we use convolution layers (CNNs) to extract features. Because the policy is something we want to learn, we parameterize our policy function, the simplest of which can be learned with a fully connected layer. The input CNN layers take in an image and then the policy network uses the image features as input and outputs an action. Thus, mapping state to action. As the model trains, we become better at mapping the input space and extracting relevant features and also optimizing the policy to get the best action for each state. Our goal here is to collect maximum cumulative rewards. To achieve that, our model weights are updated such to maximize the cumulative future reward, and in doing so, we give a higher probability to the action that leads to the higher cumulative future reward. In training neural networks previously, we used stochastic gradient descent or its variation; when training an reinforcement learning system, we seek to maximize the cumulative reward; so, instead of minimization, we look to maximize. So, we use gradient ascent to move the weights in the direction of the steepest reward signal.\n\nDeepRacer uses an advanced variation of policy optimization, called Proximal Policy Optimization (PPO), summarized in Figure 17-17.\n\n###### Figure 17-17. Training using the PPO algorithm\n\nOn the left side of Figure 17-17, we have our simulator using the latest policy (model) to get new experience (s,a,r,s'). The experience is fed into an experience replay buffer that feeds our PPO algorithm after we have a set number of episodes.\n\nOn the right side of Figure 17-17, we update our model using PPO. Although PPO is a policy optimization method, it uses the advantage actor-critic method, which we described earlier. We compute the PPO gradient and shift policy in the direction where we get the highest rewards. Blindly taking large steps in this direction can cause too much variability in training; if we take small steps training can take forever. PPO improves the stability of the policy (actor) by limiting how much the policy can update at each training step. This is done by using a clipped surrogate objective function, which prevents the policy from being updated too much, thus solving large variance, a common issue with policy optimization methods. Typically, for PPO, we keep the ratio of new and old policy at [0.8, 1.2]. The critic tells the actor how good the action taken was, and how the actor should adjust its network. After the policy is updated the new model is sent to the simulator to get more experience.\n\n## Deep Reinforcement Learning Summary with DeepRacer as an Example\n\nTo solve any problem with reinforcement learning, we need to work through the following steps:\n\n  1. Define the goal.\n\n  2. Select the input state.\n\n  3. Define the action space.\n\n  4. Construct the reward function.\n\n  5. Define the DNN architecture.\n\n  6. Pick the reinforcement learning optimization algorithm (DQN, PPO, etc.).\n\nThe fundamental manner in which a reinforcement learning model trains doesn't change if we are building a self-driving car or building a robotic arm to grasp objects. This is a huge benefit of the paradigm because it allows us to focus on a much higher level of abstraction. To solve a problem using reinforcement learning, the main task at hand is to define our problem as an MDP, followed by defining the input state and set of actions that the agent can take in a given environment, and the reward function. Practically speaking, the reward function can be one of the most difficult parts to define, and often is the most important as this is what influences the policy that our agent learns. After all the environment-dependent factors are defined, we can then focus on what the deep neural network architecture should be to map the input to actions, followed by picking the reinforcement learning algorithm (value-based, policy-based, actor-critic-based) to learn. After we pick the algorithm, we can focus on the high level knobs to control what the algorithm does. When we drive a car, we tend to focus on controls, and the understanding of the internal combustion engine doesn't influence too much the way we drive. Similarly, as long as we understand the knobs that each algorithm exposes, we can train a reinforcement learning model.\n\nIt's time to bring this home. Let's formulate the DeepRacer racing problem:\n\n  1. Goal: To finish a lap by going around the track in the least amount of time\n\n  2. Input: Grayscale 120x160 image\n\n  3. Actions: Discrete actions with combined speed and steering angle values\n\n  4. Rewards: Reward the car for being on the track, incentivize going faster, and prevent from doing a lot of corrections or zigzag behavior\n\n  5. DNN architecture: Three-layer CNN + fully connected layer (Input \u2192 CNN \u2192 CNN \u2192 CNN \u2192 FC \u2192 Output)\n\n  6. Optimization algorithm: PPO\n\n## Step 5: Improving Reinforcement Learning Models\n\nWe can now look to improve our models and also get insights into our model training. First, we focus on training improvements in the console. We have at our disposal the ability to change the reinforcement learning algorithm settings and neural network hyperparameters.\n\n### Algorithm settings\n\nThis section specifies the hyperparameters that will be used by the reinforcement learning algorithm during training. Hyperparameters are used to improve training performance.\n\n### Hyperparameters for the neural network\n\nTable 17-2 presents the hyperparameters that are available to tune the neural network. Even though the default values have experimentally proven to be good, practically speaking the developer should focus on the batch size, number of epochs, and the learning rate as they were found to be the most influential in producing high-quality models; that is, getting the best out of our reward function.\n\nTable 17-2. Description and guidance for tuneable hyperparameters for the deep neural network **Parameter** | **Description** | **Tips**  \n---|---|---  \nBatch size | The number recent of vehicle experiences sampled at random from an experience buffer and used for updating the underlying deep learning neural network weights. If we have 5,120 experiences in the buffer, and specify a batch size of 512, then ignoring random sampling, we will get 10 batches of experience. Each batch will be used, in turn, to update our neural network weights during training. | Use a larger batch size to promote more stable and smooth updates to the neural network weights, but be aware of the possibility that the training may be slower.  \nNumber of epochs | An epoch represents one pass through all batches, where the neural network weights are updated after each batch is processed, before proceeding to the next batch. Ten epochs implies we update the neural network weights, using all batches one at a time, but repeat this process 10 times. | Use a larger number of epochs to promote more stable updates, but expect slower training. When the batch size is small, we can use a smaller number of epochs.  \nLearning rate | The learning rate controls how big the updates to the neural network weights are. Simply put, when we need to change the weights of our policy to get to the maximum cumulative reward, how much should we shift our policy. | A larger learning rate will lead to faster training, but it may struggle to converge. Smaller learning rates lead to stable convergence, but can take a long time to train.  \nExploration | This refers to the method used to determine the trade-off between exploration and exploitation. In other words, what method should we use to determine when we should stop exploring (randomly choosing actions) and when should we exploit the experience we have built up. | Since we will be using a discrete action space, we should always select \"CategoricalParameters.\"  \nEntropy | A degree of uncertainty, or randomness, added to the probability distribution of the action space. This helps promote the selection of random actions to explore the state\/action space more broadly. |   \nDiscount factor | A factor that specifies how much the future rewards contribute to the expected cumulative reward. The larger the discount factor, the farther out the model looks to determine the expected cumulative reward and the slower the training. With a discount factor of 0.9, the vehicle includes rewards from an order of 10 future steps to make a move. With a discount factor of 0.999, the vehicle considers rewards from an order of 1,000 future steps to make a move. | The recommended discount factor values are 0.99, 0.999, and 0.9999.  \nLoss type | The loss type specifies the type of the objective function (cost function) used to update the network weights. The Huber and mean squared error loss types behave similarly for small updates. But as the updates become larger, the Huber loss takes smaller increments compared to the mean squared error loss. | When we have convergence problems, use the Huber loss type. When convergence is good and we want to train faster, use the mean squared error loss type.  \nNumber of episodes between each training | This parameter controls how much experience the car should obtain between each model training iteration. For more complex problems that have more local maxima, a larger experience buffer is necessary to provide more uncorrelated data points. In this case, training will be slower but more stable. | The recommended values are 10, 20, and 40.\n\n### Insights into model training\n\nAfter the model is trained, at a macro level the rewards over time graph, like the one in Figure 17-10, gives us an idea of how the training progressed and the point at which the model begins to converge. But it doesn't give us an indication of the converged policy, insights into how our reward function behaved, or areas where the speed of the car could be improved. To help gain more insights, we developed a Jupyter Notebook that analyzes the training logs, and provides suggestions. In this section, we look at some of the more useful visualization tools that can be used to gain insight into our model's training.\n\nThe log file records every step that the car takes. At each step it records the x,y location of the car, yaw (rotation), the steering angle, throttle, the progress from the start line, action taken, reward, the closest waypoint, and so on.\n\n### Heatmap visualization\n\nFor complex reward functions, we might want to understand reward distribution on the track; that is, where did the reward function give rewards to the car on the track, and its magnitude. To visualize this, we can generate a heatmap, as shown in Figure 17-18. Given that the reward function we used gave the maximum rewards closest to the center of the track, we see that region to be bright and a tiny band on either side of the centerline beyond it to be red, indicating fewer rewards. Finally, the rest of the track is dark, indicating no rewards or rewards close to 0 were given when the car was in those locations. We can follow the code to generate our own heatmap and investigate our reward distribution.\n\n###### Figure 17-18. Heatmap visualization for the example centerline reward function\n\n### Improving the speed of our model\n\nAfter we run an evaluation, we get the result with lap times. At this point, we might be curious as to the path taken by the car or where it failed, or perhaps the points where it slowed down. To understand all of these, we can use the code in this notebook to plot a race heatmap. In the example that follows, we can observe the path that the car took to navigate around the track and also visualize the speeds at which it passes through at various points on the track. This gives us insight into parts that we can optimize. One quick look at Figure 17-19 (left) indicates that the car didn't really go fast at the straight part of the track; this provides us with an opportunity. We could incentivize the model by giving more rewards to go faster at this part of the track.\n\n###### Figure 17-19. Speed heatmap of an evaluation run; (left) evaluation lap with the basic example reward function, (right) faster lap with modified reward function\n\nIn Figure 17-19 (left), it seems like the car has a small zigzag pattern at times, so one improvement here could be that we penalize the car when it turns too much. In the code example that follows, we multiply the reward by a factor of 0.8, if the car is steering beyond a threshold. We also incentivize the car to go faster by giving 20% more reward if the car goes at 2 m\/s or more. When trained with this new reward function, we can see that the car does go faster than the previous reward function. Figure 17-19 (right) shows more stability; the car follows the centerline almost to perfection and finishes the lap roughly two seconds faster than our first attempt. This is just a glimpse in to improving the model. We can continue to iterate on our models and clock better lap times using these tools. All of the suggestions are incorporated in to the reward function example shown here:\n\n    def reward_function(params):\n        '''\n        Example of penalize steering, which helps mitigate zigzag behaviors and\n        speed incentive\n        '''\n\n        # Read input parameters\n        distance_from_center = params['distance_from_center']\n        track_width = params['track_width']\n        steering = abs(params['steering_angle']) # Only need absolute steering angle\n        speed = params['speed'] # in meter\/sec\n\n        # Calculate 3 markers that are at varying distances away from the center line\n        marker_1 = 0.1 * track_width\n        marker_2 = 0.25 * track_width\n        marker_3 = 0.5 * track_width\n\n        # Give higher reward if the agent is closer to the center line and vice versa\n        if distance_from_center <= marker_1:\n            reward = 1\n        elif distance_from_center <= marker_2:\n            reward = 0.5\n        elif distance_from_center <= marker_3:\n            reward = 0.1\n        else:\n            reward = 1e-3 # likely crashed\/ close to off track\n\n        # Steering penalty threshold, change the number based on your action space \n          setting\n        ABS_STEERING_THRESHOLD = 15\n\n        # Penalize reward if the agent is steering too much\n        if steering > ABS_STEERING_THRESHOLD:\n            reward *= 0.8\n\n        # Incentivize going faster\n        if speed >= 2:\n            reward *= 1.2\n\n# Racing the AWS DeepRacer Car\n\nIt's time to bring what we've learned so far from the virtual to the physical world and race a real autonomous car. Toy sized, of course!\n\nIf you own an AWS DeepRacer car, follow the instructions provided to test your model on a physical car. For those interested in buying the car, AWS DeepRacer is available for purchase on Amazon.\n\n## Building the Track\n\nNow that we have a trained model, we can evaluate this model on a real track and with a physical AWS DeepRacer car. First, let's build a makeshift track at home to race our model. For simplicity, we'll build only part of a racetrack, but instructions on how to build an entire track are provided here.\n\nTo build a track, you need the following materials:\n\nFor track borders:\n\nWe can create a track with tape that is about two-inches wide and white or off-white color against the dark-colored track surface. In the virtual environment, the thickness of the track markers is set to be two inches. For a dark surface, use a white or off-white tape. For example, 1.88-inch width, pearl white duct tape or 1.88-inch (less sticky) masking tape.\n\nFor track surface:\n\nWe can create a track on a dark-colored hard floor such as hardwood, carpet, concrete, or asphalt felt. The latter mimics the real-world road surface with minimal reflection.\n\n## AWS DeepRacer Single-Turn Track Template\n\nThis basic track template consists of two straight track segments connected by a curved track segment, as illustrated in Figure 17-20. Models trained with this track should make our AWS DeepRacer vehicle drive in a straight line or make turns in one direction. _The angular dimensions for the turns specified are suggestive; we can use approximate measurements when laying down the track._\n\n###### Figure 17-20. The test track layout\n\n## Running the Model on AWS DeepRacer\n\nTo start the AWS DeepRacer vehicle on autonomous driving, we must upload at least one AWS DeepRacer model to our AWS DeepRacer vehicle.\n\nTo upload a model, pick our trained model from the AWS DeepRacer console and then download the model artifacts from its Amazon S3 storage to a (local or network) drive that can be accessed from the computer. There's an easy download model button provided on the model page.\n\nTo upload a trained model to the vehicle, do the following:\n\n  1. From the device console's main navigation pane, choose Models, as shown in Figure 17-21.\n\n###### Figure 17-21. The Model upload menu on the AWS DeepRacer car web console\n\n  2. On the Models page, above the Models list, choose Upload.\n\n  3. From the file picker, navigate to the drive or share where you downloaded the model artifacts and choose the model for upload.\n\n  4. When the model is uploaded successfully, it will be added to the Models list and can be loaded into the vehicle's inference engine.\n\n## Driving the AWS DeepRacer Vehicle Autonomously\n\nTo start autonomous driving, place the vehicle on a physical track and do the following:\n\n  1. Follow the instructions to sign in to the vehicle's device console, and then do the following for autonomous driving:\n\n    1. On the \"Control vehicle\" page, in the Controls section, choose \"Autonomous driving,\" as shown in Figure 17-22.\n\n###### Figure 17-22. Driving mode selection menu on the AWS DeepRacer car web console\n\n  2. On the \"Select a model\" drop-down list (Figure 17-23), choose an uploaded model, and then choose \"Load model.\" This will start loading the model into the inference engine. The process takes about 10 seconds to complete.\n\n  3. Adjust the \"Maximum speed\" setting of the vehicle to be a percentage of the maximum speed used in training the model. (Certain factors such as surface friction of the real track can reduce the maximum speed of the vehicle from the maximum speed used in the training. You'll need to experiment to find the optimal setting.)\n\n###### Figure 17-23. Model selection menu on AWS DeepRacer car web console\n\n  4. Choose \"Start vehicle\" to set the vehicle to drive autonomously.\n\n  5. Watch the vehicle drive on the physical track or the streaming video player on the device console.\n\n  6. To stop the vehicle, choose \"Stop vehicle.\"\n\n### Sim2Real transfer\n\nSimulators can be more convenient than the real world to train on. Certain scenarios can be easily created in a simulator; for example, a car colliding with a person or another car\u2014we wouldn't want to do this in the real world :). However, the simulator in most cases will not have the visual fidelity we do in the real world. Also, it might not be able to capture the physics of the real world. These factors can affect modeling the environment in the simulator, and as a result, even with great success in simulation, when the agent runs in the real world, we can experience failures. Here are some of the common approaches to handle the limitations of the simulator:\n\nSystem identification\n\nBuild a mathematical model of the real environment and calibrate the physical system to be as realistic as possible.\n\nDomain adaptation\n\nMap the simulation domain to the real environment, or vice versa, using techniques such as regularization, GANs, or transfer learning.\n\nDomain randomization\n\nCreate a variety of simulation environments with randomized properties and train a model on data from all these environments.\n\nIn the context of DeepRacer, the simulation fidelity is an approximate representation of the real world, and the physics engine could use some improvements to adequately model all the possible physics of the car. But the beauty of deep reinforcement learning is that we don't need everything to be perfect. To mitigate large perceptual changes affecting the car, we do two major things: a) instead of using an RGB image, we grayscale the image to make the perceptual differences between the simulator and the real world narrower, and b) we intentionally use a shallow feature embedder; for instance, we use only a few CNN layers; this helps the network not learn the simulation environment entirely. Instead it forces the network to learn only the important features. For example, on the race track the car learns to focus on navigating using the white track extremity markers. Take a look at Figure 17-24, which uses a technique called GradCAM to generate a heatmap of the most influential parts of the image to understand where the car is looking for navigation.\n\n###### Figure 17-24. GradCAM heatmaps for AWS DeepRacer navigation\n\n# Further Exploration\n\nTo continue with the adventure, you can become involved in various virtual and physical racing leagues. Following are a few options to explore.\n\n## DeepRacer League\n\nAWS DeepRacer has a physical league at AWS summits and monthly virtual leagues. To race in the current virtual track and win prizes, visit the league page: _https:\/\/console.aws.amazon.com\/deepracer\/home?region=us-east-1#leaderboards_.\n\n## Advanced AWS DeepRacer\n\nWe understand that some advanced developers might want more control over the simulation environment and also have the ability to define different neural network architectures. To enable this experience, we provide a Jupyter Notebook\u2013based setup with which you can provision the components needed to train your custom AWS DeepRacer models.\n\n## AI Driving Olympics\n\nAt NeurIPS 2018, the AI Driving Olympics (AI-DO) was launched with a focus on AI for self-driving cars. During the first edition, this global competition featured lane-following and fleet-management challenges on the Duckietown platform (Figure 17-25). This platform has two components, Duckiebots (miniature autonomous taxis), and Duckietowns (miniature city environments featuring roads, signs). Duckiebots have one job\u2014to transport the citizens of Duckietown (the duckies). Onboard with a tiny camera and running the computations on Raspberry Pi, Duckiebots come with easy-to-use software, helping everyone from high schoolers to university researchers to get their code running relatively quickly. Since the first edition of the AI-DO, this competition has expanded to other top AI academic conferences and now includes challenges featuring both Duckietown and DeepRacer platforms.\n\n###### Figure 17-25. Duckietown at the AI Driving Olympics\n\n## DIY Robocars\n\nThe DIY Robocars Meetup, initially started in Oakland (California), now has expanded to more than 50 meetup groups around the world. These are fun and engaging communities to try and work with other enthusiasts in the self-driving and autonomous robocar space. Many of them run monthly races and are great venues to physically race your robocar.\n\n## Roborace\n\nIt's time to channel our inner Michael Schumacher. Motorsports is typically considered a competition of horsepower, and now Roborace is turning it into a competition of intelligence. Roborace organizes competitions between all-electric, self-driving race cars, like the sleek-looking Robocar designed by Daniel Simon (known for futuristic designs like the Tron Legacy Light Cycle), shown in Figure 17-26. We are not talking about miniature-scaled cars anymore. These are full-sized, 1,350 kg, 4.8 meters, capable of reaching 200 mph (320 kph). The best part? We don't need intricate hardware knowledge to race them.\n\nThe real stars here are AI developers. Each team gets an identical car, so that the winning edge is focused on the autonomous AI software written by the competitors. Output of onboard sensors such as cameras, radar, lidar, sonar, and ultrasonic sensors is available. For high-throughput computations, the cars are also loaded with the powerful NVIDIA DRIVE platform capable of processing several teraflops per second. All we need to build is the algorithm to let the car stay on track, avoid getting into an accident, and of course, get ahead as fast as possible.\n\nTo get started, Roborace provides a race simulation environment where precise virtual models of real cars are available. Along comes virtual qualification rounds, where the winners get to participate in real races including at Formula E circuits. By 2019, the top racing teams have already closed the gap to within 5\u201310% of the best human performances in races. Soon developers will be standing on top of the podium, after beating professional drivers. Ultimately, competitions like these lead to innovations and hopefully the learnings can be translated back to the autonomous car industry, making them safer, more reliable, and performant at the same time.\n\n##### From the Creator's Desk\n\nBy Anima Anandkumar, director of research, NVIDIA, Bren Professor, California Institute of Technology\n\nFuture and Evolution of Reinforcement Learning\n\nWe would not see reinforcement learning in the usual sense where learning is assumed to be from scratch with no prior knowledge or structure. Instead, I like to use the term \"interactive learning,\" which is broader. This includes active learning where data collection is done actively based on limitations of the current model that is being trained. I envision the use of physics-infused learning in rob6tics in which we incorporate physics and classical control but add learning to improve efficiency while maintaining stability and safety.\n\n###### Figure 17-26. Robocar from Roborace designed by Daniel Simon (image courtesy of Roborace)\n\n# Summary\n\nIn the previous chapter, we looked at how to train a model for an autonomous vehicle by manually driving inside the simulator. In this chapter, we explored concepts related to reinforcement learning and learned how to formulate the ultimate problem on everyone's mind: how to get an autonomous car to learn how to drive. We used the magic of reinforcement learning to remove the human from the loop, teaching the car to drive in a simulator on its own. But why limit it to the virtual world? We brought the learnings into the physical world and raced a real car. And all it took was an hour!\n\nDeep reinforcement learning is a relatively new field but an exciting one to explore further. Some of the recent extensions to reinforcement learning are opening up new problem domains where its application can bring about large-scale automation. For example, hierarchical reinforcement learning allows us to model fine-grained decision making. Meta-RL allows us to model generalized decision making across environments. These frameworks are getting us closer to mimicking human-like behavior. It's no surprise that a lot of machine learning researchers believe reinforcement learning has the potential to get us closer to _artificial general intelligence_ and open up avenues that were previously considered science fiction.\n\n# Appendix A. A Crash Course in Convolutional Neural Networks\n\nIn keeping with the word \"practical\" in the book's title, we've focused heavily on the real-world aspects of deep learning. The goal of this appendix is meant to serve as reference material, rather than a full-fledged exploration into the theoretical aspects of deep learning. To develop a deeper understanding of some of these topics, we recommend perusing the \"Further Exploration\" for references to other source material.\n\n# Machine Learning\n\n  * Machine learning helps learn patterns from data to make predictions on unseen data.\n\n  * There are three kinds of machine learning: supervised learning (learning from labeled data), unsupervised learning (learning from unlabeled data), and reinforcement learning (learning by action and feedback from an environment).\n\n  * Supervised learning tasks include classification (output is one of many categories\/classes) and regression (output is a numeric value).\n\n  * There are various supervised machine learning techniques including naive Bayes, SVM, decision trees, k-nearest neighbors, neural networks, and others.\n\n# Perceptron\n\n  * A perceptron, as shown in Figure A-1, is the simplest form of a neural network, a single-layered neural network with one neuron.\n\n  * A perceptron calculates a weighted sum of its inputs; that is, it accepts input values, multiplies each with a corresponding weight, adds a bias term, and generates a numeric output.\n\n###### Figure A-1. An example of a perceptron\n\n  * Because a perceptron is governed by a linear equation, it only has the capacity to model linear or close-to-linear tasks well. For a regression task, the prediction can be represented as a straight line. For a classification task, the prediction can be represented as a straight line separating a plane into two parts.\n\n  * Most practical tasks are nonlinear in nature, and hence a perceptron would fail to model the underlying data well, leading to poor prediction performance.\n\n# Activation Functions\n\n  * Activation functions convert linear input into nonlinear output. A sigmoid (Figure A-2) is a type of activation function. The hyperbolic tangent function and ReLU, shown in Figure A-3 are other common activation functions.\n\n###### Figure A-2. A plot of the sigmoid function\n\n###### Figure A-3. A plot of the ReLU function\n\n  * For classification, the probability of predicting a category is usually desired. This can be accomplished by a sigmoid function (for binary classification tasks) at the end of the perceptron, which converts values in the range of 0 to 1, and hence is better suited to represent probabilities.\n\n  * Activation functions help replicate the biological concept of the firing of neurons, which, in the simplest case, can be on or off; that is, the inputs activate the neurons. A step function is one form of an activation function where the value is one when the neuron is activated and zero when not. One downside of using a step function is that it loses the magnitude. ReLU, on the other hand, outputs the magnitude of the activation if it activates, and zero otherwise. Due to its simplicity and low computational requirements, it is preferred for nonprobability-based outputs.\n\n# Neural Networks\n\n  * Combining a perceptron with a nonlinear activation function improves its predictive ability. Combining several perceptions with nonlinear activation functions improves it even further.\n\n  * A neural network, as seen in Figure A-4, consists of multiple layers, each containing multiple perceptrons. Layers are connected in a sequence, passing the output of the previous layer as the input to the subsequent layer. This data transfer is represented as connections, with each neuron connected to every neuron in the previous and subsequent layers (hence the layers are aptly named fully connected layers). Each of these connections has a multiplying factor associated with it, also known as a weight.\n\n###### Figure A-4. An example of a multilayer neural network\n\n  * Each neuron stores the associated weights and bias, and has an optional nonlinear activation function at the end of it.\n\n  * Since information flows in one direction, without any cycles, this network structure is also called a feedforward neural network or multilayer feedforward network.\n\n  * The layers between the input(s) and output(s), hidden from the end-user, are appropriately known as hidden layers.\n\n  * Neural networks have a powerful ability to model patterns in data (i.e., represent the underlying relationship between data and associated labels). The _Universal Approximation Theorem_ states that a neural network with a single hidden layer can represent any continuous function within a specific range, acting as a universal approximator. So technically, a single-layer network could model any problem in existence. The major caveat is that it might be infeasible in practice due to the sheer number of neurons required to model many problems. The alternative approach is to have multiple layers with a fewer number of neurons per layer, which works well in practice.\n\n  * The output of a single hidden layer is the result of a nonlinear function applied over a linear combination of its inputs. The output of a second hidden layer is the result of a nonlinear function applied over a linear combination of inputs, which itself are nonlinear functions applied over a linear combination of their inputs. Each layer further builds upon the predictive power of the previous layer. Thus, the higher the number of layers, the better the representative ability of the neural network to model real-world nonlinear problems. This makes the network deep, and hence such networks with many hidden layers are also called _deep neural networks_.\n\n# Backpropagation\n\n  * Training a network involves modifying the weights and biases of the network iteratively until the network starts to make predictions with minimal error.\n\n  * Backpropagation helps train neural networks. It involves making a prediction, measuring how far-off the prediction is from the ground truth, and propagating the error back into the network so that the weights can be updated in order to reduce the magnitude of the error in subsequent iterations. This process is repeated until the error of the network no longer reduces.\n\n  * Neural networks are usually initialized with random weights.\n\n  * The performance of a model is measured with loss functions. For regression tasks (with numeric predictions), mean squared error (MSE) is a commonly used loss function. Due to the squaring of the magnitude of the error, using MSE helps favor more small errors than a few large errors. For classification tasks (with category predictions), the cross-entropy loss is commonly used. While accuracy is used more as a human-understandable metric of performance, cross-entropy loss is used for training a network, since it's more nuanced.\n\n# Shortcoming of Neural Networks\n\n  * A layer of a neural network with `n` neurons and `i` inputs requires `n x i` weights and computations since each neuron is connected to its inputs. An image usually has a lot of pixels as inputs. For an image with width `w`, height `h`, and `c` channels, `w x h x c x n` computations would be required. For a standard smartphone camera's image (usually 4032 x 3024 pixels in 3 channels of red, green, and blue), if the first hidden layer has 128 neurons, over 4 billion weights and computations would result, making neural networks impractical to use for most computer-vision tasks. We want to ideally convert an image into a much smaller workable representation that can then be classified by a neural network.\n\n  * Pixels in an image usually are related to other pixels in their vicinity. Capturing this relationship can give a better understanding of the image's content. But while feeding a neural network, the relationship between these pixels is lost because the image is flattened out into a 1D array.\n\n# Desired Properties of an Image Classifier\n\n  * It is computationally efficient.\n\n  * It captures relationships between different parts of an image.\n\n  * It is resilient to differences in position, size, angle of an object present in an image, and other factors such as noise. That is, an image classifier should be invariant to geometric transformations (often referred to using terms such as translation invariance, rotational invariance, positional invariance, shift-invariance, scale invariance, etc.).\n\n  * A CNN satisfies many of these desired properties in practice (some by its design, others with data augmentation).\n\n# Convolution\n\n  * A convolutional filter is a matrix that slides across an image (row by row, left to right), and multiplies with that section of adjoining pixels to produce an output. Such filters are famously used in Photoshop to blur, sharpen, brighten, darken an image, etc.\n\n  * Convolutional filters can, among other things, be used to detect vertical, horizontal, and diagonal edges. That is, when edges are present, they generate higher numbers as output, and lower output values otherwise. Alternatively stated, they \"filter\" out the input unless it contains an edge.\n\n  * Historically, convolutional filters were hand-engineered to perform a task. In the context of neural networks, these convolutional filters can be used as building blocks whose parameters (matrix weights) can be deduced automatically via a training pipeline.\n\n  * A convolutional filter activates every time it spots the input pattern it has been trained for. Since it traverses over an entire image, the filter is capable of locating the pattern at any position in the image. As a result, the convolutional filters contribute to making the classifier position invariant.\n\n# Pooling\n\n  * A pooling operation reduces the spatial size of the input. It looks at smaller portions of an input and performs an aggregation operation reducing the output to a single number. Typically aggregation operations include finding the maximum or average of all values, also commonly known as _max pooling_ and _average pooling_ , respectively. For example, a `2 x 2` max pooling operation will replace each `2 x 2` submatrix of the input (without overlap) with a single value in the output. Similar to convolution, pooling traverses the entire input matrix, aggregating multiple blocks of values into single values. If the input matrix were of size `2N x 2N`, the resulting output would be of size `N x N`.\n\n  * In the aforementioned example, the output would be the same regardless of where in the `2 x 2` matrix the maximum occurred.\n\n  * Max pooling is more commonly used than average pooling because it acts as a noise suppressor. Because it only takes the maximum value in each operation, it ignores the lower values, which are typically characteristics of noise.\n\n# Structure of a CNN\n\n  * At a high level, a CNN (shown in Figure A-5) consists of two parts\u2014a featurizer followed by a classifier. The featurizer reduces the input image to a workable size of features (i.e., a smaller representation of the image) upon which the classifier then acts.\n\n  * A CNN's featurizer is a repeating structure of convolutional, activation, and pooling layers. The input passes through this repeating structure in a sequence one layer at a time.\n\n###### Figure A-5. A high-level overview of a convolutional network\n\n  * The transformed input is finally passed into classification layers, which are really just neural networks. In this context, the classification neural network is typically known as a fully connected layer.\n\n  * Much like in neural networks, the nonlinearity provided by activation functions help the CNN learn complex features. ReLU is the most commonly used activation function.\n\n  * Empirically, earlier layers (i.e., those closer to the input layers) detect simple features such as curves and edges. If a matching feature is detected, the layer gets activated and passes on the signal to subsequent layers. Eventual layers combine these signals to detect more complex features such as a human eye, human nose, or human ears. Ultimately, these signals might add up to detect the entire object such as a human face.\n\n  * CNNs are trained using backpropagation just like multilayer perceptron networks.\n\n  * Unlike traditional machine learning, this makes feature selection automated. Additionally, the reduction of the input into smaller features makes the process more computationally efficient. Both of these are big selling points of deep learning.\n\n# Further Exploration\n\nThe topics we cover in this book should be self-sufficient, but if you do choose to develop a deeper appreciation for the underlying fundamentals, we highly recommend checking out some of the following material:\n\n<https:\/\/www.deeplearning.ai>\n\nThis set of deep learning courses from Andrew Ng covers the foundations for a wide range of topics, encompassing theoretical knowledge and deep insights behind each concept.\n\n<https:\/\/course.fast.ai>\n\nThis course by Jeremy Howard and Rachel Thomas has a more hands-on approach to deep learning using PyTorch as the primary deep learning framework. The fast.ai library used in this course has helped many students run models with state-of-the-art performance using just a few lines of code.\n\nHands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 2nd Edition\n\nThis book by Aur\u00e9lien G\u00e9ron covers machine learning and delves into topics such as support vector machines, decision trees, random forests, etc., then bridges into deep learning topics such as convolutional and recurrent neural networks. It explains both theory as well as some hands-on examples of these techniques.\n\n# Index\n\n### Symbols\n\n  * 24\/7 availability, Real-Time Object Classification on iOS with Core ML\n  * 80legs.com, Data\n\n### A\n\n  * A\/B testing, A\/B Testing, A\/B Testing Hosted Models-Using the Experiment in Code\n  * accountability, Accountability and Explainability\n  * action space, Configure the action space-Configure the action space\n  * activation functions, Activation Functions\n  * actor-critic approach, Policy based or value based\u2014why not both?\n  * Adam optimizer, Set Training Parameters\n  * Advanced Driver Assistance Systems (ADAS), Autonomous Cars\n  * Advanced Virtual RISC (AVR), Arduino\n  * Advantage function, Policy based or value based\u2014why not both?\n  * AI Driving Olympics (AI-DO), AI Driving Olympics\n  * Airbnb, Airbnb's Photo Classification\n  * AirSim, Deep Learning, Autonomous Driving, and the Data Problem, Setup and Requirements\n  * AlchemyVision, IBM Watson Visual Recognition\n  * Algorithmia, Algorithmia\n  * Alibaba, Speaker Recognition by Alibaba\n  * alpha users, The Life Cycle of a Food Classifier App\n  * Amazon Rekognition, Amazon Rekognition\n  * Amazon SageMaker, A Brief Introduction to Reinforcement Learning\n  * AMD GPUs, Installation\n  * Anaconda Python, Installation\n  * Android Studio, Building a Real-Time Object Recognition App\n  * annotations, Data\n  * Annoy (Approximate Nearest Neighbors Oh Yeah), Annoy\n  * Apollo simulation, Deep Learning, Autonomous Driving, and the Data Problem\n  * app size, reducing, Reducing App Size-Use Create ML\n  * Apple's machine learning architecture, Apple's Machine Learning Architecture-ML Performance Primitives, Approach 2: Use Create ML\n  * approximate nearest neighbor (ANN), Approximate Nearest-Neighbor Benchmark\n  * architecture, Model Architecture\n  * Arduino, Arduino\n  * artificial general intelligence, Summary\n  * artificial intelligence (AI)\n    * browser-based, AI in the Browser with TensorFlow.js and ml5.js-Summary\n    * definition of, What Is AI?\n    * on embedded devices, Becoming a Maker: Exploring Embedded AI at the Edge-Summary\n    * examples of, An Apology, Motivating Examples-Motivating Examples (see also case studies)\n    * frequently asked questions, Frequently Asked Questions-Frequently Asked Questions\n    * history of, A Brief History of AI-How Deep Learning Became a Thing\n    * versus human accuracy, ImageNet Dataset\n    * introduction to, The Real Introduction-What Is AI?\n    * mobile-based, Real-Time Object Classification on iOS with Core ML-Summary, Shazam for Food: Developing Android Apps with TensorFlow Lite and ML Kit-Summary\n    * perfect deep learning solution, Recipe for the Perfect Deep Learning Solution-Hardware, What's in the Picture: Image Classification with Keras\n    * resources for learning about, Education and Exploration, Further Exploration\n    * responsible AI, Responsible AI-Privacy\n    * serving prediction requests, Landscape of Serving AI Predictions\n    * tools, tips, and tricks, Practical Tools, Tips, and Tricks-One Last Question\n  * aspect ratio, Effect of Change in Aspect Ratio on Transfer Learning\n  * asynchronous communication, Prefetch Data\n  * augmentations, AutoAugment\n  * AutoAugment, Tools of the Trade, AutoAugment\n  * AutoKeras, Tools of the Trade, AutoKeras\n  * Automatic Mixed Precision, Use Automatic Mixed Precision\n  * automatic tuning, AutoAugment\n  * autonomous driving\n    * data collection necessary for, Deep Learning, Autonomous Driving, and the Data Problem\n    * data exploration and preparation, Data Exploration and Preparation-Dataset Imbalance and Driving Strategies\n    * deep learning applications in, Deep Learning, Autonomous Driving, and the Data Problem\n    * further exploration, Further Exploration-Reinforcement Learning\n    * history of, A Glimmer of Hope, A Brief History of Autonomous Driving\n    * introduction to, Simulating a Self-Driving Car Using End-to-End Deep Learning with Keras\n    * landscape of, Why Learn Reinforcement Learning with an Autonomous Car?\n    * model deployment, Deploying Our Autonomous Driving Model-Deploying Our Autonomous Driving Model\n    * model training, Training Our Autonomous Driving Model-Callbacks\n    * object detection in, Autonomous Cars\n    * real-world applications of, Further Exploration\n    * SAE levels of driving automation, Simulating a Self-Driving Car Using End-to-End Deep Learning with Keras\n    * simulation and, Deep Learning, Autonomous Driving, and the Data Problem, Practical Deep Reinforcement Learning with DeepRacer, Sim2Real transfer\n    * steering through simulated environments, The \"Hello, World!\" of Autonomous Driving: Steering Through a Simulated Environment-Setup and Requirements\n  * Autonomous Driving Cookbook, Deep Learning, Autonomous Driving, and the Data Problem, Reinforcement Learning\n  * availability, High Availability, Real-Time Object Classification on iOS with Core ML\n  * AWS DeepRacer\n    * for advanced developers, Advanced AWS DeepRacer\n    * AI Driving Olympics, AI Driving Olympics\n    * creating AWS accounts, Building Our First Reinforcement Learning\n    * DeepRacer league, DeepRacer League\n    * DIY Robocars, DIY Robocars\n    * improving learning models, Step 5: Improving Reinforcement Learning Models-Racing the AWS DeepRacer Car\n    * model creation, Step 1: Create Model\n    * model evaluation, Step 4: Evaluating the Performance of the Model\n    * model training, Step 3: Model Training\n    * motivation behind, A Brief Introduction to Reinforcement Learning\n    * racing the car, Racing the AWS DeepRacer Car-Sim2Real transfer\n    * reinforcement learning algorithm, Reinforcement Learning Algorithm in AWS DeepRacer\n    * reinforcement learning in action, How Does a Reinforcement Learning System Learn?\n    * reinforcement learning summary, Deep Reinforcement Learning Summary with DeepRacer as an Example\n    * reinforcement learning terminology, Practical Deep Reinforcement Learning with DeepRacer\n    * reinforcement learning theory, Reinforcement Learning Theory\n    * reward function in, Practical Deep Reinforcement Learning with DeepRacer\n    * Roborace, Roborace\n    * training configuration, Step 2: Configure Training-Configure stop conditions\n\n### B\n\n  * backpropagation, A Glimmer of Hope, Batch Size, Backpropagation\n  * batch size\n    * effect on accuracy, Effect of Batch Size\n    * role of, Batch Size\n    * using larger, Use Larger Batch Size\n  * Bayesian Optimization, Training\n  * BDD100K, Data\n  * benchmarks, Examine Industry Benchmarks\n  * bias, Bias, Class Activation Maps, Image Captioning, Bias-Bias\n  * binary classification, Binary classification\n  * Binder, Installation\n  * Bing Image Search API, Approach 3: Web Scraper Using Bing Image Search API\n  * blind and low-vision community, Face Detection in Seeing AI\n  * bottleneck features, Image Similarity\n    * (see also embeddings)\n  * browser-based AI\n    * benchmarking and practical considerations, Benchmarking and Practical Considerations\n    * case studies, Case Studies-GAN Lab\n    * introduction to, AI in the Browser with TensorFlow.js and ml5.js\n    * JavaScript-based ML libraries, JavaScript-Based Machine Learning Libraries: A Brief History-TensorFlow.js\n    * ml5.js, ml5.js\n    * model conversion, Model Conversion for the Browser\n    * pix2pix, pix2pix-pix2pix\n    * PoseNet, PoseNet\n    * pretrained models using TensorFlow.js, Running Pretrained Models Using TensorFlow.js\n    * TensorFlow.js architecture, TensorFlow.js Architecture\n    * training, Training in the Browser-GPU Utilization\n  * brute-force algorithm, Brute Force\n\n### C\n\n  * caching, Cache Data\n  * callbacks, Callbacks\n  * Carla, Deep Learning, Autonomous Driving, and the Data Problem\n  * case studies\n    * browser-based AI\n      * Airbnb's photo classification, Airbnb's Photo Classification\n      * GAN Lab, GAN Lab\n      * Metacar, Metacar\n      * Semi-Conductor, Semi-Conductor\n      * TensorSpace, TensorSpace\n    * computer vision\n      * Giphy, Giphy\n      * InDro Robotics, InDro Robotics\n      * New York Times, The New York Times\n      * OmniEarth, OmniEarth\n      * Photobucket, Photobucket\n      * Staples, Staples\n      * Uber, Uber\n    * embedded AI devices\n      * cucumber sorter, Cucumber Sorter\n      * JetBot, JetBot\n      * squatting for metro tickets, Squatting for Metro Tickets\n    * mobile-based AI\n      * face filter\u2013like functionality, Face Contours in ML Kit\n      * HomeCourt, HomeCourt\n      * InstaSaber, InstaSaber + YoPuppet\n      * Lose It!, Lose It!\n      * Magic Sudoku, Magic Sudoku\n      * portrait mode on Pixel 3 phones, Portrait Mode on Pixel 3 Phones\n      * Seeing AI, Seeing AI\n      * speaker recognition by Alibaba, Speaker Recognition by Alibaba\n      * video segmentation, Real-Time Video Segmentation in YouTube Stories\n      * YoPuppet, InstaSaber + YoPuppet\n    * object detection\n      * autonomous cars, Autonomous Cars\n      * crowd counting, Crowd Counting\n      * face detection in Seeing AI app, Face Detection in Seeing AI\n      * smart refrigerators, Smart Refrigerator\n    * reverse image search\n      * Celebslike.me, Celebrity Doppelgangers\n      * Flickr, Flickr\n      * image captioning, Image Captioning\n      * Pinterest, Pinterest\n      * Spotify, Spotify\n  * categorical_crossentropy loss function, Start Training\n  * category unawareness, ImageNet Dataset\n  * celebrity doppelgangers, Celebrity Doppelgangers\n  * channels, Predicting an Image's Category\n  * chroma keying technique, Real-Time Video Segmentation in YouTube Stories\n  * Clarifai, Clarifai\n  * class activation maps (heat maps), Class Activation Maps-Class Activation Maps, Heatmap visualization\n  * classification problems (see also custom classifiers; image classification)\n    * binary, Binary classification\n    * computer-vision tasks, Classification\n    * multiclass, Multiclass classification\n  * classification with localization, Localization\n  * Cloud AutoML, Building a Custom Detector Without Any Code\n  * cloud computing (see also inference serving)\n    * comparing custom classification APIs, Comparing Custom Classification APIs\n    * example scenario using, Cloud APIs for Computer Vision: Up and Running in 15 Minutes\n    * getting up and running with, Getting Up and Running with Cloud APIs\n    * hosted models, Hosted Models-Using the Experiment in Code\n    * performance tuning for cloud APIs, Performance Tuning for Cloud APIs-Effect of Resizing on OCR APIs\n    * prebuilt object detection APIs, Invoking Prebuilt Cloud-Based Object Detection APIs-Invoking Prebuilt Cloud-Based Object Detection APIs\n    * reduced costs of mobile-based AI, Real-Time Object Classification on iOS with Core ML\n    * training custom classifiers, Training Our Own Custom Classifier-Training Our Own Custom Classifier\n    * troubleshooting classifiers, Top Reasons Why Our Classifier Does Not Work Satisfactorily\n    * visual recognition APIs available, The Landscape of Visual Recognition APIs-What's unique about this API?\n    * visual recognition APIs comparison, Comparing Visual Recognition APIs-Bias\n  * co-occurrences, Class Activation Maps, Approach 2: Fatkun Chrome Browser Extension\n  * COCO (Microsoft Common Objects in COntext), Datasets, Image Captioning\n  * code examples\n    * downloading, Using Code Examples\n    * permission to use, Using Code Examples\n  * code-free training, A continuously evolving landscape, Building a Custom Detector Without Any Code\n  * Cognata, Deep Learning, Autonomous Driving, and the Data Problem\n  * Cognitive Services, Microsoft Cognitive Services, Invoking Prebuilt Cloud-Based Object Detection APIs\n  * CometML, Training\n  * components, Predicting an Image's Category\n  * compression, Performance Tuning for Cloud APIs\n  * computer vision\n    * case studies, Case Studies-InDro Robotics\n    * comparing visual recognition APIs, Comparing Visual Recognition APIs-Bias\n    * custom classifer training, Training Our Own Custom Classifier-Training Our Own Custom Classifier\n    * custom classification APIs comparison, Comparing Custom Classification APIs-Comparing Custom Classification APIs\n    * example scenario using, Cloud APIs for Computer Vision: Up and Running in 15 Minutes-Cloud APIs for Computer Vision: Up and Running in 15 Minutes\n    * performance tuning for cloud APIs, Performance Tuning for Cloud APIs-Effect of Resizing on OCR APIs\n    * time estimation for achieving, How Deep Learning Became a Thing\n    * troubleshooting classifiers, Top Reasons Why Our Classifier Does Not Work Satisfactorily\n    * types of computer-vision tasks, Types of Computer-Vision Tasks-Approaches to Object Detection\n    * using cloud APIs, Getting Up and Running with Cloud APIs-Getting Up and Running with Cloud APIs\n    * visual recognition APIs available, The Landscape of Visual Recognition APIs-What's unique about this API?\n  * concept drift, Model Versioning\n  * conditional GANs , pix2pix\n  * contact information for this book, How to Contact Us\n  * continuous action space, Configure the action space\n  * contrastive loss function, Siamese Networks for One-Shot Face Verification\n  * conventions used in this book, Conventions Used in This Book\n  * ConvNetJS, ConvNetJS\n  * convolution filters, Convolution\n  * Convolutional Neural Networks (CNNs)\n    * accuracy and feature lengths, Length of Feature Vectors\n    * creating diagrams of, Model\n    * end-to-end learning example pipeline, End-to-End Deep Learning Example Pipeline-Basic Custom Network Pipeline\n    * hyperparameter effect on accuracy, How Hyperparameters Affect Accuracy-Effect of Change in Aspect Ratio on Transfer Learning\n    * image classification using, Predicting an Image's Category, A Shallow Dive into Convolutional Neural Networks\n    * improving model accuracy, From Novice to Master Predictor: Maximizing Convolutional Neural Network Accuracy-From Novice to Master Predictor: Maximizing Convolutional Neural Network Accuracy\n    * progress over past decade, How Deep Learning Became a Thing\n    * resources for learning about, Further Reading-Further Reading\n    * structure of, A Shallow Dive into Convolutional Neural Networks, Structure of a CNN\n    * techniques for ML experimentation, Common Techniques for Machine Learning Experimentation-Reproducible Experiments\n    * tools for automated tuning, Tools to Automate Tuning for Maximum Accuracy-AutoKeras\n    * tools to reduce code and effort, Tools of the Trade-tf-explain\n    * transfer learning using, Transfer Learning\n    * versus transfer learning, Approach 2: Use Create ML\n  * Coral USB accelerator, Google Coral USB Accelerator, Speeding Up with the Google Coral USB Accelerator\n  * Core ML\n    * alternatives to, Alternatives to Core ML\n    * building real-time object recognition apps, Building a Real-Time Object Recognition App-Building a Real-Time Object Recognition App\n    * conversion to, Conversion to Core ML, Model Conversion Using Core ML Tools\n    * dynamic model deployment, Dynamic Model Deployment\n    * ecosystem, Apple's Machine Learning Architecture-ML Performance Primitives\n    * history of, A Brief History of Core ML\n    * measuring energy consumption, Measuring Energy Impact-Benchmarking Load\n    * Not Hotdog app, Not Hotdog on iOS with Core ML and Create ML-Summary\n    * on-device training, On-Device Training\n    * performance analysis, Performance Analysis-Benchmarking Models on iPhones\n    * reducing app size, Reducing App Size-Use Create ML\n  * counting people, Crowd Counting\n  * Create ML, Use Create ML, Approach 2: Use Create ML-Approach 2: Use Create ML, Building a Custom Detector Without Any Code\n  * crowd counting, Crowd Counting\n  * cucumber sorting device, Cucumber Sorter\n  * curse of dimensionality, Reducing Feature-Length with PCA\n  * custom classifiers\n    * building data pipelines, Build the Data Pipeline\n    * data augmentation, Data Augmentation\n    * hosting and serving (see inference serving)\n    * model analysis, Analyzing the Results-Analyzing the Results\n    * model definition, Model Definition\n    * model testing, Test the Model\n    * model training, Set Training Parameters\n    * organizing data, Organize the Data\n    * overview of, Building a Custom Classifier in Keras with Transfer Learning\n    * using cloud APIs, Training Our Own Custom Classifier-Effect of Resizing on OCR APIs\n  * custom detectors, building without code, Building a Custom Detector Without Any Code-Building a Custom Detector Without Any Code\n    * (see also object detection)\n  * CustomVision.ai, Approach 1: Use Web UI-based Tools-Approach 1: Use Web UI-based Tools, Model Conversion Using Core ML Tools, Building a Custom Detector Without Any Code\n\n### D\n\n  * data augmentation, Data Augmentation-Data Augmentation, Data Augmentation, Data Augmentation\n  * data caching, Cache Data\n  * data generators, Drive Data Generator\n  * data inspection, Data Inspection\n  * data preparation (TensorFlow)\n    * reducing size of input data, Reduce Size of Input Data\n    * storing data as TFRecords, Store as TFRecords\n    * using TensorFlow Datasets, Use TensorFlow Datasets\n  * data reading (TensorFlow)\n    * autotuning parameter values, Autotune Parameter Values\n    * caching data, Cache Data\n    * enabling nondeterministic ordering, Enable Nondeterministic Ordering\n    * parallelizing CPU processing, Parallelize CPU Processing\n    * parallelizing I\/O and processing, Parallelize I\/O and Processing\n    * prefetching data, Prefetch Data\n    * turning on experimental optimizations, Turn on Experimental Optimizations\n    * using tf.data, Use tf.data\n  * data size\n    * effect on accuracy, Effect of Data Size on Transfer Learning\n    * reducing image size, Performance Tuning for Cloud APIs\n    * reducing in TensorFlow, Reduce Size of Input Data\n  * datamash tool, nvidia-smi\n  * dataset imbalance, Dataset Imbalance and Driving Strategies\n  * Dataset.cache() function, Cache Data\n  * DatasetList.com, Data\n  * datasets\n    * annotating data, Data\n    * approaches to collecting data, Collecting Data-Approach 3: Web Scraper Using Bing Image Search API\n    * collecting hundreds of images, Data, Approach 2: Fatkun Chrome Browser Extension\n    * downloading data in custom ways, Data\n    * finding larger, Data\n    * finding prebuilt, Data\n    * finding smaller, Data\n    * for mobile AI app development, How Do I Collect Initial Data?, Data Collection\n    * Google Dataset Search, Data\n    * ImageNet dataset, Predicting an Image's Category, ImageNet Dataset-ImageNet Dataset\n    * largest labeled, Data\n    * MS COCO (Microsoft Common Objects in COntext), Datasets, Image Captioning\n    * need for labeled, Datasets\n    * of negative classes, Data\n    * NIST and MNIST datasets, A Glimmer of Hope\n    * publicly available, Datasets, Approach 1: Find or Collect a Dataset\n    * reading popular, Data\n    * ready-to-use, TensorFlow Datasets\n    * scraping images using Bing Image Search API, Approach 3: Web Scraper Using Bing Image Search API\n    * scraping images using command line, Data\n    * sequential data, Training on Sequential Data\n    * splitting into train, validation, and test, Breaking the Data: Train, Validation, Test\n    * synthetic datasets, Which Library Should I Use?, Data\n    * TensorFlow Datasets, TensorFlow Datasets, Use TensorFlow Datasets, Data\n    * for unique problems, Data\n    * versioning tools for, Data\n    * video datasets, Data\n  * DAVE-2 system, Further Exploration\n  * DAWNBench benchmarks, Examine Industry Benchmarks\n  * debugging, Training\n  * deep learning\n    * applications in autonomous driving, Deep Learning, Autonomous Driving, and the Data Problem\n    * definition of, What Is AI?\n    * highlights of modern era, How Deep Learning Became a Thing\n    * increase in popularity of, How Deep Learning Became a Thing\n    * libraries for, Frameworks\n    * perfect deep learning solution, Recipe for the Perfect Deep Learning Solution-Hardware\n  * Deeplearning.ai, Education and Exploration\n  * detection, Detection (see also facial recognition; object detection)\n  * Digital Data Divide, Data\n  * discount factor (\u03b3), Delayed rewards and discount factor (\u03b3)\n  * discrete action space, Configure the action space\n  * Distributed Deep Reinforcement Learning for Autonomous Driving, Reinforcement Learning\n  * distributed training, Distribute Training\n  * DIY Robocars, DIY Robocars\n  * Docker, Installation\n  * driving automation, Simulating a Self-Driving Car Using End-to-End Deep Learning with Keras\n    * (see also autonomous driving)\n\n### E\n\n  * eager execution, Use tf.function\n  * early stopping, Early Stopping, Callbacks\n  * edge devices (see embedded AI devices)\n  * embedded AI devices\n    * case studies, Case Studies-Cucumber Sorter\n    * comparison of, A Qualitative Comparison of Embedded AI Devices-A Qualitative Comparison of Embedded AI Devices\n    * landscape of, Exploring the Landscape of Embedded AI Devices-Arduino\n    * performance of, Comparing the Performance of Edge Devices\n    * resources for learning about, Further Exploration\n  * embeddings \n    * benefits of, Visualizing Image Clusters with t-SNE\n    * defined, Image Similarity\n    * extracting for similarity search, Fine Tuning Without Fully Connected Layers\n    * fine tuning, Improving Accuracy with Fine Tuning\n    * Siamese networks and, Siamese Networks for One-Shot Face Verification\n    * TensorFlow Embedding projector, Visualizing Image Clusters with t-SNE\n    * visualizing, Improving Accuracy with Fine Tuning\n  * encryption, Privacy\n  * energy consumption, of mobile apps, Measuring Energy Impact-Benchmarking Load\n  * episodes, Practical Deep Reinforcement Learning with DeepRacer\n  * Euclidean distance, Similarity Search, Reducing Feature-Length with PCA\n  * experience buffer, Practical Deep Reinforcement Learning with DeepRacer\n  * experimental optimizations, Turn on Experimental Optimizations\n  * explainability, Accountability and Explainability\n    * tf-explain, tf-explain-tf-explain\n    * What-If Tool, What-If Tool-What-If Tool\n\n### F\n\n  * facial recognition, Siamese Networks for One-Shot Face Verification, Face Contours in ML Kit, Face Detection in Seeing AI\n  * failure handling, Failure Handling\n  * fairing, Fairing\n  * Faiss, Faiss\n  * Fast.ai, Education and Exploration\n  * FastProgress, Training\n  * Fatkun Batch Download Image, Data, Approach 2: Fatkun Chrome Browser Extension\n  * feature extraction, Feature Extraction-Feature Extraction, Feature Extraction\n  * feature scaling, Predicting an Image's Category\n  * feature vectors\n    * defined, Image Similarity\n    * length of, Length of Feature Vectors\n    * reducing feature length, Reducing Feature-Length with PCA\n  * federated learning, Federated Learning\n  * Field Programmable Gate Arrays (FPGAs), FPGA + PYNQ\n  * filter operations, Turn on Experimental Optimizations\n  * filter_fusion option, Filter fusion\n  * fine tuning, Fine Tuning-How Much to Fine Tune, Improving Accuracy with Fine Tuning-Fine Tuning Without Fully Connected Layers\n  * fine-grained recognition, ImageNet Dataset\n  * Firebase, ML Kit, ML Kit + Firebase\n  * fixed round-robin order, Enable Nondeterministic Ordering\n  * Flask, Flask: Build Your Own Server-Cons of Using Flask\n  * Flickr, Flickr\n  * food classifier apps\n    * building real-time object recognition apps, Building a Real-Time Object Recognition App-Building a Real-Time Object Recognition App\n    * challenges of, Shazam for Food: Developing Android Apps with TensorFlow Lite and ML Kit\n    * life cycle of, The Life Cycle of a Food Classifier App\n    * life cycle of mobile AI development, A Holistic Look at the Mobile AI App Development Cycle-How Do I Update the Model on My Users' Phones?\n    * using ML Kit + Firebase, ML Kit + Firebase-Using the Experiment in Code\n    * model conversion to TensorFlow Lite, Model Conversion to TensorFlow Lite\n    * Shazam for Food app, Shazam for Food: Developing Android Apps with TensorFlow Lite and ML Kit\n    * TensorFlow Lite architecture, TensorFlow Lite Architecture\n    * TensorFlow Lite on iOS, TensorFlow Lite on iOS\n    * TensorFlow Lite overview, An Overview of TensorFlow Lite\n    * tools for, The Life Cycle of a Food Classifier App\n  * forward pass, Batch Size\n  * FPGAs (Field Programmable Gate Arrays), FPGA + PYNQ\n  * Fritz, Fritz, Fritz-Fritz\n  * fused operations, Use Fused Operations\n\n### G\n\n  * GAN Lab, GAN Lab\n  * General Data Protection Regulation (GDPR), Real-Time Object Classification on iOS with Core ML\n  * generalization, Overtrain, and Then Generalize, Approach 2: Fatkun Chrome Browser Extension\n  * Generative Adversarial Networks (GANs), pix2pix\n  * geographic availability, Geographic Availability\n  * Giphy, Giphy\n  * Giraffe-Tree problem, Image Captioning\n  * Google Cloud ML Engine\n    * building classification APIs, Building a Classification API-Building a Classification API\n    * pros and cons of, Pros of Using Cloud ML Engine\n  * Google Cloud Platform (GCP), Uploading a hosted model\n  * Google Cloud Vision, Google Cloud Vision, Invoking Prebuilt Cloud-Based Object Detection APIs\n  * Google Coral USB accelerator, Google Coral USB Accelerator, Speeding Up with the Google Coral USB Accelerator\n  * Google Dataset Search, Data\n  * Google Edge Tensor Processing Unit (TPU), Google Coral USB Accelerator\n  * Google Seedbank, Education and Exploration\n  * Google's AI doodle, AI in the Browser with TensorFlow.js and ml5.js\n  * GPU persistence, Enable GPU Persistence\n  * GPU utilization, GPU Utilization\n  * Graphics Processing Unit (GPU) starvation\n    * displaying GPU statistics, nvidia-smi\n    * overview of, GPU Starvation\n    * visualizing program execution, TensorFlow Profiler + TensorBoard\n  * GUI-based model training tools, A continuously evolving landscape\n\n### H\n\n  * hardware\n    * installing optimized stacks for, Install an Optimized Stack for the Hardware\n    * need for deep learning, Hardware\n    * using better, Use Better Hardware\n  * heatmaps, Class Activation Maps-Class Activation Maps, tf-explain, Heatmap visualization\n  * high availability, High Availability\n  * HomeCourt, HomeCourt\n  * horizontal scaling, Distribute Training\n  * hosted models\n    * A\/B testing, A\/B Testing Hosted Models-Using the Experiment in Code\n    * accessing, Accessing a hosted model\n    * with Firebase, Hosted Models\n    * uploading, Uploading a hosted model\n  * hosting and serving (see inference serving)\n  * Hyperas, Training\n  * Hyperopt, Training\n  * hyperparameters\n    * automatic tuning, Tools of the Trade, Tools to Automate Tuning for Maximum Accuracy-AutoKeras, Autotune Parameter Values, Training\n    * effect on accuracy, How Hyperparameters Affect Accuracy-Effect of Change in Aspect Ratio on Transfer Learning\n\n### I\n\n  * IBM Watson Visual Recognition, IBM Watson Visual Recognition\n  * image captioning, Image Captioning\n  * image classification\n    * building custom classifiers in Keras, Building a Custom Classifier in Keras with Transfer Learning-Analyzing the Results\n    * category unawareness, ImageNet Dataset\n    * class activation maps (heat maps), Class Activation Maps-Class Activation Maps\n    * using CNNs, A Shallow Dive into Convolutional Neural Networks\n    * desired properties of image classifiers, Desired Properties of an Image Classifier\n    * fine-grained recognition, ImageNet Dataset\n    * how images store information, Predicting an Image's Category\n    * ImageNet dataset, ImageNet Dataset-ImageNet Dataset\n    * using Keras, Predicting an Image's Category-Predicting an Image's Category\n    * model zoos, Model Zoos\n    * preprocessing images, Predicting an Image's Category\n    * simple pipeline for, Predicting an Image's Category\n  * image labeling APIs, Effect of Resizing on Image Labeling APIs, How Do I Label My Data?\n  * image saliency, Class Activation Maps\n  * image segmentation, Image Segmentation\n  * image similarity, Image Similarity-Image Similarity\n  * image size, Performance Tuning for Cloud APIs\n  * ImageN, Data\n  * ImageNet dataset, Predicting an Image's Category, ImageNet Dataset-ImageNet Dataset\n  * ImageNet Large Scale Visual Recognition Challenge (ILSVRC), How Deep Learning Became a Thing, ImageNet Dataset\n  * ImageNet-Utils, Approach 1: Find or Collect a Dataset\n  * Imagenette, Data\n  * iMerit, Data\n  * implicit bias, Bias\n  * in-group\/out-group bias, Bias\n  * Inception architectures, Google Cloud Vision\n  * InDro Robotics, InDro Robotics\n  * inference serving\n    * desirable qualities for, Desirable Qualities in a Production-Level Serving System-Support for Multiple Machine Learning Libraries\n    * frameworks for, Frameworks\n    * frequently asked questions, Scalable Inference Serving on Cloud with TensorFlow Serving and KubeFlow\n    * Google Cloud ML Engine, Google Cloud ML Engine: A Managed Cloud AI Serving Stack-Building a Classification API\n    * KubeFlow, KubeFlow-Installation\n    * price versus performance, Price Versus Performance Considerations-Cost Analysis of Building Your Own Stack\n    * server building with Flask, Flask: Build Your Own Server-Cons of Using Flask\n    * TensorFlow Serving, TensorFlow Serving\n    * tools, libraries, and cloud services, Landscape of Serving AI Predictions\n  * inference time, Inference Time\n  * Inference-as-a-Service, Google Cloud ML Engine: A Managed Cloud AI Serving Stack, Price Versus Performance Considerations\n  * installation, Installation-Installation\n  * InstaLooter, Data\n  * instance retrieval, Building a Reverse Image Search Engine: Understanding Embeddings\n  * instance-level segmentation, Instance-level segmentation\n  * InstaSaber, InstaSaber + YoPuppet\n  * Instruction Set Architecture (ISA), FPGAs\n  * Intel Movidius Neural Compute Stick 2, Intel Movidius Neural Compute Stick\n  * Interactive Telecommunications Program (ITP), PoseNet\n  * interactivity, Real-Time Object Classification on iOS with Core ML\n  * Intersection over Union (IoU), Intersection over Union\n  * iOS models (see Core ML)\n\n### J\n\n  * JavaScript-based ML libraries\n    * ConvNetJS, ConvNetJS\n    * history of, JavaScript-Based Machine Learning Libraries: A Brief History\n    * Keras.js, Keras.js\n    * ONNX.js, ONNX.js\n    * TensorFlow.js, TensorFlow.js\n  * JetBot, JetBot\n  * Jetson Nano, NVIDIA Jetson Nano, Port to NVIDIA Jetson Nano\n  * Jupyter Notebooks\n    * machine learning examples, Education and Exploration\n    * running interactively in browsers, Installation\n    * using with KubeFlow, Fairing\n\n### K\n\n  * Keras\n    * benefits of, A continuously evolving landscape, Introducing Keras\n    * callbacks in, Callbacks\n    * conversion to Core ML, Conversion from Keras, Model Conversion Using Core ML Tools\n    * custom classifiers using transfer learning, Building a Custom Classifier in Keras with Transfer Learning-Analyzing the Results\n    * deploying models to Flask, Deploying a Keras Model to Flask\n    * fine tuning object classifiers with, Approach 3: Fine Tuning Using Keras\n    * history of, Keras\n    * image classification pipeline, Predicting an Image's Category-Predicting an Image's Category\n    * key function for feature extraction, Feature Extraction\n    * model accuracy, Investigating the Model-ImageNet Dataset\n    * pretrained models in, Model Zoos\n  * Keras functions\n    * decode_predictions, Predicting an Image's Category\n    * ImageDataGenerator, Data Augmentation\n    * preprocess_input, Predicting an Image's Category\n  * Keras Tuner, Tools of the Trade, Keras Tuner-Keras Tuner, Training\n  * Keras.js, Keras.js\n  * keypoint detection, PoseNet\n  * Knock Knock, Training\n  * KubeFlow\n    * fairing, Fairing\n    * installation, Installation\n    * introduction to, KubeFlow\n    * pipelines, Pipelines\n    * tools available, KubeFlow\n  * Kubernetes, KubeFlow\n\n### L\n\n  * labeling, Datasets, Data, How Do I Label My Data?, Labeling the Data\n  * latency, Low Latency, Real-Time Object Classification on iOS with Core ML\n  * layers, effect on accuracy, Effect of Number of Layers Fine-Tuned in Transfer Learning\n  * lazy learning, Similarity Search\n  * learning rate\n    * effect on accuracy, Effect of Learning Rate\n    * finding optimal, Find the Optimal Learning Rate\n    * role of, Set Training Parameters\n  * localization, Localization\n  * Lose It!, Lose It!\n  * loss function, Set Training Parameters, Siamese Networks for One-Shot Face Verification\n  * low latency, Low Latency\n\n### M\n\n  * machine learning (ML)\n    * Apple's machine learning architecture, Apple's Machine Learning Architecture-ML Performance Primitives, Approach 2: Use Create ML\n    * common techniques for, Common Techniques for Machine Learning Experimentation-Reproducible Experiments\n    * definition of, What Is AI?\n    * JavaScript-based ML libraries, JavaScript-Based Machine Learning Libraries: A Brief History-TensorFlow.js\n    * overview of, Machine Learning\n    * resources for learning about, Education and Exploration, Further Exploration\n    * support for multiple ML libraries, Support for Multiple Machine Learning Libraries\n    * terminology used in, Bias\n  * Magic Sudoku, Magic Sudoku\n  * magnitude-based weight pruning, Prune the Model\n  * map operations, Turn on Experimental Optimizations\n  * map_and_filter_fusion option, Map and filter fusion\n  * map_func function, Parallelize I\/O and Processing\n  * map_fusion option, Map fusion\n  * Markov decision process (MDP), The Markov decision process\n  * masks, Segmentation\n  * Mean Average Precision (mAP), Mean Average Precision\n  * Metacar, Metacar\n  * metric learning, Siamese Networks for One-Shot Face Verification\n  * metrics, Set Training Parameters\n  * MicroController Units (MCUs), TensorFlow Lite Architecture, Arduino\n  * Microsoft Cognitive Services, Microsoft Cognitive Services, Invoking Prebuilt Cloud-Based Object Detection APIs\n  * minibatches, Use Larger Batch Size\n  * MirroredStrategy, Distribute Training\n  * ML Kit\n    * benefits of, ML Kit\n    * common ML tasks supported by, ML Kit + Firebase\n    * cross-platform adaptability, TensorFlow Lite on iOS\n    * using custom trained TensorFlow Lite models, ML Kit + Firebase, Custom Models in ML Kit\n    * facial contours in, Face Contours in ML Kit\n    * using hosted models, Hosted Models-Using the Experiment in Code\n    * object classification, Object Classification in ML Kit\n  * ml5.js, ml5.js\n  * MLPerf, Examine Industry Benchmarks\n  * MMdnn, Model\n  * MNIST dataset, A Glimmer of Hope\n  * mobile-based AI\n    * Apple's machine learning architecture, Apple's Machine Learning Architecture-ML Performance Primitives, Approach 2: Use Create ML\n    * building real-time object recognition apps, Building a Real-Time Object Recognition App-Building a Real-Time Object Recognition App, Building a Real-Time Object Recognition App-Building a Real-Time Object Recognition App\n    * case studies, Case Studies-InstaSaber + YoPuppet, Case Studies-Real-Time Video Segmentation in YouTube Stories\n    * challenges of, Real-Time Object Classification on iOS with Core ML\n    * conversion to Core ML, Conversion to Core ML\n    * Core ML alternatives, Alternatives to Core ML\n    * Core ML history, A Brief History of Core ML\n    * development life cycle for, The Development Life Cycle for Artificial Intelligence on Mobile\n    * dynamic model deployment, Dynamic Model Deployment\n    * frequently asked questions, Real-Time Object Classification on iOS with Core ML, A Holistic Look at the Mobile AI App Development Cycle-How Do I Update the Model on My Users' Phones?\n    * Fritz end-to-end solution for, Fritz-Fritz\n    * measuring energy consumption, Measuring Energy Impact-Benchmarking Load\n    * on-device training, On-Device Training\n    * performance analysis, Performance Analysis-Benchmarking Models on iPhones\n    * reducing app size, Reducing App Size-Use Create ML\n    * self-evolving model, The Self-Evolving Model\n  * MobileNet, Training\n  * model accuracy\n    * improving TensorFlow performance, Inference-Enable GPU Persistence\n    * transfer learning with Keras, Analyzing the Results-Analyzing the Results\n  * model architecture, Model Architecture\n  * model cards, Model\n  * model compression, Prune the Model\n  * model definition, Model Definition\n  * model inference (TensorFlow)\n    * fused operations, Use Fused Operations\n    * GPU persistence, Enable GPU Persistence\n    * pruning, Prune the Model\n    * quantization techniques, Quantize the Model\n    * selecting efficient models, Use an Efficient Model\n  * model size, Model Size\n  * model testing, Test the Model\n  * model tips and tricks, Model\n  * model training\n    * analyzing with TensorFlow profiler, TensorFlow Profiler + TensorBoard\n    * approaches to, Training Our Model-Approach 3: Fine Tuning Using Keras\n    * browser-based AI, Training in the Browser-GPU Utilization\n    * mobile-based AI, How Do I Train My Model?\n    * optimizing in TensorFlow, Training-Examine Industry Benchmarks\n    * transfer learning with Keras, Set Training Parameters-Start Training\n    * web UI-based tools, Approach 1: Use Web UI-based Tools-Approach 1: Use Web UI-based Tools\n  * model versioning, Model Versioning\n  * model zoos, Model Zoos\n  * model-free reinforcement learning, Model free versus model based\n  * ModelDepot.io, Model\n  * ModelZoo.co, Model\n  * monitoring, Monitoring\n  * Movidius Neural Compute Stick 2, Intel Movidius Neural Compute Stick\n  * MS COCO (Microsoft Common Objects in COntext), Datasets, Image Captioning\n  * MS Paint, Model\n  * multiclass classification, Multiclass classification\n  * multilayer neural networks, A Glimmer of Hope\n  * multiple libraries support, Support for Multiple Machine Learning Libraries\n  * multiples of eight, Find the Optimal Learning Rate\n  * MultiWorkerMirroredStrategy, Distribute Training\n  * Myriad VPU, Intel Movidius Neural Compute Stick\n\n### N\n\n  * National Institute of Standards and Technology (NIST), A Glimmer of Hope\n  * NavLab, A Glimmer of Hope, A Brief History of Autonomous Driving\n  * nearest-neighbor approach\n    * efficacy of, Image Captioning\n    * scaling reverse image search using, Scaling Similarity Search with Approximate Nearest Neighbors-Faiss\n    * similarity search using, Similarity Search-Similarity Search\n  * negative classes, Data, Approach 2: Fatkun Chrome Browser Extension\n  * Neighborhood Graph and Tree (NGT), NGT\n  * Netron, Model\n  * Neural Architecture Search (NAS), Tools of the Trade, AutoKeras, Training\n  * neural networks (see also Convolutional Neural Networks)\n    * data augmentation, Data Augmentation\n    * definition of, Exciting Beginnings\n    * matrix-matrix multiplication in, Quantize the Model\n    * multilayer, A Glimmer of Hope\n    * overview of, Neural Networks\n    * purpose of validation sets, Organize the Data\n    * resources for learning about, Further Reading\n    * shortcomings of, Shortcoming of Neural Networks\n  * New York Times, The New York Times, Scalability\n  * NIST dataset, A Glimmer of Hope\n  * NN-SVG, Model\n  * Non-Maximum Suppression (NMS), Non-Maximum Suppression\n  * nondeterministic ordering, Enable Nondeterministic Ordering\n  * normalization, Predicting an Image's Category\n  * Not Hotdog app\n    * building iOS app, Building the iOS App\n    * collecting data, Collecting Data-Approach 3: Web Scraper Using Bing Image Search API\n    * model conversion using Core ML, Model Conversion Using Core ML Tools\n    * model training, Training Our Model-Approach 3: Fine Tuning Using Keras\n    * overview of steps, Not Hotdog on iOS with Core ML and Create ML\n  * number of nines, High Availability\n  * NVIDIA Data Loading Library (DALI) , tf.image built-in augmentations\n  * NVIDIA GPU Persistence Daemon, Enable GPU Persistence\n  * NVIDIA Graphics Processing Units (GPUs), Installation\n  * NVIDIA Jetson Nano, NVIDIA Jetson Nano, Port to NVIDIA Jetson Nano\n  * nvidia-smi (NVIDIA System Management Interface), nvidia-smi, Training\n\n### O\n\n  * object classification (see also food classifier apps)\n    * building iOS app, Building the iOS App\n    * building real-time apps for, Building a Real-Time Object Recognition App-Building a Real-Time Object Recognition App\n    * challenges of on mobile devices, Real-Time Object Classification on iOS with Core ML, Shazam for Food: Developing Android Apps with TensorFlow Lite and ML Kit\n    * collecting data, Collecting Data-Approach 3: Web Scraper Using Bing Image Search API\n    * in ML Kit, Object Classification in ML Kit\n    * model conversion using Core ML, Model Conversion Using Core ML Tools\n    * model training, Training Our Model-Approach 3: Fine Tuning Using Keras\n    * overview of steps, Not Hotdog on iOS with Core ML and Create ML\n    * real-time apps for, Building a Real-Time Object Recognition App-Building a Real-Time Object Recognition App\n  * object detection\n    * approaches to, Approaches to Object Detection\n    * building custom detectors without code, Building a Custom Detector Without Any Code-Building a Custom Detector Without Any Code\n    * case studies, Case Studies-Autonomous Cars\n    * evolution of, The Evolution of Object Detection\n    * image segmentation, Image Segmentation\n    * inspecting and training models, Inspecting the Model-Training\n    * key terms in, Key Terms in Object Detection-Non-Maximum Suppression\n    * model conversion, Model Conversion\n    * performance considerations, Performance Considerations\n    * prebuilt cloud-based APIs, Invoking Prebuilt Cloud-Based Object Detection APIs-Invoking Prebuilt Cloud-Based Object Detection APIs\n    * reusing pretrained models, Reusing a Pretrained Model-Deploying to a Device\n    * TensorFlow Object Detection API, Using the TensorFlow Object Detection API to Build Custom Models-Preprocessing the Data\n    * types of computer-vision tasks, Types of Computer-Vision Tasks-Approaches to Object Detection\n  * object localization, Detection\n  * object segmentation, Segmentation\n  * OCR APIs, Effect of Compression on OCR APIs\n  * OmniEarth, OmniEarth\n  * one shot learning, Siamese Networks for One-Shot Face Verification\n  * online resources\n    * code examples, Using Code Examples\n    * for this book, How to Contact Us\n  * ONNX.js, ONNX.js\n  * Open Images V4 , Data\n  * Open Neural Network Exchange (ONNX), A continuously evolving landscape, Model, ONNX.js\n  * optimizations (TensorFlow Lite)\n    * model optimization toolkit, TensorFlow Model Optimization Toolkit\n    * quantizing, Quantizing with TensorFlow Lite Converter\n  * optimizations (TensorFlow) (see also TensorFlow performance checklist)\n    * Automatic Mixed Precision, Use Automatic Mixed Precision\n    * better hardware, Use Better Hardware\n    * distributed training, Distribute Training\n    * eager execution and tf.function, Use tf.function\n    * experimental optimizations, Turn on Experimental Optimizations\n    * industry benchmarks, Examine Industry Benchmarks\n    * larger batch size, Use Larger Batch Size\n    * multiples of eight, Use Multiples of Eight\n    * optimized hardware stacks, Install an Optimized Stack for the Hardware\n    * optimizing parallel CPU threads, Optimize the Number of Parallel CPU Threads\n    * overtraining and generalization, Overtrain, and Then Generalize\n  * optimizers\n    * effect on accuracy, Effect of Optimizers\n    * role of, Set Training Parameters\n  * ordering, nondeterministic, Enable Nondeterministic Ordering\n  * overfitting, Data Augmentation, Callbacks\n  * overtraining, Early Stopping, Overtrain, and Then Generalize\n  * O'Reilly's Online Learning platform, Education and Exploration\n\n### P\n\n  * p-hacking, Reproducibility\n  * PapersWithCode.com, Model, Education and Exploration\n  * parallel processing, Parallelize CPU Processing, Optimize the Number of Parallel CPU Threads\n  * PCA (Principle Component Analysis), Reducing Feature-Length with PCA\n  * perceptrons, Exciting Beginnings, The Cold and Dark Days, Perceptron\n  * performance tuning, Performance Tuning for Cloud APIs-Effect of Resizing on OCR APIs (see also TensorFlow performance checklist)\n  * Personally Identifiable Information (PII), Real-Time Object Classification on iOS with Core ML\n  * Photobucket , Photobucket\n  * Pilot Parliaments Benchmark (PPB), Bias\n  * Pinterest, Pinterest\n  * pipelines\n    * building in custom classifiers, Build the Data Pipeline\n    * end-to-end learning example pipeline, End-to-End Deep Learning Example Pipeline-Basic Custom Network Pipeline\n    * end-to-end object recognition example, Not Hotdog on iOS with Core ML and Create ML-Summary\n    * in KubeFlow, Pipelines\n    * reading data in TensorFlow, Use tf.data\n    * ready-to-use datasets for, TensorFlow Datasets\n    * simple pipeline for image classification, Predicting an Image's Category\n    * tuning TensorFlow performance, How to Use This Checklist-Enable GPU Persistence\n  * pix2pix, pix2pix-pix2pix\n  * Pixel 3 phones, Portrait Mode on Pixel 3 Phones\n  * pixels, Predicting an Image's Category\n  * PlotNeuralNet, Model\n  * policy-based reinforcement learning, Policy based\n  * pooling operations, Length of Feature Vectors, Pooling\n  * portrait mode, Portrait Mode on Pixel 3 Phones\n  * PoseNet, PoseNet, Squatting for Metro Tickets\n  * pretrained models\n    * adapting to new frameworks, Model\n    * adapting to new tasks, Adapting Pretrained Models to New Tasks-How Much to Fine Tune\n    * from Apple Machine Learning website, ML Performance Primitives\n    * in Keras, Model Zoos\n    * locating, Model\n    * reusing for object detection, Reusing a Pretrained Model-Deploying to a Device\n    * in TensorFlow.js, Running Pretrained Models Using TensorFlow.js\n  * principal components, Reducing Feature-Length with PCA\n  * privacy, Privacy, Privacy, Real-Time Object Classification on iOS with Core ML\n  * profiling, TensorFlow Profiler + TensorBoard\n  * progress bars, Training\n  * progressive augmentation, Use progressive augmentation\n  * progressive resizing, Use progressive resizing\n  * progressive sampling, Use progressive sampling\n  * Project Oxford, Microsoft Cognitive Services\n  * proximal policy optimization (PPO) algorithm, Configure the simulation environment, Reinforcement Learning Algorithm in AWS DeepRacer\n  * pruning, Prune the Model\n  * PYNQ platform, FPGA + PYNQ\n  * PyTorch, PyTorch\n\n### Q\n\n  * Qri, Data\n  * quantization techniques, Quantize the Model, Use Quantization, Quantizing with TensorFlow Lite Converter\n  * Quilt, Data\n\n### R\n\n  * randomization, Reproducible Experiments\n  * Raspberry Pi, Raspberry Pi, Hands-On with the Raspberry Pi\n  * region of interest (ROI), Identifying the Region of Interest\n  * reinforcement learning\n    * AWS DeepRacer algorithm, Reinforcement Learning Algorithm in AWS DeepRacer\n    * AWS DeepRacer car, Racing the AWS DeepRacer Car-Sim2Real transfer\n    * AWS DeepRacer example, Practical Deep Reinforcement Learning with DeepRacer-Step 4: Evaluating the Performance of the Model\n    * crux of, Building an Autonomous Car in Under an Hour: Reinforcement Learning with AWS DeepRacer\n    * Distributed Deep Reinforcement Learning for Autonomous Driving tutorial , Reinforcement Learning\n    * further exploration, Further Exploration-Roborace\n    * improving learning models, Step 5: Improving Reinforcement Learning Models-Racing the AWS DeepRacer Car\n    * inner workings of, Reinforcement Learning in Action\n    * introduction to, A Brief Introduction to Reinforcement Learning\n    * learning with autonomous cars, Why Learn Reinforcement Learning with an Autonomous Car?-Why Learn Reinforcement Learning with an Autonomous Car?\n    * summary of, Deep Reinforcement Learning Summary with DeepRacer as an Example\n    * terminology used in, Practical Deep Reinforcement Learning with DeepRacer\n    * theory of, Reinforcement Learning Theory\n  * replay buffer, Practical Deep Reinforcement Learning with DeepRacer\n  * reporting bias, Bias\n  * reproducibility, Reproducibility, Reproducible Experiments\n  * ResearchCode, Education and Exploration\n  * resizing\n    * effect on accuracy, Effect of Resizing\n    * images for performance tuning, Performance Tuning for Cloud APIs\n  * ResNet, Training\n  * ResNet-50, Predicting an Image's Category, Predicting an Image's Category\n  * REST APIs, Making a REST API with Flask\n  * reverse image search\n    * case studies, Case Studies-Image Captioning\n    * feature extraction, Feature Extraction-Feature Extraction\n    * fine tuning for improved accuracy, Improving Accuracy with Fine Tuning-Fine Tuning Without Fully Connected Layers\n    * image similarity, Image Similarity-Image Similarity\n    * improving search speed, Improving the Speed of Similarity Search-Reducing Feature-Length with PCA\n    * introduction to, Building a Reverse Image Search Engine: Understanding Embeddings\n    * scaling with nearest-neighbor approach, Scaling Similarity Search with Approximate Nearest Neighbors-Faiss\n    * Siamese networks for, Siamese Networks for One-Shot Face Verification\n    * similarity search, Similarity Search-Similarity Search\n    * visualizing image clusters, Visualizing Image Clusters with t-SNE-Visualizing Image Clusters with t-SNE\n  * reward function, Practical Deep Reinforcement Learning with DeepRacer, Configure reward function-Configure reward function, How Does a Reinforcement Learning System Learn?\n  * Roborace, Roborace\n  * robustness, Robustness\n  * ROCm stack, Installation\n  * root-mean-square error (RMSE), Set Training Parameters\n  * round-robin order, Enable Nondeterministic Ordering\n  * Runway ML, Education and Exploration\n\n### S\n\n  * SageMaker reinforcement learning, A Brief Introduction to Reinforcement Learning\n  * SamaSource, Data\n  * scalability, Scalability\n  * scaling horizontally, Distribute Training\n  * ScrapeStorm.com, Data\n  * Scrapy.org, Data\n  * Seeing AI, Seeing AI, Face Detection in Seeing AI\n  * segmentation, Segmentation\n  * selection bias, Bias\n  * self-driving cars, A Glimmer of Hope, Autonomous Cars (see also autonomous driving; reinforcement learning)\n  * self-evolving model, The Self-Evolving Model\n  * semantic segmentation, Semantic segmentation\n  * Semi-Conductor, Semi-Conductor\n  * semi-supervised learning, Approach 3: Web Scraper Using Bing Image Search API\n  * sequential data, Training on Sequential Data\n  * Shazam for Food app, Shazam for Food: Developing Android Apps with TensorFlow Lite and ML Kit\n  * Siamese networks, Siamese Networks for One-Shot Face Verification\n  * Sim2Real transfer, Sim2Real transfer\n  * similarity search\n    * finding nearest neighbors, Similarity Search-Similarity Search\n    * improving speed of, Improving the Speed of Similarity Search-Reducing Feature-Length with PCA\n    * scaling, Scaling Similarity Search with Approximate Nearest Neighbors-Faiss\n  * simulation, Deep Learning, Autonomous Driving, and the Data Problem, Practical Deep Reinforcement Learning with DeepRacer, Sim2Real transfer\n  * simulation-to-real (sim2real) problem, Why Learn Reinforcement Learning with an Autonomous Car?, Practical Deep Reinforcement Learning with DeepRacer, Sim2Real transfer\n  * smart refrigerators, Smart Refrigerator\n  * softmax activation function, Start Training\n  * SOTAWHAT tool, Education and Exploration\n  * speaker recognition, Speaker Recognition by Alibaba\n  * Spotify, Spotify\n  * squat-tracker app, Squatting for Metro Tickets\n  * SSD MobileNetV2, Reusing a Pretrained Model\n  * Staples, Staples\n  * stop conditions, Configure stop conditions\n  * success, measuring, How Do I Measure the Success of My Model?\n  * synthetic datasets, Which Library Should I Use?, Data\n\n### T\n\n  * t-SNE algorithm, Visualizing Image Clusters with t-SNE-Visualizing Image Clusters with t-SNE\n  * Teachable Machine, Training in the Browser\n  * Tencent ML Images, Data\n  * TensorBoard, TensorBoard-What-If Tool, TensorFlow Profiler + TensorBoard\n  * TensorFlow\n    * AMD GPUs and, Installation\n    * benefits of, A continuously evolving landscape\n    * conversion to Core ML, Conversion from TensorFlow\n    * debugging scripts, Training\n    * history of, TensorFlow\n    * installation on Windows PCs, Installation\n    * testing for available GPUs, Training\n  * TensorFlow Datasets, TensorFlow Datasets, Use TensorFlow Datasets, Data\n  * TensorFlow Debugger (tfdbg), Training\n  * TensorFlow Embedding projector, Visualizing Image Clusters with t-SNE\n  * TensorFlow Encrypted, Privacy\n  * TensorFlow Extended (TFX), TensorFlow Serving\n  * TensorFlow Hub (tfhub.dev), Model\n  * TensorFlow Lite\n    * architecture, TensorFlow Lite Architecture\n    * building real-time object recognition apps, Building a Real-Time Object Recognition App-Building a Real-Time Object Recognition App\n    * history of, TensorFlow Lite\n    * model conversion to, Model Conversion to TensorFlow Lite\n    * model optimization toolkit, TensorFlow Model Optimization Toolkit\n    * on iOS, TensorFlow Lite on iOS\n    * overview of, An Overview of TensorFlow Lite\n    * performance optimizations, Performance Optimizations\n    * sample apps in repository, Building a Real-Time Object Recognition App\n    * using models with ML Kit, ML Kit + Firebase, Custom Models in ML Kit\n  * TensorFlow models repository, Building a Real-Time Object Recognition App, Obtaining the Model\n  * TensorFlow Object Detection API\n    * data collection, Data Collection\n    * data labeling, Labeling the Data\n    * data preprocessing, Preprocessing the Data\n  * TensorFlow performance checklist\n    * data augmentation, Data Augmentation-NVIDIA DALI\n    * data preparation, Data Preparation-Use TensorFlow Datasets\n    * data reading, Data Reading-Autotune Parameter Values\n    * effective use of, How to Use This Checklist\n    * GPU starvation, GPU Starvation-TensorFlow Profiler + TensorBoard\n    * inference, Inference-Enable GPU Persistence\n    * overview of, Data Preparation\n    * training, Training-Examine Industry Benchmarks\n  * TensorFlow Playground, TensorFlow.js\n  * TensorFlow Serving, TensorFlow Serving\n  * TensorFlow.js\n    * architecture, TensorFlow.js Architecture\n    * introduction to, TensorFlow.js\n    * running pretrained models, Running Pretrained Models Using TensorFlow.js\n  * TensorSpace, TensorSpace\n  * tf-coreml, Conversion from TensorFlow\n  * tf-explain, Tools of the Trade, tf-explain\n  * tf.data, Use tf.data, Enable Nondeterministic Ordering\n  * tf.data.experimental.AUTOTUNE, Autotune Parameter Values\n  * tf.data.experimental.OptimizationOptions, Turn on Experimental Optimizations\n  * tf.function, Use tf.function\n  * tf.image, tf.image built-in augmentations\n  * tf.keras, Tools of the Trade\n  * tf.test.is_gpu_available(), Training\n  * tfprof (TensorFlow profiler), TensorFlow Profiler + TensorBoard\n  * TFRecords, Store as TFRecords, Preprocessing the Data\n  * timeit command, Brute Force\n  * tools, tips, and tricks\n    * data, Data-Data\n    * education and exploration, Education and Exploration-Education and Exploration\n    * installation, Installation-Installation\n    * models, Model\n    * privacy, Privacy\n    * training, Training\n  * training parameters, Set Training Parameters\n  * Transaction Processing Council (TPC) benchmark, Examine Industry Benchmarks\n  * transfer learning\n    * adapting pretrained models to new networks, Model\n    * adapting pretrained models to new tasks, Adapting Pretrained Models to New Tasks-How Much to Fine Tune\n    * basic pipeline for, Basic Transfer Learning Pipeline\n    * building custom classifiers using, Building a Custom Classifier in Keras with Transfer Learning-Analyzing the Results\n    * Create ML, Approach 2: Use Create ML\n    * definition of, Datasets\n    * effect of hyperparmeters on accuracy, How Hyperparameters Affect Accuracy-Effect of Change in Aspect Ratio on Transfer Learning\n    * fine tuning, Fine Tuning-How Much to Fine Tune\n    * GUI-based model training tools, A continuously evolving landscape\n    * overview of, Transfer Learning\n    * versus training from scratch, Transfer Learning Versus Training from Scratch\n  * triplet loss function, Siamese Networks for One-Shot Face Verification\n\n### U\n\n  * Uber, Uber\n  * Ubuntu, Installation\n  * underfitting, Data Augmentation\n  * Uniform Manifold Approximation and Projection (UMAP), Visualizing Image Clusters with t-SNE\n  * Universal Approximation Theorem, A Glimmer of Hope\n  * updates, distributing, How Do I Update the Model on My Users' Phones?\n  * USB accelerators, Google Coral USB Accelerator, Speeding Up with the Google Coral USB Accelerator\n\n### V\n\n  * validation accuracy, Start Training\n  * value function (V), Practical Deep Reinforcement Learning with DeepRacer\n  * value-based reinforcement learning, Value based\n  * vanilla policy gradient algorithm, Reinforcement Learning Algorithm in AWS DeepRacer\n  * versioning, Training, Data, Model Versioning\n  * video datasets, Data\n  * video segmentation, Real-Time Video Segmentation in YouTube Stories\n  * visual recognition APIs (see also computer vision)\n    * accuracy, Accuracy\n    * available APIs, The Landscape of Visual Recognition APIs-What's unique about this API?\n    * bias, Bias-Bias\n    * cost, Cost\n    * service offerings, Service Offerings\n  * VisualData.io, Data\n  * visualizations\n    * real-time for training process, Training\n    * using GAN Lab, GAN Lab\n    * using t-SNE algorithm, Visualizing Image Clusters with t-SNE-Visualizing Image Clusters with t-SNE\n    * using TensorBoard, TensorBoard-What-If Tool\n    * using TensorFlow profiler, TensorFlow Profiler + TensorBoard\n    * using tf-explain, tf-explain\n    * using What-If Tool, What-If Tool-What-If Tool\n\n### W\n\n  * Waymo, Deep Learning, Autonomous Driving, and the Data Problem\n  * web UI-based tools, Approach 1: Use Web UI-based Tools-Approach 1: Use Web UI-based Tools\n  * WebScraper.io, Data\n  * weights, Model Architecture\n  * Weights and Biases, Training\n  * What-If Tool, Tools of the Trade, What-If Tool-What-If Tool\n  * wildlife conservation, Wildlife conservation\n\n### X\n\n  * Xcode, Building a Real-Time Object Recognition App\n  * Xilinx Zynq chips, FPGA + PYNQ\n\n### Y\n\n  * YFCC100M, Data\n  * YoPuppet, InstaSaber + YoPuppet\n  * YouTube Stories, Real-Time Video Segmentation in YouTube Stories\n\n# About the Authors\n\n## Primary Authors\n\n**Anirudh Koul** is a noted AI expert, a UN\/TEDx speaker, and a former senior scientist at Microsoft AI & Research, where he founded Seeing AI, often considered the most used technology among the blind community after the iPhone. Anirudh serves as the head of AI & Research at Aira, recognized by _Time_ magazine as one of the best inventions of 2018. With features shipped to a billion users, he brings over a decade of production-oriented applied research experience on petabyte-scale datasets. He has been developing technologies using AI techniques for augmented reality, robotics, speech, and productivity as well as accessibility. His work with AI for Good, which IEEE has called \"life-changing,\" has received awards from CES, FCC, MIT, Cannes Lions, and the American Council of the Blind; been showcased at events by UN, World Economic Forum, White House, House of Lords, Netflix, and National Geographic; and lauded by world leaders including Justin Trudeau and Theresa May.\n\n**Siddha Ganju** , an AI researcher who _Forbes_ featured in its \"30 under 30\" list, is a self-driving architect at Nvidia. As an AI advisor to NASA FDL, she helped build an automated meteor detection pipeline for the CAMS project at NASA, which discovered a comet. Previously at Deep Vision, she developed deep learning models for resource constraint edge devices. Her work ranges from visual question answering to generative adversarial networks to gathering insights from CERN's petabyte-scale data and has been published at top-tier conferences including CVPR and NeurIPS. She has served as a featured jury member in several international tech competitions including CES. As an advocate for diversity and inclusion in technology, she speaks at schools and colleges to motivate and grow a new generation of technologists from all backgrounds.\n\n**Meher Kasam** is a seasoned software developer with apps used by tens of millions of users every day. Currently an iOS developer at Square, and having previously worked at Microsoft and Amazon, he has shipped features for a range of apps from Square's Point of Sale to the Bing iPhone app. During his time at Microsoft, he was the mobile development lead for the Seeing AI app, which has received widespread recognition and awards from Mobile World Congress, CES, FCC, and the American Council of the Blind, to name a few. A hacker at heart with a flair for fast prototyping, he's won several hackathons and converted them to features shipped in widely used products. He also serves as a judge of international competitions including Global Mobile Awards and Edison Awards.\n\n## Guest Authors\n\n### Chapter 17\n\n**Sunil Mallya** is a principal deep learning scientist at AWS focused on deep learning and reinforcement learning at AWS. Sunil works with AWS customers in various transformation and innovation initiatives across verticals by building models for cutting edge machine learning applications. He also led the science development for AWS DeepRacer. Prior to joining AWS, Sunil cofounded the neuroscience and machine learning\u2013based image analysis and video thumbnail recommendation company Neon Labs. He has worked on building large-scale low-latency systems at Zynga and has an acute passion for serverless computing. He holds a master's degree in computer science from Brown University.\n\n### Chapter 16\n\n**Aditya Sharma** is a program manager at Microsoft where he works on the Azure Autonomous Driving team. His work focuses on helping autonomous driving companies scale their operations by leveraging the power of the cloud, greatly reducing their time to market. He is the lead PM for Project Road Runner, the team behind the Autonomous Driving Cookbook. He also leads the Deep Learning and Robotics chapter at the Microsoft Garage. Aditya holds a graduate degree from the Robotics Institute at Carnegie Mellon University.\n\n### Chapter 16\n\n**Mitchell Spryn** graduated from the University of Alabama with a dual bachelor's degree in electrical engineering and physics. While at the university, Mitchell was involved in a variety of research projects ranging from wireless power transfer to autonomous robotics. He currently works at Microsoft as a data scientist, specializing in distributed relational databases. In addition to his work on databases, Mitchell continues to contribute to projects in the robotics space such as the AirSim simulator and the Autonomous Driving Cookbook.\n\n### Chapter 15\n\n**Sam Sterckval** , a Belgium-based electronics engineer, discovered his second passion (besides electronics) during his time in college\u2014artificial intelligence. After graduation, Sam cofounded the company Edgise to combine these two passions\u2014to bring artificial intelligence to the edge by designing custom hardware that can efficiently run complex AI models. Often considered the go-to person for edge AI in Belgium, Sam helps companies to build cognitive solutions that highly depend on low latency, privacy, and cost-effectiveness.\n\n### Chapter 10\n\n**Zaid Alyafeai** is known for his popular blogs and Google Colab Notebooks on TensorFlow.js and using deep learning within the browser to power interesting scenarios. He is also a writer for the official TensorFlow blog on Medium, with many GitHub projects with more than 1000 stars combined. When he is not building fancy web applications, he is conducting research in AI as a PhD student at KFUPM in Saudi Arabia. A teacher at heart, he also wrote a reference book on advanced integration techniques.\n\n# Colophon\n\nThe animal on the cover of _Practical Deep Learning for Cloud, Mobile, and Edge_ is the American peregrine falcon ( _Falco peregrinus anatum_ ) or duck hawk, a subspecies of peregrine falcon, which is currently found in the Rocky Mountains. This subspecies has become extinct in eastern North America, but healthy hybrid populations now exist due to organized reintroductions.\n\nPeregrine falcons worldwide, no matter the subspecies, share the characteristics of being a large, fast-flying bird of prey with a dark head and wings; pale, patterned undersides; yellow beak and legs; and large eyes set in a distinctive vertical stripe of eyeblack. Falcons can be discerned at a distance by their \"bent\" but otherwise sharply outlined wings. Females are noticeably larger than males, but both average about the size of an American crow, and make a distinctive, sharp, \"kak kak kak\" call when defending territory or nestlings.\n\nPeregrines (named for their far-wandering habits outside nesting season) have established themselves on all continents except Antarctica. Their species success lies with their many adaptations for hunting birds, as well as their ability to adapt to varying nesting environments and prey. The best-known habit of the peregrine falcon is its hunting stoop or high-speed dive, in which it dives from a great height into a bird on or near the ground, reaching speeds of up to 200 mph by the time it crashes into its prey, many times killing it instantly. This makes it not only the fastest bird on earth, but the fastest member of the animal kingdom, and is why the peregrine is a favorite bird of falconers around the world.\n\nPeregrine falcons became a visible symbol of the US environmental movement in the 1960s, as widespread use of the pesticide DDT extirpated the birds across much of their former range before the chemical was banned (the birds consumed DDT through their prey, which had the effect of lethally thinning the shells of the falcon's eggs). The US Environmental Protection Agency declared this falcon an endangered species in 1970. However, the species has recovered due to the ban on DDT, followed by a captive breeding program that reintroduced the species into their former range in the eastern US. With nest boxes added into the upper reaches of city skyscrapers, populations quickly became established, as the falcons raised their chicks on pigeons and other plentiful urban birds. The peregrine falcon was removed from the Endangered Species list in 1999. While the American peregrine falcon's current conservation status is now designated as of Least Concern, many of the animals on O'Reilly covers are endangered; all of them are important to the world.\n\nThe cover illustration is by Karen Montgomery, based on a black and white engraving from _British Birds_. The cover fonts are Gilroy Semibold and Guardian Sans. The text font is Adobe Minion Pro; the heading font is Adobe Myriad Condensed; and the code font is Dalton Maag's Ubuntu Mono.\n  1. Preface\n    1. To the Backend\/Frontend\/Mobile Software Developer\n    2. To the Data Scientist\n    3. To the Student\n    4. To the Teacher\n    5. To the Robotics Enthusiast\n    6. What to Expect in Each Chapter\n    7. Conventions Used in This Book\n    8. Using Code Examples\n    9. O'Reilly Online Learning\n    10. How to Contact Us\n    11. Acknowledgments\n      1. Group Acknowledgments\n      2. Personal Acknowledgments\n  2. 1. Exploring the Landscape of Artificial Intelligence\n    1. An Apology\n    2. The Real Introduction\n    3. What Is AI?\n      1. Motivating Examples\n    4. A Brief History of AI\n      1. Exciting Beginnings\n      2. The Cold and Dark Days\n      3. A Glimmer of Hope\n      4. How Deep Learning Became a Thing\n    5. Recipe for the Perfect Deep Learning Solution\n      1. Datasets\n      2. Model Architecture\n      3. Frameworks\n      4. Hardware\n    6. Responsible AI\n      1. Bias\n      2. Accountability and Explainability\n      3. Reproducibility\n      4. Robustness\n      5. Privacy\n    7. Summary\n    8. Frequently Asked Questions\n  3. 2. What's in the Picture: Image Classification with Keras\n    1. Introducing Keras\n    2. Predicting an Image's Category\n    3. Investigating the Model\n      1. ImageNet Dataset\n      2. Model Zoos\n      3. Class Activation Maps\n    4. Summary\n  4. 3. Cats Versus Dogs: Transfer Learning in 30 Lines with Keras\n    1. Adapting Pretrained Models to New Tasks\n      1. A Shallow Dive into Convolutional Neural Networks\n      2. Transfer Learning\n      3. Fine Tuning\n      4. How Much to Fine Tune\n    2. Building a Custom Classifier in Keras with Transfer Learning\n    3. Organize the Data\n    4. Build the Data Pipeline\n      1. Number of Classes\n      2. Batch Size\n    5. Data Augmentation\n    6. Model Definition\n    7. Train the Model\n      1. Set Training Parameters\n      2. Start Training\n    8. Test the Model\n    9. Analyzing the Results\n    10. Further Reading\n    11. Summary\n  5. 4. Building a Reverse Image Search Engine: Understanding Embeddings\n    1. Image Similarity\n    2. Feature Extraction\n    3. Similarity Search\n    4. Visualizing Image Clusters with t-SNE\n    5. Improving the Speed of Similarity Search\n      1. Length of Feature Vectors\n      2. Reducing Feature-Length with PCA\n    6. Scaling Similarity Search with Approximate Nearest Neighbors\n      1. Approximate Nearest-Neighbor Benchmark\n      2. Which Library Should I Use?\n      3. Creating a Synthetic Dataset\n      4. Brute Force\n      5. Annoy\n      6. NGT\n      7. Faiss\n    7. Improving Accuracy with Fine Tuning\n      1. Fine Tuning Without Fully Connected Layers\n    8. Siamese Networks for One-Shot Face Verification\n    9. Case Studies\n      1. Flickr\n      2. Pinterest\n      3. Celebrity Doppelgangers\n      4. Spotify\n      5. Image Captioning\n    10. Summary\n  6. 5. From Novice to Master Predictor: Maximizing Convolutional Neural Network Accuracy\n    1. Tools of the Trade\n      1. TensorFlow Datasets\n      2. TensorBoard\n      3. What-If Tool\n      4. tf-explain\n    2. Common Techniques for Machine Learning Experimentation\n      1. Data Inspection\n      2. Breaking the Data: Train, Validation, Test\n      3. Early Stopping\n      4. Reproducible Experiments\n    3. End-to-End Deep Learning Example Pipeline\n      1. Basic Transfer Learning Pipeline\n      2. Basic Custom Network Pipeline\n    4. How Hyperparameters Affect Accuracy\n      1. Transfer Learning Versus Training from Scratch\n      2. Effect of Number of Layers Fine-Tuned in Transfer Learning\n      3. Effect of Data Size on Transfer Learning\n      4. Effect of Learning Rate\n      5. Effect of Optimizers\n      6. Effect of Batch Size\n      7. Effect of Resizing\n      8. Effect of Change in Aspect Ratio on Transfer Learning\n    5. Tools to Automate Tuning for Maximum Accuracy\n      1. Keras Tuner\n      2. AutoAugment\n      3. AutoKeras\n    6. Summary\n  7. 6. Maximizing Speed and Performance of TensorFlow: A Handy Checklist\n    1. GPU Starvation\n      1. nvidia-smi\n      2. TensorFlow Profiler + TensorBoard\n    2. How to Use This Checklist\n    3. Performance Checklist\n      1. Data Preparation\n      2. Data Reading\n      3. Data Augmentation\n      4. Training\n      5. Inference\n    4. Data Preparation\n      1. Store as TFRecords\n      2. Reduce Size of Input Data\n      3. Use TensorFlow Datasets\n    5. Data Reading\n      1. Use tf.data\n      2. Prefetch Data\n      3. Parallelize CPU Processing\n      4. Parallelize I\/O and Processing\n      5. Enable Nondeterministic Ordering\n      6. Cache Data\n      7. Turn on Experimental Optimizations\n      8. Autotune Parameter Values\n    6. Data Augmentation\n      1. Use GPU for Augmentation\n    7. Training\n      1. Use Automatic Mixed Precision\n      2. Use Larger Batch Size\n      3. Use Multiples of Eight\n      4. Find the Optimal Learning Rate\n      5. Use tf.function\n      6. Overtrain, and Then Generalize\n      7. Install an Optimized Stack for the Hardware\n      8. Optimize the Number of Parallel CPU Threads\n      9. Use Better Hardware\n      10. Distribute Training\n      11. Examine Industry Benchmarks\n    8. Inference\n      1. Use an Efficient Model\n      2. Quantize the Model\n      3. Prune the Model\n      4. Use Fused Operations\n      5. Enable GPU Persistence\n    9. Summary\n  8. 7. Practical Tools, Tips, and Tricks\n    1. Installation\n    2. Training\n    3. Model\n    4. Data\n    5. Privacy\n    6. Education and Exploration\n    7. One Last Question\n  9. 8. Cloud APIs for Computer Vision: Up and Running in 15 Minutes\n    1. The Landscape of Visual Recognition APIs\n      1. Clarifai\n      2. Microsoft Cognitive Services\n      3. Google Cloud Vision\n      4. Amazon Rekognition\n      5. IBM Watson Visual Recognition\n      6. Algorithmia\n    2. Comparing Visual Recognition APIs\n      1. Service Offerings\n      2. Cost\n      3. Accuracy\n      4. Bias\n    3. Getting Up and Running with Cloud APIs\n    4. Training Our Own Custom Classifier\n      1. Top Reasons Why Our Classifier Does Not Work Satisfactorily\n    5. Comparing Custom Classification APIs\n    6. Performance Tuning for Cloud APIs\n      1. Effect of Resizing on Image Labeling APIs\n      2. Effect of Compression on Image Labeling APIs\n      3. Effect of Compression on OCR APIs\n      4. Effect of Resizing on OCR APIs\n    7. Case Studies\n      1. The New York Times\n      2. Uber\n      3. Giphy\n      4. OmniEarth\n      5. Photobucket\n      6. Staples\n      7. InDro Robotics\n    8. Summary\n  10. 9. Scalable Inference Serving on Cloud with TensorFlow Serving and KubeFlow\n    1. Landscape of Serving AI Predictions\n    2. Flask: Build Your Own Server\n      1. Making a REST API with Flask\n      2. Deploying a Keras Model to Flask\n      3. Pros of Using Flask\n      4. Cons of Using Flask\n    3. Desirable Qualities in a Production-Level Serving System\n      1. High Availability\n      2. Scalability\n      3. Low Latency\n      4. Geographic Availability\n      5. Failure Handling\n      6. Monitoring\n      7. Model Versioning\n      8. A\/B Testing\n      9. Support for Multiple Machine Learning Libraries\n    4. Google Cloud ML Engine: A Managed Cloud AI Serving Stack\n      1. Pros of Using Cloud ML Engine\n      2. Cons of Using Cloud ML Engine\n      3. Building a Classification API\n    5. TensorFlow Serving\n      1. Installation\n    6. KubeFlow\n      1. Pipelines\n      2. Fairing\n      3. Installation\n    7. Price Versus Performance Considerations\n      1. Cost Analysis of Inference-as-a-Service\n      2. Cost Analysis of Building Your Own Stack\n    8. Summary\n  11. 10. AI in the Browser with TensorFlow.js and ml5.js\n    1. JavaScript-Based Machine Learning Libraries: A Brief History\n      1. ConvNetJS\n      2. Keras.js\n      3. ONNX.js\n      4. TensorFlow.js\n    2. TensorFlow.js Architecture\n    3. Running Pretrained Models Using TensorFlow.js\n    4. Model Conversion for the Browser\n    5. Training in the Browser\n      1. Feature Extraction\n      2. Data Collection\n      3. Training\n      4. GPU Utilization\n    6. ml5.js\n    7. PoseNet\n    8. pix2pix\n    9. Benchmarking and Practical Considerations\n      1. Model Size\n      2. Inference Time\n    10. Case Studies\n      1. Semi-Conductor\n      2. TensorSpace\n      3. Metacar\n      4. Airbnb's Photo Classification\n      5. GAN Lab\n    11. Summary\n  12. 11. Real-Time Object Classification on iOS with Core ML\n    1. The Development Life Cycle for Artificial Intelligence on Mobile\n    2. A Brief History of Core ML\n    3. Alternatives to Core ML\n      1. TensorFlow Lite\n      2. ML Kit\n      3. Fritz\n    4. Apple's Machine Learning Architecture\n      1. Domain-Based Frameworks\n      2. ML Framework\n      3. ML Performance Primitives\n    5. Building a Real-Time Object Recognition App\n    6. Conversion to Core ML\n      1. Conversion from Keras\n      2. Conversion from TensorFlow\n    7. Dynamic Model Deployment\n    8. On-Device Training\n      1. Federated Learning\n    9. Performance Analysis\n      1. Benchmarking Models on iPhones\n    10. Measuring Energy Impact\n      1. Benchmarking Load\n    11. Reducing App Size\n      1. Avoid Bundling the Model\n      2. Use Quantization\n      3. Use Create ML\n    12. Case Studies\n      1. Magic Sudoku\n      2. Seeing AI\n      3. HomeCourt\n      4. InstaSaber + YoPuppet\n    13. Summary\n  13. 12. Not Hotdog on iOS with Core ML and Create ML\n    1. Collecting Data\n      1. Approach 1: Find or Collect a Dataset\n      2. Approach 2: Fatkun Chrome Browser Extension\n      3. Approach 3: Web Scraper Using Bing Image Search API\n    2. Training Our Model\n      1. Approach 1: Use Web UI-based Tools\n      2. Approach 2: Use Create ML\n      3. Approach 3: Fine Tuning Using Keras\n    3. Model Conversion Using Core ML Tools\n    4. Building the iOS App\n    5. Further Exploration\n    6. Summary\n  14. 13. Shazam for Food: Developing Android Apps with TensorFlow Lite and ML Kit\n    1. The Life Cycle of a Food Classifier App\n    2. An Overview of TensorFlow Lite\n      1. TensorFlow Lite Architecture\n    3. Model Conversion to TensorFlow Lite\n    4. Building a Real-Time Object Recognition App\n    5. ML Kit + Firebase\n      1. Object Classification in ML Kit\n      2. Custom Models in ML Kit\n      3. Hosted Models\n      4. A\/B Testing Hosted Models\n      5. Using the Experiment in Code\n    6. TensorFlow Lite on iOS\n    7. Performance Optimizations\n      1. Quantizing with TensorFlow Lite Converter\n      2. TensorFlow Model Optimization Toolkit\n    8. Fritz\n    9. A Holistic Look at the Mobile AI App Development Cycle\n      1. How Do I Collect Initial Data?\n      2. How Do I Label My Data?\n      3. How Do I Train My Model?\n      4. How Do I Convert the Model to a Mobile-Friendly Format?\n      5. How Do I Make my Model Performant?\n      6. How Do I Build a Great UX for My Users?\n      7. How Do I Make the Model Available to My Users?\n      8. How Do I Measure the Success of My Model?\n      9. How Do I Improve My Model?\n      10. How Do I Update the Model on My Users' Phones?\n    10. The Self-Evolving Model\n    11. Case Studies\n      1. Lose It!\n      2. Portrait Mode on Pixel 3 Phones\n      3. Speaker Recognition by Alibaba\n      4. Face Contours in ML Kit\n      5. Real-Time Video Segmentation in YouTube Stories\n    12. Summary\n  15. 14. Building the Purrfect Cat Locator App with TensorFlow Object Detection API\n    1. Types of Computer-Vision Tasks\n      1. Classification\n      2. Localization\n      3. Detection\n      4. Segmentation\n    2. Approaches to Object Detection\n    3. Invoking Prebuilt Cloud-Based Object Detection APIs\n    4. Reusing a Pretrained Model\n      1. Obtaining the Model\n      2. Test Driving Our Model\n      3. Deploying to a Device\n    5. Building a Custom Detector Without Any Code\n    6. The Evolution of Object Detection\n      1. Performance Considerations\n    7. Key Terms in Object Detection\n      1. Intersection over Union\n      2. Mean Average Precision\n      3. Non-Maximum Suppression\n    8. Using the TensorFlow Object Detection API to Build Custom Models\n      1. Data Collection\n      2. Labeling the Data\n      3. Preprocessing the Data\n    9. Inspecting the Model\n      1. Training\n      2. Model Conversion\n    10. Image Segmentation\n    11. Case Studies\n      1. Smart Refrigerator\n      2. Crowd Counting\n      3. Face Detection in Seeing AI\n      4. Autonomous Cars\n    12. Summary\n  16. 15. Becoming a Maker: Exploring Embedded AI at the Edge\n    1. Exploring the Landscape of Embedded AI Devices\n      1. Raspberry Pi\n      2. Intel Movidius Neural Compute Stick\n      3. Google Coral USB Accelerator\n      4. NVIDIA Jetson Nano\n      5. FPGA + PYNQ\n      6. Arduino\n    2. A Qualitative Comparison of Embedded AI Devices\n    3. Hands-On with the Raspberry Pi\n    4. Speeding Up with the Google Coral USB Accelerator\n    5. Port to NVIDIA Jetson Nano\n    6. Comparing the Performance of Edge Devices\n    7. Case Studies\n      1. JetBot\n      2. Squatting for Metro Tickets\n      3. Cucumber Sorter\n    8. Further Exploration\n    9. Summary\n  17. 16. Simulating a Self-Driving Car Using End-to-End Deep Learning with Keras\n    1. A Brief History of Autonomous Driving\n    2. Deep Learning, Autonomous Driving, and the Data Problem\n    3. The \"Hello, World!\" of Autonomous Driving: Steering Through a Simulated Environment\n      1. Setup and Requirements\n    4. Data Exploration and Preparation\n      1. Identifying the Region of Interest\n      2. Data Augmentation\n      3. Dataset Imbalance and Driving Strategies\n    5. Training Our Autonomous Driving Model\n      1. Drive Data Generator\n      2. Model Definition\n    6. Deploying Our Autonomous Driving Model\n    7. Further Exploration\n      1. Expanding Our Dataset\n      2. Training on Sequential Data\n      3. Reinforcement Learning\n    8. Summary\n  18. 17. Building an Autonomous Car in Under an Hour: Reinforcement Learning with AWS DeepRacer\n    1. A Brief Introduction to Reinforcement Learning\n    2. Why Learn Reinforcement Learning with an Autonomous Car?\n    3. Practical Deep Reinforcement Learning with DeepRacer\n      1. Building Our First Reinforcement Learning\n      2. Step 1: Create Model\n      3. Step 2: Configure Training\n      4. Step 3: Model Training\n      5. Step 4: Evaluating the Performance of the Model\n    4. Reinforcement Learning in Action\n      1. How Does a Reinforcement Learning System Learn?\n      2. Reinforcement Learning Theory\n      3. Reinforcement Learning Algorithm in AWS DeepRacer\n      4. Deep Reinforcement Learning Summary with DeepRacer as an Example\n      5. Step 5: Improving Reinforcement Learning Models\n    5. Racing the AWS DeepRacer Car\n      1. Building the Track\n      2. AWS DeepRacer Single-Turn Track Template\n      3. Running the Model on AWS DeepRacer\n      4. Driving the AWS DeepRacer Vehicle Autonomously\n    6. Further Exploration\n      1. DeepRacer League\n      2. Advanced AWS DeepRacer\n      3. AI Driving Olympics\n      4. DIY Robocars\n      5. Roborace\n    7. Summary\n  19. A. A Crash Course in Convolutional Neural Networks\n    1. Machine Learning\n    2. Perceptron\n    3. Activation Functions\n    4. Neural Networks\n    5. Backpropagation\n    6. Shortcoming of Neural Networks\n    7. Desired Properties of an Image Classifier\n    8. Convolution\n    9. Pooling\n    10. Structure of a CNN\n    11. Further Exploration\n  20. Index\n\n","meta":{"redpajama_set_name":"RedPajamaBook"}}
+{"text":"#\n\n#\n\n**Du m\u00eame auteur**\n\n_Le Hasard et la volont\u00e9,_ roman, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 2012.\n\n_Le temps qui m'est donn\u00e9,_ roman, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 2010.\n\n_Cette ann\u00e9e s'envole ma jeunesse,_ r\u00e9cit, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 2009.\n\n\u2022 **Finaliste Prix du gouverneur g\u00e9n\u00e9ral 2009**\n\n_Ceci est mon corps,_ roman, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 2008.\n\n\u2022 **Finaliste Prix du gouverneur g\u00e9n\u00e9ral 2008**\n\n\u2022 **Mention d'excellence de la Soci\u00e9t\u00e9 des \u00e9crivains francophones d'Am\u00e9rique**\n\n_Quand les pierres se mirent \u00e0 r\u00eaver,_ po\u00e9sie, Le Noro\u00eet, 2007.\n\n_La Fabrication de l'aube,_ r\u00e9cit, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 2007.\n\n\u2022 **Prix des libraires du Qu\u00e9bec 2007**\n\n_Voici nos pas sur la terre,_ po\u00e9sie, Le Noro\u00eet, 2006.\n\n_Le Jour des corneilles,_ roman, Les Allusifs, 2004.\n\n\u2022 **Prix France-Qu\u00e9bec\/Jean Hamelin 2005**\n\n\u2022 **Prix du livre francophone de l'ann\u00e9e 2005, Issy-les-Moulineaux, France**\n\n\u2022 **Finaliste Prix des Cinq continents 2005**\n\n_Turkana Boy,_ roman, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 2004.\n\n_Le Petit Pont de la Louve,_ roman, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 2002.\n\n_Mon p\u00e8re est une chaise,_ roman jeunesse, Qu\u00e9bec Am\u00e9rique, coll. Titan, 2001.\n\n_Les Choses terrestres,_ roman, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 2001.\n\n_Garage Molinari,_ roman, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 1999.\n\n\u2022 **Finaliste Prix France-Qu\u00e9bec**\n\n_Comme enfant je suis cuit,_ roman, Qu\u00e9bec Am\u00e9rique, coll. Litt\u00e9rature d'Am\u00e9rique, 1998.\n\n**Collaborations**\n\n_Comment \u00e9crire un livre,_ nouvelle, dans _Ma librairie ind\u00e9pendante_ (collectif), publi\u00e9 par l'Association des libraires du Qu\u00e9bec \u00e0 l'occasion de son 40e anniversaire, 2010.\n\n_Hum...,_ texte d'humeur, Fides, dans _La Vie est belle !_ (collectif), album de textes et de photographies par Isabelle Cl\u00e9ment, 2008.\n\n_Ici Radio-Canada \u2014 50 ans de t\u00e9l\u00e9vision fran\u00e7aise,_ ouvrage command\u00e9 par la Soci\u00e9t\u00e9 Radio-Canada soulignant le 50e anniversaire de la t\u00e9l\u00e9vision publique canadienne (en collaboration avec Gil Cimon), Les \u00c9ditions de L'Homme, 2002.\n\n_Le Chien qui voulait apprendre le twist et la rumba,_ nouvelle, dans _R\u00e9cits de la f\u00eate_ (collectif), Qu\u00e9bec Am\u00e9rique, 2000.\n`Le Jour  \ndes corneilles`\n\n#\n\n**Catalogage avant publication de Biblioth\u00e8que et Archives nationales du Qu\u00e9bec et Biblioth\u00e8que et Archives Canada**\n\nBeauchemin, Jean-Fran\u00e7ois\n\nLe jour des corneilles\n\n(Collection QA Compact)\n\n\u00c9d. originale : Montr\u00e9al : Les Allusifs, 2004.\n\nISBN 978-2-7644-2252-6 (Version imprim\u00e9e)\n\nISBN 978-2-7644-2537-4 (PDF)\n\nISBN 978-2-7644-2538-1 (ePub)\n\nI. Titre.\n\nPS8553.E171J68 2013 C843'.54 C2013-940523-2\n\nPS9553.E171J68 2013\n\nNous reconnaissons l'aide financi\u00e8re du gouvernement du Canada par l'entremise du Fonds du livre du Canada pour nos activit\u00e9s d'\u00e9dition.\n\nGouvernement du Qu\u00e9bec \u2014 Programme de cr\u00e9dit d'imp\u00f4t pour l'\u00e9dition de livres \u2014 Gestion SODEC.\n\nLes \u00c9ditions Qu\u00e9bec Am\u00e9rique b\u00e9n\u00e9ficient du programme de subvention globale du Conseil des Arts du Canada. Elles tiennent \u00e9galement \u00e0 remercier la SODEC pour son appui financier.\n\nQu\u00e9bec Am\u00e9rique\n\n329, rue de la Commune Ouest, 3e \u00e9tage\n\nMontr\u00e9al (Qu\u00e9bec) Canada H2Y 2E1\n\nT\u00e9l\u00e9phone : 514 499-3000, t\u00e9l\u00e9copieur : 514 499-3010\n\nD\u00e9p\u00f4t l\u00e9gal : 2e trimestre 2013\n\nBiblioth\u00e8que nationale du Qu\u00e9bec\n\nBiblioth\u00e8que nationale du Canada\n\nProjet dirig\u00e9 par \u00c9ric St-Pierre\n\nLecture de s\u00fbret\u00e9 : Philippe Par\u00e9-Moreau\n\nConception graphique : Julie Villemaire\n\nMontage : Karine Raymond\n\nEn couverture : Photomontage r\u00e9alis\u00e9 \u00e0 partir d'\u0153uvres de Eugene Ivanov \/ Shutterstock et de andreiuc88 \/ Shutterstock\n\nConversion au format ePub : Studio C1C4\n\nTous droits de traduction, de reproduction et d'adaptation r\u00e9serv\u00e9s\n\n\u00a9 2013 \u00c9ditions Qu\u00e9bec Am\u00e9rique inc.\n\nwww.quebec-amerique.com\n**`JEAN-FRAN\u00c7OIS BEAUCHEMIN`**\n\n`Le Jour  \ndes corneilles`\n\n#\n\n_L'auteur remercie le Conseil des arts et des lettres du Qu\u00e9bec pour sa contribution financi\u00e8re \u00e0 l'\u00e9criture de ce livre._\n\n#\n\n**NOTE DE L'AUTEUR**\n\n_Le Jour des corneilles_ m\u00e9ritant, il me semble, de survivre \u00e0 la faillite de son premier \u00e9diteur, j'ai cru bon de confier aux \u00c9ditions Qu\u00e9bec Am\u00e9rique la t\u00e2che de prendre ici le relais. Cette \u0153uvre \u00e9crite en quinze jours, qui a passionn\u00e9 un vaste public tant au Qu\u00e9bec qu'en France, et dont on a tir\u00e9 un film ma foi fort \u00e9mouvant, conna\u00eet en effet un succ\u00e8s qui dure, et que je n'avais pas pr\u00e9vu. J'\u00e9tais le dernier \u00e0 me douter qu'une histoire aussi insolite, racont\u00e9e dans un langage non moins \u00e9tonnant, ferait \u00e0 ce point impression dans les esprits. J'avais craint aussi qu'on ne retienne d'elle que sa part de violence, de col\u00e8re et de chagrin. Je me r\u00e9jouis qu'on y ait surtout aper\u00e7u le t\u00e9moignage d'un homme qui, d\u00e8s lors qu'il apprend \u00e0 lire et \u00e0 \u00e9crire, \u00e0 nommer avec justesse les choses, se lib\u00e8re de son animalit\u00e9. Ce que le personnage central de cette br\u00e8ve histoire exp\u00e9rimente, c'est qu'il n'y a pas d'humanit\u00e9 sans la parole. C'est, \u00e0 mon avis, l'aspect le plus important du r\u00e9cit que fait le fils Courge devant ses juges. Aussi, chaque mot compte dans ces pages qu'on s'appr\u00eate \u00e0 lire, et tout ce que j'y ai mis est vrai, dans la mesure \u00e9videmment o\u00f9 l'on croit \u00e0 la v\u00e9rit\u00e9 de la po\u00e9sie.\n\n#\n\n_N_ _ous logions, p\u00e8re et moi, au plus \u00e9pais de la for\u00eat, dans une cabane de billes \u00e9rig\u00e9e ci-devant le grand h\u00eatre. P\u00e8re avait form\u00e9 de ses mains cette r\u00e9sidence rustique et tous ses accompagnements. Rien n'y manquait : depuis l'eau de pluie amass\u00e9e dans la barrique pour nos bouillades et mes plongements, jusqu'\u00e0 l'\u00e2tre pour la rissole du cuissot et l'\u00e9chauffage de nos membres aux rudes temps des frimasseries. Il y avait aussi nos paillasses, la table, une paire de taboureaux, et puis encore l'alambic de l'officine, o\u00f9 p\u00e8re s'affairait \u00e0 extraire, des branchottes et fruits du geni\u00e8vre avoisinant, une eau-de-vie costaude et grandement combustible._\n\n_Pour nous repa\u00eetre, nous prenions le poisson de l'\u00e9tang ou boutions hors tani\u00e8res et abris toutes b\u00eates_ _nourrici\u00e8res : garennes, g\u00e9linots, chipmonques, casteurs, putois, ratons et chevrillards. Le reste de notre p\u00e2ture se composait surtout de th\u00e9 de dalibarde, d'\u0153ufs de merles et de sarcelles, de marasmes, de racines et de baies, de souricelles assomm\u00e9es par nos soins et de rapaces doctement bombard\u00e9s de pierrettes, ou perc\u00e9s de nos fl\u00e8ches._\n\n_P\u00e8re poss\u00e9dait toutes sciences. Notions et lumi\u00e8res si\u00e9geaient sous son casque. Il concevait que Terre est plate, qu'elle stationne au milieu des cieux et que les astres tournoient \u00e0 l'entour tel le chien ancr\u00e9 au pieu. Que la d\u00e9esse Lune assure le salut de toutes choses vives : bestieuses, v\u00e9g\u00e9teuses et humaines. Que maux de corps se soignent par saign\u00e9es et autres secours modernes. Que le cauchemar engouffre la cervelle par les esgourdes. P\u00e8re traduisait aussi les all\u00e9es et venues de l'air : par simple grimpement aux arbres il \u00e9tudiait au loin le progr\u00e8s de la bourrasque ou du cyclone cheminant vers nous, et augurait ainsi de notre p\u00e9ril ou de notre qui\u00e9tude. Boussole et instruments paraissaient tenir en son pied, aussi savait-il circuler sous arbres et sur sentes sans entraves ni d\u00e9routements. Il p\u00e9n\u00e9trait le sens des astres et des \u00e9toiles, et d\u00e9tenait le don de leur lecture. Aussi, par soirs, il m'arrivait, quand il lorgnait la vo\u00fbte, de le questionner sur ma destin\u00e9e. Telle \u00e9tait ma voix : \u00ab P\u00e8re, que distingues-tu cette nuit de ce qu'il en sera de moi ? \u00bb Mais p\u00e8re n'\u00e9tait pas parleur._\n\n_D\u00e8s mon \u00e2ge le plus vert, il m'avait instruit de tout : comment prendre le poisson, d\u00e9m\u00ealer la voix de la b\u00eate, talonner le gibier, d\u00e9couper le bif, rissoler le cuissot, tailler en billettes l'arbre abattu, appr\u00eater le crevard de mouffeton, sauter la russule et autres champignes, recouvrer levant et ponant, circuler noctantement, coudre l'accoutre, \u00e9triper le chevrillard et m\u00eame juguler la vip\u00e8re qui se faufilait dans nos godillots laiss\u00e9s le soir sur le seuil._\n\n_Malgr\u00e9 qu'il f\u00fbt gorg\u00e9 d'entendement et qu'il e\u00fbt pu ais\u00e9ment susciter amples \u00e9gards, p\u00e8re go\u00fbtait une existence coite et quasiment solitaire. J'\u00e9tais, en fait, la seule humanit\u00e9 autoris\u00e9e d'avoisinance en ses parages. Ainsi coulaient ses jours, distants de tout commerce avec les gens, bourgeois ou cr\u00e9atures, qu'il qualifiait souventes fois de \u00ab racaille \u00bb, de \u00ab marauds \u00bb, de \u00ab pendards \u00bb, de \u00ab faquins \u00bb et de \u00ab gueux \u00bb. D\u00e9tournant volontiers sa face de la foule, il rebroussait toujours \u00e0 la for\u00eat, qui lui fournissait bien suffisamment tous asiles, p\u00e2tures et combustibles n\u00e9cessaires. Pr\u00e9f\u00e9rablement au discours, il \u00e9lisait les criailleries des b\u00eates, les bruissements de la bise dans les branchottes, les craquements des arbres pourris ou tordus, et m\u00eame le tonnement terrible du grain quand il cr\u00e8ve._\n\n_Non, p\u00e8re n'\u00e9tait pas parleur. Sauf quand il palabrait avec ses gens, ainsi que je le narrerai \u00e0 pr\u00e9sent, Monsieur le juge. Car s'il me faut aujourd'hui tourner pour vous les pages de mon existence, il me faudra aussi, par m\u00eame occasion et pour mieux traduire mon r\u00e9cit, ouvrir le livre de la vie de p\u00e8re, si \u00e9troitement emmaillot\u00e9e \u00e0 la mienne. Cela afin de vous instruire meilleurement des circonstances o\u00f9 je fus conduit \u00e0 achever mon prochain, puis enseign\u00e9 de vocabulaire et, enfin, men\u00e9 ci-devant vous et les membres de ce tribuneau pour trancher mon cas._\n\n#\n\nP\u00e8re \u00e9tait fort charnu. Par tous horizons, on n'avait jamais vu bourgeois aussi muscleux. Mais ce qui me laissait le plus \u00e9tonn\u00e9 \u00e9tait surtout la puissance et le nerf s\u00e9journant en ses chairs. Pour exemple, je d\u00e9peindrai premi\u00e8rement un ouvrage des plus curieux que p\u00e8re accomplit une fois. Par jour de grandes gelures, je le vis se fabriquer mitaines de cette mani\u00e8re : fourrant le bras en une tani\u00e8re, il grippa coup sur coup une paire de marmottes ventrues et enroupill\u00e9es. Les assommant par suite du marteau de son poing, il entreprit bient\u00f4t de les fendre, puis de les \u00e9vider. Une fois ce videment accompli \u00e0 l'aide de ses seuls doigts, p\u00e8re se para les mains des d\u00e9pouilles, et poursuivit son cours, les paumes bien au chaud maintenant.\n\nQuant aux jambes de p\u00e8re, c'\u00e9tait \u00e9quivalence de cuissots de rossinant par musclure aussi bien que par endurance \u00e0 la course. Aussi, nul bourgeois ne pourchassait la b\u00eate mieux que lui, ni ne s'esquivait avec plus d'allure lorsque le cyclone mena\u00e7ait. Son pied aussi impressionnait par sa surdimension. Quand p\u00e8re avan\u00e7ait sur sente de son marcher appesanti, la fourmi tressautait, le chipmonque chutait de sa branchotte, la chenillette l\u00e0-haut aussi se d\u00e9crochait de son feuillage et, en leur trou, garennes, marmottes, ratons et belets recevaient plafond sur le casque. Bref, en toutes portions de sa personne, p\u00e8re \u00e9tait important.\n\nMais ce corps, quoique baraqu\u00e9, souffrait en sa partie la plus \u00e9lev\u00e9e et souventes fois la plus utile, le casque, d'un trouble \u00e9trange : lorsqu'il \u00e9tait enti\u00e8rement \u00e9veill\u00e9 et m\u00eame affair\u00e9 \u00e0 besognes, p\u00e8re recevait parfois en r\u00eavement la visite de gens qui lui faisaient la conversation, \u00e0 laquelle il r\u00e9torquait avec des mots que je ne lui connaissais gu\u00e8re coutumi\u00e8rement. Plus alarmants \u00e9taient les grognements, gesticulades et agitations de d\u00e9moniaque accompagnant alors son parler. Mais le pire r\u00e9sidait ailleurs. En effet, les gens de p\u00e8re, quand ils s'emparaient de lui, le for\u00e7aient aux actes et missions les plus insens\u00e9s. P\u00e8re, comme sous l'empire de quelque magie d\u00e9sastreuse, formait d\u00e8s lors l'ambition d'exaucer ses gens, ce qui le menait, Monsieur le juge, au-del\u00e0 des limites raisonnables de l'agissement humain. Forc\u00e9 \u00e0 mon tour par p\u00e8re d'agir \u00e0 ses c\u00f4t\u00e9s comme second, j'ai plus d'une fois risqu\u00e9 ma vie en ces \u00e9quip\u00e9es, comme vous le concevrez bient\u00f4t par mon histoire.\n\n\u00c7a lui \u00e9tait venu, la premi\u00e8re fois, dans les heures o\u00f9 m\u00e8re nous quitta. Ma naissance termin\u00e9e, m\u00e8re commen\u00e7a \u00e0 mourir sur la paillasse, car je lui avais donn\u00e9 ample fil \u00e0 retordre avant que d'aboutir ici-bas. P\u00e8re, cependant, avait attendu \u00e0 l'ext\u00e9rieur de la cabane que m\u00e8re mette bas, profitant des bonnes heures du jour pour \u00e9visc\u00e9rer un chevrillard achev\u00e9 par haut matin. Tandis que, n\u00e9, je hurlai, p\u00e8re entra, me saisit entre ses bras muscleux et me mena bien vite devant l'\u00e2tre cr\u00e9pitant. M\u00e8re, de son c\u00f4t\u00e9, nous quittait si silencieusement que p\u00e8re ne s'avisa de rien. Ce n'est que lorsqu'il me ramena sur paillasse enaccoutr\u00e9 de ma d\u00e9froque nouvelle et qu'il se tourna finalement vers sa compagne qu'il nota : m\u00e8re, qu'il adorait telle une pierrette rarissime, avait rendu l'\u00e2me.\n\nCe fut un moment terrible. Chaque b\u00eate ou insecte terr\u00e9 en la for\u00eat e\u00fbt assur\u00e9ment le c\u0153ur cass\u00e9 en escoutant le pleur d\u00e9vers\u00e9 par p\u00e8re devant la cabane, sans compter ses plaintes et hurlades, r\u00e9pandues bien au-del\u00e0 du grand h\u00eatre. \u00ab Pourquoi ? Pourquoi ? \u00bb criaillait-il, pleuroyant, son gros poing rossant une bille, son pied infligeant savates et coups divers aux arbres \u00e0 l'entour. Mais ni b\u00eate, ni v\u00e9g\u00e9tation, ni pierraille, ni la d\u00e9esse Lune elle-m\u00eame ne pouvaient trouver r\u00e9pons \u00e0 cette question.\n\nCe m\u00eame jour, tandis que le soleil finissait son d\u00e9roulement en cieux, p\u00e8re, vaincu, vidang\u00e9 de ses larmes, empoigna son herminette et commen\u00e7a d'usiner quelques planches. Il en forma une bi\u00e8re \u00e0 dimension de m\u00e8re puis, saisissant celle-ci une finale fois en ses bras, la d\u00e9posa doucettement en cercueil. Enfin, portant sur \u00e9chine m\u00e8re embo\u00eet\u00e9e dans ce couche-mort, il chemina en for\u00eat jusqu'au pied de la grande pruche, ainsi que j'en fus instruit bien apr\u00e8s. Juste ci-dessous, de ses grosses mains, p\u00e8re approfondit un trou, qui accueillit bient\u00f4t m\u00e8re en son repos durable. Par suite il rebroussa \u00e0 la cabane, me trouvant bien \u00e9tabli sur paillasse et attendant sagement de prendre repas.\n\nL'\u0153il encore rougi, p\u00e8re se mit \u00e0 la t\u00e2che de me repa\u00eetre. Je le vis sortir un moment, puis rebrousser bient\u00f4t avec le cadavre d'un h\u00e9risson femelle, dont il tira un peu de lait. Ce fut ma premi\u00e8re pitance sur le domaine de la Terre : le lait d'une b\u00eate morte, achev\u00e9e par p\u00e8re. Ce fut par m\u00eame occasion ma premi\u00e8re encontre v\u00e9ritable avec la mort, v\u00e9ritable en ce que j'en fus p\u00e9n\u00e9tr\u00e9, puis nourri. Toute ma vie, cela devait me rester inscrit au ventre : par l\u00e0 le tr\u00e9pas avait trac\u00e9 sa sente en ma personne, comme mots se formant et s'alignant sur la page. Cependant j'avalai cette p\u00e2ture avec enthousiasme, ne soup\u00e7onnant pas de suite ce qui me guettait ici-bas, tout ce dont je p\u00e2tirais, avant peu, aupr\u00e8s de p\u00e8re.\n\nCela n'allait plus tra\u00eener.\n\n#\n\nApr\u00e8s l'enfouir de m\u00e8re et mon breuvement de lait, p\u00e8re, moulu par le chagrin, s'allongea pour la nuit, non sans avoir bien refait le capiton de ma propre paillasse et \u00e9tabli ci-dessus. C'est \u00e0 l'aube suivante que ses gens parurent en son casque pour la premi\u00e8re fois. Apr\u00e8s d\u00e9jeuner, \u00e0 peine avions-nous aval\u00e9 le gruau de joubarde que voil\u00e0 p\u00e8re qui gesticule et commence de se d\u00e9battre avec ses visiteurs cependant aussi invisibles que pet de mosquite. \u00c7a dure, \u00e7a dure, la sueur ruisselle sous la liquette de p\u00e8re, car il arpente la cabane, et s'agite, et grogne, et semonce, et rousp\u00e8te, et menace ses gens. Puis vient un moment d'acalmette, et p\u00e8re s'\u00e9tablit sur le taboureau. Sa conversation, toutefois, pers\u00e9v\u00e8re. Quoique fort vert, j'avais d\u00e9j\u00e0 l'\u0153il ouvert et l'aptitude agile. Aussi traduisis-je vitement le sens de cette \u00e9meute : quelque part sur le chemin s\u00e9parant la tombe de m\u00e8re et le seuil de la cabane, p\u00e8re avait \u00e9gar\u00e9 l'entendement. On mesurera mieux ceci quand j'aurai d\u00e9peint la mission que ses gens lui impos\u00e8rent alors, et dont je fus l'ingr\u00e9dient dominant.\n\nS'emparant de ma personne, p\u00e8re me mena par-del\u00e0 la for\u00eat jusqu'au champ de monsieur Ronce, o\u00f9 se trouvait un trou de marmotte peu aprofond. J'y fus enfourn\u00e9 puis laiss\u00e9 \u00e0 moisir, s\u00e9journant l\u00e0 pour l'\u00e9quivalence d'une course de soleil, pleuroyant extr\u00eamement de d\u00e9sarroi, de soifs et d'app\u00e9tits. Caillasses me griffaient l'\u00e9chine. Poussiers et sables m'emplissaient esgourde, \u0153il et bouche. Paille se tissait \u00e0 ma maigriotte chevelure. Larves de hannetons chutaient du plafond et m'atterrissaient sur cuissots, sur estomac et sur face. Et fourmis, et lombrics, et saute-climacie grouillaient et sautillaient formidablement \u00e0 l'entour et sous ma d\u00e9froque. Nul bourgeois, nulle cr\u00e9ature ne vint cependant \u00e0 mon aide, mon g\u00e9missement de chose faible n'atteignant sans doute m\u00eame pas leurs esgourdes. Une marmotte parut toutefois, d'abord fort intrigu\u00e9e de ma pr\u00e9sence en son logement, mais se d\u00e9sint\u00e9ressant bient\u00f4t de la question et s'\u00e9tablissant m\u00eame sur un bout de mon accoutre pour son roupil. Nous rest\u00e2mes ainsi long de temps, tapis et r\u00e9ciproquement \u00e9chauff\u00e9s. R\u00e9fugi\u00e9 de la sorte, je sentis le calme rebrousser un brin en mes chairs. Je stationnai mon blair contre celui, doux et chaud, de l'animal.\n\nLe balancement d\u00e9lass\u00e9 de son souffle me ramenait le ressouvenir du ventre de m\u00e8re, et je commen\u00e7ai \u00e0 songer \u00e0 elle, allong\u00e9e durablement en sa bi\u00e8re. Malgr\u00e9 qu'elle f\u00fbt morte et moi vif, je m'avisai que nous \u00e9tions chacun en semblables situations, reposant sous le monde et en quelque mani\u00e8re retranch\u00e9s de lui, chacun de son c\u00f4t\u00e9 aux confins de la vie. Peut-\u00eatre, si nous avions \u00e9t\u00e9 instruits de la fa\u00e7on d'accomplir une telle chose, aurions-nous pu \u00e0 ce moment nous rejoindre et nous sourire aimablement. Ce penser m'en amenait un autre, que je rem\u00e2chais ainsi : qui sait si les macchab\u00e9es ne rebroussent pas \u00e0 leur fa\u00e7on en pays de commencements, tels les petits d'humains en \u00e9tat de premi\u00e8re verdeur ? Et qui sait si les nouvellement-n\u00e9s n'ont pas, amass\u00e9e en leur besace int\u00e9rieure, toute une vie par-devers eux ? Car telle est l'outre-vie : emplie de devinettes et d'inexplicablet\u00e9s.\n\nParfois, le roupil de la marmotte paraissait troubl\u00e9 de r\u00eavements effrayants, et l'animal alors se blottissait davantage contre ma personne, comme pour y trouver renforts. Il arrivait aussi que, mon h\u00f4te bronchant en son roupillement, l'une de ses pattes se pose doucettement sur ma face, ou sur mon bras, semblablement \u00e0 la main d'un compagnon bienveillant. Ainsi avons-nous, au plus \u00e9pais de ce trou, fait \u00e9change de secours. Et tandis que les dur\u00e9es coulaient, je songeais que ma destin\u00e9e \u00e9tait peut-\u00eatre davantage de vivre comme les b\u00eates, \u00e0 tout le moins parmi elles, plut\u00f4t qu'en humanit\u00e9 affair\u00e9e, et tracass\u00e9e, et sourde \u00e0 mon g\u00e9missement souterrain. Rappelant en mon casque ces \u00e9v\u00e9nements, je m'instruis aujourd'hui seulement que ce fut l\u00e0, dans ce trou, que je re\u00e7us les t\u00e9moignages de ch\u00e9rissement les plus prolong\u00e9s. J'observe aussi que cette marmotte-l\u00e0 me prodigua davantage de chaleur et de rescousse que p\u00e8re ne m'en offrit de toute sa vie.\n\nCar p\u00e8re revint enfin me d\u00e9senfourner de l\u00e0.\n\nJuste avant, le soleil concluait sa course, et je commen\u00e7ais \u00e0 concevoir ma mort prochement. \u00ab J'aurai donc r\u00e9sid\u00e9 ici-bas le temps d'un jour \u00bb me disais-je, toisant par l'entr\u00e9e du trou un p\u00e2le agroupement d'\u00e9toiles. Je chassai une fourmi de ma face, pris en ma main une patte de la marmotte, baissai paupi\u00e8res et me pr\u00e9parai \u00e0 mourir. Il me semblait \u00eatre d\u00e9j\u00e0 \u00e0 demi en l'outre-monde lorsque je per\u00e7us, au-dehors, le marcher fruste de p\u00e8re. Levant l'\u0153il, je vis son godillot para\u00eetre, puis bient\u00f4t la face enti\u00e8re de p\u00e8re emplir l'ouverture. Je fus \u00e0 la fin extrait sans m\u00e9nagement de mon g\u00eete et emport\u00e9 sous les astres.\n\nLe sort l'avait voulu ainsi : j'\u00e9tais rebrouss\u00e9 parmi les vifs. \u00c9tabli sur l'\u00e9chine de p\u00e8re ainsi qu'une hotte emplie de tub\u00e9reuses, ma face tourn\u00e9e vers le champ de monsieur Ronce, je vis s'\u00e9vanouir dans la nuit la premi\u00e8re r\u00e9sidence digne de ce nom que j'eus en l'ici-bas et, surtout, le seul ami v\u00e9ritable que je coudoyai jamais.\n\n#\n\nDes \u00e9poques coul\u00e8rent. Les gens de p\u00e8re allaient et venaient, emplissant son casque parfois pendant de longues dur\u00e9es, s'enfuyant ensuite comme filous, puis rebroussant encore et le poss\u00e9dant tel un encha\u00een\u00e9. Mon premier \u00e2ge fut ainsi marqu\u00e9 par les d\u00e9mences r\u00e9p\u00e9t\u00e9es de p\u00e8re, auxquelles j'\u00e9tais toujours f\u00e2cheusement associ\u00e9. Encore trop frais d'\u00e2ge pour le questionner sur le sens de ses actions, je le suivais sans rousp\u00e8te dans les t\u00e2ches prescrites par ses gens. Combien de fois fus-je houspill\u00e9, affam\u00e9, appendu, enseveli, livr\u00e9 \u00e0 termiti\u00e8res ou \u00e9tabli sur gu\u00eapi\u00e8re, enduit de miell\u00e9e puis offert \u00e0 fourmis, ficel\u00e9 \u00e0 branchotte puis donn\u00e9 pour p\u00e2ture \u00e0 chenillette et quasiment noy\u00e9 sous l'\u00e9tang ? Combien de fois ? C'est l\u00e0 calculement trop extravagant pour mon casque si peu aguerri aux nombres. Mais je conserve de tout cela le sentiment d'un attentat ineffa\u00e7able.\n\nN\u00e9anmoins, malgr\u00e9 la fragilit\u00e9 de mon extr\u00eame verdeur, je surdurai \u00e0 cette p\u00e9riode. J'abordais \u00e0 pr\u00e9sent l'\u00e2ge tendre des jouvenceaux, les joues piqu\u00e9es de roussi et la t\u00eate encore boursoufl\u00e9e de canailleries. Mon r\u00f4le dans les missions d\u00e9mentielles de p\u00e8re n'en fut pas moins notable. Sans doute m\u00eame y fus-je encore davantage l'objet de rudesses.\n\nJ'ai pour exemple ressouvenir d'une folie de p\u00e8re qui me laissa fort secou\u00e9. Je terminais ce jour-l\u00e0, ci-dessous le grand h\u00eatre, le grillement d'une brochette de souricelles. P\u00e8re, qui jusqu'alors s'affairait \u00e0 fendre billettes, se montra bient\u00f4t \u00e0 mon abord avec l'\u0153il bileux, le nerf tendu et le cheveu hargneux du poss\u00e9d\u00e9. Nul doute : ses gens lui revenaient. Sa voix fut rude et commandeuse : \u00ab Parnoir ! Fils ! J'ai faim ! Sers-moi donc sur l'heure de cette viande-l\u00e0 ! \u00bb Mon r\u00e9torque fut \u00e0 peu pr\u00e8s ainsi : \u00ab Mais, P\u00e8re, c'est que je mijotai cette pitance pour mon usage personnel, sans songer que tu y poserais la lippe. Aussi y mis-je quantit\u00e9 d'assaison d'herbe-aux-rats, que je go\u00fbte fort, mais qui d'ordinaire te fait venir, \u00e0 toi, pustules et boutons vari\u00e9s ! Mais si tel est ton souhait d'avaler un peu de chair, laisse-moi assommer de mon godillot quelques-unes des souricelles qui circulent en notre cabane. Je te les embrocherai vitement et te les grillerai \u00e0 ta satisfaction. \u00bb\n\nCe r\u00e9pons ne trouva pas chez lui bon escouteur. P\u00e8re me saisit au col et entreprit de me juguler, mena\u00e7ant de me faire \u00e9garer le souffle, l'aplomb et quasiment la vie. Mon esgourde demeurant toutefois assez vive, j'entendis ces mots prononc\u00e9s pour moi par p\u00e8re, comme pierrailles fracass\u00e9es sous outil : \u00ab J'ai dit : j'ai faim ! Donne-moi de ce repas ! \u00bb Puis il d\u00e9fit lentement l'\u00e9tau de ses doigts et me rendit \u00e0 la vie. Trembleur, je fis ce qu'il attendait. P\u00e8re empoigna la brochette, s'\u00e9tablit sur le sol, puis se mit la premi\u00e8re souricelle sous la dent. \u00c7a ne lui \u00e9tait pas encore pass\u00e9 dans le gosier que boutons commenc\u00e8rent \u00e0 lui tapisser la face et les mains.\n\n\u00ab Diablerie et grain d'orage ! \u00bb fit-il aussit\u00f4t, jetant prestement au feu le reste de sa pitance. Piqu\u00e9, saisi par extr\u00eames f\u00e2cheries, il se dresse alors, vient me gripper par l'\u00e9paule et m'entra\u00eene fort mauvaisement vers la barrique. Nous y sommes. \u00ab Vide cette tonne ! \u00bb criaille-t-il. Et me voil\u00e0 renversant la barrique et d\u00e9versant sur la terre toutes les pluies amass\u00e9es ci-dedans depuis lunes et soleils. Quand c'est accompli, p\u00e8re me commande de le talonner sur la sente et de transporter par-devers moi la barrique.\n\nNous circulons en for\u00eat long de temps, silencieux de voix, mais bruiteurs de pas. La plainte de la barrique aussi, roul\u00e9e par mes soins, r\u00e9sonne sourdement sur la terre et trouble le roupil des b\u00eates. Les moins trembleuses d'entre elles sortent de leur trou et viennent nous regarder passer. Muffles et moustaches s'agitent derri\u00e8re les foug\u00e8res. Des oiseaux aussi passent, tournoient, baissent l'\u0153il un instant sur nous, puis poursuivent leur chemin.\n\nNous touchons enfin le pied du mont Tondu. C'est une colline s'\u00e9levant \u00e0 moyenne distance de la cabane, et dont le flanc et le sommet caillouteux ne laissent cro\u00eetre que pieds d'aub\u00e9pines, pr\u00eales, mousses et lycopodes, d'o\u00f9 son nom. Rompant son silence, p\u00e8re dit avec grondements dans les mots : \u00ab Maintenant, hisse la tonne tout l\u00e0-haut ! \u00bb Et me voil\u00e0 grimpant le mont Tondu, poussant puis tirant la barrique, peinant, bandant mes vertes musclures, poussant exclamations multiples, suant sels, et eaux, et poivres de corps, manquant mourir de p\u00e9nuries \u00e0 chaque pas. Maintes fois, arriv\u00e9 au bout de mon nerf, dus-je l\u00e2cher mon bagage, le laisser d\u00e9gringoler et m'en aller le reprendre au pied de la colline puis recommencer, d\u00e9bravour\u00e9, mon grimpement. P\u00e8re, cependant, stationnait non loin et posait le regard sur tout cela, sans broncher.\n\nJe parvins \u00e0 la fin \u00e0 m'\u00e9tablir sur le sommet du mont Tondu. P\u00e8re vint m'y trouver. Malgr\u00e9 que le jour f\u00fbt d'abondance chauff\u00e9 par le soleil, je fr\u00e9missais de tous mes membres. Qu'allait donc ordonner p\u00e8re \u00e0 pr\u00e9sent ? \u00c0 quelle terriblerie \u00e9tais-je promis ? Je le sus bient\u00f4t.\n\n\u00ab Faufile ta personne en barrique ! \u00bb fit-il. Je pressentis ma fin venir. Esp\u00e9rant survenue de miracle, je me glissai soumisement dans la tonne. Une fois emm\u00e9nag\u00e9 ci-dedans, je fermai l'\u0153il, me recroquevillai tel un pr\u00e9-jouvenceau en ventre de m\u00e8re, et patientai. Mon moisissement fut courtaud. Car je sentis presque aussit\u00f4t la barrique se coucher, puis commencer \u00e0 rouler, puis d\u00e9gringoler franchement le flanc du mont Tondu. P\u00e8re venait de m'y l\u00e2cher.\nJe culbutai, Monsieur le juge, comme culbute la pierraille s'\u00e9boulant aux reins des montagnes : ma chute fut vive, compliqu\u00e9e et sans m\u00e9nagement. J'allai, \u00e0 sa fin, emboutir une \u00e9pinette robuste, et m\u00eame in\u00e9branlable, et fort mal situ\u00e9e sur mon trajet. Quoique sonn\u00e9, je sortis de cette exp\u00e9rience indemne, et la barrique aussi. P\u00e8re vint \u00e0 moi, l'air toujours aussi cassant, et me fit ce discours : \u00ab \u00c0 pr\u00e9sent, rebroussons \u00e0 la cabane ! Et rapporte avec toi cette tonne ! Elle est de grand usage par temps d'averses ! \u00bb\n\nLe soir avan\u00e7ait quand nous foul\u00e2mes enfin notre seuil. Je m'\u00e9croulai sur ma paillasse sans m\u00eame avaler pitance, et commen\u00e7ai \u00e0 roupiller sur l'instant, tant \u00e9tait grande ma d\u00e9bilit\u00e9. Je n'eus point de r\u00eavements cette nuit-l\u00e0, mon casque s'affairant sans doute \u00e0 besognes plus urgentes : repl\u00e2trage d'entendement, ressemelage de notions, raccommodage de lumi\u00e8res et remises en \u00e9tat semblables.\n\nL'aube se forma. Je fus extrait de roupil par le bruit de l'herminette fendant le bois. Je m'en fus trouver p\u00e8re. J'eus vite fait de traduire, \u00e0 son \u0153il redevenu coi, que ses gens l'avaient \u00e0 pr\u00e9sent d\u00e9sert\u00e9. J'observai aussi que boutons ne s\u00e9journaient plus en sa face. P\u00e8re, cependant, ne faisait nul cas de ma pr\u00e9sence, et continuait son ouvrage. Je le questionnai enfin : \u00ab Mais, p\u00e8re, me diras-tu : pourquoi ce supplice que tu m'infligeas au mont Tondu ? \u00bb Des billettes furent encore ouvertes. Un oiseau trilla. Une vip\u00e8re glissa entre les lycopodes. Puis les cieux quitt\u00e8rent pour de bon leur accoutre de nuit, et le soleil coupa, du tranchant de ses rayons obliques, la for\u00eat de l'\u00e0 l'entour.\n\n\u00ab Ta m\u00e8re avait grand talent de cuisineuse. Jamais n'aurait-elle mis en notre pitance assaisons inappropri\u00e9s. Puisse ton \u00e9tourderie, punie par celle que tu v\u00e9cus en \u00e9boulis de barrique, te rappeler cela. \u00bb\n\nAinsi parla-t-il.\n\n#\n\nPar soirs d'\u00e9t\u00e9s, apr\u00e8s avoir grill\u00e9 puis bien aval\u00e9 le bif de putois ou le g\u00e2teau de rognons, p\u00e8re s'\u00e9tablissait ci-dessous le grand h\u00eatre aupr\u00e8s du feu et commen\u00e7ait \u00e0 tisonner pensivement. Je l'y accompagnais. Levant parfois le regard, nous toisions le sommet des arbres se fondre peu \u00e0 peu \u00e0 la nuit. Puis les astres venaient, \u00e9clairant de leur m\u00e8che cieux et \u00e9tendues, instruisant de leur boussole, peut-\u00eatre, quelque marcheur \u00e9gar\u00e9.\n\nNous restions le plus souvent, alors, la lippe close. Pour moi, ce n'\u00e9tait pas que je fusse d\u00e9pourvu de choses sur lesquelles discourir. Simplement, lorsqu'il m\u00fbrissait en ma glotte, mon discours stoppait sa course et refusait d'aller plus avant, puis de fleurir sur le terrain de ma langue. Car j'\u00e9tais pauvre de vocabulaire, aussi pauvre que le foin aux heures enfuies de l'\u00e9t\u00e9 : sec et vid\u00e9 de sa c\u00e9r\u00e9ale. Aussi fus-je ressemblant, ces soirs-l\u00e0, au hibou en sa nuit noire, pr\u00e9f\u00e9rant le silence des ombres au bruit \u00e2pre des maigres paroles. On e\u00fbt dit que j'attendais, que j'attendais d'\u00eatre instruit de vocabulaire, comme si je savais d\u00e9j\u00e0 que le jour viendrait o\u00f9 les choses et le monde trouveraient en ma bouche plus amples traductions.\n\nQui sait ce que p\u00e8re, lui, devant notre feu, m\u00e9ditait et se retenait de dire ?\n\n#\n\nEt ainsi de la course de notre existence \u00e0 la cabane. Chaque jour nouveau nous voyait assommer b\u00eates comestibles, prendre poisson, qu\u00e9rir dans les nids \u0153ufs d'oiseaux, tisser fibres pour notre v\u00eatir, coudre godillots et pelisses, et accomplir toutes t\u00e2ches ressemblantes. Lunes et soleils se succ\u00e9daient, arri\u00e8re-saisons venaient et repartaient, \u00e9poques et \u00e2ges m\u00eamement, et des laps, et des dur\u00e9es, et du temps coulaient encore.\n\nJ'arrivais \u00e0 l'\u00e8re du mi-temps de la vie. Et p\u00e8re aussi avan\u00e7ait en \u00e2ges. \u00ab Comme le temps coule ! m\u00e2chais-je parfois. Voil\u00e0 p\u00e8re m\u00fbrissant formidablement, blanc de cheveux, un brin courb\u00e9 d'\u00e9chine et fl\u00e9tri de peau, quoique encore gaillard et sans cesse visit\u00e9 par ses gens, et toujours \u00e0 palabrer avec eux. \u00bb J'avais moi aussi fort profit\u00e9, quitt\u00e9 depuis jolie lurette les premi\u00e8res verdeurs et atteint une gaillarderie extr\u00eame. Je pouvais ais\u00e9ment, \u00e0 pr\u00e9sent, enlacer le grand h\u00eatre. Quoique moins d\u00e9gourdi et moins empli de science que p\u00e8re, j'\u00e9tais adroit \u00e0 toutes besognes. P\u00e8re me confiait ainsi souventes fois mission d'aller chasser de mes fl\u00e8ches toute b\u00eate \u00e0 la chair utile \u00e0 notre pitance du soir. J'allais pareillement grimper aux branches pour y chouraver les \u0153ufs et, \u00e0 l'\u00e9tang, j'enfon\u00e7ais ma pique dans le poisson-chat. J'usinais toutes choses de mes mains : outils, planches, mobilier, d\u00e9froques, machins. Bref, c'en \u00e9tait fait de mes \u00e9poques d'enfance et de leur insuffisance. J'\u00e9tais entr\u00e9, d\u00e9sormais, en carri\u00e8re humaine.\n\nSous belle saison, quand le soir d\u00e9tr\u00f4nait le soleil, les mouches-de-feu s'emparaient de la for\u00eat. Le chant des crapauds parvenait \u00e0 notre seuil. Venue du pays, on entendait dans les lointains la bise s'avancer, jusqu'\u00e0 faire broncher le sommet des grands arbres. Par nuit aprofonde, la d\u00e9esse Lune montait \u00e0 l'azur. Parfois, aussi, la bise singeait une voix et parvenait \u00e0 mon esgourde, serinant ce discours : \u00ab Qui es-tu ? Qui es-tu ? \u00bb Puis on d\u00e9m\u00ealait le heurt des ailes du hibou qui d\u00e9campait, une souricelle dans le bec.\n\n\u00ab Qui es-tu ? Qui es-tu ? \u00bb bruissait la nuit.\n\nEt qui \u00e9tais-je, en effet ? Moi, qui roupillais ais\u00e9ment sous \u00e9toiles, qui ne se lassais gu\u00e8re du fr\u00e9missement des choses. Et je songeais encore : \u00ab Quel est ce bourgeois g\u00eetant au creux de mes chairs et se sentant chez lui sous les arbres, sous l'azur, parmi les b\u00eates ? \u00bb Par nuits, quand nous restions \u00e0 \u00e9chauffer nos membres aupr\u00e8s du feu, il m'arrivait de m'ouvrir \u00e0 p\u00e8re de ces pensers, et d'autres encore. Tel \u00e9tait mon dire : \u00ab P\u00e8re, toi qui v\u00e9cus, dis-moi : sous cieux et sur Terre, qui sommes-nous v\u00e9ritablement ? Oui, quelle sorte de b\u00eate est donc l'humanit\u00e9 ? Et d'o\u00f9 vient que, lorsque les soirs paraissent, notre casque s'embue de la songerie de ces choses-l\u00e0, ci-devant l'immensit\u00e9 nombreuse des astres en cieux ? \u00bb\n\nMais c'\u00e9tait langage ignor\u00e9 de p\u00e8re. Aussit\u00f4t que je terminais mon discours, il se d\u00e9sint\u00e9ressait des flammes, allongeait sa personne sur la foug\u00e8re et baissait paupi\u00e8res. Son r\u00e9torque \u00e9tait ainsi : \u00ab Demain, il nous faudra chasser puis \u00e9triper chevrillards. Nous aurons ample besoin de forces. Aussi, voici venue l'heure du roupil. Adieu donc, Fils ! \u00bb\n\nCar p\u00e8re n'\u00e9tait pas songeur.\n\n#\n\nIl \u00e9tait terrifi\u00e9 par le tr\u00e9pas. Le seul penser qu'il allait un jour rendre l'\u00e2me lui inspirait frissons et saisissements extraordinaires. Aussi p\u00e8re usait-il d'un \u00e9trange man\u00e8ge pour retarder le moment de sa transformation en cadavre : il d\u00e9vorait chaque jour quantit\u00e9 de viandes grasses et de lard nouveau. \u00ab Il faut que le corps s'engraisse, Fils ! Parnoir ! Le tr\u00e9pas s'int\u00e9resse-t-il aux mangeurs et aux pansus ? Jamais ! \u00bb\n\nTel \u00e9tait son entendement. Moi, il me paraissait que le mourir n'\u00e9tait pas si grand malheur. Le pire, pour les p\u00e9ris, devait \u00eatre de se r\u00e9signer \u00e0 ne plus r\u00e9sider en l'ici-bas, \u00e0 ne plus entendre le doux bruissement des choses vives. Pour s\u00fbr, je concevais qu'aux premiers moments coul\u00e9s dans le tr\u00e9pas, la frisquetterie devait \u00eatre un brin trop forte. Mais apr\u00e8s, chaleurs devaient assur\u00e9ment rebrousser, et avec elles, apaisements. Au bout d'une dur\u00e9e on devait s'habituer, s'habituer \u00e0 tout.\n\nUn jour, p\u00e8re s'instruisit que je pouvais lorgner les d\u00e9funts. Mais on pourrait dire aussi que c'est moi qui traduisis que lui ne le pouvait pas. Depuis que j'\u00e9tais sorti de m\u00e8re et que j'avais commenc\u00e9 \u00e0 voir le jour, j'avais con\u00e7u que tous, bourgeois, cr\u00e9atures, jouvenceaux et m\u00eamement b\u00eates, oiseaux et poissons, \u00e9taient dot\u00e9s de cette facult\u00e9. Ce jour-l\u00e0, p\u00e8re me fit voir qu'il n'en \u00e9tait rien.\n\nNous \u00e9tions grimp\u00e9s sur le toit, affair\u00e9s \u00e0 r\u00e9tablir le couvert de la cabane qui gouttait, lorsque j'aper\u00e7us m\u00e8re, venant \u00e0 moi en silence, comme il est coutume pour les morts. Parvenue \u00e0 mon abord, elle s'assied sur les billes et m'examine avec bont\u00e9. Peut-\u00eatre afin de m'accorder le plus possible au mutisme des d\u00e9funts, j'avais adopt\u00e9 la r\u00e8gle de ne prononcer que peu de paroles en leur pr\u00e9sence. Simplement, quand ils paraissaient, je poursuivais ma besogne du moment et me contentais d'aller et venir \u00e0 l'entour d'eux, jusqu'\u00e0 ce qu'ils se d\u00e9cident \u00e0 rebrousser en l'outre-monde.\n\nM\u00e8re, comme toujours depuis son tr\u00e9pas, et pareillement \u00e0 tous macchab\u00e9es, paraissait attrist\u00e9e. Je m'approchai et vint poser ma main sur son \u00e9paule, effleurant le tissu de son accoutre. P\u00e8re, pendant ce temps, cognait avec la mailloche sur les billes. Interrompant subitement son labeur, il m'aper\u00e7ut d\u00e9visager le vide et former des gestes apparemment insens\u00e9s. Il m'entendit m\u00eame bruire ces mots : \u00ab Qu'as-tu donc, petite M\u00e8re ? Tu parais toujours si d\u00e9sol\u00e9e depuis ton passage chez l'outre-peuple ! Dis-moi : pourquoi donc les morts affichent-ils tous pareille m\u00e9lancole ? \u00bb Il en \u00e9prouva quelque trouble. Tel fut son langage, \u00e9mis d'une voix mena\u00e7ante : \u00ab Avec qui donc devises-tu de la sorte, Fils ? Le Diable ? Et pour qui sont ces mouvements de mains que tu \u00e9lucubres l\u00e0 ? \u00c9garerais-tu l'entendement ? \u00bb Me tournant vers lui, je d\u00e9clarai : \u00ab Mais, P\u00e8re, ne vois-tu pas que m\u00e8re p\u00e2tit de tristesse ? \u00bb Dans la seconde, sa face blafardit, sa chevelure se h\u00e9rissa. La mailloche lui \u00e9chappa des mains et s'en alla rouler tout du long du couvert, jusqu'\u00e0 d\u00e9gringoler dans la barrique. P\u00e8re chancela, \u00e9gara le pied, s'\u00e9croula, puis commen\u00e7a \u00e0 rouler sur les billes. M\u00e8re et moi le v\u00eemes quitter le domaine s\u00fbr du toit, effectuer un vol lourdaud et choir sur un lit de lycopodes. Je me pr\u00e9cipitai \u00e0 sa suite, non sans user de l'\u00e9chelle.\n\nJe dus, pour le ranimer, lui mettre sur face nombre d'\u00e9cuelles d'eau frisquette et am\u00e9lior\u00e9e de poivre de bazzanie. Vint le temps o\u00f9 il \u00e9mergea. Aussit\u00f4t son sens reconquis, il me grippa par l'encolure et m'interrogea en ces mots : \u00ab Ai-je ou\u00ef correctement, Fils ? Parbleu ! Tu vois les macchab\u00e9es ? \u00bb Je lui fis ce r\u00e9torque : \u00ab Mais, P\u00e8re, n'en est-il pas de m\u00eame pour toute personne pourvue de sens ? Oui, je vois les d\u00e9funts, qui me visitent assid\u00fbment. Ne les toises-tu pas aussi ? Ce n'est l\u00e0 que banalit\u00e9. Pourquoi t'alarmer comme si putois te pourchassait ? \u00bb\n\nDe son c\u00f4t\u00e9, m\u00e8re descendit doucettement du couvert, me mit une derni\u00e8re fois caresse sur joue et s'estompa en silence entre les taillis, une petite lueur bleut\u00e9e emmaillotant sa fr\u00eale personne.\n\n#\n\n\u00ab Mais que vois-tu de ce monde-l\u00e0, que vois-tu ? Des cadavres ? Des squelets ? Les feux de l'enfer ? Des vapeurs ? Des monstrueux ? \u00bb Bl\u00eame d'\u00e9pouvante, p\u00e8re me chauffait de questions. Concevant soudainement qu'il n'entendait rien aux choses outre-terrestres, j'entrepris de l'en instruire, au mieux de ma comp\u00e9tence. \u00ab \u00c0 dire vrai, j'y distingue peu de choses, commen\u00e7ai-je. Puisque ce sont les tr\u00e9pass\u00e9s qui nous c\u00f4toient et non l'inverse, je n'aper\u00e7ois jamais que leurs personnes, et pas du tout la contr\u00e9e en laquelle ils circulent. Aussi ne connais-je rien de leur logement, ni non plus de leurs paysages. Chacun porte \u00e0 son entour une lueur bleut\u00e9e, ce qui permet \u00e0 coup s\u00fbr de le d\u00e9m\u00ealer de nous, les vifs. Tous vont dans le mutisme le plus entier, sans doute parce qu'il en est ainsi en outre-monde : en ces parages, peut-on imaginer l'oiseau triller, la branchotte craqueler, la vache meugler, la cloche tinter ? Assur\u00e9ment, non. C'est le tr\u00e9pas, quoi. Et puis, tous affichent triste mine, dont je ne saurais traduire la cause. Je me prends souventes fois \u00e0 m\u00e9diter cela. Telle est ma songerie : \"Quelles sont les raisons de cette chagrinerie-l\u00e0 ? Pourquoi cette d\u00e9pression sur la face de m\u00e8re, et aussi sur celle de ses semblables d\u00e9funts ?\" C'est l\u00e0 mati\u00e8re que je m\u00e2che toujours. Mais j'ai beau fouir en tous coins de mon casque, rien n'y luit. Car, enfin, pour les p\u00e9ris, la mort ne semble pas si d\u00e9testable : on y va et vient ais\u00e9ment, on y coudoie aimablement, sans doute, d'autres d\u00e9funts. On rebrousse m\u00eame parfois parmi les vifs afin de visitement. Et puis, en outre-vie, les embarras coutumiers \u00e0 l'ici-bas ne paraissent plus exister. Gripperie, acc\u00e8s, branle-\u00e0-bas d'estomac, souffrance de molaire, \u00e9panchement de blair, brisure de cheville et autres f\u00e2cheries ne sont plus que mis\u00e9rables ressouvenirs, ce me semble. En somme, le tr\u00e9pas ne se montre r\u00e9ellement contrariant que pour les vifs qui surdurent aux d\u00e9funts et qui ont connu pour eux quelque sentiment. Car, n'est-ce pas, nous ne pouvons, le plus souvent, talonner nos aim\u00e9s en contr\u00e9e de mort. Alors quoi ? Quelle est donc la cause de la d\u00e9sole des outrepass\u00e9s ? Oui, plus je me creuse le casque, moins la lumi\u00e8re y d\u00e9bouche. Je ne traduis rien de l'abattement des morts. \u00bb\n\nP\u00e8re atteignait l'effarement. Appliquant les mains sur les flancs de son casque tels des bouchons, il s'exprima ainsi : \u00ab Tais-toi, Fils ! Mes esgourdes n'en peuvent plus ! J'ignore qui t'as accord\u00e9 cette facult\u00e9, mais que je t'avise : ne fais plus jamais mention de ces choses abominables ci-devant moi ! \u00bb Puis, arc et fl\u00e8ches en \u00e9chine, il s'en fut vitement sur la sente, feignant l'urgence de quelque chasse.\n\nJe cessai donc \u2014 mais pour un temps seulement, ainsi que vous le verrez, Monsieur le juge \u2014 de l'entretenir de cette mati\u00e8re. Je continuai toutefois \u00e0 recevoir r\u00e9p\u00e9titivement la visite de m\u00e8re et d'autres d\u00e9funts, souventes fois m\u00eame sous les propres yeux de p\u00e8re, mais sans que celui-ci le suspecte jamais. Car p\u00e8re \u00e9tait fort affair\u00e9 \u00e0 ne pas expirer.\n\n#\n\nJe crus pendant longues \u00e9poques que p\u00e8re n'\u00e9tait effray\u00e9 que par la mort. Or, il arriva une fois que je pus mesurer qu'une autre peur le gardait sous son empire. P\u00e8re tremblait en effet des mouvements inconsid\u00e9r\u00e9s de l'air, et s'alarmait maximalement d\u00e8s lors que le grain crevait. Quand le cyclone mena\u00e7ait, p\u00e8re \u00e9garait ainsi toute hardiesse, et s'enfuyait jambes \u00e0 son col afin d'\u00e9chapper \u00e0 la col\u00e8re des cieux.\n\nCe fut un de ces jours de courroux a\u00e9rien. Par haut matin, nous f\u00fbmes tir\u00e9s de notre roupil par le toquement du vent contre notre fen\u00eatre, par les craqueries r\u00e9p\u00e9t\u00e9es des billes de la cabane et par la secouade de notre porte semblablement heurt\u00e9e. Lorsque nous accour\u00fbmes \u00e0 la vitre afin de traduire ce branle-\u00e0-bas, nous v\u00eemes les herbes couch\u00e9es par les vents puissants, et les branches des arbres ployant sous l'attaque, et m\u00eamement plusieurs casser et partir en vol sans toucher le sol avant longs parcoursements. Les oiseaux aussi luttaient contre l'air rageur et, s'obstinant pour la plupart \u00e0 lui faire face, volaient cependant \u00e0 reculons, emport\u00e9s \u00e0 leur tour comme les branches de tant\u00f4t. C'\u00e9tait comme si la patinoire de l'air sur laquelle tous glissaient coutumi\u00e8rement s'\u00e9tait \u00e9tablie ce matin deboutement, et formait \u00e0 pr\u00e9sent un mur vivant, for\u00e7ant son avanc\u00e9e sur le monde au m\u00e9pris de ses habitants, de ses reliefs.\n\nLe sang de p\u00e8re se figea, sa chevelure se dressa, et son corps tout entier fut en un instant comme frapp\u00e9 de stupeur. Il eut alors cette parole, bizarrement rendue d'une voix \u00e0 la fois trembleuse et cat\u00e9gorique : \u00ab Fichtrons le camp, Fils ! Voici que cieux et mondes se f\u00e2chent ! \u00bb Je m'\u00e9lan\u00e7ai d\u00e8s lors pour le retenir, car il s'appr\u00eatait d\u00e9j\u00e0 \u00e0 ouvrir la porte et \u00e0 se jeter en p\u00e2ture au cyclone. Mon r\u00e9torque fut preste : \u00ab Retiens ton pas, P\u00e8re ! Car, enfin, o\u00f9 irions-nous dans ces turbulences ? \u00c0 levant ? \u00c0 ponant ? \u00c0 sud, \u00e0 nord ? Sous pareille ouragane, nous ne tiendrions nulle part, et serions bien vite transport\u00e9s ainsi que ces oiseaux : avan\u00e7ant reculoirement et ne menant nos personnes qu'\u00e0 rebours. Nous ne serions que jouets de bourrasque, et risquerions, plut\u00f4t que de sauver nos vies, d'en achever le cours. Car les vents, assur\u00e9ment, ignoreront nos corps gaillards et nous traiteront ainsi que brins et autres l\u00e9g\u00e8ret\u00e9s : ils n'h\u00e9siteront pas \u00e0 projeter nos personnes sur les arbres, ou sur les rochers, ou sur tout obstacle se dressant en notre route, nous fracassant et surbrisant les os. Est-ce l\u00e0 ce que tu ambitionnes ? \u00bb\n\nJe vis son \u0153il s'assombrir, la sueur former ruisselets sur sa face, sa m\u00e2choire se crisper, les musclures de son col se tendre comme l'arc. Sa denture, aussi, crissa, comme lorsqu'on gratte le gr\u00e8s vernis avec la pointe du coutelas. Sa voix grimpa, et forma ceci : \u00ab Alors, demeurons ! Demeurons et prions ! Prions la d\u00e9esse Lune, Fils ! Sans quoi je ne donne pas cher de notre peau ! \u00bb Puis, pleuroyant d'effroi, il se jette \u00e0 genoux, joint les mains, entrem\u00eale les doigts et commence invocations, appels, cloches et trompes.\n\nIl resta prieur et frissonnant pendant un court laps. Puis, comme saisi, sa face s'\u00e9claira soudainement et je le vis se remettre sur ses jambes, en proie manifestement \u00e0 quelque vive \u00e9meute. Simagr\u00e9es et contorsions le tordirent bient\u00f4t. Nul doute : ses gens lui faisaient visitement. Le voil\u00e0 par suite qui file en l'officine et en rebrousse aussit\u00f4t avec une cruche d'eau-de-geni\u00e8vre sous le bras. \u00ab Saisis le grelin et talonne-moi, Fils ! \u00bb me dit-il, me grippant l'\u00e9paule.\n\nBravant la bourrasque, esquivant de mon mieux branches, mati\u00e8res, volatiles et m\u00eame chipmonques pouss\u00e9s par les vents, je mis mes pas dans les siens jusque ci-dessous le grand bouleau croissant non loin de chez nous. Parvenu l\u00e0, stupeur ! Je l'entends me commander de ficeler sa personne au tronc de l'arbre ! Ses gens n'avaient donc ni piti\u00e9 ni sens ? Nul moyen pour moi de protester : d\u00e9j\u00e0 p\u00e8re s'\u00e9tablissait contre l'\u00e9corce et attendait que je l'y fixe. Le cyclone, du reste, couvrait ma voix et m'interdisait de r\u00e9pliquer, et la bousculade incoh\u00e9rente de toutes choses \u00e0 l'entour for\u00e7ait \u00e0 l'action, non aux palabres. L'esgourde pleine de la dispute des vents, j'attachai donc solidement au tronc toutes parties de p\u00e8re, non sans m'\u00eatre auparavant assur\u00e9 que la liane qui me servait de grelin r\u00e9sisterait \u00e0 la puissance de sa musclure. Futile souci : p\u00e8re se laissait brider telle la caille \u00e9visc\u00e9r\u00e9e avant la grillade. Une fois cette ligote bien ex\u00e9cut\u00e9e, p\u00e8re, d\u00e9signant \u00e0 ses pieds la cruche de tord-boyaux, me criailla ces paroles \u00e0 faire dresser chevelure : \u00ab Et maintenant, Fils, jette en quantit\u00e9 sur le sol, en cercle autour de moi, de cette eau-de-geni\u00e8vre ! \u00bb Des tressaillements me parcoururent. Quelle folie nouvelle p\u00e8re projetait-il donc ? J'ex\u00e9cutai cependant son mandement, sachant fort bien que toute opposition ne me vaudrait que claques au casque, enfermements, privations de soupiasse et autres d\u00e9triments. Je vidangeai la cruche autour du grand bouleau. Puis vinrent de la lippe de p\u00e8re les mots que je redoutais : \u00ab Parnoir ! \u00c0 pr\u00e9sent, course \u00e0 la cabane, saisis-y vite le feu, et viens le r\u00e9pandre sur la terre mouill\u00e9e de cette gn\u00f4le ! \u00bb Docile tel le chien de compagnie, je rebroussai donc \u00e0 l'\u00e2tre, y cueillis un tison et, reparvenu au grand bouleau, mis le feu \u00e0 son pied. La hargne des vents fit bient\u00f4t et vitement gonfler la flamme.\n\nSit\u00f4t le liquide enflamm\u00e9, je vis ce spectacle affligeant : p\u00e8re l\u00e9ch\u00e9 par le feu, \u00e9vacuant de son corps multiples sueurs, sa face ex\u00e9cutant abondance de grimaces ! Car, assur\u00e9ment, l'enfer n'\u00e9tait pas moins \u00e9chauff\u00e9. Cependant, hormis simagr\u00e9es et sudations, p\u00e8re subissait ce supplice sans broncher, apparemment sans se soucier de son calcinement tout proche. \u00c9tait-ce l\u00e0 r\u00e9ellement son ambition que d'\u00eatre roussi corps et casque, ligot\u00e9 au grand bouleau ? Qu'importe ! C'\u00e9tait trop pour un fils sensible comme faon : je n'allais pas concourir \u00e0 ce tr\u00e9pas organis\u00e9. Sans compter que les vents, fous de furie, mena\u00e7aient d\u00e9j\u00e0 de s'emparer des flammes et de propager l'incendie. Aussi entrepris-je sur-le-champ d'\u00e9teindre le tout \u00e0 grands coups de semelles, avant que l'\u00e9l\u00e9ment ne consume la for\u00eat et p\u00e8re tout entiers. J'y parvins non sans immolade de mes godillots, ni sans noircissement de ma face et de tout mon accoutre, \u00e0 force de pi\u00e9tiner, de taper et de cracher tel un poss\u00e9d\u00e9 sur le feu.\n\n\u00c0 la fin, tandis qu'il ne subsistait du brasier que fum\u00e9es et parfums de carnes r\u00f4ties, je me jetai aux pieds de p\u00e8re et, vaincu, sans souffle, le questionnai de cette mani\u00e8re : \u00ab Mais pourquoi, pourquoi cette folie ? \u00bb Mais lui, encore tout fumant et toussotant, ne r\u00e9pondit ni goutte ni n\u00e8fles. Il semblait encore en proie \u00e0 l'envo\u00fbtement de ses gens. Aussi h\u00e9sitais-je \u00e0 le lib\u00e9rer de ses liens : ces d\u00e9mons-l\u00e0 allaient-ils, si je le d\u00e9ligotais, ordonner \u00e0 p\u00e8re de m'administrer une bross\u00e9e, peut-\u00eatre m\u00eame de m'exp\u00e9dier en contr\u00e9e de mort ? Je restais coi, \u00e9tabli immobilement aux pieds de p\u00e8re, le guettant du coin de l'\u0153il. Du temps coula. Le cyclone, \u00e0 pr\u00e9sent, perdait en menace, ses souffles mourant d\u00e9j\u00e0 parmi les arbres lointains. Je vis bient\u00f4t dans le regard de p\u00e8re que ses gens avaient fui. Je coupai donc le grelin, puis nous rentr\u00e2mes en cabane, chancelants, surpris de pouvoir encore imprimer nos pas sur la terre, nous supportant l'un l'autre.\n\n#\n\nNous rest\u00e2mes long de temps comme sans vie, \u00e9tablis sur taboureaux, vides de pensers et de mouvements.\n\nLe jour fit son habituel tricotet d'heures. Puis vint le soir.\n\nJuste avant que la lampe des premiers astres ne s'allume, p\u00e8re sortit un moment de sa torpeur et me fit ce bref discours : \u00ab Il est dit que le feu \u00e9loigne la peur pour celui qui s'y frotte. Et ta m\u00e8re, Fils, \u00e9tait emplie de courage. Elle \u00e9tait semblable aux pumas qui peuplent cette for\u00eat, ignorant l'alarme, et le p\u00e9ril, et la col\u00e8re des cieux ! \u00bb\n\nPuis il se leva doucettement et, parvenu \u00e0 sa paillasse, s'y allongea, la face contre le mur. \u00c0 peine un demi-laps coula-t-il que d\u00e9j\u00e0 cornets r\u00e9sonnaient en notre logement. Car p\u00e8re \u00e9tait fort ronfleur.\n\nAinsi en \u00e9tait-il chaque fois que le grain crevait. Parfois, les gens de p\u00e8re venaient, parfois, ils ne venaient pas. Cependant, p\u00e8re c\u00e9dait toujours au paniquement le plus terrible. Je songeais : \u00ab Mais, enfin, pourquoi cette si forte crainte du vent ? \u00bb Certes, p\u00e8re \u00e9tait lecteur d'astres, et donc sans doute aussi un brin d\u00e9m\u00ealeur d'excentricit\u00e9s c\u00e9lestes. Mais sa lecture du grain \u00e9tait inexacte, pareillement \u00e0 celle qu'il se faisait du tr\u00e9pas. Car en quoi f\u00e2cheries de ciel et \u00e9ternuements de nuages pouvaient-ils inqui\u00e9ter ? Ce n'\u00e9tait l\u00e0, \u00e0 mon estime, que riens et broutils.\n\nMais, vous le verrez, Monsieur le juge, et vous aussi, Bourgeois et Cr\u00e9atures membres du jury de ce tribuneau, un jour allait venir o\u00f9 je traduirais l'effroi de p\u00e8re \u00e0 l'endroit du cyclone. Et j'allais aussi, par l\u00e0, embrasser meilleurement nombre d'inexplicablet\u00e9s le concernant.\n\n#\n\nLe sort d\u00e9cida une fois que nous s\u00e9journerions pour un temps au village. Ce jour-l\u00e0, p\u00e8re s'affairait par matin \u00e0 cueillir \u0153ufs de merle au sommet d'un fr\u00eane, non loin de la cabane. Il mit \u00e0 un moment le pied sur une branchotte un peu trop fluette, qui cassa, et lui fit \u00e9garer l'aplomb. Priv\u00e9 de son assiette, p\u00e8re chut, talonn\u00e9 par la branchotte, les \u0153ufs et amples d\u00e9bris d'\u00e9corce vari\u00e9s, racl\u00e9s par lui au passage. Arriv\u00e9 en bas, il hurla si fortement que les b\u00eates \u00e0 l'entour s'en effray\u00e8rent et fil\u00e8rent ventre \u00e0 terre dans un tonnerre de galopade. Car sa cheville paraissait disloque, et en son godillot la grosseur profitait. Comme j'avais patient\u00e9, au pied du fr\u00eane, que p\u00e8re redescende avec les \u0153ufs, j'\u00e9tais pr\u00eat \u00e0 intervenir. M'approchant de p\u00e8re d\u00e9gringol\u00e9, je dis, surpassant ses hurlades : \u00ab P\u00e8re, voici que ton godillot boursoufle avec empressement. J'incline \u00e0 croire que l'os est rompu, ci-dessous. Patiente ici, je course au village qu\u00e9rir le secours ! \u00bb Puis je d\u00e9talai, filant comme la fl\u00e8che lib\u00e9r\u00e9e de l'arc.\n\nApr\u00e8s une rude course en for\u00eat, je parvins donc \u00e0 l'abord du bourg, presque vid\u00e9 de mon souffle. Je n'\u00e9tais pas entr\u00e9 souventes fois au village. Bien s\u00fbr y avais-je pos\u00e9 les pieds par occasions, mais toujours assez sourdinement, afin d'y chouraver quelque fruit, morcelle de lard ou michotte de pain, et toujours sans en aviser p\u00e8re. Car p\u00e8re n'aimait pas l'humanit\u00e9, et encore moins que j'y exerce avoisinance.\n\nQue faire \u00e0 pr\u00e9sent ? Je songeai qu'il existait des bourgeois, appel\u00e9s doctes, dont le m\u00e9tier est de remettre en \u00e9tat les corps lorsqu'ils se rompent. Comment trouver ici une telle personne ? Je r\u00e9solus d'aborder le premier bourgeois, la premi\u00e8re cr\u00e9ature \u00e0 passer par-devant ma face. Car je ne pouvais bien s\u00fbr rep\u00e9rer la r\u00e9sidence du docte \u00e0 son panonceau, puisque j'\u00e9tais aussi analphabique qu'un putois. Ma man\u0153uvre fit merveille. Le bourgeois que j'avertis en attira d'autres, auxquels s'ajout\u00e8rent des cr\u00e9atures et m\u00eamement des jouvenceaux. Tous me questionn\u00e8rent. Je d\u00e9peignis de mon mieux le malheur de p\u00e8re : la branchotte c\u00e9dant sous son poids, sa descente, sa hurlade, les \u0153ufs cass\u00e9s sur sa liquette, les b\u00eates apeur\u00e9es, ma course jusqu'ici pour d\u00e9goter le secours. Comme chaque fois que je m'\u00e9tais m\u00eal\u00e9 \u00e0 l'humanit\u00e9 par le pass\u00e9, on me toisait avec curiosit\u00e9, comme si j'avais \u00e9t\u00e9 une bizarrerie. \u00ab Vous avez vu ses v\u00eatements ? Et ses chaussures ? Et comme il est sale ! \u00bb entendais-je sortir chuchotamment des bouches.\n\nUne cr\u00e9ature se d\u00e9gagea de l'agroupement. \u00ab Et votre p\u00e8re, o\u00f9 est-il en ce moment ? \u00bb me questionna cette charmante. \u00ab P\u00e8re est jonch\u00e9 sur la terre au pied du fr\u00eane, non loin de notre cabane ! \u00bb r\u00e9torquai-je. Puis elle me prit la main. J'en fus comme tout esbaudi, n'ayant jusque-l\u00e0 jamais fait l'objet d'un tel \u00e9gard. L'espace d'une dur\u00e9e, j'en oubliai quasiment le pourquoi de ma pr\u00e9sence au village. Quelque chose me piqueta dans le gosier, un peu comme quand p\u00e8re me faisait avaler la soupiasse aux fourmis pour me punir de quelque canaillerie. Je sentis mon c\u0153ur empli de cabriolades. Le souffle me fit d\u00e9faut un brin, quoique je fusse \u00e0 ce moment aussi stationn\u00e9 que l'ourse enroupill\u00e9e. Je l\u00e2chai : \u00ab Diable ! Quand vous me touchez la main tel qu'ainsi, c'est comme si farfadettes me chatouillaient sous le pied ! Cela me met hilarit\u00e9s au corps ! \u00bb Souriant doucettement, elle me dit : \u00ab D\u00e9p\u00eachons-nous d'aller chercher le docteur. Votre p\u00e8re est en difficult\u00e9. Et apr\u00e8s, il faudra vous d\u00e9barbouiller un peu. \u00bb\n\nMe _d\u00e9barbouiller_ ? Que concevait-elle par l\u00e0 ?\n\nMais j'oubliai vitement ces paroles \u00e9tranges, car je fus aussit\u00f4t distrait par les enseignements et sommations qu'elle distribuait maintenant aux curieux : \u00ab Le p\u00e8re Courge est bless\u00e9 ! \u00bb d\u00e9clara-t-elle. Puis, s'adressant plus directement \u00e0 l'un et l'autre bourgeois : \u00ab Toi et toi, courez jusqu'\u00e0 la cabane et voyez dans quel \u00e9tat se trouve ce pauvre homme, et rassurez-le : d\u00e8s que possible, nous arriverons avec le docteur ! \u00bb Et l\u00e0-dessus, ils commencent \u00e0 d\u00e9camper vers la for\u00eat.\n\nJusque-l\u00e0, je n'avais jamais con\u00e7u que p\u00e8re p\u00fbt poss\u00e9der un nom. J'ignorais aussi qu'il \u00e9tait connu au village et que toutes ces personnes \u00e9taient au fait de l'adresse de notre cabane. Je fus instruis bien plus tard que, si p\u00e8re poss\u00e9dait ici quelque r\u00e9putation, c'est qu'il s'\u00e9tait mesur\u00e9, en d'anciennes p\u00e9riodes, \u00e0 gens du village. Mais je rebrousserai en temps prochain, Monsieur le juge, sur cette portion de mon r\u00e9cit.\n\nD\u00e9j\u00e0 nous foncions vers le s\u00e9jour du docte. Voici que nous parvenons \u00e0 son seuil et que nous toquons. Le docte para\u00eet, et s'\u00e9tonne : \u00ab Manon ? Que faites-vous ici ? Vous \u00eates souffrante ? \u00bb Puis, m'apercevant ci-derri\u00e8re elle et plissant le blair : \u00ab Tiens ? Le fils Courge ! Comme il sent mauvais ! D'o\u00f9 diable sort-il encore ? Du d\u00e9potoir municipal ? \u00bb Je fus enseign\u00e9 ainsi du pr\u00e9nom de la bienveillante cr\u00e9ature que j'accompagnais. \u00ab Je vous en prie, fit-elle, imp\u00e9rieuse. Ne perdons pas de temps en vaines paroles, Docteur ! Le p\u00e8re Courge semble avoir eu un accident, il vous faut tout de suite venir avec nous le trouver dans la for\u00eat ! \u00bb Et l\u00e0-dessus, il d\u00e9clare : \u00ab Oh ! \u00bb Puis il s'\u00e9clipse h\u00e2tement \u00e0 l'int\u00e9rieur. Il revient sur le seuil avec sa valisette, et dit encore : \u00ab Je vous suis, Manon ! \u00bb Alors je repars en course, talonn\u00e9 par ma charmante et le docte.\n\nApr\u00e8s galops et parcoursements, nous voil\u00e0 \u00e0 l'abord de p\u00e8re presque crevant dans son godillot et passant une guirlande aux bourgeois de tant\u00f4t, arriv\u00e9s par-devant nous. Car p\u00e8re n'\u00e9tait pas commode avec l'humanit\u00e9. Mais d\u00e9j\u00e0 le docte d\u00e9saccoutre la cheville et commence son \u00e9tude. Il toise, il potasse, il enqu\u00eate. Puis il veut palper la grosseur afin de la mieux traduire. Il pose indextre dessus. P\u00e8re surtressaute et gueule de souffrance. De son pied encore valide, il met une savate dans l'estomac de son bourreau. Le docte se plie comme roseau en ouragane et sort la langue, et m\u00eame l'\u0153il un brin. Sa face \u00e9gare couleur, puis la recouvre, mais anormalement : la voici pareille \u00e0 bleuets de cornouiller. Pour s\u00fbr, il avoisine suffocation. Manon et les bourgeois accourent et s'inqui\u00e8tent de sa vie. Mais leur alarme est vaine : sous peu le docte reprend ressort et, se tenant le bedain, le souffle encore courtaud, d\u00e9clare : \u00ab C'est l'os. L'os de la cheville de Courge : cass\u00e9 ! \u00bb\n\n#\n\nIl nous fallut usiner une brancarde de branchottes et transporter p\u00e8re jusqu'\u00e0 l'hospite du village. L\u00e0-bas, on commen\u00e7a de suite \u00e0 lui coincer la cheville dans une pl\u00e2trade, afin, me dit-on, de permettre \u00e0 l'os de relever.\n\nCependant que p\u00e8re \u00e9tait livr\u00e9 au docte-pl\u00e2treur, Manon entreprit de me _d\u00e9barbouiller_. Je fus d'abord enseign\u00e9 du sens de ce mot-l\u00e0, qui d\u00e9signe immersion en barrique et savonnade vigoureuse. Dans une chambrette de l'hospite, je fus ainsi r\u00e9cur\u00e9 pour la premi\u00e8re fois de ma vie, \u00e0 l'aide de brosse et savonnette, que Manon maniait avec industrie. Tandis que je sentais la brosse man\u0153uvrer, il me paraissait que ma charmante ne faisait pas qu'enlever cro\u00fbtes et \u00e9tages de crasse sur ma peau, mais aussi qu'elle atteignait de plus aprofondes zones, jusqu'\u00e0 l'abord d'une contr\u00e9e encore ignor\u00e9e. Comme si elle se faufilait en ma personne, y d\u00e9frichait une for\u00eat nouvelle et y venait s'\u00e9tablir. Je songeais \u00e0 l'\u00e9tranget\u00e9 que voici : souventes fois, nous nous concevons reclus en nous-m\u00eames comme en accoutre \u00e9tanche. Puis, un jour, le commerce aimable des autres nous p\u00e9n\u00e8tre et abolit cette solitude de captif.\n\nIl y avait aussi dans la chambrette un instrument des plus prodigieux. Je crus d'abord qu'il s'agissait d'une fen\u00eatre, puisque j'y vis bient\u00f4t s'y fixer un bourgeois, qui n'avait de cesse de me lorgner et de singer chacune de mes postures. S'\u00e9gayant de ma surprise, Manon m'instruisit que l'instrument s'appelait un _miroir_ , ce qui \u00e9tait fort sens\u00e9 puisqu'on pouvait en effet s'y mirer. Moins grande fut mon aise quand je traduisis enfin que la face aper\u00e7ue l\u00e0-dedans n'\u00e9tait autre que la mienne. Je fus pris d'antipathie. Quoi ? Ce bourgeois \u00e0 la mine de vilain et de simple, c'\u00e9tait moi ? Quoi ? Ce d\u00e9sagr\u00e9able ? Cet \u0153il de pouacre ? Ce chevelu ? D\u00e9saplomb des choses ! Comment ce monde pouvait-il abriter si vives dissemblances : laiderons tel le fils Courge, et joliesses pareilles \u00e0 Manon ? Je d\u00e9tournai le regard et tentai de rebrousser en la contr\u00e9e que je d\u00e9couvrais par les soins de ma charmante. J'y parvins, cependant que Manon maniait avec hardiesse la brosse sur mon \u00e9chine. Mon consolement fut \u00e0 son comble quand, lorsque je sortais de la barrique, on me fit porter \u00e0 manger une savoureuse soupiasse de v\u00e9g\u00e9tation.\n\nEt tandis que j'avalais cette pitance, je m\u00e9ditai que Manon m'avait appris ce jour-l\u00e0 une paire de choses : le chatouillis du ch\u00e9rissement et, aussi, comme ligot\u00e9 \u00e0 cela, le sens de ces mots nouveaux pour moi : _d_ _\u00e9barbouiller_ et _miroir_. Sans doute, Monsieur le juge, fut-ce l\u00e0 la premi\u00e8re curiosit\u00e9 que je con\u00e7us non seulement pour amour, mais \u00e9galement pour vocabulaire. C'est \u00e0 ce moment-l\u00e0 qu'amour \u00e9tablit sa paillasse en ma personne. C'est l\u00e0 aussi que je pressentis que parole donne vie \u00e0 toutes choses en les baptisant d'un nom. J'appris le nom de p\u00e8re, puis celui de Manon, et ce fut pour moi comme si ces personnes commen\u00e7aient \u00e0 vivre v\u00e9ritablement : je les vis pour la premi\u00e8re fois. Je toisai en ces noms-l\u00e0 comme je toisai en miroir ma face d\u00e9livr\u00e9e de ses crasses : ce fut r\u00e9v\u00e9lation, et saisissement.\n\nP\u00e8re fut mis \u00e0 stationnement pour quelque temps. Malgr\u00e9 ses protestes, on nous assigna une chambrette, car le docte d\u00e9sirait profiter de notre pr\u00e9sence pour _faire des tests_ sur lui. C'est que p\u00e8re, depuis le d\u00e9but, avait \u00e9t\u00e9 surpris au moins une fois \u00e0 palabrer avec ses gens, \u00e9veillant ainsi perplexit\u00e9s parmi doctes et garde-souffrants de l'hospite. Suivant ses palabres, robustesse de bras avait \u00e9t\u00e9 n\u00e9cessaire afin de le retenir de partir en d\u00e9mente mission.\n\nEnqu\u00eates furent donc faites sur sa personne. C'est par suite de ces fouilles, men\u00e9es non sans difficult\u00e9s, qu'on lui d\u00e9nicha dans le casque le trouble dont je discoure depuis le d\u00e9but de mon histoire et qu'il contracta, je crois, le jour de l'inhumation de m\u00e8re.\n\nNotre s\u00e9jour dura nombre de courses de lunes et de soleils. Nous demeur\u00e2mes, p\u00e8re et moi, clo\u00eetr\u00e9s dans la chambrette. Chaque jour, un garde-souffrant venait avec l'intention de nous administrer une savonnade. Je me laissais brosser ainsi qu'il l'entendait, jusque dans le creux de l'ortelle et ci-derri\u00e8re l'esgourde. P\u00e8re n'\u00e9tait pas si docile. Chaque fois le garde-souffrant re\u00e7ut en face et sur casque savonnette et chaudron \u00e0 d\u00e9crassade, projet\u00e9s avec force par p\u00e8re. Presque chaque jour aussi, le docte venait me trouver et me disait : \u00ab Mais pourquoi diable ne sortez-vous pas un peu ? De gr\u00e2ce, changez-vous les id\u00e9es, allez au moins marcher dans les rues ! \u00bb Mais j'\u00e9lisais pr\u00e9f\u00e9rablement de demeurer aupr\u00e8s de p\u00e8re, presque toujours tenu stationn\u00e9 sur la paillasse par sa pl\u00e2trade. Car je sentais cro\u00eetre myst\u00e9rieusement comme ficelle me liant \u00e0 lui. Je ne traduisais gu\u00e8re cela encore, mais je con\u00e7ois aujourd'hui que c'\u00e9tait l\u00e0, d\u00e9j\u00e0, mon ch\u00e9rissement envers p\u00e8re qui ouvrait sa route en mes machineries. Et puis, j'aimais bien rester \u00e0 moisir moi aussi, go\u00fbtant repos et mollesses, avalant \u00e0 heures pr\u00e9cises soupiasses exquises et, surtout, guettant \u00e0 la fen\u00eatre le passage de Manon. Mais jamais, jamais, ainsi que vous le d\u00e9couvrirez bient\u00f4t, Monsieur le juge, ne pus-je rediscourir avec elle, ni non plus jouir de ses bienveillances.\n\nPuis, un jour, on lib\u00e9ra la cheville de p\u00e8re. Tandis que nous nous pr\u00e9parions \u00e0 quitter l'hospite, le docte remit \u00e0 p\u00e8re une fiole de granules et un papier lui permettant de s'en munir par suite autant que voulu aupr\u00e8s d'un docte-droguiste. Voici son dire d'alors, qu'il professa \u00e0 p\u00e8re avec sourcile bas : \u00ab Votre \u00e9tat est grave. Vous veillerez \u00e0 prendre ces m\u00e9dicaments tous les jours. Il en va de votre sant\u00e9 mentale et, partant, de la bonne marche de toute votre existence. Si vous n'avez pas les moyens financiers n\u00e9cessaires, faites une demande aupr\u00e8s de la S\u00e9curit\u00e9 sociale, ils vous inscriront. Et puis... vous devriez songer \u00e0 vivre autrement. Croyez-moi, pour vous comme pour votre fils, cette vie retir\u00e9e dans les bois comme un animal, \u00e7a n'a rien de bon. Les b\u00eates sauvages font ainsi. Pas les gens. \u00bb P\u00e8re prit rudement les granules et le papier, \u00e9mit un grognement, me grippa le coude et m'entra\u00eena par-devers lui.\n\nDehors, le soleil nous griffa l'\u0153il.\n\nSit\u00f4t que nous f\u00fbmes hors hospite, la fiole alla choir sur une pelouse, exp\u00e9di\u00e9 par le pied de p\u00e8re tout revigor\u00e9 apr\u00e8s son rel\u00e8vement dans la pl\u00e2trade. Puis nous rebrouss\u00e2mes vers la cabane. En chemin, bourgeois, cr\u00e9atures et jouvenceaux se retournaient sur notre passage et nous toisaient \u00e9trangement par-dessus l'\u00e9paule, comme si nous \u00e9tions de la race des b\u00eates. \u00ab Maudite racaille ! \u00bb clamait p\u00e8re \u00e0 leur endroit. \u00ab Faquins ! Pendards ! Marauds ! \u00bb serinait-il, simulant pour chaque villageois savates et mornifles, battant l'air de ses poings mena\u00e7ants, crachant sur les godillots des plus avoisinants.\n\n#\n\nJe continuai d'\u00eatre qui j'\u00e9tais : le fils Courge, besogneux sous les arbres de la for\u00eat, esclav\u00e9 \u00e0 p\u00e8re comme la fleur \u00e0 la pluie, le suivant en sa folie de casque comme arbre sous foudre : me br\u00fblant par feu de ciel, me d\u00e9chirant puis tombant enfin, renon\u00e7ant \u00e0 me tenir dress\u00e9. Cependant quelque chose avait commenc\u00e9 \u00e0 voir le jour au milieu de moi. Je sentais en mon poitrin battre un tambour nouveau et carillonner formidableries de cloches. Par nuits silencieuses, et aussi par moments d'accalmettes, je traduisais ce bruit, mont\u00e9 de mon c\u0153ur ainsi que la fum\u00e9e d'un feu de joie. En mes murs int\u00e9rieurs r\u00e9sonnait d\u00e9sormais ce nom aussi d\u00e9licieux que soupiasse de poisson-chat : Manon, Manon, Manon. Mais je rebrousserai bient\u00f4t, Monsieur le juge, sur cette mati\u00e8re-l\u00e0, b\u00e2tisseuse de tout l'\u00e9difice du malheur qui m'a men\u00e9 ci-devant vous.\n\nP\u00e8re, quant \u00e0 lui, ne se modifia gu\u00e8re, et continua d'\u00eatre visit\u00e9 par ses gens. Aussi partions-nous souventes fois dans la for\u00eat, parfois tout le jour, pour accomplir obligatoirement l'une ou l'autre action d\u00e9mente. Sur sentes et monts, par vaux, gorges et en ravins, je talonnais p\u00e8re tr\u00e9pignant, persuad\u00e9 par ses visiteurs imagin\u00e9s. J'allais sur son pas, me pliant \u00e0 son empire absolument, tenant donc le r\u00f4le de complice esclav\u00e9, faute de quoi j'\u00e9tais assur\u00e9 de subir ch\u00e2timents et exemples.\n\nJ'observais toutefois que les actions que ses gens lui mandaient semblaient toujours li\u00e9es par m\u00eame fil : toutes se rapportaient \u00e0 m\u00e8re, et \u00e9taient accomplies afin d'en honorer le ressouvenir. Comme si, \u00e0 d\u00e9faut de la coudoyer en ce monde des vifs, p\u00e8re la rappelait \u00e0 lui par le moyen de ses gens. Peut-\u00eatre ainsi, m\u00e2chais-je en moi-m\u00eame, la folie est-elle une sorte de corridor reliant provisoirement vifs et tr\u00e9pass\u00e9s. Ainsi la d\u00e9raison de p\u00e8re n'\u00e9tait-elle pas aussi folle qu'il y paraissait : m\u00e8re, par-del\u00e0 m\u00eame les limites de l'ici-bas, permettait \u00e0 p\u00e8re d'entretenir en sa vie quelque direction.\n\nMais comme il \u00e9tait ardu de traduire cette direction ! Un jour, nous \u00e9tions \u00e0 ramasser de la dalibarde pour le th\u00e9, quand p\u00e8re commen\u00e7a \u00e0 mouvoir les bras tel le moulin en pleine bourrasque. Puis le voil\u00e0 qui gratte l'air, et criaille, et exhibe denture, et retrousse lippe, et plisse blair. Il crache invectives et mises en demeure. Par suite il choit, roule sur le sol et ass\u00e8ne dans le vide coups de pied et crochets. M'est avis qu'il t\u00e2tait ainsi de griffer la face et d'enfoncer le ventre de ses gens afin de les pousser \u00e0 rebrousser en leur contr\u00e9e de diablotins. Car sit\u00f4t que ses d\u00e9mons lui venaient, p\u00e8re paraissait toujours guerroyer contre eux, luttant d'abord avec extr\u00eame z\u00e8le afin de garder son casque froid et vide de ces monstrueux-l\u00e0. Mais toujours cette lutte s'estompait, et toujours p\u00e8re capitulait puis commen\u00e7ait \u00e0 se soumettre aux desseins de ses gens. Il semblait d\u00e8s lorsqu'il \u00e9tait fendu en une paire de parties, ainsi qu'une billette sous le coup de l'outil : la terre de l'ici-bas ne paraissait plus supporter qu'un seul de ses pieds rondelets, cependant que l'autre foulait le sol d'un outre-monde.\n\nMais voil\u00e0 que son d\u00e9m\u00eal\u00e9 se finit. P\u00e8re se ram\u00e8ne sur ses pieds, secoue sa culotte pour en d\u00e9loger la terre et entreprend de discourir avec ses visiteurs chim\u00e9riques. \u00c0 la fin, il me tend sa hotte de dalibarde et commande : \u00ab Parvert ! Fils, emporte cette herbe \u00e0 la cabane. Nous avons \u00e0 faire ! \u00bb Alarm\u00e9 maximalement, je chargeai donc nos hottes sur mon \u00e9chine et talonnai p\u00e8re sur la sente.\n\nParvenu avec lui \u00e0 notre logement, je dus proc\u00e9der \u00e0 la plus d\u00e9testable composition culinaire de mon existence. P\u00e8re m'ordonna en effet de poser sur le feu la grosse marmite, puis d'y fourrer les ingr\u00e9dients que voici : quantit\u00e9 d'eau de la barrique, une couleuvre, les tripes d'un garenne, un godillot, abondance de petits cailloux, une poign\u00e9e de fourmis, une hotte compl\u00e8te de dalibarde, une famille de lombrics et toutes humeurs extraites d'un chipmonque : sang, flegme, s\u00e8ve, bile et atrabile. J'avais \u00e9t\u00e9 soumis, d\u00e9j\u00e0, \u00e0 avalement d'autres sordides repas : crevard de mouffeton, troubl\u00e9 de bif, r\u00e9pugnant de poularde ou piteux de f\u00e9tuque. Mais le rata que me fit appr\u00eater p\u00e8re ce jour-l\u00e0 outrepassa, en infamie, toute empiffrade d'avant.\n\nUne fois mes ingr\u00e9dients jet\u00e9s dans la marmite, p\u00e8re s'assied devant l'\u00e2tre et moisit jusqu'\u00e0 ce que \u00e7a bouillonne. Puis il prend la louche, brasse un brin, et gorge de cette affreuse mixture nos \u00e9cuelles \u00e0 soupiasse. Entamant la sienne, il dit : \u00ab Avale, Fils ! \u00bb D'un bond, je me r\u00e9pands \u00e0 son genou, pleurniche presque et rousp\u00e8te : \u00ab Mais pourquoi ? Pourquoi me faire avaler ce tord-bedain ? \u00bb Seulement, nul discours ne traverse sa lippe. Plut\u00f4t, il ing\u00e8re sa m\u00e9prisable bouillade goul\u00fbment, l\u00e8che la cuill\u00e8re. \u00ab Pourquoi ? Pourquoi ? \u00bb serinais-je, toujours \u00e0 son genou, tandis qu'il se servait d\u00e9j\u00e0 une nouvelle portion. \u00c0 la fin, il se dresse brusquement, pointe son doigt mena\u00e7ant vers moi et relance avec courroux : \u00ab Avale, Fils ! \u00bb R\u00e9sign\u00e9, je rebrousse sur taboureau et commence \u00e0 m'enfourner cette chose. Je manque vomir encore aujourd'hui rien que de l'\u00e9voquer. Pour bien satisfaire p\u00e8re j'avale tout et racle m\u00eame l'\u00e9cuelle, non sans grimaces. P\u00e8re \u00e9tant exauc\u00e9, je file par suite finir le jour et entamer le soir sous le grand h\u00eatre, allong\u00e9 et g\u00e9missant, me grippant les tripes.\n\nPlus tard, quand la nuit commen\u00e7a \u00e0 s'\u00e9taler et que nous pr\u00e9parions silencieusement nos paillasses, p\u00e8re pronon\u00e7a ces mots avec sto\u00efquerie, comme \u00e0 sa coutume quand il \u00e9voquait m\u00e8re : \u00ab Ta m\u00e8re ch\u00e9rissait fortes soupiasses. Celle d'aujourd'hui le fut. \u00bb Puis il mouilla la m\u00e8che.\n\nToute la nuit je fus tourment\u00e9 de crampes. L'aube me trouva, priv\u00e9 de sommeil, la face blanche comme galette de sel, le corps tout entier f\u00e2ch\u00e9 au simple penser d'engloutir quelque nourriture nouvelle.\n\n#\n\nMais il me faut \u00e0 pr\u00e9sent discourir plus longuement \u00e0 propos d'amour.\n\nManon avait donc mis le feu \u00e0 la poudre de mes chairs et, par reflet, \u00e0 celle de mon c\u0153ur. J'entrerai toutefois en cela par une autre porte : celle qui conduit au c\u0153ur de p\u00e8re. Car ce sont l\u00e0 mati\u00e8res avoisinantes.\n\n\u00ab Bien s\u00fbr, p\u00e8re ch\u00e9rit m\u00e8re, me mis-je \u00e0 songer. Mais m'aime-t-il, moi ? Nourrit-il \u00e0 mon endroit au moins quelque sentiment ? \u00bb Toujours, depuis mon encontre avec la douce Manon, cette m\u00e9ditation me venait. P\u00e8re m'aimait-il ? Rien ne me le laissait concevoir. Il me rossait. Il me soumettait \u00e0 des enfermements prolong\u00e9s dans la cabane. Il me for\u00e7ait au labeur le plus ingrat, sous climats de pluie ou de froid extraordinaires. Il m'extrayait du roupil d\u00e8s l'aube avec grandes criailleries, ne m'abandonnait jamais au repos avant l'apparition de la premi\u00e8re \u00e9toile du soir. Il me ravitaillait d'insectes grouillants, de pitances faisand\u00e9es, m'emp\u00eachait de revigorer le capiton de ma paillasse. Me refusait tout abord du village, sous peine de ch\u00e2timent redoutable. Me contraignait d'\u00e9gorger b\u00eates et oiseaux pour le seul plaisir d'achever, pas m\u00eame \u00e0 des fins d'avalement. Et, surtout, il m'entra\u00eenait avec lui dans sa folie, sous la gouverne de ses gens. P\u00e8re m'aimait-il ? Nul signe d'une telle chose depuis l'heure de mon premier jour.\n\nMe vint une fois la question que voici : aurait-il enfoui son ch\u00e9rissement pour moi dans la tombe avec m\u00e8re ? Je m\u00e2chais ce penser : \u00ab Serait-ce donc faisable de mettre en terre le sentiment humain ? \u00bb Bien qu'avoisinant en \u00e2ge la longueur d'une demi-vie au moins, je ne connaissais pas encore tr\u00e8s bien les lois conduisant le monde. Pouvait-on inhumer le sentiment, comme on le faisait coutumi\u00e8rement des chairs ? Produits de c\u0153ur seraient ainsi autant palpables que visc\u00e8res, que musclures ? Aussi enfermables que gangstaires ? Cela, en tout cas, traduirait la s\u00e9cheresse de p\u00e8re \u00e0 mon endroit. Il me fallait \u00e9prouver cette vue.\n\nJe m'en fus une nuit, grippant derri\u00e8re la porte notre pelle et usant de maintes pr\u00e9cautions afin de ne pas \u00e9veiller p\u00e8re. La d\u00e9esse Lune ne paraissait point, mais il me fut ais\u00e9 au seul \u00e9clairage des \u00e9toiles de mettre mes pas sur la sente, tant j'\u00e9tais form\u00e9 au dessin de ce pays. J'arpentai sous peu assez de terrain pour aboutir \u00e0 l'embranchement du grand m\u00e9l\u00e8ze, laissant entre la cabane et moi consid\u00e9rable distance. Je dus alors m'\u00e9tablir sur une billette et d\u00e9lib\u00e9rer : par o\u00f9 prendre ? \u00c0 levant, \u00e0 ponant ? Devant, derri\u00e8re ? Car, enfin, je ne concevais aucunement le lieu de l'ensevelissement de m\u00e8re, hormis qu'il existait quelque part en la vaste for\u00eat. Que je le sache, p\u00e8re n'avait jamais rebrouss\u00e9 \u00e0 la tombe. Jamais non plus ne m'y avait-il pilot\u00e9. La piste y menant \u00e9quivalait donc \u00e0 cheveu de mouche. Comment, en semblable circonstance, d\u00e9goter le repos de m\u00e8re ? Diable ! Et au plus \u00e9pais de la nuit !\n\nJ'errai parmi les arbres long de temps, furetant le sol \u00e0 la recherche d'un s\u00e9pulcre. N'aboutissant \u00e0 rien d'autre qu'\u00e0 lever ici et l\u00e0 g\u00e9linots et garennes, je commen\u00e7ai \u00e0 d\u00e9courager lorsque la fortune me vint en assistance. Le ciel pr\u00eata-t-il esgourde \u00e0 mes geignements ? Mais le ciel s'embarrasse-t-il au moins de nous ? Ou n'est-il simplement que b\u00e2che sur nos casques fi\u00e9vreux et peu instruits ? Un bourgeois parut, et je connus sans tarder \u00e0 la petite lueur bleue qui l'entourait qu'il s'agissait d'un tr\u00e9pass\u00e9. Il m'aborda, posa sa main sur mon \u00e9paule et, puisque les macchab\u00e9es ne discourent jamais, ne fit plus rien d'autre pendant un moment que me toiser. Je notai, en lui retournant son regardement, qu'il avait cette m\u00eame mine chagrin\u00e9e que m\u00e8re. Hormis cela, son chef \u00e9tait recouvert d'une chevelure abondante et d\u00e9sordr\u00e9e, et sa lippe habill\u00e9e de moustaches. Il paraissait p\u00e9n\u00e9trer mon entendement. M'est avis qu'il con\u00e7ut \u00e0 ce moment que j'\u00e9tais en qu\u00eate de la tombe de m\u00e8re. Le voil\u00e0 qui recule d'un pas, stationne, et pointe l'indextre en direction de ponant. Puis il rebrousse encore un peu, me salue de la main et s'\u00e9clipse lentement en la for\u00eat. D\u00e9j\u00e0 il n'est plus qu'ombre. Coudoyer un mort est toujours une exp\u00e9rience troublante. Que faut-il lire de leur coude touchant le n\u00f4tre ? Et si le tr\u00e9pas nous prenait l\u00e0, tandis que nous sommes affair\u00e9s enti\u00e8rement \u00e0 vivre ? Mais je boute hors mon casque cette m\u00e9ditation, car d\u00e9j\u00e0 la nuit passe et il me faut poursuivre mon dessein. J'empoigne la pelle que j'avais pos\u00e9e sur le sol et, m'en remettant au sort, m'avance sur la sente du c\u00f4t\u00e9 indiqu\u00e9 par le mort.\n\nJe progresse, respectant le cap. Puis voici la grande pruche. Je cesse mon cours, consid\u00e8re l'\u00e0 l'entour. Sur l'arbre, \u00e0 hauteur de genou, le dessin d'un c\u0153ur anciennement creus\u00e9 dans l'\u00e9corce. \u00c0 son pied, un monticule de terreau o\u00f9 croissent beno\u00eetes : le repos de m\u00e8re, assur\u00e9ment.\n\n#\n\nDans les branchottes, un hibou me toisait. \u00ab Que fabrique donc ici, et sous pareille heure, le fils Courge ? \u00bb semblait-il peser. Peut-\u00eatre l'empire de la fatigue commen\u00e7ait-il \u00e0 me faire \u00e9garer l'entendement, mais je go\u00fbtai \u00e0 ce moment le besoin de discourir \u00e0 cet oiseau. M'adressant \u00e0 lui comme s'il s'agissait d'un bourgeois, je lan\u00e7ai : \u00ab J'ai voyag\u00e9 jusqu'ici cette nuit pour m'instruire d'une chose : oui ou non, peut-on enfouir le sentiment humain, le peut-on ? Aussi, si le ch\u00e9rissement de p\u00e8re \u00e0 mon endroit repose sous ce terreau aux c\u00f4t\u00e9s des restes de m\u00e8re, puis-je, moi, ramasser ce sentiment, le ramener comme gibier \u00e0 la cabane et le r\u00e9introduire en p\u00e8re, aux fins qu'il m'aime d\u00e9sormais tel qu'un p\u00e8re doit aimer son rejeton ? \u00bb Ainsi parlai-je \u00e0 l'oiseau-ma\u00eetre-de-la-nuit.\n\nJe grippai par suite la pelle et entrepris de fouiller la terre. Je cognai bient\u00f4t sur le bois de la bi\u00e8re. Aussit\u00f4t, tous mes membres chevrot\u00e8rent et, en toutes parties de ma personne, secousses se firent sentir. Qu'\u00e9tait-ce donc que cette tr\u00e9pidation ? L'\u00e9motion de retrouver m\u00e8re en sa contr\u00e9e d'outre-vie ? Non, puisque, d\u00e9j\u00e0, m\u00e8re paraissait \u00e0 mon regard souventes fois, et m\u00eame deboutement plut\u00f4t qu'allong\u00e9e en couche-mort. Mais alors ? Le trouble de trouver au flanc de m\u00e8re l'affection dont p\u00e8re m'avait frustr\u00e9 jusqu'ici ? Oui, sans nul doute. D'un coup de pelle, je trouai les premi\u00e8res planches. Me jetant \u00e0 genoux et \u00e9cartant encore la terre, j'arrachai les suivantes, pourries par le passage des \u00e9poques. Un loup hurlupina.\n\nM\u00e8re \u2014 du moins ses restes \u2014 gisait l\u00e0, rompue, nettoy\u00e9e de ses chairs et humeurs. Je ramassai un \u00e0 un les os de sa carcasse, en formai un petit monceau au pied de la pruche. Quand tout le squelet fut extrait de la bi\u00e8re, j'examinai attentivement l'int\u00e9rieur. \u00ab \u00c0 quoi reconna\u00eet-on le sentiment ? \u00bb songeai-je alors. \u00ab Quel espace occupe-t-il ? Est-il sombre, lumineux, l\u00e9ger, lourdaud ? \u00bb Rien ne paraissait. R\u00e9solu, je d\u00e9cidai d'inspecter la bi\u00e8re de fond en comble en y introduisant ma personne. M'y glissant enti\u00e8rement, je m'y d\u00e9pliai les membres. Ainsi \u00e9tabli, j'eus bient\u00f4t le point de vue d'un macchab\u00e9e. Je toisais le monde des vifs tel qu'on devait le faire depuis l'outre-existence. Tout n'\u00e9tait qu'ombres et silences. Une chouette passait, parfois, ainsi que le fait la fl\u00e8che : trouant l'air, sifflant \u00e0 peine, impatiente de son dessein. En cieux, d'obscurs nuages se rassemblaient et formaient le plan de pluies \u00e0 venir.\n\nMais, autour de moi, nulle trace du ch\u00e9rissement de p\u00e8re.\n\nJe stationnai de la sorte long de temps, la carcasse de m\u00e8re \u00e0 mon c\u00f4t\u00e9. La for\u00eat \u00e0 pr\u00e9sent \u00e9tait coite, les b\u00eates, toutes assoupies, la nuit, frisquette et achevante. Haussant le regard, je per\u00e7us la branche d\u00e9shabit\u00e9e : le hibou avait fui. Je m'\u00e9branlai pour faire semblablement. Je rangeai m\u00e8re dans la bi\u00e8re et rebricolai son repos. Puis je remis sous mon pas le ruban de la sente, et m'en rebroussai vers la cabane. D\u00e9roulant mon chemin, j'emplis mon casque de cette obs\u00e9dante m\u00e9ditation : \u00ab Ainsi, me dis-je, p\u00e8re a conserv\u00e9 son sentiment pour moi \u00e0 l'int\u00e9rieur de lui-m\u00eame, puisque je ne le d\u00e9gotai pas ici cette nuit. \u00bb Cependant, un autre penser me venait et me suppliciait, que je brassais en moi-m\u00eame et que je ne cessais d'exprimer \u00e0 voix haute tout en marchant. Son \u00e9cho allait heurter les arbres \u00e0 l'entour et rebroussait \u00e0 moi, tel un r\u00eavement mauvais qui ne veut plus nous quitter et \u00e9tablit sa r\u00e9sidence en nos appareils. \u00ab Mais p\u00e8re m'aime-t-il, m'aime-t-il seulement ? \u00bb Ainsi parlais-je.\n\nD\u00e8s lors, je d\u00e9cidai de ceci : il me faudrait savoir, savoir absolument, d\u00e9sormais, si p\u00e8re me ch\u00e9rissait. Plus encore : il me faudrait _voir_ son sentiment, le toiser comme on toise toutes choses en domaine de la Terre. Car il devait \u00eatre possible, n'est-ce pas, que l'\u0153il saisisse cela, il le fallait, il le fallait. Oui, je serais \u00e0 compter d'ici tel l'aigle chassant la souricelle du haut de l'azur et ne pouvant s'en remettre enti\u00e8rement \u00e0 son esgourde, ou \u00e0 son blair : il me faudrait la preuve que procure le regard. De toute mani\u00e8re, pouvait-on concevoir de vivre, de circuler d'un bout \u00e0 l'autre de l'existence sans que le regard encontre le sentiment ? Peste soit de cette introuvablet\u00e9 ! On voit bien l'arbre, la pluie, le poisson-chat, et ce n'est que juste ph\u00e9nom\u00e8ne : ces choses-l\u00e0 nous sont indispensables au vivre. Ainsi de l'amour.\n\nMais, pour l'heure, je ne pouvais me d\u00e9router de ce penser d\u00e9solant : la nuit s'achevait et mon labeur \u00e9tait rest\u00e9 vain. En tous horizons, nulle preuve encore que je fus aim\u00e9 de p\u00e8re. Ah ! moi qui nourrissais pour lui de si tendres feux ! Pourrais-je un jour go\u00fbter de lui, en retour, semblables chaleurs ? Le pourrais-je ? Comme ce jour me semblait loin !\n\nL'aube commen\u00e7ait \u00e0 effacer les ombres quand je me glissai sur ma paillasse. Pleuroyai-je, Monsieur le juge ? Il me semble bien que non, tant \u00e9tait aprofond mon chagrin, et aprofondes les larmes qui auraient pu monter \u00e0 mon \u0153il.\n\n#\n\nCette mati\u00e8re avoisinait en moi l'espoir de retoiser Manon. Ma charmante n'avait-elle pas, la premi\u00e8re, sem\u00e9 en ma personne le besoin de ch\u00e9rissement ?\n\nJe profitais des siestes de p\u00e8re, ou de ce qu'il cuvait l'eau-de-geni\u00e8vre, pour filer en douce par la sente et gagner le flanc de la for\u00eat, d'o\u00f9 je pouvais lorgner \u00e0 loisir le va-et-vient des villageois. J'ascensionnais alors le grand orme qui croissait l\u00e0 et m'y logeais sur la plus haute branche. Je jouissais de la sorte d'une vue panoramique de la grande rue, o\u00f9 grouillait toute une population d'humanit\u00e9. Souventes fois, le soleil pouvait courser fort longuement en l'azur que je restais en mon mirador, ancr\u00e9 aux rameaux, \u00e0 lorgner chaque promeneur dans l'espoir d'y reconna\u00eetre l'exquise Manon. Mais jamais sa glorieuse face ne paraissait, et c'est le c\u0153ur empli de m\u00e9lancole que je me r\u00e9signais chaque fois \u00e0 redescendre sur terre, puis \u00e0 rejoindre p\u00e8re.\n\nUn jour que je rebroussais \u00e0 la cabane au terme d'une dur\u00e9e dans l'orme, je notai que p\u00e8re avait \u00e9court\u00e9 sa sieste et m'attendait, l'\u0153il mauvais, deboutement sur le seuil. \u00c9tait-il \u00e0 nouveau la proie de ses gens ? Je con\u00e7us en tout cas que j'aurais sous peu \u00e0 lui fournir traduction de mon absence. Car, hormis pour les missions de chasse qu'il me confiait, p\u00e8re aimait \u00e0 me garder \u00e0 distance de regard et me couvrait de ses foudres quand je me risquais \u00e0 d\u00e9porter sans raison recevable ma personne hors du champ de son \u0153il. Cette obsession tenait \u00e0 sa crainte, justement, de me voir gagner le village, de me m\u00ealer aux populations et adopter \u00e0 la fin leurs us et conduites.\n\nQuoi qu'il en soit, aussit\u00f4t que je fus \u00e0 son avoisinance, il me saisit l'esgourde et me la d\u00e9vissa sans quartier, jusqu'\u00e0 ce que je crie gr\u00e2ce. Il m'intima ensuite de p\u00e9n\u00e9trer par-devers lui dans la cabane, non sans m'administrer au passage une bross\u00e9e de godillots sur le train. \u00ab O\u00f9 \u00e9tais-tu terr\u00e9 ? \u00bb tonna-t-il. Je fis le na\u00eff et mentis formidablement, escomptant qu'il marcherait dans pareille dissimulation. \u00ab Euh... Je grappillais un peu de polytric pour bourrer ma paillasse, dis-je. Il me semble que le capiton est pi\u00e9tin\u00e9, depuis un moment ! \u00bb Cette f\u00e9lonie me valut mornifles sonores et branlements de casque. \u00ab Grain d'orage ! Ne me pigeonne pas, Fils ! \u00bb brama p\u00e8re. Je dus me confesser : \u00ab J'\u00e9tais au sommet du grand orme, \u00e0 l'imminence du village. Je fondais espoir d'y lorgner Manon, puis de l'aborder ensuite. \u00bb\n\nJ'aurais avou\u00e9 d'\u00eatre acoquin\u00e9 avec d\u00e9mons que p\u00e8re n'aurait pas \u00e9t\u00e9 moins secou\u00e9. Son courroux fut g\u00e9ant. Il me roua de coups jusqu'\u00e0 ce que j'en \u00e9gare les sens et m'\u00e9vanouisse au pied du taboureau. Par suite, il dut m'empoigner et me mener dans les buissons, car c'est l\u00e0 que j'\u00e9mergeai de l'anesth\u00e9sie o\u00f9 il m'avait plong\u00e9 pendant je ne sais quelle dur\u00e9e. J'y stationnai long de temps parmi les feuillages, allong\u00e9, raide et comme pierraille refroidie. Je fus tant bris\u00e9 que je ne sais comment mes parties se ranim\u00e8rent \u00e0 la fin et recommenc\u00e8rent leur emploi. Je notai bient\u00f4t cependant que le geste rebroussait en mes chairs et m'autorisait \u00e0 nouveau de bronchement.\n\nJe rentrai m'allonger sur ma paillasse. Il me fallut une course compl\u00e8te de soleil, puis une autre encore avant que je puisse \u00e0 peu pr\u00e8s ressentir les annonces du rel\u00e8vement. Pendant tout ce temps, p\u00e8re ne parut pas. Sans doute \u00e9tait-il \u00e0 battre la for\u00eat, reniflant la piste de quelque b\u00eate, ou emplissant sa hotte de vibournoume et de muguette-\u00e0-soupiasse.\n\nLe jour expirait lorsqu'il rebroussa \u00e0 la cabane, un garenne tr\u00e9pass\u00e9 en sa hotte. Muet comme bille, il \u00e9ventra la b\u00eate, l'\u00e9visc\u00e9ra, en fit une bouillade et m'en servit une \u00e9cuelle. Apr\u00e8s mon s\u00e9jour \u00e0 je\u00fbner sur paillasse, ce fut pour moi comme repas de monarque. Je restai long de temps \u00e0 go\u00fbter en silence, paupi\u00e8res baiss\u00e9es, la r\u00e9paration de mes chairs que me fournit cette nourriture.\n\nAux premiers astres, je questionnai enfin p\u00e8re faiblottement : \u00ab Me diras-tu, P\u00e8re : pourquoi cette si haute d\u00e9testation en ta personne \u00e0 l'endroit des villageois ? \u00bb Mais ma question fut ignor\u00e9e. Pour tout r\u00e9torque, p\u00e8re retira sa liquette, ajusta sa culotte et s'allongea, puis il mouilla la m\u00e8che. Je flairais cependant que l'\u0153il ne lui fermait gu\u00e8re. Et en effet, bien plus tard, quand la nuit fut \u00e0 son comble, il me dit, depuis sa paillasse : \u00ab Je ne blaire pas ces villageois. Ils nous ont fait outragure, \u00e0 ta m\u00e8re et moi. \u00bb\n\nJe restai longuement \u00e9tabli dans le noir, \u00e9chine contre mur et bras longeant le corps, l'\u0153il fixant les t\u00e9n\u00e8bres, \u00e0 m\u00e9diter ces paroles. Peu avant que l'aube ne se dresse, je risquai de questionner le sens de ses mots : \u00ab Mais, P\u00e8re, enfin : quelle fut cette outragure dont tu parles ? \u00bb Le r\u00e9pons qu'il me fit me parvint en ces termes, troublant un moment le silence \u00e9pais de la for\u00eat et les premiers mouvements du vent : \u00ab Parjaune ! Ne me pique plus avec le penser de ces pendards, et roupille donc, je te l'ordonne ! \u00bb Puis j'entendis la paillasse craquer sous le poids de p\u00e8re tournant sa personne contre le mur. La d\u00e9esse Lune, p\u00e2lissante d\u00e9j\u00e0, n'eut pas le temps de faire un saut de crapaud que tout p\u00e8re commen\u00e7ait \u00e0 tr\u00e9pider sous l'empire de son ronflement.\n\n#\n\nLorsque le soleil \u00e9tait bien enfonc\u00e9 \u00e0 ponant, il m'arrivait, quand il s'essayait \u00e0 traduire le ciel et son ornement d'\u00e9toiles, de questionner p\u00e8re sur ma destin\u00e9e. Car p\u00e8re \u00e9tait lecteur d'astres et, par m\u00eame occasion, d\u00e9chiffreur des avenirs inscrits en eux. Et moi, j'\u00e9tais en cette mati\u00e8re aussi curieux que le loutron : je n'avais de cesse de me cuisiner et de m'interviewer sur le sort me guettant et sur la tournure de ma personne. J'\u00e9tais ainsi fait, Monsieur le juge : je ne me rassasiais gu\u00e8re du jour coulant, et m'\u00e9tait besoin de creuser les \u00e9poques prochaines. Ah ! comme j'aurais go\u00fbt\u00e9 de me transporter en avant, en quelque machine ou brouette avaleuse de temps ! Pourquoi ? Il ne m'est pas ais\u00e9 de le traduire. Peut-\u00eatre cherchais-je en demains ce qu'aujourd'huis ne m'offraient que m\u00e9diocrement. On e\u00fbt dit que l'\u00e9poque pr\u00e9sente ne me suffisait jamais, et qu'il me fallait embrasser, afin de parfaire cette \u00e9poque, le projet et m\u00eamement la conclusion de mon existence. J'incline \u00e0 croire qu'il me fallait, pour mieux vivre, entrevoir la destination des choses, et ainsi imprimer signification \u00e0 tout ce qui pr\u00e9c\u00e9dait cette conclusion, un peu \u00e0 la mani\u00e8re de la fourmi qui rapporte en sa fourmili\u00e8re la goutte de miel assurant la survie de ses s\u0153urs insectes. M'\u00e9tait besoin de savoir que m'attendait quelque part une fourmili\u00e8re, et que ce que j'y promettais en mon trajet lui \u00e9tait n\u00e9cessaire. Et peut-\u00eatre \u00e9tais-je moi-m\u00eame une sorte d'insecte rapporteur, cherchant en ce monde \u00e0 se lier \u00e0 sa soci\u00e9t\u00e9 de semblables afin de lui fournir contribution. Quelle contribution ? Je n'avais en v\u00e9rit\u00e9 que peu de choses \u00e0 offrir, hormis la besogne de mon c\u0153ur, mon ouvrage de sentiment. Je ressentais souventes fois cela, lorsque je grimpais le grand orme et que je guettais en l'horizon l'apparition de ma charmante Manon. Je lui aurais alors volontiers fait cadeau de mes jours. Oui, il me fallait voir lointainement, afin de me mieux mesurer aux choses de maintenant.\n\nLe train bien \u00e9tabli sur la foug\u00e8re, le bedain plaisamment empli de rago\u00fbte de putois, de terrine de pimprenelle ou de brouet d'herbe-\u00e0-l'\u00e2ne, telle \u00e9tait donc ma voix : \u00ab P\u00e8re, que distingues-tu cette nuit de ce qu'il en sera de moi ? \u00bb Quoiqu'il ne f\u00fbt de coutume gu\u00e8re parleur, vint une fois o\u00f9 la glotte lui d\u00e9sengourdit. Lorgnant pensivement les mondes en cieux, il composa alors ce propos : \u00ab Parbeige ! Les astres cette nuit montrent de bien \u00e9tranges choses, Fils ! J'y lorgne qu'un jour viendra o\u00f9, par quelque diablerie, tu seras comme instruit de mots, et qu'alors lumi\u00e8res t'appara\u00eetront. \u00bb\n\nUn aprofond silence suivit ce pr\u00e9sage. P\u00e8re lui-m\u00eame en parut tout retourn\u00e9 et me consid\u00e9ra de son \u0153il mauvais, comme si j'abritais en ma personne une lign\u00e9e de d\u00e9mons. Aussi coupa-t-il court, avec les mots que voici, \u00e0 notre veill\u00e9e devant le feu : \u00ab Mais j'observe pareillement que nous entamons le laps n\u00e9cessaire \u00e0 la r\u00e9paration de nos forces. Je te l'ordonne, Fils : d\u00e9roule tes membres sur cette foug\u00e8re, et roupille \u00e0 pr\u00e9sent ! \u00bb\n\n\u00ab... et qu'alors lumi\u00e8res t'appara\u00eetront. \u00bb Paroles \u00e9nigmatiques, que je ne p\u00e9n\u00e9trai que bien plus tard ! Oh ! comme j'aurais voulu, cette m\u00eame nuit, creuser davantage le langage tenu par p\u00e8re, et en traduire sur l'heure toute l'esp\u00e8ce ! Quel \u00e9tait donc cette clart\u00e9 dont il me d\u00e9peignait l'av\u00e8nement, et dont je sentais d\u00e9j\u00e0 confus\u00e9ment la machine s'\u00e9branler en moi-m\u00eame ? Mais p\u00e8re lui-m\u00eame traduisait-il avec s\u00fbret\u00e9 toute l'importance de son oracle ? Telle \u00e9tait ma jonglerie, cependant qu'il \u00e9mettait, encore une fois, ronfleries et autres bruits de corps.\n\nIl me para\u00eet que les astres op\u00e9raient sur sa langue quelque besogne secr\u00e8te et prodigieuse. Car sous vo\u00fbte noire, la glotte lui d\u00e9la\u00e7ait parfois fortement, lui d'ordinaire si p\u00e9tri de silences. Il fut ainsi une autre fois o\u00f9 la parure des \u00e9toiles fit p\u00e8re plus parleur qu'en la tradition.\n\nC'\u00e9tait en heure d'aube, nous nous \u00e9branlions afin de courir le garenne. Il nous tardait, en effet, non seulement d'assurer le repas du soir, mais aussi de regarnir notre magasin d'accoutres. Car nos cache-esgourdes, excuse-train, mitaines, godillots-de-poil, tapisse-parties, escorte-blair et pousse-cuisses habituels mena\u00e7aient d'usure. Les cieux commen\u00e7aient d'\u00f4ter leur housse d'ombre, mais les \u00e9toiles toutefois r\u00e9sistaient encore et allongeaient leur m\u00e8che. Ayant \u00e9tabli sa besace sur son \u00e9chine, p\u00e8re ouvrit la porte et posa son pas sur le pr\u00eale et les premi\u00e8res foug\u00e8res. La face baiss\u00e9e, je le talonnai insoucieusement. Aussi ne notai-je pas qu'il stoppa bient\u00f4t son cours et stationna tout net ci-devant moi. Je le percutai, mais lui ne broncha brin, tant il \u00e9tait plant\u00e9 comme l'\u00e9pinette sur la terre. Rebroussant d'une longueur, je m'\u00e9tonnai de sa station et le questionnai ainsi : \u00ab Mais, P\u00e8re : pourquoi cette stopperie, lorsque nous entamons \u00e0 peine d'\u00e9branler ? P\u00e2tirais-tu de crampe \u00e0 l'ortelle ? \u00bb Mais p\u00e8re n'offrait point de r\u00e9torque et n'affichait nul tourment d'ortelle. Simplement, il moisissait l\u00e0, la face men\u00e9e \u00e0 l'azur. Puis, comme pensif, il plia genoux, peut-\u00eatre pour s'approcher de la voix des astres que semblait lui renvoyer la terre, plissa paupi\u00e8res et continua de plus belle \u00e0 toiser les choses ci-haut.\n\nJe con\u00e7us \u00e0 ce moment que le livre p\u00e2lissant des \u00e9toiles s'ouvrait \u00e0 son \u0153il et que p\u00e8re y lisait quelque r\u00e9cit nouveau. Tandis que j'\u00e9tais, moi, tout nanti de nos fl\u00e8ches, piques et collets, que j'avais le godillot impatient, le nerf prompt et l'\u0153il avis\u00e9, p\u00e8re dit, d'une voix \u00e9trangement coite : \u00ab Par hauts matins comme celui-ci, il me semble apercevoir la face de ta m\u00e8re en ce plafond de nuits et de clart\u00e9s m\u00eal\u00e9es. Car ainsi \u00e9tait ta m\u00e8re, lumineuse et ombreuse tout \u00e0 la fois. Un moment emplie de liesse, le moment d'ensuite opprim\u00e9e par quelque chagrinerie. Que contenait donc ce c\u0153ur, toujours balan\u00e7ant entre l'\u00e9lan et l'arr\u00eatoir ? Parbleu ! Ta m\u00e8re, Fils, \u00e9tait ainsi que l'aube : avan\u00e7ant au jour, regrettant la nuit. \u00bb\n\nSur ce mot, p\u00e8re se dresse lentement, toise une derni\u00e8re fois meutes et compagnies d'\u00e9toiles. Puis il engage son pas sur la sente et ses promesses de pelisses.\n\nJe le talonnai long de temps, le pied r\u00eaveur, le casque r\u00e9sonant de ses paroles. Surtout, je remuais la question pos\u00e9e par grand matin. En effet, que contenait donc le c\u0153ur de m\u00e8re ? Plus tard, lorsque je per\u00e7ais notre premi\u00e8re b\u00eate puis que j'en retirais distraitement la fl\u00e8che, je crus concevoir le r\u00e9pons que j'appelais. Oui, m\u00e8re \u00e9tait ainsi que le d\u00e9chirement de l'aube : son corps pouvait bien passer et tomber sous la bross\u00e9e de la mort. Mais son c\u0153ur ne se r\u00e9signait gu\u00e8re au tr\u00e9pas.\n\n#\n\nAvec les \u00e9poques, p\u00e8re, apparemment plus encore que moi, avan\u00e7ait \u00e0 grands pas en son m\u00fbrissement. Quel \u00e2ge pouvait-il compter ? Je ne saurais l'estimer, d'autant que j'ignorais le sens des nombres. Il me paraissait en tout cas qu'il \u00e9tait de plus en plus ancien, je le notais \u00e0 l'arc de son dos, aux rouillades de ses mains, au chaulage de sa chevelure, au labour de sa face. \u00c0 sa voix aussi, de plus en plus tremblante, approchant le b\u00e9gu\u00e8tement de la ch\u00e8vre.\n\nUn jour que nous courions le putois, p\u00e8re eut une g\u00eane. Lui de coutume si adroit \u00e0 la course, il dut \u00e0 ce moment stationner au pied d'un bouleau et s'empoigner le poitrin \u00e0 l'altitude du c\u0153ur, tant il paraissait p\u00e2tir. Inqui\u00e9t\u00e9 maximalement, je le questionnai : \u00ab Qu'as-tu, P\u00e8re, \u00e0 rompre ainsi notre cours et \u00e0 te tordre le corps ? \u00bb Mais il n'eut gu\u00e8re le temps de prononcer r\u00e9pons, car vint le moment o\u00f9 il s'effondra de toute sa dimension, le blair enfoui en terrain. \u00ab P\u00e8re ! P\u00e8re ! \u00bb criaillai-je alors, stupide, me jetant \u00e0 son chevet, le retournant comme galette et me collant l'esgourde \u00e0 son c\u0153ur pour lui appr\u00e9cier le pouls. Mais rien ne remuait sous sa liquette, et p\u00e8re demeurait aussi stationn\u00e9 que le poisson-chat gel\u00e9 sous l'\u00e9tang.\n\n\u00ab P\u00e8re tr\u00e9passe ! P\u00e8re tr\u00e9passe ! \u00bb fis-je dans mon alarme, me dressant et m'adressant \u00e0 voix forte \u00e0 la for\u00eat, prenant comme t\u00e9moins toutes choses v\u00e9g\u00e9teuses qui croissaient \u00e0 notre avoisinance.\n\nLe grippant par l'accoutre, aspirant \u00e0 je ne sais quel prodige, je me mis \u00e0 le remuer avec fr\u00e9n\u00e9sie en tous c\u00f4t\u00e9s tel un marionnet, \u00e0 lui battre le flanc et \u00e0 lui ployer jambes et bras, \u00e0 lui tirailler chevelure et blair, peut-\u00eatre au fond lui pr\u00eatais-je de cette mani\u00e8re un peu de la bougeotte des vifs, je ne sais, je ne sais. Mais, malgr\u00e9 tout ce branle-\u00e0-bas, nulle r\u00e9plique ne vint de son endroit. La bise s'essoufflant, il y eut un apaisement dans la ramure. Sur une branchotte, un cassenoix cessa son gazouillement et pencha l'\u0153il sur nous. Une mosquite passa.\n\nP\u00e8re ne respirait plus.\n\nLe c\u0153ur disloque, je me campai mollement \u00e0 son flanc, m'ensevelis la face dans sa liquette de fibrilles et commen\u00e7ai \u00e0 r\u00e9pandre mon pleur.\n\nM\u00e8re vint. Elle s'assit sur une billette et, de sa mine triste encore, nous contempla longuement. Puis elle s'avan\u00e7a jusqu'\u00e0 notre abord. Je pus fr\u00f4ler sa d\u00e9froque. M\u00e8re mit alors sa main sur mon c\u0153ur, puis sur celui de p\u00e8re. J'en fus si troubl\u00e9 que j'abaissai paupi\u00e8res, comme pour go\u00fbter, sans la distraction de l'\u00e0 l'entour, le sentiment que ce geste me causait. Je restai ainsi longue dur\u00e9e, semblablement qu'en \u00e9tat de pri\u00e8re : coi, \u00e9trangement distant du monde, de la for\u00eat et m\u00eamement de m\u00e8re, mais fort proche cependant de p\u00e8re cueilli tant\u00f4t par le tr\u00e9pas. Inexplicablet\u00e9 de nos conceptions ! Lorsque la mort cueille nos aim\u00e9s, nous voici pench\u00e9s \u00e0 leur chevet, comme s'il nous fallait saisir en eux ce que tr\u00e9pas y laisse, croyons-nous. Mais que laisse donc la mort en nos cadavres, hormis squelets et viandes promises au pourrir ? Que surdure-t-il donc en chairs refroidies des d\u00e9funts que nous ne discernons pas clairement, mais que quelque part de nous-m\u00eames semble flairer pourtant ?\n\n\u00ab Ouvre tes esgourdes et ton blair, Fils ! Le gibier me para\u00eet \u00eatre de ce c\u00f4t\u00e9 ! \u00bb\n\nAinsi parla p\u00e8re.\n\nJ'ouvris l'\u0153il. D\u00e9contenu, me dressant prestement, je le vis faire de m\u00eame et prendre par la sente, vert d'allure et leste d'enjamb\u00e9e comme s'il n'avait jamais tr\u00e9pass\u00e9, offrant blair aux vents et recevant \u00e0 distance l'odeur du putois. Pendant ce temps, m\u00e8re avait d\u00fb regagner l'outre-monde, car elle n'\u00e9tait plus visible \u00e0 pr\u00e9sent.\n\nOui, p\u00e8re m\u00fbrissait, et approchait l'\u00e2ge des enterrements. Mais on e\u00fbt dit que m\u00e8re, \u00e0 la fa\u00e7on \u00e9trange qu'ont les outrepass\u00e9s de veiller sur les vifs, le pr\u00e9servait du tr\u00e9pas. \u00c0 moins que ce ne f\u00fbt moi-m\u00eame qu'elle prot\u00e9geait, usant de p\u00e8re comme d'un outil ?\n\nUn soir, nous \u00e9tions affair\u00e9s \u00e0 bourrer de polytric le fondement de nos godillots. Les insectes, ces petits m\u00e9nestrels de la nuit, emplissaient l'air de leurs chants obstin\u00e9s. Des mouches-de-feu passaient. Tandis que je guettais l'embrasement de leurs ventres extravagants, je go\u00fbtais semblablement un feu en moi-m\u00eame. Je ne saurais traduire tr\u00e8s bien quel \u00e9tait ce feu-l\u00e0. Aujourd'hui que je me tiens deboutement ci-devant vous, Monsieur le juge, et que vous et les membres de ce tribuneau estimerez bient\u00f4t de mon histoire, puis d\u00e9ciderez de ma vie ou de ma fin, il me para\u00eet que cet incendie br\u00fblant en mes plomberies et terrains n'\u00e9tait autre que la joie d'\u00eatre vif. Surprise des jours ! Lors m\u00eame que le malheur s'abat sur nous comme gr\u00ealons, voici que brille en nos ventres une joie, courtaude et t\u00e9nue, mais pourtant suffisante pour pardonner au sort ses rudesses ! Ph\u00e9nom\u00e8ne des fortunes ! Miracle du monde ! Chef-d'\u0153uvre du ventre !\n\nCependant, ce feu, ce sentiment si grand, et m\u00eamement tous les sentiments qui s\u00e9journaient en ma personne, ne m'apportaient pas que liesse. Ils me rappelaient aussi ma tristesse de ne toujours point trouver pour moi chez p\u00e8re de sentiment, justement. Oh ! comme il me tardait d'apercevoir enfin, comme \u00e9manant de lui, un peu de ce ch\u00e9rissement-l\u00e0 !\n\nOui, oui, p\u00e8re m\u00fbrissait en sa vie et mena\u00e7ait, ainsi que le fait le fruit vieillissant sur l'arbre, de quitter la branche de l'ici-bas. Et plus les jours fuyaient, plus j'ambitionnais de voir amour sortir de lui, et s'exprimer de ses machineries, et se montrer, se montrer enfin ! Oui, o\u00f9 donc se cachait amour chez p\u00e8re, et comment le pouvait-on apercevoir en vrai ? Oh ! comme j'impatientais d'une telle vision, et comme je craignais de ne la pouvoir concr\u00e9tiser avant que, \u00e0 la faveur des jours fuyant, p\u00e8re ne tr\u00e9passe !\n\n#\n\n\u00c0 peu de temps de l\u00e0, p\u00e8re eut une nouvelle visite de ses gens. C'\u00e9tait au mi-temps de la nuit. S'\u00e9tant lev\u00e9 prestement, p\u00e8re alluma la m\u00e8che et s'avan\u00e7a, pr\u00eat \u00e0 entamer l'une de ses folies. \u00ab Sur pied, Fils ! \u00bb furent ses premi\u00e8res paroles. Me dressant alors sur ma paillasse, je n'aper\u00e7us d'abord, Monsieur le juge, que sa silhouette g\u00e9ante se dessinant devant la fen\u00eatre. Mon \u0153il s'accoutumant bient\u00f4t \u00e0 la lueur faible de la lampe, j'embrassai plus ais\u00e9ment le d\u00e9tail de sa personne et notai la d\u00e9mence sur son masque. De sit\u00f4t, j'estimai : \u00ab Par le diable ! Ce sont encore ses gens ! Mais ces affreux-l\u00e0 n'ont-ils aucun respect pour le roupil humain ? \u00bb M'exp\u00e9diant \u00e0 la face mon accoutre, p\u00e8re dit par suite : \u00ab Parnoir ! Enjambe ta culotte, et suis-moi ! \u00bb\n\nCependant, p\u00e8re \u00e9tait sorti et commen\u00e7ait \u00e0 \u00e9visc\u00e9rer le premier de la paire de faons achev\u00e9s le jour d'avant, puis remis\u00e9s contre la barrique. Je m'en fus le trouver, et le questionnai : \u00ab Du diable, P\u00e8re ! Pour quel motif t'escrimes-tu sous pareille heure \u00e0 l'\u00e9videment de ces b\u00eates ? Estimes-tu donc qu'elles recouvreront vie puis d\u00e9marreront avec allure par-devers les boqueteaux ? \u00bb Pour tout r\u00e9pons, p\u00e8re d\u00e9gaina un coutelas de son ceinturon et l'envoya se ficher dans le sol \u00e0 mon pied. \u00ab \u00c9tripe celui-l\u00e0 ! \u00bb fit-il mauvaisement, d\u00e9signant de son menton le second animal. Fr\u00e9missant tel le bouleau trembleur, je m'\u00e9tablis sans plus un mot.\n\nJ'entaillai pour commencer la peau sous le col, engageai le tranchant dans les chairs, puis formai un sillon jusqu'au bas-ventre. \u00c9cartant par suite couenne, musclures et ossements, je glissai mon coutelas dans la panse et me mis \u00e0 racler la cloison, d\u00e9grafant au passage boyaux, cr\u00e9pine, fraises, conduits et tubulures. Je compl\u00e9tai en ramassant les humeurs qui s'en \u00e9chapp\u00e8rent. Vint par suite l'\u00e9tape de videment du casque, que j'entrepris en tailladant depuis le col jusqu'au menton, puis par \u00e9corchade minutieuse jusqu'au sommet du front. Je pr\u00e9levai bient\u00f4t cervelle, plomberies de carafon, mati\u00e8res et si\u00e8ges d'entendement. Le dos, le flanc et les pattes p\u00e2tirent du m\u00eame sort. J'obtins \u00e0 la fin un monceau d'ossements et d'entrailles expertement s\u00e9par\u00e9s du cuir. Puis je commen\u00e7ai \u00e0 hacher, \u00e0 trancher, \u00e0 tailler et \u00e0 former en rosbives, ronds-de-bif, gigottes, g\u00e2teaux et pulpes-\u00e0-rago\u00fbtes les fruits obtenus de ma besogne.\n\nL'\u00e9toile de l'aube paraissait quand nous termin\u00e2mes nos \u00e9tripements. N\u00e9anmoins je d\u00e9m\u00ealai dans l'\u0153il de p\u00e8re qu'il \u00e9tait toujours poss\u00e9d\u00e9 par ses maudits visiteurs. Il \u00e9mit alors ce mandement, qui aujourd'hui encore me refroidit l'\u00e9chine quand j'y rumine : \u00ab \u00c0 pr\u00e9sent, passe cette d\u00e9pouillade, Fils ! \u00bb Je protestai prestement. \u00ab Quoi ? Glisser ma personne sous ce cuir encore moite de ses beurres ? Mais, P\u00e8re, aussi bien exiger de m'allonger dans quelque crottin de vache ! \u00bb Sa repartie, d\u00e9taill\u00e9e et tonitruante, tonna jusqu'aux confins de la for\u00eat. \u00ab Fais ce que je prescris ! \u00bb tonna-t-il crasseusement.\n\nJ'entrai donc sous la peau du chevrillard. P\u00e8re fit semblablement avec la b\u00eate vid\u00e9e par ses soins. Puis, j'entendis le son encavern\u00e9 de sa voix qui, sous le cuir de l'autre animal, ordonna : \u00ab Talonne-moi, et singe-moi ! \u00bb\n\nNous err\u00e2mes par les sentes un long moment, fagot\u00e9s de nos b\u00eates vidang\u00e9es et pendouillantes, mimant les attitudes du chevrillard. Je ne savais o\u00f9 p\u00e8re nous menait ainsi, mais je me gardais de rousp\u00e9ter, me contentant de l'exaucer et de le serrer de pr\u00e8s. \u00ab P\u00e8re aurait-il bascul\u00e9 pleinement dans la d\u00e9mence ? Son intermittente indisposition se serait-elle modifi\u00e9e en d\u00e9raison pure et continue ? \u00bb Tels \u00e9taient mes pensers, tandis que je trottais par-derri\u00e8re lui dans l'obscurit\u00e9 p\u00e2lissante du haut matin.\n\nNous arriv\u00e2mes \u00e0 la fin au mont Tondu. Nous le grimp\u00e2mes, toujours v\u00eatus de nos peaux, sous le trait rose des premiers rayons. Parvenus au sommet, nous stopp\u00e2mes notre trot et contempl\u00e2mes par les trous d'yeux des d\u00e9pouilles la for\u00eat s'\u00e9tendant en contrebas, telle une chevelure. P\u00e8re dit \u00e0 ce moment : \u00ab Pose l\u00e0 ta d\u00e9pouillade ! \u00bb J'ob\u00e9is de sit\u00f4t, par trop ravi d'all\u00e9ger ma personne d'un tel poids, et aussi de blairer enfin la bise plut\u00f4t que les effluves faisand\u00e9s du cuir. P\u00e8re fit de m\u00eame, et ajouta : \u00ab \u00c0 pr\u00e9sent, r\u00e9unis les flancs de ta b\u00eate, et couds-les pour qu'on ne s'avise plus de leur d\u00e9sunion. \u00bb Sans rouspet ni r\u00e2lement de nulle mani\u00e8re, sans question ni sondage sur le dessein de p\u00e8re, je m'en fus qu\u00e9rir une pique d'aub\u00e9pine, formai ample fil de sphaigne, puis m'affairai \u00e0 raccommoder le ventre de mon animal. P\u00e8re me toisait comme le ge\u00f4lier l'e\u00fbt fait d'un reclus. Je percevais \u00e0 mon pied la cha\u00eene de son empire. Mais j'avais peine \u00e0 d\u00e9m\u00ealer son \u00e2me. Mon labeur fini, p\u00e8re ordonna ce qui suit : \u00ab Paille-le ! \u00bb J'allai donc d\u00e9planter quantit\u00e9 de polytric, de dicrane et de sphaigne, rebroussai au chevrillard puis entamai de le bourrer par les trous de blair, d'yeux et de train. Il fut bient\u00f4t tant et si bien gorg\u00e9 qu'il put tenir debout et quasiment adopter la posture de l'animal vif. Par suite, p\u00e8re me commanda de r\u00e9p\u00e9ter les m\u00eames gestes sur l'autre d\u00e9pouille. Ainsi je reprisai le cuir qu'il avait port\u00e9 en for\u00eat, puis le bourrai jusqu'\u00e0 l'empli. J'obtins de la sorte une paire d'animaux certes paill\u00e9s, mais paraissant proche de d\u00e9camper tant leur d\u00e9gaine \u00e9voquait celle de leurs semblables vifs.\n\nP\u00e8re pronon\u00e7a alors ces paroles d\u00e9testables : \u00ab Tu stationneras seul, ici, dissimul\u00e9 dans les mousses, jusqu'\u00e0 ce qu'un faon de chair et d'os s'avance, attir\u00e9 par ces guignols. Puis tu le juguleras de la seule force de tes mains, et t'en rebrousseras \u00e0 la cabane avec son cadavre. \u00bb Ma r\u00e9plique fut ainsi : \u00ab Quoi ? Mais, P\u00e8re, tu con\u00e7ois comme moi qu'il peut se couler des jours et des lunes avant que semblable affaire n'advienne ! Il se pourrait que je stationne ainsi sur le mont Tondu des laps, des dur\u00e9es, des \u00e9poques et m\u00eame des \u00e2ges ! Qu'adviendra-t-il de ma personne, sans recours \u00e0 me loger, m'\u00e9chauffer, me repa\u00eetre et me breuver d\u00e9cemment ? Oh ! P\u00e8re, pourquoi tant de fureur et d'aveuglement si\u00e9geant sous ton casque ? \u00bb\n\nMais la bise seule r\u00e9pondit.\n\n\u00ab Tu souhaites donc m'achever ? \u00bb ajoutai-je faiblottement, non sans afficher mis\u00e9rable mine. Mais d\u00e9j\u00e0 p\u00e8re s'\u00e9branlait, s'affairant \u00e0 son rebours \u00e0 la cabane. C'est alors que me naquit une id\u00e9e : et si je t\u00e2tais de le ranimer, de le rebrousser soudainement \u00e0 l'entendement ordinaire ? Un heurt, une stupeur ne pourraient-ils le rendre \u00e0 une conduite plus proche du commun ?\n\nJe l'abordai.\n\nLe toisant dans l'ivoire de l'\u0153il, je comprimai les poings et les lui exp\u00e9diai avec force l'un suivant l'autre dans la face. P\u00e8re, malgr\u00e9 son \u00e2ge important, re\u00e7ut ma bross\u00e9e comme une tarte de jouvenceau. Il ne broncha que brin.\n\nSa riposte fut terrible.\n\nIl m'enfon\u00e7a premi\u00e8rement son pied dans l'estomac. Quand je fus pli\u00e9 sur la terre \u00e0 rechercher mon souffle, il me fit embrasser le mont Tondu en me m\u00ealant les gencives au sol par forte pression de sa main sur mon col. Je go\u00fbtai par suite de sa savate, mon flanc en accusant l'encontre pendant une dur\u00e9e qui me parut interminable.\n\nUn laps coula. Recouvrant peu \u00e0 peu le sens, je lorgnai p\u00e8re au travers du brouillard r\u00e9pandu en mon casque. Je m'effor\u00e7ai de m'instruire si au moins \u00e0 pr\u00e9sent ses gens l'avaient d\u00e9sert\u00e9. Je notai \u00e0 sa prunelle qu'il n'en \u00e9tait rien : p\u00e8re \u00e9tait plus que jamais en proie aux requ\u00eates d\u00e9mentes de ses d\u00e9mons. Je l'aper\u00e7us bient\u00f4t regagner la sente, puis se fondre \u00e0 la for\u00eat.\n\n#\n\nAujourd'hui encore, je ne sais \u00e9voquer ce s\u00e9jour sur le mont Tondu qu'avec vive frissonnade. J'y coulai des jours et des jours, tapi dans les lits de pr\u00eales avoisinants, guettant \u00e0 toute heure l'abord d'un faon. J'avais \u00e0 c\u0153ur d'exaucer p\u00e8re, car je concevais qu'il ne me recevrait jamais plus \u00e0 la cabane que si j'y rebroussais avec le cadavre qu'il escomptait. Me montrant \u00e0 lui sur notre seuil sans le faon tr\u00e9pass\u00e9, j'aurais assur\u00e9ment go\u00fbt\u00e9 de sa foudre. Peut-\u00eatre m\u00eame n'y aurais-je pas surdur\u00e9. Aussi bien alors n'y plus jamais refluer, et m'en aller subsister solitairement et pour perp\u00e9tue, tel le loup endeuill\u00e9.\n\nOr, j'ambitionnais de retrouver p\u00e8re. D'o\u00f9 me venait que, malgr\u00e9 ses cruels mouvements \u00e0 mon endroit, je le ch\u00e9rissais cependant plus que l'existence m\u00eame ? \u00c9tait-ce l\u00e0 l'effet puissant et imp\u00e9n\u00e9trable de la lign\u00e9e ? Le sang qui course dans nos veines est-il \u00e0 ce point porteur de sentiment ? Myst\u00e8re de nos jours ! Diablerie de la naissance, de la souche et de la famille !\n\n\u00c0 ses heures les plus hautes, le soleil me cuisait le cuir. J'avais beau \u00eatre abrit\u00e9 de ma chevelure, le trait s'y faufilait jusqu'\u00e0 me torr\u00e9fier le sommet du casque. Mon v\u00eatement non plus ne me secourait gu\u00e8re, le feu du soleil s'y m\u00ealant avec tant d'ardeur que je paraissais rissoler sous la fibre de la d\u00e9froque. J'adoptais alors la nudit\u00e9, ce qui condamnait l'enti\u00e8ret\u00e9 de mon corps \u00e0 plus funeste sort encore. Mais il arrivait aussi que le grain cr\u00e8ve et que je moisisse sous sa trombe ruisselante des p\u00e9riodes durant. Ou sinon, c'\u00e9tait la frisquetterie de la nuit qui me mordillait les membres, et je devais pour y parer me briquer tant et tant la peau que j'y imprimais plaies et bleus. Me repa\u00eetre n'\u00e9tait semblablement d'aucune sin\u00e9cure, et souventes fois je fus r\u00e9duit \u00e0 croquer le hanneton, la chenille et le ver, ou pire encore \u00e0 l\u00e9cher avidement la fourmili\u00e8re. Je devais aussi, pour m'\u00e9viter le dess\u00e8chement, breuvoir de ma propre pisse, puis p\u00e2tir par suite de mille crampes et conflits ventraux. Enfin, noctantement surtout, des b\u00eates sauvages venaient me visiter avec le projet de me d\u00e9vorer. Le bruit de leur pas et de leur reniflade int\u00e9ress\u00e9e me tirait d'un roupil fi\u00e9vreux, affreux, peupl\u00e9 de songes effrayants, o\u00f9 le visage mena\u00e7ant de p\u00e8re paraissait, crachant les asticots que j'avais moi-m\u00eame aval\u00e9s durant le jour. Par suite de mon \u00e9veil brusque et affol\u00e9 s'engageait un bref combat avec l'animal, au terme duquel il s'enfuyait, me laissant sans souffle et sans force, plus mis\u00e9rable encore qu'ant\u00e9rieurement.\n\nJ'implorais m\u00e8re de para\u00eetre. Elle seule, de son tranquille regard, aurait su me fournir r\u00e9confort. Je l'appelais, ma voix affaiblie s'\u00e9levant \u00e0 peine ci-dessus le chaume des pr\u00eales. Et quand le ressort me quittait et que je ne pouvais plus former ni parole, ni son, j'orientais mon toisement et mon discours vers l'int\u00e9rieur de ma personne et continuais \u00e0 convoquer m\u00e8re. Mais jamais elle ne parut.\n\n#\n\nPuis, un soir, j'aper\u00e7us non loin un faon grimper puis arpenter le mont Tondu. Le voil\u00e0 dans l'avoisinance des guignols. Il allait avec l'impr\u00e9voyance de la verdeur, reniflant plaisamment les d\u00e9pouilles comme si elles avaient \u00e9t\u00e9 ses s\u0153urs vives. Allong\u00e9, \u00e0 demi nu et mourant presque sur mon lit de mousses, j'agroupai mes forces restantes afin de mettre au point mon man\u00e8ge. Ma m\u00e9ditation allait ainsi : \u00ab Comment diable capturer, puis juguler ce chevrillard avec le brin de fougue qu'il me reste encore ? \u00bb Rien ne me venait, et tout mon casque r\u00e9sonnait de ce vide-l\u00e0, comme lorsque la cuill\u00e8re heurte la marmite d\u00e9semplie de sa soupiasse.\n\nPeut-\u00eatre au fond est-ce seulement \u00e0 cet instant que m\u00e8re parut, ou du moins me rendit mes appels. Certes non pas visiblement, par la m\u00e9diation de l'\u0153il \u2014 car je ne la toisai nullement \u2014, mais bien plut\u00f4t par sa simple pr\u00e9sence t\u00e9n\u00e9breuse et son conseil secret. Car je formai subitement ce dessein : ne pas m\u00eame remuer d'un pet de mosquite, mais singer le roupil et escompter que l'animal vert et fouinard vienne me renifler.\n\nMon escompte s'av\u00e9ra : je sentis bient\u00f4t la chaleur d'une truffe sur ma chevelure. Aussit\u00f4t, je pris vie et rassemblai promptement volont\u00e9, nerf et s\u00e8ve et grippai avec violence le faon au col.\n\nDe toute mon existence, ce fut le laps le plus surhumain \u00e0 mener. Couch\u00e9 sur le flanc, enserr\u00e9 au plus pr\u00e8s par mes bras, l'animal se braquait et se rebiffait \u00e0 grands coups de tordures. L'\u00e9treignant de la sorte, esp\u00e9rant son rapide tr\u00e9pas, je souffrais par mon corps innombrables douleurs. Je d\u00e9routais de mon mieux mes esgourdes des bruits secs de d\u00e9chirures et de bris \u00e9mis par mes entrailles et ma carcasse. Je fermais les yeux semblablement, par effroi de percevoir pendant l'assaut mes humeurs se r\u00e9pandre et se m\u00ealer \u00e0 la terre. Ma m\u00e2chure se serrait, chacune de mes musclures se tendait tel le nerf propulsant la fl\u00e8che, ma denture m\u00eame crissait horriblement. Cependant le chevrillard tenait bon sa vie, tandis que je me vidangeais peu \u00e0 peu de la mienne. Ses sabots labouraient et soulevaient la terre, ses yeux quittaient quasiment leurs trous, et je pouvais sentir le souffle hargneux s'\u00e9chapper de sa truffe. \u00ab Vivre ! Vivre encore ! \u00bb paraissait-il criailler. Et moi : \u00ab Tenir ! Tenir ! \u00bb Ce mot-l\u00e0 emplissait mon casque et ma bouche. \u00ab Tenir ! \u00bb Et ce n'\u00e9tait plus moi qui pronon\u00e7ais cela, mais mes visc\u00e8res elles-m\u00eames, mes charpentes, toutes parties me composant.\n\n\u00c0 l'instant o\u00f9 j'allais renoncer, j'entendis ces paroles prodigieuses, venant de la for\u00eat proche : \u00ab M\u00e8re ! M\u00e8re ! Pr\u00eate \u00e0 ton fils la puissance dont il est d\u00e9pourvu ! \u00bb Puis quelque chose craqua. Le faon connut d\u00e8s lors un ultime soubresaut, puis expira enfin, le col bris\u00e9 par mon \u00e9tau.\n\nJe stationnai longuement tel qu'ainsi, \u00e0 mi-chemin du tr\u00e9pas, mon corps jouxtant le cadavre, ma chevelure troubl\u00e9e par la bise. Je ne songeais de rien. J'\u00e9tais d\u00e9sempli. Le soir vint tout \u00e0 fait, puis la nuit, puis le jour, et encore le soir. Le sens regagna peu \u00e0 peu mon casque, et toutes machineries me concernant recommenc\u00e8rent \u00e0 s'activer. J'observai les \u00e9toiles poindre, la d\u00e9esse Lune s'avancer.\n\nJe songeai alors mollement : \u00ab Qui donc, tapi hier dans la for\u00eat, s'adressa \u00e0 m\u00e8re de la sorte pour l'implorer ? Fr\u00e8re ? Aurais-je donc un fr\u00e8re ici-bas ? \u00bb Je me levai avec ce que mes forces permettaient encore d'\u00e9branlement, descendis d\u00e9bilement en for\u00eat et fis, pour la premi\u00e8re fois depuis long de temps, un repas soutenant, form\u00e9 de racines et de baies. Je fus ainsi pendant une dur\u00e9e, immobile de gestes mais non de pensers. Peu \u00e0 peu mon sang fit son ouvrage et redonna quelque ressort \u00e0 mon corps. Ainsi recourag\u00e9, j'entrepris par suite d'explorer l'environ afin de d\u00e9busquer ce fr\u00e8re que je soup\u00e7onnais et qui, sans doute, m'avait sauv\u00e9 la vie. Je circulai en for\u00eat toute la nuit, h\u00e9lant sans rel\u00e2che mon parent. Mais le jour naquit sans que j'aie pu le conna\u00eetre.\n\nJe mijotai une bouillade de garenne, que j'avalai, l'esprit gorg\u00e9 de d\u00e9lib\u00e9rations. Je rebroussai au mont Tondu, consid\u00e9rai un moment le cadavre du faon. Puis je l'\u00e9tablis sur mon \u00e9chine et entamai mon rebours aupr\u00e8s de p\u00e8re. Sur la sente, je concevais \u00e0 pr\u00e9sent l'inexistence de fr\u00e8re, oui, je la concevais. Puisque c'\u00e9tait moi, uniquement moi, dans un acc\u00e8s de d\u00e9mence et dans un ultime effort pour \u00e9viter de toquer pour ainsi dire \u00e0 l'huis de l'outre-monde, qui avais prononc\u00e9 l'autre fois ces paroles \u00e0 l'endroit de m\u00e8re.\n\n#\n\nLe jour \u00e9tait \u00e0 son mi-temps quand j'aboutis \u00e0 la cabane. P\u00e8re cuisait le repas ci-devant le grand h\u00eatre. Je l'abordai, le lorgnai un moment dans le regard. Il fit de m\u00eame. De tout son corps, nul sentiment ne transpirait. Tout en pr\u00e9servant mon \u0153il dans le sien, j'inclinai l'\u00e9chine et fit choir l'animal d\u00e9sanim\u00e9 sur le sol. Mon discours fut ainsi ramass\u00e9 : \u00ab Me voici rebrouss\u00e9, P\u00e8re. \u00bb Il quitta mon regard et commen\u00e7a \u00e0 consid\u00e9rer pr\u00e9f\u00e9rablement le cadavre. D\u00e9fouraillant son coutelas, il transper\u00e7a le col de l'animal. Des humeurs s'\u00e9pandirent. P\u00e8re trempa les doigts ci-dedans, puis s'en mouilla la lippe. Relevant l'\u0153il, il me dit par suite : \u00ab Tu as bien fait, Fils ! \u00bb Et le voil\u00e0 maniant encore son coutelas et d\u00e9membrant d\u00e9j\u00e0 le chevrillard.\n\nUne dur\u00e9e coula, o\u00f9 j'\u00e9tudiai p\u00e8re \u00e9triper l'animal et en extraire mati\u00e8res \u00e0 cuire. Je risquai bient\u00f4t cette demande : \u00ab Qu'as-tu form\u00e9 pendant mon absence ? \u00bb Et lui, ne cessant en rien son labeur d'\u00e9carnage, ne quittant gu\u00e8re des yeux non plus son ouvrage : \u00ab J'ai d\u00e9vor\u00e9 la viande et le lard nouveau des chevrillards guignols. Tu devrais faire semblablement de ces carnes-ci. Ainsi serais-tu toi aussi d\u00e9tourn\u00e9 du tr\u00e9pas. \u00bb\n\nJe restai sans causerie, la face \u00e9gar\u00e9e dans le songe. Puis je le questionnai de telle mani\u00e8re : \u00ab Pourquoi m'as-tu contraint \u00e0 presque rendre l'\u00e2me au mont Tondu ? \u00bb Sa r\u00e9plique tarda, cependant qu'il d\u00e9coupait une rosbive. Une mouche vint sur ma main. Les feuilles du tremble bruirent. L'alouette \u00e9chauffa le nid. P\u00e8re dit enfin : \u00ab Ta m\u00e8re avait la gr\u00e2ce et la verdeur du faon. Il fallait, usant de man\u00e8ges et de trucs, capturer cette gr\u00e2ce et cette verdeur. \u00c0 pr\u00e9sent, Fils, pose ton train sur ce taboureau et engloute avec moi cette carne et ce lard nouvellement d\u00e9coup\u00e9s. \u00bb\n\n#\n\nMais avaler carnes et beurres ne lui suffisaient point, et toujours rebroussaient ses effrois quant au tr\u00e9pas. Une nuit, je fus extrait du roupil par une criaillerie \u00e0 faire dresser chevelure. J'estimai d'abord que p\u00e8re \u00e9tait la proie de cauchemars. J'eus alors ressouvenir de son enseignement, aux temps de ma verdeur, concernant cette sorte de songes malvenus : \u00ab Tout cauchemar p\u00e9n\u00e8tre la cervelle par les esgourdes, Fils ! Aussi, lorsque ces songes de d\u00e9mons te saisissent, scelle bien ces esgourdes-l\u00e0 de tes mains, ainsi que bouchons de cruche ! \u00bb De ma paillasse, je lan\u00e7ai donc cette directive, escomptant que mes paroles perceraient le roupil de p\u00e8re et atteindraient son casque : \u00ab P\u00e8re ! Te voil\u00e0 le jouet de songes d\u00e9testables ! Scelle tes esgourdes ! Scelle tes esgourdes ! \u00bb Mais mes efforts rest\u00e8rent sans fruit. P\u00e8re persistait dans sa gueulerie. Cependant, je le vis bient\u00f4t quitter sa paillasse et commencer d'arpenter la cabane, non sans \u00e9mettre encore couinements et hurlades. P\u00e9n\u00e9tr\u00e9 d'alarme, je me repr\u00e9sentai alors que ses gens reparaissaient. Me dressant sur ma couche, je m'exclamai : \u00ab P\u00e8re ! Tes gens sont de rebours ! Voici qu'ils te poss\u00e8dent ! \u00bb J'enfilai par suite ma culotte, enflammai la m\u00e8che et m'empressai d'aborder p\u00e8re, lui grippant les poignes et le for\u00e7ant \u00e0 me toiser dans la prunelle. \u00ab Bataille ! Guerroie contre ces importuns-l\u00e0 ! Ne les laisse pas conqu\u00e9rir ton casque ! \u00bb hurlais-je. Mais p\u00e8re, col\u00e9reux, se d\u00e9tournant un courtaud instant de son \u00e9tat, r\u00e9pliqua tel qu'ainsi : \u00ab \u00c9garerais-tu l'entendement, Fils ? Parbleu ! De quels gens causes-tu donc ? Contre qui m'engages-tu de quereller ? Ne p\u00e9n\u00e8tres-tu pas que mes criailleries ne sont que le fruit d'une ample frayeur ? \u00bb \u00c9prouvant son dire, je le questionnai en cette mani\u00e8re : \u00ab Mais, alors, quelle est au juste cette terreur formidable que tu exprimes ? \u00bb Et lui de repartir, tout \u00e0 la fois maussade et hagard : \u00ab C'est que, cette nuit, le roupil m'esquive et je ne cesse de me repr\u00e9senter en l'\u00e9tat de macchab\u00e9e, errant dans l'outre-monde effrayant ! Oh ! quelle infortune que de devoir sortir un jour d'ici-bas pour p\u00e9n\u00e9trer le domaine des monstrueux ! \u00bb\n\nJe rel\u00e2chai doucettement les poignes de p\u00e8re. Lui, de sit\u00f4t lib\u00e9r\u00e9, s'en fut s'effondrer sur le taboureau puis commen\u00e7a \u00e0 d\u00e9verser son pleur. Ce tableau me remua jusqu'au plus \u00e9pais de ma personne : \u00e0 part au moment du tr\u00e9pas de m\u00e8re, c'\u00e9tait la premi\u00e8re fois que j'apercevais p\u00e8re pleuroyer avec autant d'abondance.\n\nJe m'en fus tirer l'autre taboureau et m'\u00e9tablir aupr\u00e8s de lui. Nous rest\u00e2mes une longue p\u00e9riode ainsi post\u00e9s, lui se r\u00e9pandant comme rivi\u00e8re, et moi recherchant quelque discours d'apaisement. Durant ce temps, des avis me vinrent. Je songeais aux tr\u00e9pass\u00e9s que j'apercevais parfois. \u00ab \u00c0 la v\u00e9rit\u00e9, aucun de ces quitte-la-vie n'\u00e9veille quelque crainte que ce soit \u00bb me disais-je. Puis : \u00ab Et si p\u00e8re s'instruisait \u00e0 voir les tr\u00e9pass\u00e9s ? Peut-\u00eatre moi-m\u00eame pourrais-je t\u00e2ter sur lui semblable enseignement ? Ainsi, il entendrait bien que l'outre-monde et ses occupants n'ont rien de plus terrifiant que l'ici-bas. \u00bb Je stationnai encore long de temps sur le taboureau \u00e0 balancer ce dessein, tandis que p\u00e8re s'\u00e9tait remis sur jambes et avait repris ses rotations autour de la table, non sans geignements et autres marques d'affres.\n\nSortant \u00e0 la fin de songerie, je me campai ci-devant lui et d\u00e9clarai : \u00ab P\u00e8re, m'est avis que tu te dois, par mon \u00e9cole, de percevoir les tr\u00e9pass\u00e9s ! \u00bb En r\u00e9pons \u00e0 ce discours, p\u00e8re serra la denture et \u00e9mit par elle amples crissements, et son visage aussi s'\u00e9nerva, affichant d\u00e9bravourement. T\u00e2tant par suite de le mener \u00e0 raison, je pr\u00e9cisai : \u00ab P\u00e8se un peu, P\u00e8re ! Moi qui toise les d\u00e9funts depuis l'\u00e2ge de verdeur, en aucun jour n'ai-je entraper\u00e7u chez eux quelque tableau excitant l'effroi ! Jamais ne me suis-je vu trembleur au moment de mes encontres avec les macchab\u00e9es ! Et m\u00eame, j'ai presque senti \u00e0 leur abord grande calmerie. Presquement, aussi, une assurance de qui\u00e9tude pour l'heure promise de mon propre repos. Et \u00e0 dire vrai, une seule devinette pers\u00e9v\u00e8re toujours ci-dedans moi \u00e0 leur propos : pourquoi affichent-ils pareille tristesse puisque, tout pes\u00e9, l'outre-monde ne para\u00eet pas un lieu si terrible o\u00f9 s\u00e9journer ? Mais j'escompte un jour venir \u00e0 bout de cette devinette-l\u00e0, assur\u00e9ment, assur\u00e9ment. \u00bb\n\nQuoique grandement apeur\u00e9, p\u00e8re pr\u00eatait esgourde \u00e0 mon pr\u00eache. Il allait et avan\u00e7ait, rebroussait, puis red\u00e9marrait avec fi\u00e8vre de tous c\u00f4t\u00e9s de la cabane, cognant au passage un membre sur une cruche d'eau-de-geni\u00e8vre, chavirant la pelle, ou emboutissant taboureau et paillasse. Ses mains se briquaient l'une l'autre par trop d'\u00e9meute int\u00e9rieure. Sa face se tordait, sa lippe souffrait nombre de morsures, son sourcile s'\u00e9levait puis ravalait de sit\u00f4t. Au bout d'un moment, sa tremblerie alentit. Il parla : \u00ab Mais comment m\u00e8neras-tu donc pareil dessein, Fils ? Con\u00e7ois-tu r\u00e9ellement pouvoir me professer semblables diableries ? \u00bb Puis, imprimant pour ainsi dire sa face sur la mienne, il devint plus enqu\u00eateur encore : \u00ab Et toi-m\u00eame, au reste, peut-\u00eatre es-tu le d\u00e9mon en personne, ou sinon son estafette ? Peut-\u00eatre cet enseignement que tu pr\u00eaches n'est-il qu'intrigue pour m'exp\u00e9dier au plus vite en patrie des monstrueux. Qui me dit que mon propre fils, qui peut ordinairement voir macchab\u00e9es et suites, n'est pas comparse de ces horribles-l\u00e0, et ne recherche pas \u00e0 recruter parmi les vifs ? \u00bb\n\nJe me h\u00e2tai de le d\u00e9tourner de cette conception : \u00ab Mais, P\u00e8re, si j'\u00e9tais, tel que tu le proposes, valet du diable, ne percevrais-je pas le diable lui-m\u00eame, avec sa queue roussie, tra\u00eenant par-devers lui ses marmites o\u00f9 cuire damn\u00e9s, ha\u00efssables et r\u00e9prouv\u00e9s ? Or, je le serine : en aucun jour n'ai-je entraper\u00e7u pareille chose, ni quelque motif d'affres. \u00bb\n\nCe discours commen\u00e7a d'amollir le frein de p\u00e8re. La nuit continua d'avancer. Je fis semblablement dans mon argumenterie. Le jour pointait quand j'eus raison de ses derni\u00e8res r\u00e9sistances : il fut tranch\u00e9 que ce jour m\u00eame serait celui de sa premi\u00e8re le\u00e7on.\n\n#\n\nCe n'est assur\u00e9ment pas sans r\u00e9pugne qu'il s'accorda \u00e0 mes mandements. Depuis l'instant de mon d\u00e9clenchement en cette vie, je n'avais cess\u00e9 d'\u00eatre sous l'empire de p\u00e8re, de jouer l'apprenti. Voil\u00e0 que les emplois s'inversaient et que je brocantais ce r\u00f4le pour celui du ma\u00eetre, cependant que p\u00e8re prenait celui de l'\u00e9l\u00e8ve. Il lui fallut quelque sursis avant que de concevoir ce personnage nouveau. Les premi\u00e8res journ\u00e9es de mon enseignement me valurent souventes fois bross\u00e9es et mornifles, embrassades de sol par gencives et m\u00eame savates au corps, car p\u00e8re ne blairait gu\u00e8re qu'on lui dicte conduites. Beaucoup d'occasions se renouvel\u00e8rent o\u00f9 je fus forc\u00e9 de lui ramener en casque mon dessein. Je le suppliais : \u00ab P\u00e8re, il faut te plier \u00e0 mon \u00e9cole, malgr\u00e9 ta d\u00e9plaise. Sans quoi ton effroi de l'outre-monde ne tarira qu'\u00e0 la saint-glin-glin. Est-ce l\u00e0 ce que tu ambitionnes ? D'aborder ton moment dernier encore tout aussi trembleur que fillette \u00e9gar\u00e9e noctantement en for\u00eat ? \u00bb \u00c0 la longue, mes argues firent profit, et p\u00e8re, quoique bourrel\u00e9 d'acrimone, se rangea \u00e0 mon gr\u00e9.\n\nComment per\u00e7oit-on les macchab\u00e9es ? Pourquoi paraissent-ils ? O\u00f9 donc se situe l'outre-monde ? Peut-on y entrer et en sortir comme d'un moulin ? Pourquoi les vifs ne poss\u00e8dent-ils pas tous unanimement le don de percevoir les outrepass\u00e9s ? Quel est le motif de la tristesse lue sur la face de ceux qui s\u00e9journent en contr\u00e9e de disparus ? Peut-\u00eatre \u00e9taient-ce objets qu'il m'aurait fallu d\u00e9m\u00ealer avant toute chose. Mais il me paraissait que p\u00e8re \u00e9tait en n\u00e9cessit\u00e9 d'exercice concret plut\u00f4t que de m\u00e9ditation. Je l'instruisais donc en cette direction, lui serinant une loi persistante : \u00ab Il importe en premi\u00e8re vue, lui affirmais-je, de concevoir que les tr\u00e9pass\u00e9s ne paraissent jamais qu'au moment le plus improbable. Tu les apercevras souventes fois \u00e0 l'instant o\u00f9 tu usines quelque soupiasse, sinon cependant que tu abordes le roupil ou que tu circules insoucieux sur la sente. \u00bb \u00c0 ce discours, p\u00e8re me lorgnait curieusement. \u00ab Forbanterie ! s'exclama-t-il une fois. Ces affligeants-l\u00e0 n'ont-ils aucunement le sens du savoir-arriver ? Ainsi devrais-je augurer de leur visite m\u00eamement que je ne la souhaite point ? Qu'est-ce donc que pareils gueux ? Des marchands de savonnettes ? \u00bb Et le voil\u00e0 tout orageux, allant et venant ci-devant le grand h\u00eatre, aplatissant aralies et dalibardes sur son passage et imprimant piste circulaire de ses pas.\n\nJe le pliais \u00e0 des essais, auxquels il se r\u00e9duisait toujours avec scepticisme. Mais peut-\u00eatre aussi cette d\u00e9fiance n'\u00e9tait-elle pour lui que truc pour berner son alarme, et de la sorte n'apercevoir jamais les \u00eatres qu'il continuait cependant de redouter. J'applaudissais toutefois que p\u00e8re ne f\u00fbt somme toute que dans de si placides dispositions. Qu'e\u00fbt-ce \u00e9t\u00e9 si ses gens lui \u00e9taient venus en ces jours-l\u00e0 ? Car ma tourmenterie \u00e9tait que, par quelque guigne, p\u00e8re aper\u00e7oive du m\u00eame coup et ses gens et ceux de l'outre-monde. Qui sait quel aurait \u00e9t\u00e9 son mouvement ? Aurait-il d'entr\u00e9e subi acc\u00e8s ou attaque fatale de corps, et rejoint l'outre-peuple en m\u00eame temps qu'il y aurait tra\u00een\u00e9 ses gens ? Et les gens de p\u00e8re, une fois sortis d'ici-bas avec lui, comment se seraient-ils comport\u00e9s ? Se seraient-ils montr\u00e9s par suite aux vifs, ainsi que tout d\u00e9funt normal ? Je nourrissais pareilles tracasseries.\n\nMais ces m\u00e9ditations ne freinaient gu\u00e8re mon enseignement. Une matin\u00e9e, je parlai ainsi : \u00ab Les morts viennent lorsque nous n'y songeons point. Aussi n'y songerons-nous pas aujourd'hui. \u00bb L'ayant avis\u00e9 de cela, j'enjoignis p\u00e8re de commencer son labeur ordinaire et de ne se soucier en rien d'autre chose que de la coutume. Il m'exau\u00e7a, affichant soulagement sur sa face, r\u00e9joui de d\u00e9laisser enfin la mati\u00e8re des morts. Je le vis par suite s'\u00e9tablir sur le seuil et s'affairer \u00e0 la confection d'un attrape-loutre. D\u00e9j\u00e0 p\u00e8re \u00e9tait rebrouss\u00e9 en ses secr\u00e8tes contr\u00e9es de silence. Grippant la pelle, je le singeai aussit\u00f4t et commen\u00e7ai de chercher racines.\n\nUn temps coula, que nous ne v\u00eemes gu\u00e8re, tant le labeur nous confisqua \u00e0 nous-m\u00eames. Dans un trait de soleil, m\u00e8re parut soudainement, se frayant chemin \u00e0 travers branchottes d'arbouse et raisins d'ours. Laissant l\u00e0 ma besogne, je toisai m\u00e8re dans les yeux. J'y vis encore m\u00e9lancole, et fus pr\u00e8s de la questionner sur cette mati\u00e8re. Peut-\u00eatre aurait-elle pu alors, par gestes et signes, m'enseigner de ces choses-l\u00e0. Mais il me revint que j'avais plus imp\u00e9rieux \u00e0 faire : \u00ab P\u00e8re ! P\u00e8re ! soufflai-je. Lorgne un peu du c\u00f4t\u00e9 de ce boqueteau d'arbouses ! \u00bb P\u00e8re d\u00e9consid\u00e9ra sa rati\u00e8re, souleva le regard et embrassa le boqueteau, puis pareillement tout l'\u00e0 l'entour. \u00ab Quoi ? fit-il par suite. Aurais-tu renifl\u00e9 la course d'un raton ? \u00bb\n\nAinsi, malgr\u00e9 nos exercices encore frais, son \u0153il ne servait de rien ! Des jours durant j'avais pein\u00e9, tentant d'aiguiser ses facult\u00e9s d'observation, le for\u00e7ant \u00e0 pressentir autres sujets que gibier s'approchant, \u00e0 s'absorber d'autres mati\u00e8res que bourrasques et grains grossissant, eau-de-geni\u00e8vre gargouillant dans l'alambic, d\u00e9broussaillement de sente, ingestion de barbaque et de lard nouveau, \u00e9tripement de b\u00eates ou ab\u00eemement dans le roupil ! Des jours durant, pour l'\u00e9prouver, j'avais command\u00e9 \u00e0 p\u00e8re exercices, man\u0153uvres et devoirs, lui faisant ainsi le\u00e7on de l'art d'entrevoir l'outre-peuple. Autant de besogne inutilement men\u00e9e !\n\nCependant m\u00e8re s'avan\u00e7ait. Elle ne fut gu\u00e8re longue \u00e0 gagner l'abord de p\u00e8re. Je la vis s'\u00e9tablir sur la v\u00e9g\u00e9tation croissant par-devant notre seuil, et commencer avec ample qui\u00e9tude d'\u00e9tudier son compagnon d'anciennement. Mais malgr\u00e9 qu'il e\u00fbt m\u00e8re camp\u00e9e sous le blair, p\u00e8re n'en distinguait que paille et continuait d'escompter la venue sur nos terres de quelque rongeur. \u00ab Et alors ? clamait-il, exhibant \u00e0 bout de bras sa bricolade. Par o\u00f9 vient-elle donc cette b\u00eate, que je l'encoffre dans mon grippe-putois ? \u00bb Dominant avec peine mon bouillonnement, je criaillai quasiment, allongeant la main vers lui : \u00ab L\u00e0, l\u00e0, ci-devant toi, P\u00e8re ! N'aper\u00e7ois-tu pas m\u00e8re s\u00e9ante ? \u00bb Mais il ne percevait que vide, trou et d\u00e9sert. S'instruisant toutefois par moi de la pr\u00e9sence d'un outrepass\u00e9 dans son environ, il se dressa tout \u00e0 coup, empoign\u00e9 de panique. \u00ab Un macchab\u00e9e ? Un macchab\u00e9e ? O\u00f9 \u00e7a ? Racaillerie ! Grain d'orage ! Et tu pr\u00e9tends que cet atroce-l\u00e0 m'avoisine ? \u00bb Et le voil\u00e0 l\u00e2chant tout et commen\u00e7ant \u00e0 longer la cabane, l'\u00e9chine l\u00e9chant le mur. L'effroi, Monsieur le juge, se lisait sur sa face blanchie supr\u00eamement. Ses mains aussi exprimaient une alarme majeure par la trembloterie qui les agitait comme ailes de papillon. Je crois m\u00eamement que p\u00e8re pissa en culotte. Cependant, il clamait : \u00ab O\u00f9 est-il, Fils ? Parnoir ! O\u00f9 est le macchab\u00e9e dont tu m'entretiens l\u00e0 ? O\u00f9 est-il que je ne tr\u00e9buche sur lui quand je d\u00e9talerai dans un moment ? \u00bb Ainsi concevait-il les morts : comme obstacles ne servant qu'\u00e0 faire tr\u00e9bucher les vifs dans leur course.\n\nPuis il se transporta hors de regard aussi prestement que garenne traqu\u00e9.\n\n#\n\nM\u00e8re et moi rest\u00e2mes seuls. Je m'approchai paisiblement, et m'\u00e9tablis \u00e0 son abord sur l'herbe. J'\u00e9mis quelques plaintes \u00e0 propos de p\u00e8re. Je dis : \u00ab Malgr\u00e9 mes soins, p\u00e8re demeure toujours aussi effar\u00e9 ci-devant la mort. En ceci il est ind\u00e9crottable. C'est comme lui prescrire de ne plus avaler d'eau-de-geni\u00e8vre : aussi bien \u00e9dicter au hibou de cesser de circuler noctantement dans ces bois. Chose infaisable. \u00bb\n\nM\u00e8re posa long de temps son regard triste sur ma face apr\u00e8s que j'eus prononc\u00e9 ces mots. On e\u00fbt dit qu'elle aspirait \u00e0 parler \u00e0 son tour. Mais m\u00e8re \u00e9tait morte, et je savais fort bien, depuis le temps que je coudoyais les morts, qu'aucun d'entre eux ne devise jamais. Aussi poursuivis-je : \u00ab Et puis, p\u00e8re a beau toiser et retoiser, il n'aper\u00e7oit rien de rien du peuple de l'outre-monde, dont tu fais depuis jolie lurette si muettement partie. \u00bb Par suite, je restai un laps dans le silence, tout songeur et p\u00e9n\u00e9tr\u00e9 de mes tentatives des derniers temps aupr\u00e8s de p\u00e8re. M\u00e8re paraissait attendre la suite de mon discours. Je repris donc : \u00ab Voil\u00e0 mati\u00e8re bien \u00e9nigmatique. Car, enfin, malgr\u00e9 qu'il f\u00fbt arriv\u00e9 ici-bas muni de regards per\u00e7ants, de sens pointus comme bec d'aigle et d'entendement remarquable, je m'instruis aujourd'hui seulement combien p\u00e8re souffre d'autre part d'aveuglade et de sourdisme. Je con\u00e7ois ainsi que, bien que poss\u00e9dant toutes sciences, p\u00e8re fait figure de jouvenceau en mati\u00e8re de d\u00e9chiffrement de l'invisible. En cette mati\u00e8re, \u00e0 dire vrai, et par quelque prodige formidable, seul le d\u00e9chiffrement des astres, en lui-m\u00eame pourtant fort myst\u00e9rieux, lui para\u00eet ais\u00e9. P\u00e8re est un ab\u00eeme. Je ne puis le sonder. \u00bb\n\nJe vis alors l'\u0153il de m\u00e8re creuser le mien plus intens\u00e9ment, puis sa main venir se poser finement sur ma face. Je scellai mes paupi\u00e8res. Des visions et des pensers me vinrent. J'aper\u00e7us en mon casque, se jouxtant, l'image de p\u00e8re, puis la mienne. Je vis ainsi p\u00e8re dot\u00e9 de ses nombreux talents de chasseur, de preneur de poisson, d'allumeur de feu et de lecteur d'astres. Je le vis aussi d\u00e9membrer expertement quelque b\u00eate et se mettre \u00e0 l'\u00e9chauffement d'une escalopette ou d'un bif de putois. Je le toisai long de temps allant et venant lestement par sentes, puis ascensionnant h\u00eatres, ch\u00eanes et ormes pour \u00e9tudier le grain lointain. Et puis je le voyais s'\u00e9pouvantant \u00e0 l'id\u00e9e de l'outre-monde, et prier avec force la d\u00e9esse Lune de le retenir ici-bas encore un brin, de lui accorder sursis et force suffisante pour vivre encore quelque dur\u00e9e. Je le voyais prendre jambes au col lorsque le cyclone s'avan\u00e7ait. Enfin, je saisis l'image de p\u00e8re palabrant avec ses gens et effectuant pour eux les actes les plus bizarres, tels qu'ils ont \u00e9t\u00e9 narr\u00e9s par-devant vous, Monsieur le juge, et ci-devant votre face aussi, Membres du jury de ce tribuneau.\n\nJe voyais tout cela, pour ainsi dire, d'un seul c\u00f4t\u00e9 de mon casque. De l'autre, je m'apercevais moi-m\u00eame, tout contraire : adroit, quoique modestement dou\u00e9 tant pour chasse qu'emparement de poisson, certes sachant lire points cardinaux et \u00e9coulement de saisons, mais analphabique complet quant au livre c\u00e9leste et ignorant tout des coutumes du vent. Je m'instruisais \u00e9galement que j'\u00e9tais peu habit\u00e9 par l'entendement ou par d'autres formes de lumi\u00e8res. Mais en compensation, le dessin int\u00e9rieur que je me faisais de ma personne \u00e9tait celui d'un bourgeois mis\u00e9ricordieux, d\u00e9bordant d'une p\u00e9tulante sensiblet\u00e9 pour l'humanit\u00e9, observateur habitu\u00e9 de l'invisible, familier de la mort et ne redoutant en rien les gens de l'outre-vie. Me rappelant la joliesse de Manon encontrant la disgr\u00e2ce de ma face, je pensais : \u00ab Par quelle diablerie une paire de natures si adverses que la mienne et celle de p\u00e8re peuvent-elles s'avoisiner de si pr\u00e8s ? \u00bb Farce de la race ! Myst\u00e8re du sang transmis !\n\nMais pourquoi m\u00e8re me donnait-elle \u00e0 voir pareils contrastes ? J'incline aujourd'hui \u00e0 ce penser : ce fut pour elle occasion de me peindre le monde en son enti\u00e8ret\u00e9, et d'en changer l'image fausse que j'en poss\u00e9dais. Car la Terre est ainsi faite : abritant sous son ciel et portant sur son plancher humanit\u00e9 innombrable, et b\u00eates disparates, et paysages divers, et toutes choses bigarr\u00e9es formant, une fois agroup\u00e9es, le domaine de la vie. Aussi m\u00e8re me faisait-elle entendre que j'\u00e9tais \u00e0 ma fa\u00e7on aussi pi\u00e8tre toiseur que p\u00e8re, et qu'il me faudrait encore m'aiguiser l'\u0153il. Car, enfin, ne faut-il pas poss\u00e9der claire pupille pour voir amour, ainsi que je l'ambitionnais ? M\u00e8re, \u00e0 l'\u00e9vidence, concevait cela et me le rappelait aimablement.\n\n\u00c0 la fin, n'\u00e9prouvant plus ses doigts sur mes paupi\u00e8res, j'ouvris le regard. M\u00e8re n'\u00e9tait plus s\u00e9ante. O\u00f9 avait-elle fui ? La for\u00eat ne bruissait pas, non plus que l'air. Une averse vint, qui baigna promptement la terre, puis cessa aussit\u00f4t. Le soleil per\u00e7a les cieux, ouvrant sa route \u00e0 l'alouette.\n\nLe monde pleuroyait et chantait tout \u00e0 la fois.\n\n#\n\nNombre de soleils coul\u00e8rent. Rien ne survenait plus des folies de p\u00e8re : ses gens semblaient avoir d\u00e9sert\u00e9 son casque. Je le trouvais quasiment chaque jour de fort bon poil, accomplissant nos t\u00e2ches le plus paisiblement du monde, tout p\u00e9n\u00e9tr\u00e9 de son silence coutumier, vaquant \u00e0 toutes choses de son geste rustre mais judicieux, ce qui me paraissait le signe assur\u00e9 de son retour \u00e0 la sant\u00e9. Rares \u00e9taient les fois o\u00f9 je devais endurer ses foudres. Et \u00e0 la v\u00e9rit\u00e9, \u00e0 part quelques broutils, pas une seule fois en ces jours-l\u00e0 ne fus-je astreint \u00e0 avaler la soupiasse aux fourmis, ou ne re\u00e7us-je le pied de p\u00e8re sur le train, ou, de m\u00eame mani\u00e8re, ne subis-je de lui quelque grand punissement que ce soit. Les jours \u00e0 la cabane prenaient une couleur guillerette. Je respirais. J'allais, \u00e9tourdi et superficiel.\n\nMais c'\u00e9tait l\u00e0 montrer un optimisme insoutenable et avouer une fois encore le peu de perspicacit\u00e9 s\u00e9journant en ma personne.\n\nLe jour vint en effet o\u00f9 p\u00e8re bascula de nouveau dans le trou de la d\u00e9mence. Je m'esquintais encore, ce matin-l\u00e0, \u00e0 lui faire apercevoir les morts. Je lui parlais avec d\u00e9tails des outrepass\u00e9s, et p\u00e8re, \u00e9tabli sur une souche, devait ouvrir grandes ses esgourdes et r\u00e9sister \u00e0 son envie de les bouchonner. Tel \u00e9tait l'exercice auquel je l'avais clou\u00e9. Certes, la sueur lui gouttait sur la tempe. Ses mains se tordaient et ses genoux, parfois, venaient toquer l'un sur l'autre \u00e0 la fa\u00e7on du pic sur l'arbre. Sa denture s'agitait semblablement. Quand il \u00e9tait d'avis que mon discours outrepassait les bornes, p\u00e8re quittait la souche et, d'un bond de garenne, se mettait sur pied puis venait me mettre une mornifle. Mais ce n'\u00e9tait que raret\u00e9 : le plus souvent il tenait bon.\n\nAdvint un \u00e9v\u00e9nement. Tandis que je discourais par-devant p\u00e8re, nos esgourdes furent emplies d'un bruit singulier. C'\u00e9tait une criaillerie s\u00e8che et br\u00e8ve, assur\u00e9ment sortie du gosier d'un bourgeois, serr\u00e9e de pr\u00e8s par le bruit d'un cordage soudainement tendu. P\u00e8re et moi laiss\u00e2mes s\u00e9ance tenante notre exercice et commen\u00e7\u00e2mes \u00e0 ruer nos personnes vers le lieu d'o\u00f9 avait \u00e9man\u00e9 le bruit, avoisinant du grand pin \u00e0 peu pr\u00e8s.\n\nNous y arriv\u00e2mes bient\u00f4t, soufflants et curieux comme loutres. Ce que nous v\u00eemes nous laissa tout abasourds : accroch\u00e9 \u00e0 une branche par le col, un bourgeois oscillait sous le grand pin telle une pendeloque, sans vie, sa langue d\u00e9gain\u00e9e de bouche en une simagr\u00e9e d\u00e9testable. Surmontant sa surprise, p\u00e8re s'en fut d\u00e9crocher le mort. D'un coup de coutelas, la corde fut bris\u00e9e et le bourgeois chut, s'effondrant, mort et m\u00eame surmort, sur un lit de russules en une attitude ridicule. Ce qui, un laps plus t\u00f4t, \u00e9tait encore un quidam vif et habit\u00e9 n'\u00e9tait plus \u00e0 pr\u00e9sent que sac creux.\n\nC'est \u00e0 ce moment fort mal choisi que les gens de p\u00e8re rappliqu\u00e8rent. Je le con\u00e7us \u00e0 sa conduite m\u00eame : c'\u00e9tait comme s'il cherchait tout \u00e0 coup \u00e0 chasser de son environ quelque mosquite importune. Ses mains battaient l'air \u00e0 l'entour des esgourdes puis s'abattaient \u00e0 plat, subitement et sans annonce, sur sa face \u00e9tonn\u00e9e. Pareillement, son \u00e9chine se courbait, puis de sit\u00f4t se redressait, et ses jambes aussi tressautaient en un fr\u00e9n\u00e9tique rigodon, paraissant vouloir \u00e9viter la bastonnade d'un ennemi immat\u00e9riel. Enfin, p\u00e8re avait quitt\u00e9 son silence et commenc\u00e9 \u00e0 distribuer injures et pouilles \u00e0 ses adversaires cependant introuvables. Nul doute, ses gens rebroussaient.\n\nAussi, une fois le premier laps coul\u00e9 et son \u00e9querre retrouv\u00e9e, p\u00e8re manda : \u00ab Fils, empoigne ce d\u00e9funt, et talonne-moi jusqu'\u00e0 la cabane ! \u00bb Redoutant le pire, je ne dis mot, installai le pendu sur mon \u00e9chine, ses bras formant foulard autour de mon col, et commen\u00e7ai \u00e0 serrer p\u00e8re au plus pr\u00e8s sur la sente menant chez nous.\n\nSit\u00f4t parvenu l\u00e0, que vois-je ? P\u00e8re fouillant sous sa paillasse \u00e0 la recherche de quelque chose. De cet explorement r\u00e9sulte la trouvaille du vieux v\u00eatement de m\u00e8re. Et voici la voix de p\u00e8re : \u00ab Accoutre le pendu de cette d\u00e9froque ! \u00bb Je me roidis tel un gourdin. \u00ab Quoi ? clame-je. Tu voudrais v\u00eatir un d\u00e9funt de l'habit de m\u00e8re ? Sacril\u00e8ge ! H\u00e9r\u00e9sie ! Blasph\u00e9mation ! Aussi bien cracher \u00e0 la face de la d\u00e9esse Lune. M\u00e8re nous maudira de commettre pareil outrage et, de sa tombe, prescrira de suite notre m\u00e9tamorphose en crapauds ! \u00bb P\u00e8re ne prisa pas ce discours. Me grippant la clavicule et m'enfon\u00e7ant ses doigts griffus et puissants entre l'os et la pulpe, il me m\u00e8ne dehors par-devant la barrique et m'y submerge le casque tout entier. Je reste ainsi, maintenu en poste par lui, disputant ma vie avec bougui-wougui et danses de Saint-Guy, jusqu'\u00e0 manquer d'\u00e9touffer maximalement. Puis p\u00e8re me ram\u00e8ne \u00e0 air, et me re-fourre aussit\u00f4t sous flotte. Je suspends tout gigotement, r\u00e9sign\u00e9 \u00e0 pr\u00e9sent \u00e0 crever l\u00e0. Lorsqu'il me sait \u00e0 demi p\u00e9ri, p\u00e8re me sort de l\u00e0 enfin et me reconduit, semblablement gripp\u00e9 par l'os, jusqu'au taboureau. J'entends p\u00e8re seriner imp\u00e9rieusement, avec du mauvais dans la voix : \u00ab Parnoir ! Accoutre le pendu, Fils ! \u00bb\n\nReconqu\u00e9rant mollement mon ressort, je commen\u00e7ai donc \u00e0 costumer le mort. Tandis que je m'affairais \u00e0 cette t\u00e2che, p\u00e8re \u00e9tait sorti filer et nouer quelques brins de lycopode, entrem\u00eal\u00e9s de foug\u00e8re-\u00e0-l'autruche et de sceau-de-salomon. Quand il rebroussa, j'avais conclu ma besogne. P\u00e8re posa de suite par-dessus la chevelure du pendu son assemblage de lycopodes, ce qui singea avec effrayante ressemblance la coiffure de m\u00e8re. Il finit l'ouvrage en parant les poignets et le col du mort de menus bijoux ayant autrefois d\u00e9cor\u00e9 m\u00e8re. Je p\u00e2tis d\u00e8s lors d'un malaise incalculable : ce mort-l\u00e0 prenait d'un coup l'aspect de m\u00e8re, du moins en ses g\u00e9n\u00e9ralit\u00e9s. C'\u00e9tait comme si, fid\u00e8le \u00e0 sa coutume, elle paraissait ci-devant moi en provenance de l'outre-monde. Hormis le regard de poisson r\u00f4ti, le pendu adoptait en presque tous points les traits que je ch\u00e9rissais tant chez mon macchab\u00e9e favori.\n\nMais \u00e0 peine m'instruisais-je de cette ressemblance que p\u00e8re interrompait mon penser. \u00ab Place donc, dit-il, le taboureau ici, puis \u00e9tablis le mort dessus, train sur si\u00e8ge et \u00e9chine contre mur. Par suite, dispose ses membres afin de mettre tout ce macchab\u00e9e en posture d'avaleur de pitance, singeant l'attitude d'un bourgeois affair\u00e9 \u00e0 prendre repas. Enfin, mitonne pour nous tous quelque mangeaille ! \u00bb Je murmurai : \u00ab Pour _nous tous_ , P\u00e8re ? \u00bb Son r\u00e9pons, r\u00e9union du grain et du tonnerre, fut tranchant comme la fl\u00e8che d\u00e9chirant l'air : \u00ab Oui ! \u00bb Je m'affairai donc ensuite \u00e0 la cuisson d'une rago\u00fbte de marasmes li\u00e9e de jus d'if, relev\u00e9e de menues racines et de ca\u00efeux. \u00ab Et mets-y beaucoup de lard nouveau, parbleu ! \u00bb me criaille p\u00e8re, tandis que je suis pench\u00e9 sur la marmite.\n\nPuis vient le moment o\u00f9 c'est cuit. Trembleur, m'effrayant du dessein de p\u00e8re, j'emplis les \u00e9cuelles, que je place \u00e0 table ci-devant chaque avaleur. Mais aussit\u00f4t que je m'\u00e9tablis sur taboureau et commence d'engouffrer ma pitance, p\u00e8re cogne sur la table de son poing lourd comme pierre des champs. \u00ab Par cieux, Fils ! rugit-il. Qui te mande de manger cette bouffetance ? Reste plut\u00f4t stationn\u00e9 et en recueillement par-dessus elle ! Ne bronche ni ne cille ! Fais pri\u00e8re, voil\u00e0 tout ce que je prescris ! \u00bb Je mis donc ma cuill\u00e8re par c\u00f4t\u00e9 de l'\u00e9cuelle et commen\u00e7ai \u00e0 me recueillir. Je crus d'abord qu'il s'agissait d'\u00e9voquer l'image de m\u00e8re et de ramener ainsi \u00e0 notre m\u00e9moire sa douce face, encore que je m'\u00e9tonnais qu'une telle pratique d\u00fbt se faire en pr\u00e9sence d'un immigr\u00e9, tr\u00e9pass\u00e9 de surcro\u00eet. Mais, lorsque p\u00e8re \u00e9tait visit\u00e9 par ses gens, j'avais appris \u00e0 ne pas le questionner sur ses actes. Aussi baissai-je bient\u00f4t paupi\u00e8res pour m'enfoncer plus ais\u00e9ment en \u00e9tat de pi\u00e9t\u00e9. Mais voici que p\u00e8re cogne de nouveau avec force sur la table, provoquant saut de crapaud chez nos \u00e9cuelles et cuill\u00e8res. \u00ab Parjaune ! J'ai dit : ne cille pas, Fils ! \u00bb Telles furent ses paroles fulminantes, auxquelles j'objectai faiblottement : \u00ab Mais, P\u00e8re, j'\u00e9tais pourtant entr\u00e9 en pi\u00e9t\u00e9 et ne mouvais quoi que ce soit sur ma personne. O\u00f9 as-tu vu que quelque portion de moi-m\u00eame a boug\u00e9 ? \u00bb P\u00e8re pr\u00e9cisa son penser en ces mots : \u00ab J'ai dit : ne cille pas ! Ton \u0153il ne doit pas se clore, tant que je ne l'ordonnerai ! \u00bb\n\nDe ma vie, j'avais cru conna\u00eetre les plus rudes malheurs. Mais ce nouveau mandement rivalisait avec tout ce que j'avais travers\u00e9 jusque-l\u00e0. Pardonnez mon toupet, Monsieur le juge, mais je tiens pour certain qu'il vous est impossible de concevoir l'extr\u00eame torture de ne pouvoir bouger le cil pendant des dur\u00e9es et des dur\u00e9es. Pour moi, roussir \u00e0 maigre feu sur b\u00fbcher e\u00fbt \u00e9t\u00e9 moins ardu. Le plus \u00e9trange est que p\u00e8re, lui, arrivait sans peine \u00e0 pareil prodige. Quant au pendu, on concevra qu'il lui \u00e9tait ais\u00e9 de ne broncher cil, non plus que rien d'autre de sa personne. J'\u00e9tais ainsi le seul de cette sapr\u00e9e tabl\u00e9e \u00e0 souffrir picotements horribles, ass\u00e8chement d'yeux et crises de machineries tandis que dura l'exercice ordonn\u00e9 par p\u00e8re. Combien de temps ce supplice persista-t-il ? Je ne saurais l'attester. Il me semble retrouver ressouvenir d'un parcours complet de soleil en azur pendant que nous restions l\u00e0, l'\u0153il b\u00e9ant, stationn\u00e9s et r\u00e9citant int\u00e9rieurement pri\u00e8res d'usage.\n\nAu bout d'une p\u00e9riode l'exercice se finit. P\u00e8re lui-m\u00eame sortit de l'esp\u00e8ce de transe o\u00f9 il s'\u00e9tait fourr\u00e9. Lorsque je con\u00e7us que la s\u00e9ance \u00e9tait termin\u00e9e, je me mis \u00e0 ciller d\u00e9mesur\u00e9ment. \u00c0 chaque battement de paupi\u00e8res correspondait une br\u00fblure sadique. On aurait cru que tisons me briquaient pupilles. Pour couper le poivre de cet incendie, j'enfouis bient\u00f4t ma face, yeux grands ouverts, en mon \u00e9cuelle de rago\u00fbte. Le lait refroidi des marasmes fut comme pansement sur la plaie vive de mon \u0153il. Je restai ainsi post\u00e9 long de temps, tandis que j'entendais p\u00e8re vaquer autour de moi, retirant au mort sa perruque de lycopodes, le d\u00e9parant de ses bijoux, le d\u00e9saccoutrant.\n\nQuand je levai enfin la t\u00eate et remis mes yeux en face des choses, j'aper\u00e7us que p\u00e8re empoignait le mort et l'\u00e9tablissait sur son \u00e9chine pour le mener dehors. \u00ab Que ferons-nous de ce d\u00e9funt, P\u00e8re ? \u00bb dis-je, la face encore enduite de lait de champigne et de jus d'if. Mais pour tout r\u00e9pons, p\u00e8re dit : \u00ab Viens allumer du feu, Fils ! \u00bb Je le suivis donc jusque sous le grand h\u00eatre et fis s'enflammer nombre de nos planches. Quand le brasier fut \u00e0 son comble, p\u00e8re ne fit ni pompes ni appr\u00eats, et jeta en ce foyer le corps nu du pendu.\n\n#\n\nQuand il ne resta plus du mort que cendres, p\u00e8re et moi nous en retourn\u00e2mes \u00e0 notre ordinaire. Vers le d\u00e9clin du jour, tandis que nous \u00e9tions \u00e0 prendre p\u00e2ture en l'\u00e9tang, je constatai que les gens de p\u00e8re avaient pris cong\u00e9 de lui. En effet, renfrognement avait reconquis pour de bon ses chairs. Aussi, apr\u00e8s que j'eus clou\u00e9 d'un coup de pique un poisson-chat, je risquai cette question : \u00ab Mais, P\u00e8re, \u00e0 la fin : pourquoi, pourquoi ces pri\u00e8res, ce p\u00e9nible stationnement de cils, cette invitation \u00e0 table d'un pendu accoutr\u00e9 ? Pourquoi ? Pourquoi ? \u00bb\n\nEn cieux, le soleil passait sa course mollement, et s'en allait d\u00e9c\u00e9der derri\u00e8re la for\u00eat. Une pie vint se poser sur une quenouille et nous lorgna. Dans les herbes un raton couina, puis d\u00e9tala.\n\nEnfin, l'\u0153il fix\u00e9 sur l'onde, le bras lev\u00e9 et arm\u00e9 de sa pique, attendant roide comme tronc le passage d'un poisson, p\u00e8re fit ce r\u00e9torque : \u00ab Ta m\u00e8re \u00e9tait d\u00e9vote. Souventes fois, \u00e0 table, elle offrait pitance \u00e0 la d\u00e9esse Lune puis entrait en recueillement, fixant long de temps le vide sans ciller jamais, pareille \u00e0 un outrepass\u00e9. On aurait dit qu'elle entrait alors en \u00e9tat d'extase. C'\u00e9tait une cr\u00e9ature de forte religion. Gr\u00e2ce lui fut rendue aujourd'hui. \u00bb\n\nJ'en fus, Monsieur le juge, tout \u00e9mu, tout vaincu. J'enfouis par suite ma prise dans ma besace et me r\u00e9tablis en posture. D'autres poissons nag\u00e8rent sous l'arceau de mes jambes. Distrait de mon ouvrage, je n'en piquai pourtant plus aucun, m\u00e9ditant plut\u00f4t les r\u00e9centes paroles de p\u00e8re. Une fois sorti de ce r\u00eavement, je rentrai \u00e0 la cabane, suivi de p\u00e8re. J'allumai un feu, fabriquai quelques braises et y jetai le poisson-chat. Quand il fut mangeable, p\u00e8re et moi l'aval\u00e2mes, puis nous raviv\u00e2mes la flamme et rest\u00e2mes encore tout le soir et jusqu'\u00e0 l'aube suivante, traversant la nuit sans discours et presque sans geste. Au d\u00e9tour d'un penser, quittant provisoirement ma songerie, je rebroussais au r\u00e9el et guettais avec trac notre m\u00e9tamorphose en crapauds.\n\nRien ne vint toutefois.\n\n#\n\nMalgr\u00e9 le p\u00e9ril des foudres de p\u00e8re, je rebroussais toujours ascensionner le grand orme. Mais le spectacle qui s'offrait \u00e0 l'\u0153il du haut de mon perchement ne m'exau\u00e7ait gu\u00e8re. Un jour, j'outrepassai donc les limites de la for\u00eat et entrai en village avec l'intention d'y aborder Manon. On ne tarda pas \u00e0 me reconna\u00eetre, comme chaque fois que j'avais pilot\u00e9 ma personne par les rues du village. J'entendais \u00e0 ma suite ces mots, sortis comme bruissements des bouches : \u00ab C'est le fils Courge ! C'est le fils Courge ! \u00bb \u00c9tais-je donc \u00e0 ce point un \u00e9v\u00e9nement ? Cependant, j'allais, chaloupant et l\u00e9ger.\n\nJamais ne fus-je moins parleur que ce jour-l\u00e0. J'escoutais, esgourdes b\u00e9antes, chercheur du moindre pr\u00e9sage du trouvement de Manon, l'humanit\u00e9 qui passait l\u00e0. Au bout d'une p\u00e9riode d'errements, lass\u00e9 de ne rien traduire ni d\u00e9goter, je r\u00e9solus d'engager le discours. J'abordai un bourgeois. J'adoptai la fa\u00e7on du tra\u00eenard puis, singeant l'indiff\u00e9rence, je le questionnai : \u00ab Dites : je fur\u00e8te depuis jolie lurette afin de trouver la cr\u00e9ature Manon. Quel est son s\u00e9jour, que je course y toquer ? \u00bb\n\nLe r\u00e9pons que je r\u00e9coltai m'enfarina d'\u00e9bahissement.\n\n\u00ab Manon ? fit avec surprise le bourgeois. Manon Grandjour ? Mais elle est morte, mon pauvre vieux ! Voil\u00e0 des mois et des mois ! C'est Henri Ronce et sa femme qui l'on trouv\u00e9e, un soir, dans leur pr\u00e9, le cou bris\u00e9, un trou gros comme \u00e7a entre les omoplates. Le taureau l'avait encorn\u00e9e ! Une b\u00eate \u00e9norme, forte comme vingt hommes ! Pauvre Manon ! Ah, \u00e7a oui : pauvre, pauvre Manon ! \u00bb\n\nAvez-vous jamais exp\u00e9riment\u00e9, Monsieur le juge, qu'on vous tranche les jambes ? C'est ce que je go\u00fbtai int\u00e9rieurement, ce jour-l\u00e0, dans la grande rue du village, au sortir de ce narrement cuisant. Au dernier mot l\u00e2ch\u00e9 par le bourgeois, je commen\u00e7ai mollement de rebrousser \u00e0 la for\u00eat. J'allais, hagard\u00e9, tel un engourdi. \u00ab Manon est morte ! Manon est morte ! \u00bb serinais-je comme pour moi-m\u00eame, quoique n'\u00e9tant plus bien assur\u00e9 d'\u00eatre encore habitant de ma personne. J'avan\u00e7ais parmi l'humanit\u00e9 buissonnante des rues. Comment retrouvai-je la sente menant \u00e0 la cabane ? Myst\u00e8re ! Je ne saurais dire non plus, encore \u00e0 pr\u00e9sent, en quels endroits mes pas me pilot\u00e8rent durant le long parcoursement qui s'ensuivit, et quelles pistes je pris. Quelle force obscure s\u00e9journait donc alors en moi ? Sans doute quelque entendement prodigieux am\u00e9nag\u00e9 en mon pied et traduisant mieux que moi-m\u00eame les lois de l'orientation. Qu'y avait-il en mon casque ? Rien d'humain, ce me semble : un gouffre, o\u00f9 mon sens et ma clart\u00e9 venaient se suicider. Que tenais-je en mon c\u0153ur ? Un brisement de vitre.\n\nJe dus en tout cas errer en for\u00eat longuement, car je ne parvins \u00e0 notre seuil qu'en fin de jour. P\u00e8re m'attendait et, traduisant l'objet de ma fugue, me re\u00e7ut comme en coutume avec savates et claques au casque. J'encaissai cette correction d'une bien \u00e9trange mani\u00e8re : comme en r\u00eavement. Mais, aussi, sans doute n'habitais-je pas v\u00e9ritablement ma personne \u00e0 ce moment-l\u00e0. On e\u00fbt dit que je logeais plut\u00f4t en celle de Manon. Tandis que je paraissais par-devant p\u00e8re et que j'accusais r\u00e9ception de sa d\u00e9gel\u00e9e, je vis en penser quelque image d'elle, je souffris sur ma propre \u00e9chine la charge du taureau, je p\u00e2tis en ma carcasse du heurt de la b\u00eate furieuse, je sentis mes viandes s'ouvrir et mon sang commencer \u00e0 me quitter. Et ce n'\u00e9tait plus dans la cabane que j'\u00e9tais, mais dans le pr\u00e9 de monsieur Ronce, \u00e0 p\u00e9rir sous un pi\u00e9tinement de sabots, \u00e0 sentir sur mon col le souffle enrag\u00e9 d'une b\u00eate muscleuse et encol\u00e9r\u00e9e.\n\nOui, sans doute, sans doute quittai-je un moment ce jour-l\u00e0 le logement de mes propres chairs. Car, en v\u00e9rit\u00e9, j'ai peu de ressouvenir des blessures que m'infligea p\u00e8re. On objectera \u00e0 mes paroles que, par ph\u00e9nom\u00e8ne ou diablerie, je confondais alors une paire de douleurs : celle de la perte de Manon et celle de la bross\u00e9e de p\u00e8re. Pourtant, je puis l'attester sans tortiller : pour la premi\u00e8re fois, les coups de p\u00e8re ne m'atteignaient pas, ou brin. Ce n'\u00e9tait que poux de mouches et autres riens. Mais terrible \u00e9tait le souffrir que je go\u00fbtais de concevoir que jamais plus je n'effleurerais la main de Manon, ni ne serais d\u00e9terg\u00e9 par elle en barrique.\n\nSes coups distribu\u00e9s, p\u00e8re me laissa \u00e0 moisir sur les foug\u00e8res. J'y fus comme mort, n'entretenant plus avec l'existence que ch\u00e9tives relations. Le soleil parcourut en firmament nombre de trajets avant que je recouvre mon sens. Mais je repris assiette finalement, et recommen\u00e7ai peu \u00e0 peu ma coutume. Seulement, toujours et plus encore qu'avant je questionnais cieux et Terre, et pierrailles et v\u00e9g\u00e9tations, et b\u00eates aussi : o\u00f9, mais o\u00f9 donc se terrait amour ?\n\n#\n\nEt toujours des saisons paraissaient, s'\u00e9tablissaient puis repliaient, abandonnant \u00e0 la for\u00eat leurs pluies, leurs b\u00eates nouvelles, leurs soci\u00e9t\u00e9s d'oiseaux, leurs brigades de tani\u00e8res, leurs branches engross\u00e9es. Par printemps, l'air s'\u00e9chauffait et gonflait de s\u00e8ve arbres et boqueteaux. En arri\u00e8re-saison, les cieux ornaient le monde du rideau souple des averses. Ramures saignaient puis l\u00e2chaient leur cargaison de feuilles comme pages d\u00e9chir\u00e9es. Bourrasques s'en emparaient, et c'\u00e9tait tout le r\u00e9cit de l'\u00e9t\u00e9 qui s'envolait. Venaient ensuite neigettes, d\u00e9posant couvercle sur l'\u00e9tang et capiton d'ouate sur toutes choses. En leurs trous, ratons, putois, belets, marmottes et ours entamaient ample roupil, et patientaient sous chairs ensiest\u00e9es que rebroussent herbettes. La for\u00eat elle-m\u00eame stoppait sa vie en attendant que lombrics, faufil\u00e9s en leurs couloirs, recommencent \u00e0 manger la terre.\n\nEt, en effet, la terre un jour se d\u00e9figeait, et oisillons s'envolaient des nids, et cieux nouveaux se recomposaient.\n\n#\n\nVint une \u00e9poque o\u00f9 les nuages se tarirent, ne transportant plus en leur \u00e9tendue que bises archis\u00e8ches, poussiers et volatiles amaigri\u00e9s. La barrique b\u00e9ait, ne contenant plus \u00e0 pr\u00e9sent que filetteries d'araign\u00e9es. Et m\u00eamement, nous avions beau fouiller de notre pelle le sol de la for\u00eat, rien n'y sourdait jamais que quelques lombrics an\u00e9miques. L'\u00e9tang lui-m\u00eame, d'ordinaire si ourl\u00e9 de clapotis, ne rendait plus gu\u00e8re qu'un bruit boueux sous le pas des b\u00eates et offrait au jour ses limons \u00e0 s\u00e9cher. Partout, ce n'\u00e9tait que sables et poudres. Mais d'humidit\u00e9, point non. Bref, nous manqu\u00e2mes d'eau.\n\nP\u00e8re poss\u00e9dait la science des sourciers et savait, par simple maniement de sa badine de bois, faire avouer \u00e0 la terre o\u00f9 se dissimulaient ses liquides les plus aprofonds. Je le vis ainsi un matin tailler le sumac puis, \u00e0 la fin, tenir en ses mains la branchotte r\u00eaveuse d'eau. Du levant jusqu'au milieu du jour, p\u00e8re pilota doctement sa personne de buissons en fourr\u00e9s, sa badine point\u00e9e vers le sol, \u00e0 distance d'une genoul\u00e9e de la terre. Tel un aveugle, il \u00e9tudiait ainsi, non pas de son \u0153il pour cela inutile, mais du bout de ses doigts, le domaine le plus cach\u00e9 de la for\u00eat, le plus enfoui, sa subconscience pour ainsi dire. Et bien qu'il men\u00e2t son pas \u00e0 l'heure pleine du jour, p\u00e8re allait tel un t\u00e2tonneur, cherchant comme en nuit \u00e9paisse.\n\nJe m'avisai alors pour la premi\u00e8re fois que c'\u00e9tait \u00e0 peu pr\u00e8s la seule aprofondeur \u00e0 laquelle il \u00e9tait capable de se mesurer et pour laquelle il montrait quelque comp\u00e9tence. Obscurit\u00e9 de l'industrie humaine ! Pour emplir la barrique, pour assurer nos bouillades, pour breuver nos chairs, p\u00e8re \u00e9tait capable de d\u00e9goter toutes pistes souterraines. Mais pour creuser ses semblables, pour en embrasser le prix, en peser l'essence et en visiter le c\u0153ur, p\u00e8re \u00e9tait ainsi que marmotte aux heures de l'arri\u00e8re-saison : p\u00e9tri d'engourds, l'\u0153il clos et le blair \u00e9teint, ne r\u00eavant de rien. Car en mati\u00e8re d'humanit\u00e9 p\u00e8re n'\u00e9tait pas chercheur, et d\u00e9viait volontiers son penser de tout ce qui r\u00e9clamait usage extraordinaire de casque. Ce n'\u00e9tait pas qu'il f\u00fbt d\u00e9pourvu d'\u00e9clairage. Seulement, il allait, semblable au poisson-chat : man\u0153uvrant juste sous la surface des choses, comme si plonger en leur fondement risquait de l'engluer de limon. Son apparence elle-m\u00eame \u00e9voquait quelque b\u00eate sous-marine : quoique charnu et r\u00e2bl\u00e9, p\u00e8re avait quelque chose de glissant et d'ondulant, une inexplicablet\u00e9 qui lui permettait d'aller parmi les choses sans se heurter \u00e0 elles, les \u00e9vitant plut\u00f4t, les contournant, s'y d\u00e9robant. Oui, si par malheur quelque obstacle venait pourrir son trajet et s'objecter \u00e0 son pas, toujours il congravitait autour et \u00e9lisait pr\u00e9f\u00e9rablement quelque d\u00e9robade. Pourtant, nul n'allait sur sente, ni ne gravissait monts, ni ne circulait parmi les arbres plus adroitement que lui, m\u00eamement au milieu de la nuit, en son heure la plus \u00e9paisse. C'est que, pour piloter adroitement sa personne, p\u00e8re s'en remettait enti\u00e8rement \u00e0 son pied, qui en toutes circonstances lui servait pour ainsi dire d'\u0153il. Et sa main, son blair, ses esgourdes unanimement lui \u00e9taient alors une mani\u00e8re de guide. Voil\u00e0 pour son guidement au sein des lieux. Toute diff\u00e9rente \u00e9tait sa fa\u00e7on d'aborder les ph\u00e9nom\u00e8nes. Et, ainsi que je l'ai dit, c'\u00e9tait pire encore quand il coudoyait ma personne, ou toute humanit\u00e9. Car pour lui, ses semblables \u00e9taient incirculable g\u00e9ographie, et s'il se hasardait jamais en leur paysagement int\u00e9rieur, il s'y d\u00e9routait \u00e0 coup s\u00fbr lamentablement. Aussi inclinait-il plut\u00f4t \u00e0 ne jamais s'aventurer en ces \u00e9paisseurs-l\u00e0. Peut-\u00eatre tremblait-il d'y encontrer cyclones et paysages d'outre-tombe. Qui le saura ?\n\nJ'ai encore ce qui suit pour exemple de son peu d'enclin \u00e0 creuser les choses. Un soir que nous finissions d'avaler un cuit de raton, p\u00e8re s'allongea aupr\u00e8s du feu et commen\u00e7a \u00e0 mirer les astres. Il demeura ainsi long de temps, lorgnant songeusement la vo\u00fbte. Qu'y avait-il en ce corps stationn\u00e9, en ce casque d'ordinaire si peu chercheur, mais soudainement \u00e9trangement pensif ? Une tige d'herbe \u00e0 la lippe, je t\u00e2tai ma chance et le questionnai : \u00ab P\u00e8re, te voil\u00e0 bien enferm\u00e9 en toi-m\u00eame. Que vois-tu de la sorte sur firmament qui te rende \u00e0 ce point rem\u00e2chant ? La cl\u00f4ture de ta vie ? Est-ce bien cela que tu d\u00e9chiffres en cieux et qui te trouble au point de te laisser tout paralytique ? \u00bb Je comptais secr\u00e8tement qu'il d\u00e9voile son fond et qu'il fasse mentir l'opinion de lui que j'avais forg\u00e9e jusque-l\u00e0, \u00e0 savoir qu'il n'\u00e9tait que poisson glissant sur toutes choses. J'ambitionnais de l'entendre exprimer, pour une fois, le fruit d'une m\u00e9ditation v\u00e9ritable, rendant un \u00e9cho diff\u00e9rent que celui de notre barrique lorsqu'elle est \u00e0 sec et qu'un chipmonque y chute. Son r\u00e9pons d\u00e9\u00e7ut mon attente. Ne d\u00e9laissant point de ses yeux la vo\u00fbte, mais se grippant alors d'une main le ventre : \u00ab Parbleu ! Ce cuit de raton m'incendie les tripes ! Es-tu assur\u00e9, Fils, que tu y mis juste dose de poivre de bazzanie ? \u00bb\n\nAinsi parla-t-il.\n\nJe m'en instruisis ce m\u00eame soir : p\u00e8re soulageait ses aigreurs digestives de pareille mani\u00e8re qu'il \u00e9vitait usage extraordinaire de casque : en regardant ailleurs et en songeant \u00e0 autre chose.\n\nLes membres de ce tribuneau et vous-m\u00eame, Monsieur le juge, objecterez peut-\u00eatre que p\u00e8re, pourtant, \u00e9tait capable de singuli\u00e8res et v\u00e9ritables m\u00e9ditations lorsqu'il terminait les folles missions de ses gens. Aussi je dirai : en ces moments-l\u00e0, j'incline \u00e0 croire que ce n'\u00e9tait gu\u00e8re p\u00e8re lui-m\u00eame qui discourait, mais plut\u00f4t ses gens, toujours tapis non loin, qui lui soufflaient les mots. Quant aux observations, fort valables aussi, que lui inspirait parfois sa lecture des astres, j'aurai bient\u00f4t occasion de vous en entretenir plus \u00e9clairamment.\n\nIl m'arrive parfois, Monsieur le juge, de m'interviewer sur ces mati\u00e8res et de risquer ceci : p\u00e8re compensait-il son incapacit\u00e9 \u00e0 creuser autrui par son extr\u00eame talent \u00e0 d\u00e9pecer les b\u00eates ? Son \u00e9tonnante inhabilet\u00e9 \u00e0 commercer avec bourgeois et cr\u00e9atures de tous horizons ne se serait-elle pas mu\u00e9e, par quelque tour \u00e9trange, en un extraordinaire don pour le videment de tripes, l'\u00e9carnage et le d\u00e9collement du lard nouveau ? En la for\u00eat, il ne se passait en effet gu\u00e8re de jour sans qu'une b\u00eate go\u00fbte du tranchant de son coutelas. Et n'est-ce pas l\u00e0, au fond, la forme de creusement la plus directement et ais\u00e9ment r\u00e9alisable ? N'est-ce pas, au fond, une mani\u00e8re comme une autre de conna\u00eetre son vis-\u00e0-vis, d'en explorer la vie int\u00e9rieure, de le toucher le plus r\u00e9ellement et de s'en ravitailler, de se repa\u00eetre de lui, comme on se repa\u00eet parfois du commerce humain ?\n\n\u00c0 ma fa\u00e7on, peut-\u00eatre ai-je par ailleurs h\u00e9rit\u00e9 de p\u00e8re cet app\u00e9tit de fouir les choses, de les retourner, d'en diss\u00e9quer et s\u00e9parer les parties. Cela non pas pour en faire rosbives ou lardons, mais afin plut\u00f4t d'en trouver le sens et la direction, d'en d\u00e9tacher une autre sorte de chair, puis de m'en nourrir. Et peut-\u00eatre bien me suis-je nourri de cette chair-l\u00e0, la chair du sentiment humain, pour reculer moi aussi la venue de ma mort. Car m\u00eame si je ne crains gu\u00e8re le moment de mon simple tr\u00e9pas, je tremble comme p\u00e2querette sous brise au seul penser de quitter un jour l'ici-bas et d'y avoir v\u00e9cu en vain. Et quoi de plus vain, Monsieur le juge, qu'une existence de bourgeois ou de cr\u00e9ature sans ch\u00e9rissement, c'est-\u00e0-dire sans ouverture menant au c\u0153ur ? C'est l\u00e0, en ma carri\u00e8re humaine, l'objet de ma plus tendre peur.\n\nJ'observais donc p\u00e8re circuler \u00e0 l'entour muni de sa badine. Survint alors un \u00e9v\u00e9nement li\u00e9 \u00e0 ce que je viens d'exposer, et dont je garde encore \u00e0 pr\u00e9sent un ressouvenir fort troublant. Un macchab\u00e9e parut, qui vint s'\u00e9tablir \u00e0 mon c\u00f4t\u00e9 sur une billette. Ce mort-l\u00e0 portait barbe et longue chevelure, et robe \u00e0 manches et \u00e0 capuchette. Le v\u00eatement de son pied n'\u00e9tait gu\u00e8re semblable \u00e0 nos godillots, ni m\u00eame \u00e0 ceux des villageois : ce n'\u00e9tait que sandale, que semelle assujettie \u00e0 ortelle et cheville par cordons de cuir. On e\u00fbt dit un mis\u00e9rable, et cependant une lampe brillait par trop brillamment en son \u0153il pour s'accorder \u00e0 mis\u00e8re.\n\nIl stationna long de temps, semblant d\u00e9lib\u00e9rer, lorgnant attentivement p\u00e8re dans son all\u00e9e et venue. Par instants, il tournait le regard vers moi, et je notais sur sa face la tristesse des p\u00e9ris. Au bout d'une dur\u00e9e, il sortit de sa torpeur et fouilla en une poche de son v\u00eatement, d'o\u00f9 il sortit une michotte de pain et une gourde d'eau-de-raisin. Usant de sa robe comme d'une table, il rompit le pain et en avala une morcelle. Il m'offrit par suite \u00e0 manger, ce que je fis volontiers, car, \u00e0 cette heure, j'avais l'estomac qui commen\u00e7ait \u00e0 grincer. Puis il d\u00e9boucha la gourde, y breuva et m'invita \u00e0 le singer, ce que je fis encore.\n\nC'est ensuite que \u00e7a se passa, Monsieur le juge.\n\nGlissant alors la main sous son accoutre \u00e0 l'altitude du poitrin, il commen\u00e7a \u00e0 fouiller ce secteur f\u00e9brilement. Puis le voici qui sort avec lenteur sa main, enserrant \u00e0 pr\u00e9sent quelque chose que je ne d\u00e9m\u00eale pas encore. J'agrandis l'\u0153il, soul\u00e8ve le sourcile, allonge le col, et que vois-je ? Vision fabuleuse ! Spectacle incroyable ! Impensable tableau ! Le mort tenait son c\u0153ur, son c\u0153ur, son c\u0153ur encore tout remuant de mouvements et de vies ! Clou\u00e9 sur ma billette, j'\u00e9tais si stupide que j'en \u00e9garai sur le moment la parole, et presque le respire. J'avais l'\u0153il pour ainsi dire hors trou, la bouche comme \u00e9perdue.\n\nEn silence, me toisant au plus aprofond du regard, le mort me tendit doucettement son c\u0153ur, comme on tend une main secourable. Trembleur mais comme envo\u00fbt\u00e9, j'\u00e9tendis le bras et touchai ce c\u0153ur-l\u00e0. Ce que j'y sentis battre me trouble encore aujourd'hui, et je n'en puis parler sans ressentir \u00e0 nouveau tout le secouement de mon \u00e9difice. Qu'\u00e9tait-ce que ce battement-l\u00e0, ce tressaillement pareil aux chants des b\u00eates so\u00fbl\u00e9es de vie dans l'\u00e9t\u00e9 ? De toute mon existence, je ne fus jamais instruit de mon pr\u00e9nom. Sans doute n'en poss\u00e9dais-je aucunement. Pourtant, le fr\u00e9missement de ce c\u0153ur, cette ineffa\u00e7able m\u00e9lodie r\u00e9percut\u00e9e ce jour-l\u00e0 en tous bouts de moi-m\u00eame, Monsieur le juge, r\u00e9sonnait en ma personne comme l'aurait fait le son de mon pr\u00e9nom !\n\nJ'en restai tout paralytique, ma main comme soud\u00e9e au c\u0153ur du p\u00e9ri. Au bout d'un laps, le voil\u00e0 qui ram\u00e8ne tout \u00e7a sous sa robe. Puis il se l\u00e8ve calmettement et commence \u00e0 marcher vers les arbres tout pr\u00e8s. Il dispara\u00eet en for\u00eat. Les b\u00eates qui nous guettaient, tapies dans les bosquets \u00e0 l'entour, auraient assur\u00e9ment pu en t\u00e9moigner : jamais on ne vit en ces bois bourgeois plus secou\u00e9 que je ne le fus \u00e0 ce moment. Oui, je me trouvai, ce jour-l\u00e0, troubl\u00e9 comme peu le furent. Et tandis que je rapporte ci-devant vous ces choses, Monsieur le juge, je ne puis me d\u00e9fendre de songer combien ce transport qui s'empara de moi \u00e9tait parent de celui que j'avais ressenti pour Manon. Car on aurait dit que le c\u0153ur de ce mort-l\u00e0 contenait cargaisons de ch\u00e9rissement, tout le ch\u00e9rissement du monde. Mieux : que ce ch\u00e9rissement-l\u00e0 m'\u00e9tait promis ! \u00c9nigme de nos destin\u00e9es ! Nous croissons pendant des lunes en les chairs ombreuses d'un ventre, nous arrivons ensuite en cette vie faite de p\u00e9nombres, puis nous marchons sur sentes brouillardeuses, et voil\u00e0 qu'amour nous aborde et nous appelle par notre nom, que lumi\u00e8re surgit des t\u00e9n\u00e8bres !\n\nJe levai le regard jusqu'au-dessus du sommet des arbres. Dans les cieux, la d\u00e9esse Lune, encore p\u00e2le comme galette de sel, patientait pour commencer \u00e0 briller que s'\u00e9vanouissent les heures du jour. En est-il ainsi de nous, Monsieur le juge ? Nous faut-il atteindre l'ombre \u00e9paisse des nuits pour que s'\u00e9clairent enfin nos corps et nos c\u0153urs ?\n\nAbaissant l'\u0153il et remettant le monde en ma mire, j'entendis soudainement p\u00e8re s'\u00e9crier : \u00ab Hourraille ! Houraille ! De l'eau ! De l'eau, parvert ! \u00bb Et, en effet, on d\u00e9m\u00ealait \u00e0 distance le bruit de ses godillots clapotissant dans la flaque nouvelle qu'il venait de mettre au jour.\n\n#\n\nFut-ce l'effet apaisant de l'eau sur ses soifs ? P\u00e8re, en tout cas, me signifia un matin, au sortir de paillasse, qu'il ne souhaitait plus recevoir quelque enseignement que ce soit en mati\u00e8re de toisement de macchab\u00e9es. \u00ab C'est assez de ce sorcellement ! \u00bb lan\u00e7a-t-il, se dressant vitement et posant son pied sur le sol de la cabane. \u00ab On ne doit pas s\u00e9duire diabletons et esprits terribles ! Sinon, qui sait si ces monstrueux-l\u00e0 ne nous attirerons pas d'avance en outre-fin, interrompant du coup notre s\u00e9jour ici-bas et nous transformant en d\u00e9funts avant notre heure ! Pire encore : peut-\u00eatre, profitant de nos fr\u00e9quentations avec les morts, nous enr\u00f4leront-ils en leurs diableries ! Parjaune ! Qu'on me pende casque en bas durant une lunaison compl\u00e8te si, de mon vif, je me frotte \u00e0 d\u00e9mons ! Et puis, les macchab\u00e9es ne doivent pas s'entrem\u00ealer aux vifs. Ce sont individus prodigieux, tandis que nous sommes coutumiers et insipides. On ne peut brasser en m\u00eame marmite raisins d'ours et tub\u00e9reuses. \u00bb\n\nTel fut son dire. Je fus tent\u00e9 de l'instruire qu'en mon opinion, sapristi, ses gens l'enr\u00f4laient d'ores et d\u00e9j\u00e0 en diableries et prenaient toute apparence de d\u00e9mons. Mais je flairais que p\u00e8re ne traduisait en rien la nature satanique des actes qu'il accomplissait sous l'empire de ses gens. Je con\u00e7ois que ce n'\u00e9tait, pour lui, que conduite commune et convenue. Aussi me tus-je plut\u00f4t, m'\u00e9pargnant pour s\u00fbr salves de savates, et pr\u00e9f\u00e9rant m'affairer \u00e0 l'appr\u00eat de notre d\u00e9jeuner : biscottes de parsnip et gruau de bazzanie. Ce fut donc le bout de ses efforts afin de rep\u00e9rer les d\u00e9funts.\n\nPeut-\u00eatre par ailleurs pressentait-il le grand \u00e2ge avancer sur lui avec tant d'audace qu'il ne souhaitait plus, \u00e0 pr\u00e9sent, se consacrer qu'\u00e0 des besognes plus simples. Car il est vrai que sa saison de vie avan\u00e7ait et r\u00e9tr\u00e9cissait grandement sa carri\u00e8re. P\u00e8re avait beau avaler d'abondance viandes grasses et lards nouveaux, les \u00e9poques s'empilaient n\u00e9anmoins sur lui et le tiraient chaque jour davantage vers la grande anciennet\u00e9. Mais il persistait dans le vivre, comme un garenne \u00e0 demi pris au collet, toisant la mort \u00e0 proximit\u00e9 et se d\u00e9battant pour en reporter encore un brin le commencement. Maintes fois, par matins au r\u00e9veil, p\u00e8re \u00e9tait pris de toux impressionnantes. On e\u00fbt cru qu'il s'appr\u00eatait \u00e0 cracher sur table poumons et suites. Mais toujours il surdurait, reprenait couleurs et s'\u00e9tablissait enfin sur taboureau pour le d\u00e9vorement de nos gruaux. Parfois aussi, sa carcasse tout enti\u00e8re \u00e9tait la proie de rhumatismes vari\u00e9s, qui le for\u00e7aient \u00e0 rester sur paillasse pendant plusieurs courses de soleil. Goutte et lumbagos lui gauchissaient le corps. Fi\u00e8vres, picotements, engourds, incontinades, courantes, pouss\u00e9es, nages, migrains, \u00e9c\u0153urements et coliques le houspillaient sans vacances. Oui, \u00e0 n'en pas douter, p\u00e8re d\u00e9clinait sous les coups que l'\u00e2ge lui portait. En sa personne, toutes chairs accusaient fatigues et pourrissements.\n\n#\n\nLes macchab\u00e9es, comme toujours, passaient chez nous, s'attardaient un temps, puis rebroussaient en leur contr\u00e9e d'outre-jour. L'un d'eux, toutefois, s\u00e9journa \u00e0 la cabane plus que de coutume. C'\u00e9tait un m\u00e9nestrel \u00e0 longue chevelure et aux yeux entour\u00e9s de lunettes, qui trimbalait en permanence un gratte-cordes. Il venait, s'assoyait sur une billette et empoignait son instrument dont il pin\u00e7ait le cordage, produisant de la sorte des m\u00e9lodies exquises.\n\nCe furent les seuls sons que j'entendis jamais \u00e9manant d'un mort. \u00ab \u00c0 d\u00e9faut de vocabulaire, rem\u00e2chais-je, les morts useraient donc de musique ? \u00bb Je cherchais \u00e0 traduire ce myst\u00e8re. Et je songeais, non sans trouble : \u00ab Peut-\u00eatre d\u00e9couvrirons-nous, \u00e0 l'heure de notre fin, que parole, au fond, est par trop pauvre et insignifiante pour traduire notre domaine int\u00e9rieur. Mais que m\u00e9lodies constituent non seulement un langage plus appropri\u00e9, mais aussi plus rassembleur, et immortel, et formant passerelle entre les mondes. Et comme, le plus souvent, musique est exquise, peut-\u00eatre trouverons-nous finalement que beaut\u00e9 est seule grammaire qui vaille. \u00bb N'est-il pas singulier qu'une telle m\u00e9ditation m'ait \u00e9t\u00e9 inspir\u00e9e par un d\u00e9funt ? C'est mati\u00e8re que j'avais voulu enfoncer en p\u00e8re : que mort puisse \u00eatre bon ma\u00eetre et servir aux vifs comme boussole aux \u00e9gar\u00e9s.\n\nCette m\u00eame nuit, lorsqu'il se fut assoupi, j'ai ressouvenir d'avoir encore longuement \u00e9tudi\u00e9 p\u00e8re \u00e0 la faveur du feu. Oui, qu'y avait-il en ce corps \u00e9quarri, comme taill\u00e9 \u00e0 m\u00eame le bois du grand h\u00eatre ? Y avait-il en ce mortel-l\u00e0 quelque app\u00e9tit de beaut\u00e9 ? Ou p\u00e8re ne fr\u00e9missait-il qu'\u00e0 l'enfui de l'\u00e9t\u00e9, lorsque s'abattaient sur la cabane neiges et frisquetteries ? Que contenait donc son c\u0153ur, \u00e0 lui ? Certes, m\u00e8re y s\u00e9journait toujours. Mais m\u00e8re \u00e9tait morte depuis longues \u00e9poques. N'y avait-il rien de vif en ce c\u0153ur-l\u00e0 que l'image d'un tr\u00e9pass\u00e9 ?\n\nCependant p\u00e8re allait et venait et ne percevait rien des m\u00e9lodies du mort. Pourtant, le m\u00e9nestrel usait abondamment de son gratte-cordes, emplissant de ses notes tout l'\u00e0 l'entour de la cabane durant longues p\u00e9riodes du jour. C'\u00e9tait, ainsi que je l'ai dit, charmantes musiques, que j'escoutais r\u00eaveusement tandis que je m'affairais \u00e0 besognes. Mais il arriva que, sans m'en rendre compte, je cessai mon labeur et fermai le regard afin de me mieux p\u00e9n\u00e9trer de ces beaut\u00e9s. P\u00e8re me surprit en cette attitude. \u00ab Fils ! tonna-t-il. Quitte un peu ce r\u00eavement de tra\u00eenard et rebrousse \u00e0 ton emploi ! \u00bb Car p\u00e8re \u00e9tait besogneur, et condamnait vivement toutes vaines songeries. Ouvrant l'\u0153il, je r\u00e9pliquai de cette sorte : \u00ab Mais, P\u00e8re, \u00e0 d\u00e9faut de voir le m\u00e9nestrel, n'entends-tu pas le murmure mirifique de son gratte-cordes ? Mais aussi, peut-\u00eatre, ton grand \u00e2ge te bouche-t-il d\u00e9sormais l'esgourde par trop d'\u00e9poques accumul\u00e9es en sa machinerie. \u00bb Son r\u00e9torque fut prompt, et viril : \u00ab \u00c2net\u00e9s, na\u00efveries ! Mon esgourde traduit encore toutes choses bruiteuses sur terre ! Et il n'est de murmure en ici que celui de l'arri\u00e8re-saison qui arrive sur nous \u00e0 grands pas ! Reprends ton emploi, j'ai dit ! Et fends ces billes, et fais provision de racines, et recouds nos liquettes, et usine-nous tuques et calots de pelisses ! \u00bb Ainsi r\u00e9pliquait p\u00e8re \u00e0 la beaut\u00e9, y compris \u00e0 la beaut\u00e9 des inexplicablet\u00e9s : par fabriquement de bonnets, stockage de racines et tricotet d'accoutre.\n\nAdvint le jour o\u00f9 le m\u00e9nestrel quitta notre abord. Je le vis par matin suspendre son r\u00e9cital, poser son gratte-cordes sur l'\u00e9paule, quitter mollement la billette puis dispara\u00eetre dans un taillis de barbe-de-bouc. Il ne parut plus. Les pluies, enfin, reprirent.\n\nL'hiver venait.\n\n#\n\nBourrasques, tombements de neige et verglace formaient \u00e0 pr\u00e9sent les jours. La terre, d\u00e9sormais, \u00e9tait aussi cro\u00fbteuse qu'\u00e9chine de tortue. Surgel\u00e9es accoutraient les arbres, qui craquaient au moindre soufflement de bise. Or, ces divers manquements de chaleurs ne manquaient pas de grossir mon chagrin d'avoir \u00e0 jamais perdu Manon.\n\nMa peine me fit-elle \u00e9garer lumi\u00e8re ? Toujours est-il que je formai en ces jours une singuli\u00e8re manie : je me mis \u00e0 m'immerger en barrique et \u00e0 me frotter peaux et endroits. Quand l'image de Manon me tenaillait excessivement, je d\u00e9saccoutrais ma personne, la plongeais sous eau et la d\u00e9tergeais \u00e0 grands coups de pain de sable. Je conservais pour cela notre tonne par-devant l'\u00e2tre et r\u00e9utilisais son eau de fois en fois. \u00c9nigmerie de nos \u00e9lans : mon gr\u00e9 n'\u00e9tait pour rien dans cette manie-l\u00e0. D\u00e8s que cela m'\u00e9tait possible, il me fallait me plonger jusque par-dessus chevelure, puis me briquer intens\u00e9ment. P\u00e8re, quoique perplexe devant mon man\u00e8ge de mouillements, ne me foudrait pas pour autant. Tout au plus me toisait-il avec bizarrerie, comme si j'avais \u00e9t\u00e9 ph\u00e9nom\u00e8ne ou \u00e9v\u00e9nement. \u00ab Parvert ! Pourquoi t'astiquer de la sorte, Fils ? \u00bb s'\u00e9tonnait-il parfois. Je restais silencieux et comme \u00e9tanche.\n\nJe ne soufflais mot des origines de la manie qui m'habitait. Car je n'ambitionnais qu'une chose : de pouvoir continuer \u00e0 loisir mes plongements. P\u00e8re, cependant, besognait mollement, ou se cuitait \u00e0 l'eau-de-geni\u00e8vre, ou posait le regard sur la vitre, se tracassant de l'\u00e9tat du grain ou attendant le rebours des jours chauds sur la for\u00eat. J'avais peu de r\u00f4le en son occupation. Aussi n'eus-je pas \u00e0 user aupr\u00e8s de lui de carottes, de menteries et de sc\u00e9narios.\n\nIl m'est ardu de traduire l'apaisement que m'apportaient ces plongements. Sans doute construisais-je par eux quelque pont prodigieux et secret permettant notre r\u00e9union, \u00e0 ma charmante et \u00e0 moi. Oh ! comme j'impatientais, certains jours, d'entrer en barrique ! Comme j'ambitionnais de revivre l'exquis moment o\u00f9 je fus r\u00e9cur\u00e9 par Manon ! Il me semblait alors que je m'approchais de mon v\u0153u de toujours : voir, de mon \u0153il voir, le sentiment, oui, le toiser vraiment. Je ne d\u00e9sesp\u00e9rais jamais de l'apercevoir un jour s'\u00e9lever en air, puis de suivre sa course \u00e0 l'entour des choses, et de le voir \u00e0 la fin venir se poser sur mon \u00e9paule ou ma chevelure, ou encore p\u00e9n\u00e9trer mon esgourde ou mon blair pour aller s'\u00e9tablir en toute ma personne. Car j'en \u00e9tais convaincu : il devait y avoir, quelque part en nos chairs, tani\u00e8re ou alc\u00f4ve recelant amour et rel\u00e2chant au-dessus de nous son image, oui, cela se devait.\n\nMais, bien s\u00fbr, concernant Manon, ce n'\u00e9tait que r\u00eavement : peut-on voir le sentiment des fant\u00f4mes, Monsieur le juge ? Car, contrairement \u00e0 m\u00e8re ou \u00e0 d'autres macchab\u00e9es, Manon ne paraissait jamais. C'\u00e9tait, v\u00e9ritablement, un fant\u00f4me. Telle \u00e9tait ma songerie : peut-\u00eatre en est-il ainsi de certains morts. Pour motifs inconnus, ils pr\u00e9f\u00e8rent ne plus commercer avec les vifs, et stationnent plut\u00f4t solitairement en leur patrie d'outre-vie.\n\nO\u00f9 trouver, d\u00e8s lors, ceux qui quitt\u00e8rent l'ici-bas et qui se refusent \u00e0 nos yeux ? O\u00f9 donc, sinon dans le r\u00eavement ?\n\n#\n\nArriva enfin le temps o\u00f9 je connus pourquoi p\u00e8re fut, de toute sa vie, si f\u00e9rocement ennemi de l'humanit\u00e9. J'avais eu vent, d\u00e9j\u00e0, de querelles occasionnelles entre lui et les villageois. Il en fut une cependant, bien ant\u00e9rieure \u00e0 toutes les autres et comme fondatrice, pourrait-on dire, qui d\u00e9clencha pour de bon sa d\u00e9testation. M\u00e8re y avait quelque r\u00f4le.\n\nJe fus instruit de tout cela fortuitement, une nuit o\u00f9 le roupil de p\u00e8re fut troubl\u00e9 et tout transpirant de paroles r\u00e9v\u00e9latrices. Celles-ci parvinrent en effet \u00e0 mon esgourde, creus\u00e8rent jusqu'en ma cervelle et s'y grav\u00e8rent, me laissant, depuis, fort songeur.\n\nC'\u00e9tait au temps lointain mais encore joyeux de la premi\u00e8re verdeur de p\u00e8re, lorsqu'il promenait sur casque bouclettes de jouvencelle. Sa face, m\u00eamement, affichait les \u00e9claboussures de roussi typiques aux garnements. Son \u0153il, espi\u00e8gle et d\u00e9livr\u00e9, refl\u00e9tait la candeur de sa vie encore neuve.\n\nM\u00e8re et lui, d\u00e9j\u00e0, \u00e9taient \u00e9pris l'un de l'autre.\n\nChacun habitait, aupr\u00e8s de ses parents, une paire de maisonnettes voisines en quelque rue du village. Chaque jour on les voyait, mains m\u00eal\u00e9es, promener leurs petites personnes le long des trottoirs, devant les \u00e9tals ou \u00e0 l'abord de la grande for\u00eat, composant bouquets d'aralie et bottes de val\u00e9riane. L'existence, en ces temps-l\u00e0, leur \u00e9tait l\u00e9g\u00e8re, inoffensive.\n\nSurvint, un jour de fin d'\u00e9t\u00e9, le drame qui fit \u00e0 demi basculer p\u00e8re.\n\nCependant que lui et m\u00e8re \u00e9taient sortis, un incendie naquit en la grangeote ci-derri\u00e8re l'une des maisonnettes. Le vent soufflant \u00e0 grands poumons ce jour-l\u00e0, les flammes gagn\u00e8rent vite en importance. D\u00e9j\u00e0 on entendait, venu des champs, le criaillement des vaches lorgnant tout cela, et on lisait aussi leur peur dans les replis de l'air furibond, \u00e9chauff\u00e9 par le feu. Ci-dedans la grangeote, des poulardes enroupill\u00e9es un moment plus t\u00f4t et maintenant prises au pi\u00e8ge cuisaient sur place, ne tentant m\u00eame plus de fuir tant la chaleur tenaillait toutes choses. Les fourrages engrangeot\u00e9s flambaient et, sous l'empire de la forge, s'envolaient en noires fum\u00e9es pareilles \u00e0 de lugubres nu\u00e9es de corneilles. Le lait bouillonnait dans les barriques, qui finissaient par \u00e9clater \u00e0 cause de l'ulc\u00e8re trop vif de leurs flancs. Cela se r\u00e9pandait ensuite, semblablement au tapis que d\u00e9roule sur l'\u00e9tang, par soirs, la d\u00e9esse Lune.\n\nLes parents de p\u00e8re et de m\u00e8re accoururent, suivis bient\u00f4t par d'autres villageois ahuris, tressautants et l'\u0153il agrandi. Tous criaillaient, certains leur d\u00e9sarroi ou leur effroi, d'autres leurs pri\u00e8res. Et les parents de p\u00e8re et de m\u00e8re, sans doute momentan\u00e9ment priv\u00e9s de r\u00e9flexion, se mirent alors follement en casque de vouloir sauver quelques malheureuses poulardes. Ce geste, Monsieur le juge, leur fut fatal. Car sit\u00f4t qu'ils pass\u00e8rent le seuil de la grangeote, on les vit comme aval\u00e9s par les noires fum\u00e9es, et possiblement m\u00eame par les flammes.\n\nOr, il arriva que p\u00e8re et m\u00e8re, alert\u00e9s par les g\u00e9missements de la foule, surgirent comme courseurs sur les lieux. Ci-devant un tel th\u00e9\u00e2tre, m\u00e8re s'effondra de douleur, \u00e9gara l'assiette et, peu apr\u00e8s, la conscience. On la cueillit, puis elle fut conduite plus loin pour quelques ablutions. P\u00e8re fut plus hardi. Malgr\u00e9 son \u00e2ge encore fort vert, il entreprit de sauver du p\u00e9ril ses propres parents et aussi ceux de m\u00e8re. Enaccoutr\u00e9 de sa culotte courte et de sa liquette \u00e0 mi-bras, il commen\u00e7a \u00e0 courser vers le brasier.\n\nUn bourgeois se mit sur son chemin. Voici sa voix : \u00ab Faut pas insister, Petit. Sont d\u00e9j\u00e0 tous cuits \u00e0 point, l\u00e0-dedans ! \u00bb\n\nC'\u00e9tait compter sans l'extr\u00eame combativit\u00e9 de p\u00e8re. Contrari\u00e9, celui-ci grippa la fourche qui tra\u00eenait l\u00e0 et l'enfon\u00e7a dans l'estomac de l'autre, qui s'affaissa sur les genoux, interdit, l'\u0153il et la bouche d\u00e9versant d\u00e9j\u00e0 sangs, fluides et liqueurs vives. La mort le cueillit bient\u00f4t telle une citrouille blette.\n\nD'autres bourgeois et cr\u00e9atures vinrent, affol\u00e9s, qui s'empar\u00e8rent de p\u00e8re puis le ma\u00eetris\u00e8rent, cependant qu'il rugissait son d\u00e9sespoir, pleuroyait et criaillait encore, ass\u00e9nant de ses jambes d\u00e9j\u00e0 muscleuses d'innombrables coups \u00e0 ceux qui le tenaient ainsi. On tenta de le mener plus loin, aupr\u00e8s de m\u00e8re, afin qu'il ne f\u00fbt plus le t\u00e9moin agit\u00e9 et d\u00e9cervel\u00e9 du feu emportant hors de cette vie ses parents et voisins. Il r\u00e9sista tant et tant, se tortillant, griffant, se cabrant et crachant \u00e0 la face de ses opposants, qu'\u00e0 la fin on lui accorda de stationner l\u00e0, quoique bien tenu \u00e0 distance.\n\nSurvint \u00e0 ce moment cette chose terrible, oh ! combien terrible : on vit soudainement sortir en courant de la grangeote les parents de p\u00e8re, encore tout enlac\u00e9s et se consumant, oui, se consumant, d\u00e9vor\u00e9s par les flammes qui leur d\u00e9tachaient la peau des membres, qui leur faisaient \u00e9clater les yeux, qui leur fondaient les esgourdes, qui leur grugeaient les chairs et d\u00e9v\u00eataient leurs os. Fascin\u00e9s, paralytiques, muets et traumatis\u00e9s, tous, y compris p\u00e8re, virent ces promis-\u00e0-la-tombe achever ainsi leur existence. Car, s'\u00e9tant \u00e0 la fin sinistrement allong\u00e9s sur la terre, m\u00ealant leurs chairs calcin\u00e9es et leurs destin\u00e9es jusque dans le tr\u00e9pas, il ne resta bient\u00f4t plus d'eux qu'une paire de carcasses enlac\u00e9es, enlac\u00e9es encore et toujours.\n\nC'est ainsi que, en ce jour maudit et inoubliable, p\u00e8re, r\u00e9duit \u00e0 impuissance et immobilit\u00e9 par les villageois, fut le spectateur stup\u00e9fait de la grillade de sa famille, la seule v\u00e9ritable qu'il e\u00fbt jamais coudoy\u00e9e. Nul doute : c'est ce jour-l\u00e0 que s'envola sa verdeur, sur les ailes noires des grandes corneilles de fum\u00e9e dont j'ai parl\u00e9 tout \u00e0 l'heure. Oui, ce fut assur\u00e9ment, pour lui, un jour noir, et tout empli de volutes funestes, et r\u00e9sonant des croassements de troublants oiseaux de malheur. Le jour des corneilles.\n\n#\n\nP\u00e8re fut donc emp\u00each\u00e9 d'apporter secours \u00e0 ses parents. Cette attitude des villageois fut pour lui comme outragure, comme tra\u00eetrise. Quoi ! cette existence doucette et paisible, emplie jusque-l\u00e0 de promesses d\u00e9licieuses, de bouquets d'aralie et de soupiasses chaudes, cette existence men\u00e9e si plaisamment n'\u00e9tait donc en fait que repaire de trembleurs, de sans-courage et de grilleurs de parent\u00e9 ? Quoi ! il lui faudrait \u00e0 pr\u00e9sent vivre parmi cette racaille, ces marauds, ces pendards, ces faquins, ces gueux ?\n\nJamais ! Plus jamais !\n\nP\u00e8re et m\u00e8re s'en furent en for\u00eat, et y v\u00e9curent pareillement aux ermites, malgr\u00e9 qu'ils fussent encore \u00e0 l'\u00e2ge des primes verdeurs. On ne fit pas proc\u00e8s \u00e0 p\u00e8re pour le meurtre du bourgeois transperc\u00e9 en bedain. Peut-\u00eatre jugea-t-on qu'il avait d\u00e9j\u00e0 suffisamment p\u00e2ti de la perte de ses parents, et aussi des parents de m\u00e8re. On n'opposa gu\u00e8re, non plus, de r\u00e9sistance aux desseins de mes parents. Peut-\u00eatre les eut-on \u00e0 l'\u0153il, de loin, pendant quelque temps. Puis, les \u00e9poques et saisons coulant, peut-\u00eatre furent-ils laiss\u00e9s \u00e0 eux-m\u00eames, un peu oubli\u00e9s et consid\u00e9r\u00e9s, \u00e0 la fin, ces orphelins solitaires, comme une sorte d'attraction, de curiosit\u00e9. Je n'en puis rien savoir. Il est ais\u00e9 de concevoir toutefois que p\u00e8re se jura, \u00e0 compter du jour des corneilles, d'aimer m\u00e8re \u00e0 tout jamais et plus que tout. Car on traduira de tout cela qu'il ne lui est rest\u00e9, apr\u00e8s cette affaire, que cet unique et ultime tr\u00e9sor \u00e0 ch\u00e9rir en l'ici-bas. En unissant sa vie \u00e0 celle de m\u00e8re, savait-il qu'il marquait son destin d'un si tragique pr\u00e9sage ? Je ne sais, je ne sais.\n\nCar h\u00e9las ! tant de fois h\u00e9las ! vint le temps du tr\u00e9pas de m\u00e8re, dans les circonstances que j'ai d\u00e9peintes. P\u00e8re, d\u00e8s lors, termina de basculer et sombra dans les derni\u00e8res aprofondeurs de l'enfer, du plus redoutable de tous les enfers : celui paraissant parfois aux vifs cependant m\u00eame qu'ils circulent encore sur la Terre.\n\nAinsi fut le r\u00e9cit de p\u00e8re, qu'il fit malgr\u00e9 lui, alors qu'il \u00e9tait enserr\u00e9 dans son roupil mauvais et qu'il narrait ces choses \u00e0 quelque t\u00e9moin invisible. Mais quel fut ce t\u00e9moin ? \u00c0 qui p\u00e8re parla-t-il durant toutes ces heures, tandis que la d\u00e9esse Lune passait, que les astres accompagnaient le monde dans son cours ? Il me semble poss\u00e9der le r\u00e9pons \u00e0 cette question. Car dans la cohue des paroles se bousculant sur sa lippe cette nuit-l\u00e0, je crus l'entendre prononcer, au plus fort de son narrement, ces mots poignants et qui stationnent, depuis, en mon casque et mon c\u0153ur palpitants : \u00ab \u00d4 P\u00e8re, \u00d4 M\u00e8re, vous qui entendez mon histoire, pardonnez-moi, pardonnez-moi ! Pardonnez-moi de n'avoir pas su vous secourir ! \u00bb\n\nAussi ne puis-je sans me troubler accueillir en moi-m\u00eame ce penser : p\u00e8re fut donc, comme je le fus moi-m\u00eame et le suis toujours, visit\u00e9 par les macchab\u00e9es, au moins par ces deux-l\u00e0, au moins en son roupil, au moins cette nuit-l\u00e0.\n\n#\n\nToute sa vie, p\u00e8re fut travers\u00e9 d'\u00e9pouvante d\u00e8s lors que les vents soufflaient fortement. Toute sa vie aussi, la seule \u00e9vocation du tr\u00e9pas lui avait glac\u00e9 les os et roidi l'\u00e9chine. Sans doute ces peurs-l\u00e0 trouv\u00e8rent-elles leur commencement en cette journ\u00e9e d'ambiance de fin du monde et de rafales mauvaises, o\u00f9 p\u00e8re vit famille et voisins p\u00e9rir tels des g\u00e9linots embroch\u00e9s. Sans doute aussi toute sa science de lecteur d'astres remonte-t-elle \u00e0 cette journ\u00e9e maudite o\u00f9 les corneilles de fum\u00e9es lui prirent parents : de ce jour, p\u00e8re n'a cess\u00e9 de chercher en cieux quelque trace restante de ses aim\u00e9s. La nuit, surtout, fut propice \u00e0 sa qu\u00eate. La vo\u00fbte, sous heures obscures, ne rev\u00eat-elle pas le noir accoutre des corneilles ? Je le traduis \u00e0 pr\u00e9sent, Monsieur le juge : p\u00e8re a tant foui le plafond des nuits qu'il en a connu le secret de ses \u00e9toiles.\n\n#\n\nL'arri\u00e8re-saison s'avan\u00e7ant et \u00e9tendant \u00e0 pr\u00e9sent son domaine, minceur de menu s'agrandissait, faminerie mena\u00e7ait. Car, depuis les premi\u00e8res neigettes, b\u00eates \u00e0 viandes avaient fui ou devenaient aussi rarissimes que poil de vip\u00e8re. Or, \u00e0 chaque jour coulant, p\u00e8re et moi \u00e9garions un peu plus de poids et d'importance, et commencions \u00e0 flotter en nos accoutres. P\u00e8re s'en alarmait maximalement, et clamait \u00e0 toute heure : \u00ab Parnoir ! Le tr\u00e9pas nous guette ! Voici que nous nous efflanquons et nous d\u00e9poss\u00e9dons p\u00e9rilleusement de nos beurres ! Si cela se maintient, nous ne serons plus avant long de temps que squelets bons \u00e0 enfouir et qui iront rejoindre monstrueux en patrie de morts ! Il nous faut urgemment d\u00e9goter cargaisons de viandes grasses et de lards nouveaux, sinon je ne donne pas cher de nos vies ! \u00bb Aussi, souventes fois, quittions-nous d\u00e8s l'aube la cabane avec en \u00e9chine arcs et fl\u00e8ches, et attrape-loutre, et lance-caillou, et divers traquenardiers \u00e0 gibier bourr\u00e9s de crevard de mouffeton. Mais la fortune ne nous souriait gu\u00e8re, et nombre de fois d\u00fbmes-nous rebrousser \u00e0 notre seuil, le soir, avec pour seul butin quelque raton victime de notre chasse. Bien qu'il ne f\u00fbt pas du m\u00eame sang que piquepoqu\u00e8tes et larrons, p\u00e8re alla un jour jusqu'\u00e0 d\u00e9trousser une poularde en la grangeote d'un villageois. Nous en f\u00eemes une bouillade avec assaisons de sel de racine et poudraille d'araign\u00e9e, que nous d\u00e9vor\u00e2mes tels des insatiables. Mais ce ne fut qu'\u00e9ph\u00e9m\u00e8re ripaille, et bient\u00f4t nos estomacs recommenc\u00e8rent \u00e0 se tordre et \u00e0 bruiter, \u00e0 exiger secours.\n\nEn ces temps de difficult\u00e9s, p\u00e8re t\u00e2chait d'\u00e9touffer son alarme en se pintant plus que de coutume. En l'officine, l'alambic chauffait et parchauffait, produisant cruches et cruches d'eau-de-geni\u00e8vre, que p\u00e8re breuvait goul\u00fbment.\n\nIl arriva que, gris comme pierre, m\u00ealant un pied \u00e0 l'autre, il \u00e9gara l'assiette, chut, et s'en alla toquer le mur avec son casque. Sangs et s\u00e8ves jaillirent, cependant que conscience le fuyait. Catastrophique, j'abandonnai ma besogne du moment et me pr\u00e9cipitai sur p\u00e8re afin de questionner son \u00e9tat. Sa liqueur gouttait, suivait le dessin de la face et du col et venait mouiller le sol. P\u00e8re se vidangeait, et mena\u00e7ait de se tarir ! Appelant toute ma poigne, je le soulevai, le menai sur sa paillasse et commen\u00e7ai \u00e0 le secourir.\n\nIl me fallut d'abord juguler le saignement, ce que j'accomplis de la fa\u00e7on que voici : je confectionnai une pommade d'herbe aux goutteux, que j'\u00e9talai sur la plaie. Par suite, \u00e0 l'aide d'un pansement de fibrille, je ceinturai le casque. Cela me satisfit fort, puisque l'\u00e9coulement de p\u00e8re, d\u00e8s lors, cessa tout net. Je pr\u00e9parai aussi un crachin de couleuvr\u00e9e, que je mis \u00e0 d\u00e9gorger. Quand ce fut pr\u00eat et bien ti\u00e8de, je for\u00e7ai le tout en la bouche de p\u00e8re et le fit d\u00e9glutir, enjoignant sa m\u00e2choire au labeur \u00e0 l'aide de ma main. Afin de pr\u00e9server la souplesse de ses musclures et la solidit\u00e9 de ses chairs, j'usai aussi sur lui de frictionnements divers. Je dus pour cela composer une p\u00e2t\u00e9e d'amanite tue-mouche, que je liai au suc d'une esgourde-de-Judas. Cette derni\u00e8re \u00e9tant, en pareille saison, fort dess\u00e9ch\u00e9e, j'en su\u00e7ai long de temps la pulpe et obtins ainsi que ses s\u00e8ves revivent. Puis je poissai mes doigts de cette mixture et la r\u00e9pandit en massant sur toutes parties du malade. J'\u00e9puisai, pour la confection de toutes ces drogues, ce qui restait de nos r\u00e9serves d'herbes et de champignes. Il nous faudrait, dor\u00e9navant, couler la rude saison sans elles et esp\u00e9rer que nous n'en aurions plus n\u00e9cessit\u00e9. Mais il faut traduire de ma conduite que nul prix n'\u00e9tait trop fort pour assurer le rel\u00e8vement de p\u00e8re. Car on n'en doit pas douter, Monsieur le juge : j'aurais voyag\u00e9, si requis, jusqu'aux confins du monde afin d'y trouver le rem\u00e8de au mal de p\u00e8re.\n\nP\u00e8re, cependant, ne recouvrait pas le sens. Je sus malgr\u00e9 tout qu'il vivait toujours car son poitrin se soulevait et son pouls lui battait le poignet. Il ne me restait qu'\u00e0 stationner aupr\u00e8s de lui et veiller \u00e0 son confort. Je restai ainsi toute la nuit, repla\u00e7ant le capiton de sa paillasse, m'assurant que f\u00e9brilit\u00e9s et pouss\u00e9es ne le tenaillaient pas trop, lui administrant ici et l\u00e0 un brin de crachin, fra\u00eechissant aussi sa face d'une poign\u00e9e de neige. Les heures furent longuettes. Recueilli sur le taboureau, je priais la d\u00e9esse Lune de ramener p\u00e8re en son \u00e9tat d'heureuse sant\u00e9. Au c\u0153ur des t\u00e9n\u00e8bres, la m\u00e8che de notre lampe m'\u00e9tait comme boussole : en sa lueur reposaient toutes esp\u00e9rances et r\u00eaves fi\u00e9vreux de retrouvailles. Dehors, frimasseries mordaient les choses, tourmentes gauchissaient les arbres, toutes b\u00eates restaient tapies et silencieuses, attendant en leur tani\u00e8re un r\u00e9pit, le rebours de quelque chaleur. J'\u00e9tais semblablement \u00e0 ces b\u00eates : seul en ma tani\u00e8re, affreux, escoutant le bruit du monde, esp\u00e9rant de meilleurs temps.\n\nJ'observais p\u00e8re. Oh ! comme je le ch\u00e9rissais, ce bourgeois inventeur de mes jours ! Comme je le v\u00e9n\u00e9rais, ce chef de la for\u00eat, ce prince des sentes, cet attrapeur de b\u00eates et juguleur de vip\u00e8res, ce ma\u00eetre tailleur de billettes, cet ogre mangeur de lards, ce bandeur d'arcs, ce lecteur d'astres ! Qu'il \u00e9tait beau avec sa chevelure tels des pyl\u00f4nes de farine blanche, son sourcile en taillis, son blair \u00e9pat\u00e9, sa joue mafflue et sa lippe moueuse ! Oh oui ! comme je le ch\u00e9rissais ! Et comme en \u00e9cho \u00e0 cela, plus tard dans les heures me vinrent encore ces questions qui ne me l\u00e2chaient plus, d\u00e9sormais : \u00ab P\u00e8re m'aime-t-il ? Et si oui, o\u00f9 se terre donc son amour ? \u00bb Ah ! si au moins j'avais pu apercevoir cet amour-l\u00e0 !\n\nNon, amour ne nous doit pas \u00eatre invisible, non plus qu'immat\u00e9riel. Quoi, amour serait comme vapeurs, comme riens : intouchable, inblairable et introuvable ? Je ne peux m'y r\u00e9soudre. Je dis : amour est comme nous-m\u00eames, b\u00e2ti de chairs et de substances flagrantes et observables. Mais peut-\u00eatre aussi notre \u0153il lui-m\u00eame est-il par trop aveugle, et incomp\u00e9tent \u00e0 saisir mati\u00e8re aussi fuyante. Voil\u00e0 pourquoi je me questionnais tant : p\u00e8re m'aimait-il, m'aimait-il seulement ?\n\n#\n\nLes cieux commen\u00e7aient \u00e0 rosir quand il \u00e9mergea enfin. Comment traduire ma joie d'alors ? Ceux qui furent naufrag\u00e9s, puis rescap\u00e9s, peuvent sans doute le mieux trouver les mots pour cela. \u00c0 l'\u00e9veil de p\u00e8re, je me jetai \u00e0 genoux et, bien que je ne fusse pas encore enseign\u00e9 de vocabulaire, commen\u00e7ai \u00e0 peindre mon merci \u00e0 la d\u00e9esse Lune par profusion de paroles. J'\u00e9tais encore tout pli\u00e9 et remerciant quand p\u00e8re \u00e9mit d'une voix caillouteuse ces mots, les premiers depuis des \u00e2ges, ce me semblait : \u00ab Tordvert ! O\u00f9 suis-je ? Fils ! Que formes-tu l\u00e0, pli\u00e9 \u00e0 mon chevet et discourant avec un tel abond ? \u00bb Je me pr\u00e9cipitai sur sa personne, le serrant si fort qu'il en \u00e9garait le souffle. \u00ab P\u00e8re ! P\u00e8re ! g\u00e9mis-je. Tu faillis cette nuit passer le seuil de l'ici-bas, et n'e\u00fbt \u00e9t\u00e9 de mon secours, tu ne serais d\u00e9j\u00e0 plus qu'un d\u00e9funt en ton accoutre ! \u00bb\n\nIl leva le sourcile.\n\n\u00ab Et comment me soignas-tu donc ? \u00bb fit-il, tout empli de d\u00e9fiance. Ainsi fut ma voix : \u00ab J'usinai crachin, pommade et p\u00e2t\u00e9e, pour stopper l'\u00e9coulement de ta liqueur, te frictionner et te garder robuste et vital ! \u00bb Je le vis d\u00e8s lors se renfrogner. Ce qui suit sortit de sa bouche : \u00ab Comment ? Tu usas de nos ultimes herbes et champignes pour m'administrer tes soins de fillette ? Alors que la rude-saison n'en est encore qu'\u00e0 sa demie ? Et comment tiendrons-nous, maintenant, lorsque viendront plus grandes disettes encore et, avec elles, graves grincements de tripes et autres vrais troubles de corps ? \u00bb\n\nLe courroux de p\u00e8re fut immense. Se dressant, il grippa l'une de nos cruches et se mit en gosier nombre d'\u00e9cuelles d'eau-de-geni\u00e8vre. Il m'empoigna ensuite par chevelure et me remorqua dans la neige jusqu'au grand h\u00eatre, o\u00f9 je fus ligot\u00e9 solidement. Puis, p\u00e9tri de furie, p\u00e8re d\u00e9chira ma liquette, me laissant ventre au vent par temp\u00e9rature r\u00e9frig\u00e9rante. Je le vis l'instant d'apr\u00e8s usiner nombre de boulets de neige fort tass\u00e9e. Prenant alors ma personne pour cible, il m'en d\u00e9cocha dessus par myriades. Je recevais chacun pareillement \u00e0 une morsure.\n\nAu bout de cette torture, p\u00e8re vint me d\u00e9livrer de mes liens et me for\u00e7a \u00e0 l'accompagner jusqu'\u00e0 l'\u00e9tang. Il nous fallut long de temps pour y parvenir, notre circulation \u00e9tant grandement g\u00ean\u00e9e par la neige amoncel\u00e9e sur la sente. Mais j'allais docilement ci-devant p\u00e8re, craignant, sinon, grossissement de sa foudre. Mon pas \u00e9tant p\u00e9nible et contrari\u00e9, je devais parfois progresser telle une b\u00eate, c'est-\u00e0-dire avec pieds et mains pour assurer l'aplomb. Doigts et ortelles roidissaient leurs gonds en mes chairs. Mon c\u0153ur toquait follement, comme voulant sortir \u00e0 tout prix de sa ge\u00f4le d'os. Sous accoutre, mes cuissots se figeaient, tant \u00e9tait grande la piq\u00fbre de la frisquetterie. Le souffle me fuyait, et maintes fois, par souci de recouvrer l'assiette, je dus alentir ma course. P\u00e8re me punissait alors d'une savate, ou pire encore d'un coup de bille sur le train. En la for\u00eat, surfroid p\u00e9trifiait toutes choses, et j'\u00e9tais pareil \u00e0 toutes choses : comme \u00e0 demi mort et ne me questionnant plus sur cela. J'allais, telle une poularde qu'on m\u00e8ne au couperet.\n\nNous abord\u00e2mes enfin l'\u00e9tang. Sa surface \u00e9tait couverte d'un couvercle de gelure. Je tremblais au penser du sort qui m'attendait l\u00e0. J'en fus bient\u00f4t instruit. P\u00e8re alla qu\u00e9rir, non loin, une lourde pierre. Soulevant celle-ci \u00e0 bout de bras, il la projeta ensuite sur l'\u00e9tang. Un trou en naquit, qui fit sourdre de sous le couvercle quantit\u00e9 d'eau temp\u00e9tueuse et comme surprise, troubl\u00e9e dans son roupil. P\u00e8re dit alors cette phrase, qui me fit entrevoir la fin de ma carri\u00e8re de vif : \u00ab \u00c0 pr\u00e9sent, Fils, marche \u00e0 ce trou et jettes-y ta personne ! \u00bb\n\nJ'avan\u00e7ai donc lentement jusqu'\u00e0 l'abord du trou. Juste avant de m'y plonger, j'abaissai paupi\u00e8res et priai la d\u00e9esse Lune. Je lui mandai de m'accorder doucette fin et de faire en sorte que j'aboutisse sur le seuil de l'outre-monde sans avoir par trop p\u00e2ti de glaceries et souffrances \u00e9quivalentes. Par m\u00eame occasion, j'appelai int\u00e9rieurement m\u00e8re. Je fus exauc\u00e9 : elle parut, mais, pour la premi\u00e8re fois, sa silhouette ne me fut visible que ci-dedans ma personne, et nullement \u00e0 l'entour comme en la coutume. Sa pr\u00e9sence me fit grand bien, apaisant un brin le toquement de mon c\u0153ur et l'affre de ma chair. M\u00e8re \u00e9tait l\u00e0, en moi, ceinte de la petite lumi\u00e8re bleue typique des p\u00e9ris. \u00c0 force de toiser sa douce face, je sentis le calme me gagner. Peut-\u00eatre souris-je m\u00eame. Oui, la vie pouvait me quitter \u00e0 pr\u00e9sent, j'\u00e9tais par\u00e9 \u00e0 rejoindre l'outre-race. Je m'adressai de la sorte \u00e0 m\u00e8re : \u00ab Patiente encore un peu, petite M\u00e8re. Tu me verras te coudoyer bient\u00f4t. Ainsi pourrons-nous circuler de concert sur les sentes du pays des d\u00e9funts. Ainsi pourrai-je \u00e0 loisir lorgner l\u00e0-bas les grands arbres que j'imagine, bleus et stationn\u00e9s, puisque ni vent ni bourrasque n'en trouble jamais les branchottes. Ainsi pourrai-je mesurer ce grand silence qui semble r\u00e9gner en ce pays. Ainsi pourrai-je enfin reposer mon casque et le faire taire, lui qui ne cesse jamais, en l'ici-bas, de chercher \u00e0 traduire toutes choses. Car sans doute obtient-on l\u00e0-bas r\u00e9pons \u00e0 chaque questionnement, sans doute, sans doute. Et peut-\u00eatre ainsi conna\u00eetrai-je enfin, puisque nul d\u00e9c\u00e9d\u00e9 ne peut me l'expliquer aujourd'hui par mots, ce que signifie cette tristesse sur ta face et sur celle de tous tes camarades de mort. Oh ! comme j'impatiente de p\u00e9n\u00e9trer ce monde-l\u00e0, si douce\u00e2tre et si placide ! \u00bb\n\nPuis je sentis mon pied tout \u00e0 la fois glisser, quitter le couvercle de l'\u00e9tang et toucher l'eau ci-dessous. P\u00e8re venait de me pousser en trou.\n\n#\n\nCe fut comme si mon corps \u00e9tait enti\u00e8rement soumis \u00e0 \u00e9corchure. Bien que l'\u00e9tang repos\u00e2t en une froiderie sans nom, l'onde en me touchant me br\u00fbla plut\u00f4t, tant fut foudroyante son \u00e9treinte. Mais, d\u00e9j\u00e0, engourds et roideurs me d\u00e9tournaient de ce supplice et commen\u00e7aient \u00e0 m'emmener en outre-monde. D\u00e9j\u00e0, nuits et surnuits m'enrubannaient. Des bras gelur\u00e9s refermaient leur pince sur moi, des chevelures atroces venaient fr\u00f4ler ma face, emmaillottaient jambes et ventre, me grippaient par clavicule et m'indiquaient ainsi le passage pour outre-vie. J'avalais sablons, bourbes, gravillons, carcasses d'\u00e9crevisses et charognes de t\u00eatards. J'effleurais des poissons d\u00e9testables et laids, couch\u00e9s sur \u00e9chine, bouche b\u00e9ate, ab\u00eem\u00e9s dans leur roupil d'hiver, en apparence de cadavres. Des grenouilles \u00e0 demi mortes aussi flottaient \u00e0 l'entour, comme coagul\u00e9es, statuettes de chair gris\u00e2tre attendant chaleurs nouvelles. Je sombrais, Monsieur le juge, ainsi que sombre dans l'\u00e9tang le poisson-chat arriv\u00e9 au bout de son \u00e2ge : les flancs meurtris, promis \u00e0 l'effrayant silence des gouffres, \u00e0 l'enlacement visqueux des limons.\n\nM\u00e8re, cependant, sortit de moi et s'\u00e9tablit \u00e0 quelque distance, observant mon engloutissement. Lui tendant les bras, je pronon\u00e7ai \u00e0 son intention d'ultimes paroles, escort\u00e9es hors de ma bouche par nombres de bulles. Plus haut, \u00e0 genoux sur le couvercle et scrutant les profonds, p\u00e8re devait traduire de ces bulles-l\u00e0 l'approche de mon tr\u00e9pas. Toujours calme comme hibou sur branche, je dis \u00e0 m\u00e8re : \u00ab Avec ces bullettes s'envole ma vie : voici venir la d\u00e9sinvention de mes jours ! Accueille-moi donc \u00e0 ton c\u00f4t\u00e9, petite M\u00e8re ! \u00bb Puis je touchai son \u00e9paule. Sa d\u00e9froque ondulait au rythme tra\u00eenant des eaux. Je vis bient\u00f4t qu'elle versait quelque pleur, car des gouttes bleues, du m\u00eame bleu que l'aur\u00e9ole des d\u00e9funts, avaient commenc\u00e9 \u00e0 flotter \u00e0 l'entour d'elle, semblables \u00e0 des pierrettes de prix. Ainsi donc, les larmes des morts ne se m\u00ealaient pas \u00e0 l'eau vive. Mais y avait-il, \u00e0 la fin, quelque chose du mourir qui se puisse brasser au vivre ? Ou vifs et macchab\u00e9es \u00e9taient-ils condamn\u00e9s \u00e0 ne jamais \u00e9changer v\u00e9ritablement ? Fallait-il donc absolument quitter cette vie pour nous relier aux tr\u00e9pass\u00e9s ? N'existait-il donc pas quelque seuil mitoyen o\u00f9 vie et mort puissent enfin communiquer ?\n\nMais ma fin s'avan\u00e7ait, aussi fermai-je l'\u0153il et demandai ceci, je ne sais \u00e0 qui ni bien pourquoi : \u00ab \u00c0 pr\u00e9sent que le vivre me d\u00e9serte, le temps vient de se questionner. Fus-je convenable r\u00e9sident de la Terre ? Aurai-je \u00e9t\u00e9 bon fils ? Aimai-je suffisamment et d'ad\u00e9quate mani\u00e8re ? Et, surtout, fus-je aim\u00e9, oh ! le fus-je seulement ? \u00bb Tels furent mes pensers de ce moment-l\u00e0. Nuls autres ne me vinrent.\n\n#\n\nUne fois allong\u00e9 sur le limon, je me revis reposant l'autre fois dans la bi\u00e8re de m\u00e8re, et toisant le monde avec l'\u0153il du tr\u00e9pass\u00e9. Mais le monde, \u00e0 pr\u00e9sent, n'\u00e9tait que brouillards et poisseries troubles. Mais sur le monde, \u00e0 pr\u00e9sent, flottaient cadavres de grenouilles, et viandes gangren\u00e9es, et v\u00e9g\u00e9tations pourries et sans racines. Comment avais-je pu vivre dans ce monde-l\u00e0 ? Encore m\u00eame aujourd'hui je ne peux \u00e9viter de songer : p\u00e8re se repr\u00e9sentait la mort comme domaine de terribleries, d'alarmes, de disgr\u00e2ces et de grimaces. Mais \u00e0 la v\u00e9rit\u00e9, n'est-ce pas plut\u00f4t le fief des vifs lui-m\u00eame qui est le plus terrible ?\n\nJe n'eus pas loisir de finir mon toisement sous-marin. Dans le brouillard bourbeux des profonds, je vis tout \u00e0 coup le bout d'une pique, allong\u00e9e de son crochet. Cela fouilla habilement et prestement l'eau d'\u00e0 l'entour, puis trouva \u00e0 la fin l'encolure de ma liquette. Sit\u00f4t mon v\u00eatement gripp\u00e9, je me sentis soulev\u00e9 avec force de ma couche de vases. Une poigne puissante venait de tirer sur la pique et me ramenait vers l'air. Le temps d'un saut de crapaud, j'\u00e9tais hiss\u00e9 tel un loutron sur le couvercle de gelure. Levant avec faiblesse la face et risquant un \u0153il, j'aper\u00e7us p\u00e8re pos\u00e9 \u00e0 mes c\u00f4t\u00e9s, sa pique encore d\u00e9goulinante. Il me toisa mauvaisement, l'espace d'une courte dur\u00e9e. Je l'entendis alors prononcer ce dire : \u00ab Ta m\u00e8re avait souci de pr\u00e9venance, et n'usait jamais \u00e0 la l\u00e9g\u00e8re de nos herbes. Cette nuit, j'aurais recouvr\u00e9 mes sens m\u00eame sans tes pr\u00e9parations et soins de fillette. Saign\u00e9es et autres secours modernes auraient suffi \u00e0 mon rel\u00e8vement. M\u00e9dite cela, la prochaine fois que tu voudras usiner quelque m\u00e9decine ! \u00bb\n\nPar suite, de son pas lourd et de tranquille allure, il s'en fut prendre la sente pour rebrousser \u00e0 la cabane.\n\nN'e\u00fbt \u00e9t\u00e9 de la morsure des froideries r\u00e9gnant \u00e0 l'entour, j'aurais stationn\u00e9 long de temps couch\u00e9 et avoisinant le trou. Mais alors, le tr\u00e9pas e\u00fbt vite fait de me cueillir comme fleur de glace. Quoique d\u00e9bile, douleurant en chaque repli de mon corps, j'\u00e9lus pr\u00e9f\u00e9rablement de m'emplir le poumon de tout l'air qui m'\u00e9tait \u00e0 pr\u00e9sent redonn\u00e9, puis de m'\u00e9branler avec ce que m'autorisait encore le peu de nerf qui me restait. Ce n'est qu'\u00e0 ce moment que je notai l'absence de m\u00e8re, que je vis aussi que le vent ne soufflait plus, que les arbres, comme r\u00e9sign\u00e9s, avaient suspendu leur plainte de craquements. Aucun signe non plus de quelque animal. J'\u00e9tais seul, bien plus seul, ce me semblait, que je ne l'aurais \u00e9t\u00e9 en outre-monde.\n\n\u00c0 demi congelur\u00e9, je me dressai donc p\u00e9niblement, puis pris \u00e0 mon tour le chemin de notre seuil. Derri\u00e8re moi, glaceries commen\u00e7aient d\u00e9j\u00e0 \u00e0 refermer le trou de l'\u00e9tang, telle une plaie sur la peau d'une b\u00eate. Car la nature n'est pas pareille \u00e0 nous : nul c\u0153ur et nul sentiment ne toquent en son poitrin. Aussi son rel\u00e8vement ne tarde-t-il gu\u00e8re.\n\n#\n\nP\u00e8re avait-il agi ainsi sous l'empire de ses gens ? Ou alors ce d\u00e9testable \u00e9pisode avait-il \u00e9t\u00e9 un \u00e9lan de son propre gr\u00e9, peut-\u00eatre d\u00fb \u00e0 son grand \u00e2ge ? Les \u00e9poques s'accumulant en sa vie, peut-\u00eatre ses gens ne lui \u00e9taient-ils plus n\u00e9cessaires, \u00e0 pr\u00e9sent, pour atteindre au d\u00e9r\u00e8glement de son sens. Je ne pouvais le d\u00e9m\u00ealer. Mais une certitude m'emplissait tout entier : je ne pouvais plus, d\u00e9sormais, agir comme second et comme esclave dans ses folles entreprises. Il me fallait outrepasser ce funeste \u00e9tat et freiner le bras cruel de p\u00e8re. C'en \u00e9tait assez, assez : j'avais par trop essuy\u00e9 d'infortune et p\u00e2ti de ses man\u00e8ges. Et puis, surtout, je devais savoir enfin : p\u00e8re m'aimait-il, m'aimait-il seulement ?\n\n\u00c9garai-je l'aplomb de mon casque, et pareillement la sant\u00e9 de mes gestes ? J'abordai en tout cas moi aussi, avec cette r\u00e9cente folie de p\u00e8re, les rives d'une sorte de d\u00e9mence.\n\nJe patientai quelques jours que mes forces rebroussent tout \u00e0 fait. Puis, par soir venu, quand la d\u00e9esse Lune fut \u00e0 son comble, que la for\u00eat stationnait sous frimasseries et que p\u00e8re reposait en son roupil le plus calmement du monde, je quittai le s\u00e9jour de ma paillasse et m'en fut empoigner l'un de nos coutelas. J'abordai p\u00e8re. Assist\u00e9 en mon ouvrage par la vive clart\u00e9 des astres, je visai au travers de la liquette et, d'un coup de bras incontestable, fis traverser le c\u0153ur de p\u00e8re au m\u00e9tal brillant. Le sort voulut que p\u00e8re survive au coup. Alors, renon\u00e7ant enti\u00e8rement au roupil, le voil\u00e0 qui se dresse sur paillasse avec fortes et terribles criailleries et jette sur ma personne des yeux horribles, comme traduisant quelque cauchemar. Vif encore malgr\u00e9 tout (peut-\u00eatre par la s\u00e8ve du d\u00e9sespoir, je ne sais), il m'assaille et me met sur face nombre d'ardentes mornifles puis, d'une savate, m'enfonce l'estomac, et me gauchit le col, et me d\u00e9racine cheveux et poils, et me sarcle de ses ongles, et manque me disjoindre la clavicule, et pendant tout ce temps s'\u00e9gosille et tonitrue \u00e0 crever l'esgourde. Mais sur la paillasse le sang de p\u00e8re goutte de plus en plus, aussi chancelle-t-il \u00e0 la fin, jusqu'\u00e0 manquer d'assiette. Quand il choit, je le grippe par la chevelure et le m\u00e8ne dehors, tra\u00eenant sur le sol de la chambrette son grand corps \u00e0 demi vif et culbutant au passage table et taboureaux.\n\nParvenu au grand air je jette p\u00e8re sur le sol, me penche sur lui et, tandis qu'il meurt presque, lui murmure : \u00ab Regarde, P\u00e8re. Regarde bien. Toise encore une fois ces astres, ces \u00e9toiles : que lis-tu dans cet ouvrage c\u00e9leste, que lis-tu ? Toi qui ne per\u00e7ois que si peu l'image des morts, aper\u00e7ois-tu au moins, ci-haut, celle des vifs ? Vois-tu l'image des vifs, P\u00e8re ? Les vois-tu, eux et leur c\u0153ur tout palpitant et impatient de trouver en leurs semblables la marque du sentiment ? Ne vois-tu pas qu'ils ne peuvent convenablement exister sur terre si ce sentiment ne transpara\u00eet et ne se rend jusqu'\u00e0 ce c\u0153ur ? Ne vois-tu pas que bourgeois et cr\u00e9atures rendent l'\u00e2me petit \u00e0 petit s'ils ne sont abreuv\u00e9s de la s\u00e8ve du sentiment ? Et ne vois-tu pas que d\u00e9c\u00e9der ainsi, morceau par morceau, dur\u00e9e par dur\u00e9e, ass\u00e9ch\u00e9 par manquement d'affection, est pour eux un sort bien plus funeste que celui r\u00e9serv\u00e9 \u00e0 ceux que la mort prend rondement ? Ne le vois-tu pas, ne le vois-tu pas ? Aussi fais-je preuve \u00e0 ton endroit du plus grand amour en posant aujourd'hui le geste que voici. \u00bb\n\nEt sur ces paroles, la vue troubl\u00e9e par mon pleur, je commen\u00e7ai \u00e0 d\u00e9pecer p\u00e8re.\n\n#\n\nJ'enfon\u00e7ai d'abord le coutelas dans le col et, par suite, le fis courir tout le long du ventre jusqu'\u00e0 ouverture totale et s\u00e9paration des peaux en parties \u00e9gales. Apr\u00e8s, fouillant lards, beurres, musclure et nervades, je raclai la cloison et entrepris de d\u00e9tacher le cuir de la charpente. J'allais efficacement : la d\u00e9esse Lune avait boug\u00e9 d'un saut de crapaud \u00e0 peine quand j'entamai le d\u00e9grafage des boyaux, cr\u00e9pine, carnes et ficelles. Je vis se crever sous mes doigts les rognons, s'\u00e9pandre la liqueur des boudins, se fendre les \u00e9toffes et tissures lardeuses jouxtant la carcasse, se briser l'arbre du poumon. D\u00e9j\u00e0 ce n'\u00e9tait plus p\u00e8re que je creusais, mais un cloaque, mais un \u00e9gout.\n\nJe m'attelai par suite \u00e0 le d\u00e9pelurer enti\u00e8rement. Chaque doigt fut vitement fendu par sa longueur, permettant ainsi \u00e0 la pointe du coutelas de se faufiler sous le cuir et de d\u00e9saccoutrer la chair. La main p\u00e2tit du m\u00eame sort et go\u00fbta, par le milieu, du tranchant de mon outil, autorisant aussi bien le d\u00e9garnissement des paumes, des revers et des poignements. J'op\u00e9rai semblablement sur toute la route du bras, divisant autant viandes molles que mati\u00e8res raides en parties semblables et nues. J'abordai les soies du menton avec aplomb : jamais on ne vit tranchant avancer plus hardiment en la substance d'une b\u00eate. Cela me mena \u00e0 la base de la face, que je d\u00e9coupai non sans traverser le blair par son centre. Tout le casque suivit, le coutelas tra\u00e7ant l\u00e0-dessus itin\u00e9raires depuis le front jusqu'\u00e0 la nuque, ouvrant chemin sous chevelure. Je retournai alors p\u00e8re telle une tranche de choix, puis dessinai une sente de sang de la nuque jusqu'au train. L\u00e0, je formai en perpendicule une ligne semblable, d\u00e9sunissant le tronc des zones jambi\u00e8res, ce qui permit le d\u00e9garnissement de toute la pelure drapant la moiti\u00e9 sup\u00e9rieure de la personne de p\u00e8re. Vinrent les jambes, pour lesquelles j'usai de la m\u00eame technique d'ouverture que pour les doigts : chacune fut entaill\u00e9e comme petit pain et lib\u00e9r\u00e9e de ses int\u00e9rieurs. Par \u00e9corchade, je d\u00e9sengodillai enfin de leurs peaux les pieds de p\u00e8re. Il \u00e9tait nu, \u00e0 pr\u00e9sent. J'en fus tout esbaudi : de ma vie, pour la premi\u00e8re fois, je pouvais voir enti\u00e8rement en ses machineries, et le connaissais authentiquement.\n\nJe consid\u00e9rai un moment la d\u00e9pouille. Puis, je me penchai sur le territoire du c\u0153ur, demeur\u00e9 accroch\u00e9 depuis tant\u00f4t \u00e0 la r\u00e9sille des couennes. Tremblotteur, je d\u00e9pris ce c\u0153ur, le d\u00e9fis en tranches fines au creux de ma main, pour voir, pour voir enfin. Mais mon espoir d\u00e9senfla aussit\u00f4t, et je con\u00e7us d\u00e8s lors \u00e0 quel point je n'avais \u00e9t\u00e9, sur la Terre, qu'un fils mal aim\u00e9 de son parent comme dans le repos de m\u00e8re autrefois, je ne d\u00e9gotai nulle trace de sentiment dans les creux, replis et galeries du c\u0153ur de p\u00e8re.\n\nJetant au sous-bois cette chair morte, les pleurs me piquant l'\u0153il, je continuai ma besogne. Je la terminai enfin par l'\u00e9tape du videment du casque, dont la teneur s'\u00e9pandit bient\u00f4t, se m\u00ealant aux barbaques vari\u00e9es jonchant le sol. P\u00e8re n'\u00e9tait plus que cela, \u00e0 pr\u00e9sent : que viandes jonchant le sol, que mati\u00e8res se m\u00ealant \u00e0 la for\u00eat. Partout ce n'\u00e9tait que bidoches ti\u00e8des et sirupeuses, sans la moindre trace du plus infime sentiment. Car je n'aper\u00e7us pas plus de ch\u00e9rissement dans le reste de l'inf\u00e2me rago\u00fbte des chairs de p\u00e8re que dans son c\u0153ur.\n\n\u00c0 travers les branches nues du grand h\u00eatre, la vo\u00fbte p\u00e2lissait quand je finis mon ouvrage.\n\nC'est ainsi que les bourgeois du village nous trouv\u00e8rent. Dans la nuit, les criailleries effrayantes de p\u00e8re avaient alarm\u00e9 les habitants. \u00c0 l'aube, quelques-uns \u00e9taient accourus \u00e0 la cabane, curieux de s'instruire de la nature de ces bruits montant noctantement du plus \u00e9pais de la for\u00eat et de notre cabane, la cabane des Courge. \u00c0 leur arriv\u00e9e, p\u00e8re reposait sous le grand h\u00eatre en une paire de parties, si j'ose dire : son squelet, sa couenne, ses tripes, musclures et tubulures jonchant le tapis de neigette. Et sa pelure encore ti\u00e8de de ses substances reposant entre mes bras. Car depuis tout \u00e0 l'heure j'\u00e9tais rest\u00e9 ainsi, caressant la peau vidang\u00e9e de p\u00e8re, pleuroyant comme jouvenceau, d\u00e9versant sur ce reste mou et bilieux tout le pleur jaillissant de mes yeux.\n\n#\n\nEt ainsi fus-je men\u00e9, par ces quelques bourgeois, \u00e0 l'hospite dans lequel p\u00e8re et moi avions s\u00e9journ\u00e9 d\u00e9j\u00e0. Je n'offris nulle r\u00e9sistance, me soumettant par suite tout mollement aux doctes et \u00e0 leurs cures. Car on concevait volontiers que j'avais besoin d'\u00eatre cur\u00e9, en effet, mon apparence et mes mani\u00e8res rappelant \u00e0 pr\u00e9sent celles des insens\u00e9s, mes quelques rarissimes paroles \u00e9tant enti\u00e8rement d\u00e9pourvues de clart\u00e9 et la direction m\u00eame de ma vie ne faisant plus partie de ma pr\u00e9occupation. D'une lune \u00e0 l'autre, je demeurais ainsi allong\u00e9 sur la paillasse de l'hospite, en parfait \u00e9tat de l\u00e9gumerie, apparent\u00e9 en mon penser au lycopode, \u00e0 la foug\u00e8re ou \u00e0 quelque plante croissant en for\u00eat.\n\nApr\u00e8s m'avoir fait avaler maintes granules et astreint aux plus grands examens de casque et de corps, on me posta comme \u00e0 demeure dans une chambrette, o\u00f9 je restai une dur\u00e9e qui me parut fort longue. Chaque jour venait sur l'autre et rempla\u00e7ait l'ancien avec les m\u00eames apparences semblablement : rien ne paraissait diff\u00e9rent qu'en hier. Il ne me venait m\u00eame plus en casque l'id\u00e9e de me plonger en barrique et de m'y briquer sous eau ti\u00e9dasse. Manon n'\u00e9veillait plus en moi le moelleux transport d'avant. Ces jours-l\u00e0 s'\u00e9taient envol\u00e9s, \u00e0 pr\u00e9sent.\n\nPar matin, le soleil s'\u00e9branlait puis marchait en azur, jusqu'\u00e0 p\u00e9rir \u00e0 ponant. Se composait alors la vo\u00fbte nocturne avec sa cohue d'astres et son silence tonitruant. Car les \u00e9toiles, m\u00e9ditais-je faiblottement, sont \u00e0 la fois coites et volubiles. On ne les entend gu\u00e8re, mais on y devine une signification. Comme j'aurais d\u00e9sir\u00e9 lire en elles, comme p\u00e8re le faisait aussi ais\u00e9ment qu'en livre ! Et \u00e0 la v\u00e9rit\u00e9, aucun autre moment que celui-l\u00e0 ne me rappelait davantage p\u00e8re. Hors lui, y eut-il jamais sur terre bourgeois au silence plus entrem\u00eal\u00e9 de fracas ? Car si p\u00e8re n'usait que peu des paroles, sa conduite en revanche criaillait, oui, elle criaillait.\n\nCes pensers et ressouvenirs emplissaient mes nuits et me trouvaient \u00e0 l'aube \u00e9puis\u00e9, bris\u00e9 comme f\u00e9tu, pleuroyant extr\u00eamement et p\u00e9n\u00e9tr\u00e9 de l'image obs\u00e9dante de p\u00e8re. Un docte arrivait alors dans la chambrette, escort\u00e9 par quelque garde-souffrant expert en l'art de piq\u00fbre. On me piquait effectivement, on me faisait avaler encore rem\u00e8des et solutions, puis je d\u00e9lassais. On quittait la chambrette, et quand je retrouvais balance on rebroussait, on me questionnait. Je r\u00e9pondais maigrement, ne sachant moi-m\u00eame pas tr\u00e8s assur\u00e9ment comment se comportaient ni mes chairs, ni mon casque, ni toute autre portion de ma personne.\n\nPuis au fil des jours me sont rebrouss\u00e9es, au milieu de mes d\u00e9r\u00e8glements, diverses et fragiles songeries de Manon. Je revoyais alors sa douce face, je croyais fr\u00f4ler \u00e0 nouveau sa main et entendre encore ses paroles impr\u00e9gn\u00e9es de secours. Je rego\u00fbtais en pensers, quoique petitement, les joies de mes plong\u00e9es en barrique. Il me paraissait que, de ma vie, jamais je n'avais \u00e9t\u00e9 autant ch\u00e9ri qu'en ces brefs instants o\u00f9 la bonne Manon m'avait men\u00e9 par les rues du village puis d\u00e9terg\u00e9 vigoureusement. Jamais, hormis par la marmotte chez qui je s\u00e9journai \u00e0 mon premier jour, n'avais-je en somme autant senti l'enchantement de la proximit\u00e9. J'observais ainsi que, par une autre bizarrerie du sort, j'avais \u00e9t\u00e9 men\u00e9 \u00e0 conna\u00eetre ce sentiment, ce mouvement, ce saisissement mirobolant et insolite \u00e0 cause de p\u00e8re lui-m\u00eame, qui pourtant m'en avait toujours soustrait. Obscurit\u00e9 du sort ! S'instruira-t-on jamais du chemin imp\u00e9n\u00e9trable qu'empruntent les circonstances, du tour fumeux qu'adopte si souventes fois l'existence ? Ainsi, Monsieur le juge, qui aurait seulement pu entrevoir que je puisse, moi, le fils Courge, si semblable \u00e0 la feuille du bouleau dans sa sensiblet\u00e9 et son fr\u00e9missement, fendre les chairs de mon parent et lui infliger d'aller rejoindre l'outre-monde qu'il redoutait tant ? Comment songer qu'un fils aussi affam\u00e9 de sentiment, de rapprochement et d'\u00e9lan en arrive \u00e0 achever son propre p\u00e8re ?\n\nMais \u00e0 force de m\u00e9ditation et de solitaire d\u00e9lib\u00e9ration, je puis fort ais\u00e9ment, aujourd'hui, trouver r\u00e9pons \u00e0 cette question. Car, enfin, o\u00f9 nous reste-t-il \u00e0 chercher chez notre semblable, alors que nous ne trouvons ni sur sa face, ni sur ses mains, ni en sa conduite, ni m\u00eame en son silence le ch\u00e9rissement tant convoit\u00e9 ? O\u00f9 fouiller sinon en son c\u0153ur m\u00eame, en ses chairs les plus secr\u00e8tes ? Aussi je vous questionne tous, ci-devant moi rassembl\u00e9s avec mission de trancher mon cas : quel fut mon tort ? \u00c0 quelle coupablerie suis-je aujourd'hui r\u00e9duit ? D'avoir voulu \u00eatre ch\u00e9ri au point de sonder les profonds les plus \u00e9pais de p\u00e8re ? D'avoir ambitionn\u00e9 de conna\u00eetre en quelle tani\u00e8re amour se terre ? Mais alors, qui, de nous tous, r\u00e9sidents humains de la Terre, est \u00e0 ce jour blanc de faute ? De la premi\u00e8re jusqu'\u00e0 la finale heure de l'existence, ne sommes-nous pas, tous unanimement, chercheurs de ch\u00e9rissement, et comme sourciers de c\u0153urs battants ? Ne d\u00e9sirons-nous pas ou\u00efr le timbre exquis d'un gratte-cordes ci-dedans nos aim\u00e9s ? Voici ma voix : c'est analphabisme d\u00e9solant que de ne pouvoir lire en cieux par soirs d'\u00e9toiles. Mais c'en est un bien plus d\u00e9solant encore que celui de ne pouvoir lire en nos semblables humains.\n\n#\n\nL'\u00e9poque vint o\u00f9 je me sentis relever. Cela fut notable par l'app\u00e9tit que j'eus un matin de me plonger \u00e0 nouveau en barrique. J'en fis le v\u0153u aupr\u00e8s d'un garde-souffrant. Ce jour-l\u00e0, je fus donc infus\u00e9 puis d\u00e9terg\u00e9 longuement. Il me sembla d\u00e8s lors doucettement retrouver en moi le bourgeois que j'avais cru \u00e0 jamais \u00e9gar\u00e9. Recouvrant facult\u00e9 d'enthousiasme ci-devant l'image int\u00e9rieure de Manon, je reconqu\u00e9rais par m\u00eame occasion une sorte de verdeur. Je m\u00e2chais cette inexplicablet\u00e9 : tandis m\u00eame que nous nous concevons sur le seuil du tr\u00e9pas, il arrive que des forces obscures et nouvelles nous rebroussent \u00e0 la vie, nous forcent \u00e0 rena\u00eetre.\n\nUne dur\u00e9e coula encore. J'\u00e9mergeai, \u00e0 la fin, de l'\u00e9tat v\u00e9g\u00e9teux o\u00f9 j'avais chu. On con\u00e7ut d\u00e8s lors que j'\u00e9tais propre \u00e0 me rem\u00ealer \u00e0 l'humanit\u00e9. Menottes au poing et entrave \u00e0 cheville, je fus donc bient\u00f4t voitur\u00e9 par policemen jusqu'\u00e0 une ge\u00f4le, ce logement que certains ont baptis\u00e9 _taule_. J'y coudoyai nombre de bourgeois stationnant l\u00e0 depuis jolie lurette, partageant avec moi semblable destin de coquin, de fripouille et d'astucieux. On m'y boucla en chambrette barrel\u00e9e, avec fen\u00eatre semblablement grill\u00e9e, paillasse moderne et bourrel\u00e9e de faux polytric. J'y fus fort confortablement log\u00e9.\n\nL\u00e0, pendant que je patientai longues p\u00e9riodes que la justice puisse attaquer mon cas, et afin d'aider au passage du temps, on m'instruisit de vocabulaire. Car il y avait, en ce lieu, \u00e9cole accord\u00e9e \u00e0 ceux d'entre les fripons qui ambitionnaient de s'\u00e9lever au-dessus de leur condition vulgaire. Je con\u00e7us de cette sorte, au fil des jours, non seulement nombre de mots, mais pareillement leur sens et leur mesure, et mieux encore le pouvoir de leur r\u00e9sonance en esgourdes d'autrui. Qui d\u00e9mentira en effet que paroles comp\u00e9tentes rapportent davantage que mots \u00e9chevel\u00e9s et trou\u00e9s de silence ? Telle \u00e9tait ma condition avant que d'aboutir en ge\u00f4le : j'allais, comparable en cela \u00e0 p\u00e8re, ainsi qu'un sous-parleur. Ce n'\u00e9tait pas, comme lui, que je discourais peu. Mais mon discours \u00e9tait bancroche et b\u00e9quillard. Si j'avais \u00e9t\u00e9 plus t\u00f4t p\u00e9n\u00e9tr\u00e9 de vocabulaire, peut-\u00eatre aurais-je pu mieux aborder p\u00e8re, et ainsi m'en mieux faire comprendre. Et peut-\u00eatre aurais-je pu, par la seule puissance fouillante des mots, le creuser de meilleure fa\u00e7on et, \u00e0 la fin, d\u00e9goter le lieu secret en sa personne o\u00f9 amour se dissimulait, et voir cet amour-l\u00e0 jaillir en air \u00e0 l'entour comme alouette quittant le nid.\n\nCar forte parole est outil. Mais moi, je ne sus qu'empoigner coutelas. Et forte parole est lorgnette tonique de regard. Mais moi, je ne sus user de mon \u0153il qu'\u00e0 la mani\u00e8re des taupes : myopiquement, et sans m\u00e8che allum\u00e9e en sa pupille.\n\nAh ! si j'avais con\u00e7u plus t\u00f4t que les mots sont comme cl\u00e9s de glotte, et que par eux se d\u00e9font les obscurit\u00e9s du secret ! Si j'avais connu avant d'\u00e9vider p\u00e8re les paroles emplisseuses d'entendement, et avoueuses de sentiment, et traductrices de gouffre humain ! Si j'avais su que discours est comme cl\u00f4ture \u00e0 troupeau : qu'il restreint la dispersion de soi, et en autorise l'agroupement, et en favorise le guet et la meilleure connaissance. Oh ! si j'avais pu conna\u00eetre \u00e0 plus h\u00e2tive heure la parole qui nomme toutes choses, y compris amour, et ainsi leur donne corps et forme concr\u00e8te, si concr\u00e8te qu'on peut d\u00e9sormais voir, voir toutes choses ! Ah ! si j'avais su que vocabulaire est ainsi que le drap pos\u00e9 sur le fant\u00f4me, lui donnant apparences et dehors, et lui retirant enfin sa d\u00e9testable invisiblet\u00e9 !\n\nAussi je dis : heureux soit le parleur, car avec ses mots il est pareillement au bourgeois qui voit \u00e0 l'int\u00e9rieur de ses semblables, qui les d\u00e9membre, les \u00e9visc\u00e8re, mais sans user pour autant de son coutelas, et d\u00e9chiffre leurs teneurs, et traduit la substance de leurs c\u0153urs, et voit ainsi v\u00e9ritablement l'amour. Heureux, les parleurs ! Car ils ont l'\u0153il clairvoyant et traverseur de corps, et fouilleur d'\u00e9paisseurs ! Et, de m\u00eame, celui qui poss\u00e8de forte parole \u00e9vite que le tranchant de l'outil d'autrui ne lui d\u00e9coupe ses propres viandes et beurres : car celui-l\u00e0 use de son discours comme d'un livre ouvert, et tourne les pages de son histoire \u00e0 qui les veut lire. Oui, heureux, les parleurs ! Car ils voient de leur \u0153il le sentiment, ce raton terr\u00e9 en nos tani\u00e8res et logements de chairs ! Heureux, heureux, les parleurs ! Car ils avancent sur la vie et reculent l'horizon de la mort, peupl\u00e9e de taiseurs !\n\nEnfin, au bout d'une \u00e9poque, on me renseigna que justice \u00e9tait par\u00e9e \u00e0 me recevoir. Ce fut le bout de mes le\u00e7ons de mots, \u00e9tait venu le temps du quittement de tout cela. Je fis adieux \u00e0 mes compagnons de bagne ainsi qu'\u00e0 mes garde-coquins. Ne poss\u00e9dant que clous et n\u00e9ant, je n'emportai par-devers moi aucun bien, hormis le v\u00eatement de captif dont on avait accoutr\u00e9 ma personne \u00e0 mon arriv\u00e9e. Je portais toutefois et \u00e0 jamais au plus \u00e9pais de moi-m\u00eame l'avoir le plus pr\u00e9cieux : la parole, la parole qu'on m'avait autoris\u00e9 \u00e0 prendre en me faisant enseignement de vocabulaire, et qui, \u00e0 pr\u00e9sent, \u00e9clairait mon pas dans les t\u00e9n\u00e8bres du monde. Car forte parole est outil, et lampe dans le soir.\n\nPuis on me mena bient\u00f4t ci-devant vous, Monsieur le juge, et pareillement sous vos regards, Bourgeois et Cr\u00e9atures composant le jury de ce tribuneau, afin que je narre mon histoire, que je reforme pour vous le parcoursement de mes pas sur la Terre, et qu'on en puisse juger.\n\nC'est, \u00e0 pr\u00e9sent, mati\u00e8re accomplie.\n\n#\n\n_J_ _e terminerai mon r\u00e9cit, Monsieur le juge, en risquant ceci : j'incline \u00e0 croire que si parfois les morts paraissent aux vifs, c'est afin de continuer de leur t\u00e9moigner quelque sentiment et bont\u00e9s. Ainsi vis-je toujours m\u00e8re para\u00eetre, ainsi p\u00e8re lui-m\u00eame vit-il para\u00eetre, une nuit, ses propres p\u00e8re et m\u00e8re : pour qu'ils puissent, par-del\u00e0 les limites de l'ici-bas, manifester leur ch\u00e9rissement, m\u00eame si ce ch\u00e9rissement-l\u00e0 ne peut jamais supplanter celui des vifs. Car les morts sont avec les morts, et les vifs, avec les vifs : on ne saurait passer d'un monde \u00e0 l'autre si commod\u00e9ment._\n\n_Mais c'est un tour singulier du sort que, depuis son tr\u00e9pas, p\u00e8re n'ait jamais paru ci-devant ma personne. Comment traduire cela ? Je le puis dire aujourd'hui : simplement, la hotte int\u00e9rieure de p\u00e8re ne contenait plus assez de sentiment pour moi. N'avait-il pas tout donn\u00e9 \u00e0 m\u00e8re ? Ce fut l\u00e0, assur\u00e9ment, toute mon infortune : de n'\u00eatre apparu sur cette Terre qu'apr\u00e8s m\u00e8re, et lorsque les machineries de p\u00e8re, exsangues et calcin\u00e9es, fumaient encore des suites du grand sinistre de son enfance. Apr\u00e8s le tr\u00e9passement de m\u00e8re, p\u00e8re aura employ\u00e9 toute sa vie \u00e0 tenter de r\u00e9parer cette mort h\u00e2tive, imm\u00e9rit\u00e9e et douloureuse. Aura-t-il abouti ? Sera-t-il parvenu, \u00e0 la fin, \u00e0 quelque adoucissement de sa douleur ? Peut-\u00eatre ses gens auront-ils servi \u00e0 cela : \u00e0 le faire p\u00e9n\u00e9trer en un monde avoisinant celui o\u00f9 m\u00e8re existait d\u00e9sormais, un monde r\u00e9el, quoique d\u00e9pourvu, \u00e0 notre \u0153il, de sens et de direction._\n\n_Et par m\u00eame raisonnement je con\u00e7ois aussi, \u00e0 pr\u00e9sent que je puis par les mots cerner les choses et les d\u00e9chiffrer meilleurement, que si, apr\u00e8s sa mort, m\u00e8re ne parut jamais \u00e0 p\u00e8re, c'est qu'elle m'aura, depuis l'outre-monde, fait don entier de son sentiment. Oui, cette chose que m'apportait toujours m\u00e8re lorsqu'elle paraissait, ce tremblement qui faisait cogner mon c\u0153ur en poitrin, cela se nommait amour. Oui, je fus assur\u00e9ment et fortement aim\u00e9, je le fus, je le fus. Seulement je ne l'aurai \u00e9t\u00e9 que d'une morte, d'une morte seulement._\n\n_Aussi est-ce aujourd'hui finalement que je m'instruis de la signification du chagrin sans cesse lisible sur la face de m\u00e8re et sur celle de chaque macchab\u00e9e. Cette m\u00e9lancole tient en ce que vie et mort demeurent \u00e9ternellement contraires, et qu'entre elles nul discours, nulle communication durable et nul feu aprofond ne sont possibles. Aucun couloir ne les r\u00e9concilie jamais. Ainsi, m\u00e8re est triste parce que vie est timbre, et note, et musique, et embrasement, cependant que mort est silence, et mutisme, et d\u00e9laissement, et d\u00e9solation._\n\n_En fruit de tout cela, je dis : fasse donc la d\u00e9esse Lune, veillant sur salut humain, que bient\u00f4t potence me rompe le col et me m\u00e8ne aupr\u00e8s de p\u00e8re et m\u00e8re dans le silence des choses, et qu'ainsi reli\u00e9s nous devenions premiers macchab\u00e9es \u00e0 sourire enfin._\n\nPour joindre l'auteur :\n\njfbeauchemin@aol.com\n\net sur Facebook  \n\n#\n\n","meta":{"redpajama_set_name":"RedPajamaBook"}}
+{"text":"\n#  \n\n#  \nCopyright \u00a9 2018 by Steven Rosenfeld\n\nForeword Copyright \u00a9 2017 by David Talbot\n\nAll rights reserved. No part of this book may be reproduced in any manner without the express written consent of the publisher, except in the case of brief excerpts in critical reviews or articles. All inquiries should be addressed to Skyhorse Publishing, 307 West 36th Street, 11th Floor, New York, NY 10018.\n\nHot Books may be purchased in bulk at special discounts for sales promotion, corporate gifts, fund-raising, or educational purposes. Special editions can also be created to specifications. For details, contact the Special Sales Department, Skyhorse Publishing, 307 West 36th Street, 11th Floor, New York, NY 10018 or info@skyhorsepublishing.com.\n\nHot Books\u00ae and Skyhorse Publishing\u00ae are registered trademarks of\n\nSkyhorse Publishing, Inc.\u00ae, a Delaware corporation.\n\nVisit our website at www.hotbookspress.com\n\n10 9 8 7 6 5 4 3 2 1\n\nLibrary of Congress Cataloging-in-Publication Data is available on file.\n\nCover design by Brian Peterson\n\nPrint ISBN: 978-1-5107-2945-2\n\nEbook ISBN: 978-1-5107-2946-9\n\nPrinted in the United States of America\nI am trying to write history while it is happening\n\nand I don't want to be wrong.\n\n\u2014John Steinbeck\nCONTENTS\n\nFOREWORD BY DAVID TALBOT\n\nINTRODUCTION\n\nSECTION I:\n\nThe Democrats\n\n 1 BERNIE SANDERS\n\n 2 WHO WON IOWA?\n\n 3 LAS VEGAS RULES\n\n 4 NEW QUESTIONS, NEW BARRIERS\n\n 5 THE NOT-SO-GOLDEN STATE\n\n 6 THE DEMOCRATIC NATIONAL COMMITTEE\n\nSECTION II:\n\nThe Republicans\n\n 7 GAMING THE RULES OF POLITICS\n\n 8 CREATING PARTISAN GHETTOS\n\n 9 LOOKING FORWARD, NOT BACKWARD\n\n10 THE VOTING WARS\n\n11 THE STARTING LINE\n\n12 THE FINISH LINE\n\n13 THE PLAYBOOK: DO EVERYTHING\n\nSECTION III:\n\nThe Recounts\n\n14 WHAT HAPPENED?\n\n15 THE FINAL SWING STATES\n\n16 THE ELECTORAL COLLEGE\n\nAFTERWORD\n\nACKNOWLEDGMENTS\n\nENDNOTES\n\nINDEX\nFOREWORD\n\nBY DAVID TALBOT\n\nTHE WORLD IS BURNING, AND YET the firelight illuminates the way out. The times are dire, even catastrophic. Nonetheless we can sense a grand awakening, a growing realization all around the globe that \"people have the power, to dream, to rule, to wrestle the world from fools\" in the prophetic words of Patti Smith.\n\nBut in order to rouse ourselves from the nightmares that hold us in their grip, we need to know more about the forces that bedevil us, the structures of power that profit from humanity's exploitation and from that of the earth. That's the impetus behind Hot Books, a series that seeks to expose the dark operations of power and to light the way forward.\n\nSkyhorse publisher Tony Lyons and I started Hot Books in 2015 because we believe that books can make a difference. Since then the Hot Books series has shined a light on the cruel reign of racism and police violence in Baltimore (D. Watkins' The Beast Side); the poisoning of U.S. soldiers by their own environmentally reckless commanding officers (Joseph Hickman's The Burn Pits); the urgent need to hold U.S. officials accountable for their criminal actions during the war on terror (Rachel Gordon's American Nuremberg); the covert manipulation of the media by intelligence agencies (Nicholas Schou's Spooked); the rise of a rape culture on campus (Kirby Dick and Amy Ziering's The Hunting Ground); the insidious demonizing of Muslims in the media and Washington (Arsalan Iftikhar's Scapegoats); the crackdown on whistleblowers who know the government's dirty secrets (Mark Hertsgaard's Bravehearts); the disastrous policies of the liberal elite that led to the triumph of Trump (Chris Hedges' Unspeakable); the American wastelands that gave rise to this dark reign (Alexander Zaitchik's The Gilded Rage); the energy titans and their political servants who are threatening human survival (Dick Russell's Horsemen of the Apocalypse); the utilization of authoritarian tactics by Donald Trump that threaten to erode American democracy (Brian Klaas's The Despot's Apprentice); the capture, torture, and detention of the first \"high-value target\" captured by the CIA after 9\/11 (Joseph Hickman and John Kiriakou's The Convenient Terrorist); and the deportation of American veterans (J Malcolm Garcia's Without a Country). And the series continues, going where few publishers dare.\n\nHot Books are more condensed than standard-length books. They're packed with provocative information and points of view that mainstream publishers usually shy from. Hot Books are meant not just to stir readers' thinking, but to stir trouble.\n\nHot Books authors follow the blazing path of such legendary muckrakers and troublemakers as Upton Sinclair, Lincoln Steffens, Rachel Carson, Jane Jacobs, Jessica Mitford, I.F. Stone and Seymour Hersh. The magazines and newspapers that once provided a forum for this deep and dangerous journalism have shrunk in number and available resources. Hot Books aims to fill this crucial gap.\n\nAmerican journalism has become increasingly digitized and commodified. If the news isn't fake, it's usually shallow. But there's a growing hunger for information that is both credible and undiluted by corporate filters.\n\nA publishing series with this intensity cannot keep burning in a vacuum. Hot Books needs a culture of equally passionate readers. Please spread the word about these titles\u2014encourage your bookstores to carry them, post comments about them in online stores and forums, persuade your book clubs, schools, political groups and community organizations to read them and invite the authors to speak.\n\nIt's time to go beyond packaged news and propaganda. It's time for Hot Books... journalism without borders.\nINTRODUCTION\n\nIT'S MID-DECEMBER 2016 IN NEW YORK City. The sting of the presidential election has not worn off. There's a harsh wind blowing as I'm cutting across Lincoln Center for a meeting I am drawn to, but also wary about. I have been a national political reporter since the 1990s at various outlets, some mainstream like NPR and others in progressive media. Over the years, I've covered elections deeply. This is not just candidates and speeches, but who can vote, who can't, how that happens, and whether the count can be trusted. I've been pulled back into the vortex of our elections where, despite efforts to expand the vote and improve our system, the more ruthless and smarter forces have won again. You and I\u2014and the vast majority of our fellow citizens\u2014are the losers.\n\nI've been here before. In 2004, I spent weeks, then months, then two years with others tracing and writing about what happened in Ohio where Republican President George W. Bush beat Democrat John Kerry. We produced a catalog of dirty tricks that Republicans have resurrected with predictable uniformity ever since. Donald Trump's surprise victories and the US Senate staying in Republican hands felt eerily familiar. Many explanations in the aftermath did not add up. In 2004, we were told a wave of rural, white southern Ohio evangelicals had reelected Bush. We looked but didn't find them. Now a different white flock\u2014economically struggling and lacking higher education, we were told\u2014elected Trump. And the election was not even entirely over.\n\nI was to meet Lulu Friesdat, a filmmaker who had been taking time off from her job as a CBS-TV producer to film the 2016 presidential recount in Wisconsin. We had talked on the phone about what she saw and shot. These were breakdowns where no one in an official role could verify to the satisfaction of observers and computer scientists what constituted the actual final count. Lulu witnessed it, filmed it, and was mortified. I posted her clips on AlterNet.org. She already concluded that the Democratic Party had kept the nomination from Bernie Sanders. I agreed, though we differed on particulars of its antidemocratic culture and playbook. Now Friesdat was discovering another facet of the way elections really work in America.\n\nHer most memorable clip showed a process filled with earnest citizens; banal officials; and maddening, unnecessary complications rooted in arcane election rules. Friesdat had gone to the Racine County government center, not far from Lake Michigan. Outside the center, people were rooting for the Green Bay Packers or shopping for Christmas\u2014but inside, the center was a world apart. Most people thought the election was over. They turned away from states certifying results, the Electoral College convention, and then Congress ratifying that vote. Interrupting this progression was the Green Party's feisty candidate, Jill Stein, who had a legal right to file for recounts in the last three states\u2014Wisconsin, Michigan, and Pennsylvania\u2014giving Trump his apparent victory.\n\nStein supporters unexpectedly had donated millions over the Thanksgiving holiday, after her legal team could not convince Hillary Clinton's campaign to enlist some of their donors. That grassroots response was completely unexpected. Like much of America, Stein's donors questioned 2016's official results. They were part of 2016's grassroots surge seeking real change, especially millennial youths. They wanted to know what happened. I'd seen the Greens file for a recount in 2004 in Ohio. That effort did not change its outcome. But it did help expose a spectrum of tactics in the fine print of voting that confirmed the GOP was committed to strategically targeting and suppressing known Democratic Party blocs. In other words, the voting process had been turned against its rightful participants.\n\nInside the government center, which could be any election office in America, the slow methodical business of recounting an election churned. A team of middle-aged women in sweaters parsed and fed paper ballots one-by-one into an electronic scanner. Three Green Party observers, led by Liz Whitlock from nearby Mount Pleasant, stood back and patiently watched. These Greens were well-groomed citizens in their middle years, anything but wild Bernie bros from the spring campaign. They wanted Racine County to hand count its paper ballots, citing academics that said electronic scanners like those in use have implicit error rates from misreading ink marks.\n\nThey wanted to know if Trump had really won, by how much, where his voters were, or what wasn't accurate in the process. But Joan Rennert, a veteran local official overseeing that day's recount, was losing patience in a plodding exercise where the final outcome was unlikely to change. Across the state in liberal Madison, a judge gave each Wisconsin county the option of doing a hand recount or not. Racine declined.\n\nSo Whitlock and her colleagues devised a simple work-around. They bought manual counters, mechanical hand clickers. They clicked away as each ballot slipped into the electronic scanner, tallying how many votes were for Clinton, Trump or blank\u2014a so-called undervote. They kept noticing their totals varied from the scanners. When Whitlock's team saw the scanner miscount fifteen votes in a three hundred\u2013voter precinct in Elmwood Park, she politely asked officials to do a hand count of that precinct. That was an error rate of 5 percent in a contest, where statewide, Trump's margin of victory was less than 1 percent. Whitlock's request was swiftly denied.\n\nThe Greens huddled. They were upset. Whitlock then filed a challenge, a paper document that stopped the process. A face-off ensued with Rennert taking charge. Her bevy of women watched and waited, arms folded across chests, hands on their hips. Whitlock's team bowed their heads but stayed firm. It looked like a standoff between teachers and the principal's office. All appeared well-meaning. But it spiraled downhill as Friesdat's camera kept rolling.\n\n\"Three observers click counted votes. The Clinton and Trump counters clicked considerable more votes than a scanner counted,\" said Rennert, wearily reading aloud the challenge. \"I have no idea what you're saying except you are requesting a hand count.\"\n\nRennert put down the paper. She looked at the room, palms up, exasperated. She said nothing.\n\n\"Do you want to ask me if you don't understand?\" asked Whitlock.\n\n\"No,\" Rennert tartly replied. \"Is this the purpose of a hand count? Yes or No?\"\n\n\"The purpose of a hand count is to get to the truth,\" Whitlock replied.\n\n\"No,\" Rennert said, her hands up like stop signs. \"How much is it?\"\n\n\"Three-hundred,\" a worker said, referring to the number of ballots.\n\n\"I don't care if it's five,\" Rennert declared. \"I am not going to do a hand count for anybody.\"\n\nI don't have to tell you who had the final word. Nor whether the official vote count matched the hand clickers or computer scanners. Stubbornness and institutional inertia won.\n\nThis small scene points to one of the great ironies of our time. Our system of voting has increasingly subverted American democracy. How that has happened, how it was seen in 2016 and its precursors this decade, and why that matters is the subject of this book. The antidemocratic features in our elections are betraying the country, its citizens, and our collective future. This is a bigger problem than which party or factions win or lose\u2014although those most responsible for this crisis, the Republicans, have triumphed.\n\nThe scene in Racine's government center was a microcosm of the fragility, imperfections, opaqueness, and frustrations that all too often comprises American elections. The volunteers and workers weren't bent on bad will, but they weren't to up to the job, given its seriousness and its stakes. Their ineptitude and stonewalling were less excusable because these weren't overwhelmed poll workers, but arms of local government that should have known and done better.\n\nNotably, this snapshot didn't even involve hard-core political types\u2014the partisans, in our era almost exclusively Republicans, who deliberately undermine their opponents' base at many key steps along the way before the vote count. I had been in Friesdat's shoes, discovering slights that undermined the public's expectation that every vote mattered, every voter was equal and all ballots would be counted. When I went to meet her, the unofficial results\u2014another GOP sweep\u2014were raw. I had not delved into the minutiae of elections for several years. I wasn't a newcomer, but 2016 was different\u2014and not just Trump's ugly path.\n\nAfter reporting on Ohio, I covered the 2008 election as a voting rights beat and saw how President Obama's landslide breached the anti-voter barriers we documented. These were micro-aggressions that would have favored Republicans in a closer race had the economy not crashed that fall. These attacks are embedded in the fine-print rules of voting. They are almost always presented by politicians and regulators as neutral changes or safeguards or colorblind. In reality they shape who votes and is likely to win. They favor a status quo that is almost always the GOP's shrinking, aging, largely white, and suburban base.\n\nFor example, restricting the forms of state-issued ID required to get a ballot at a polling place. Or adding an extra requirement on a state voter registration form that's not on the federal form, found in every post office. Those moves block students, people of color, and the poor, just as they are intended to do. Or curtailing early voting on the weekend before Election Day, when clergy urge congregations to go to the polls. Or rejecting ballots turned in at the wrong table in a polling place. Or darker tactics such as top state officials knowingly using error-prone databases to purge infrequent voters, but not telling anyone until they show up to vote. Or deliberately segregating Democratic voters into a smaller number of urban districts and spreading out those in suburbs and rural areas when drawing political maps so that the Democrats will never win.\n\nIn 2009, I joined a team at the Pew Center on the States to modernize voter registration. After Obama's blue wave, I felt improving turnout was the key to changing our politics. Pew's ensuing data center included identifying all the eligible voters in a state and then contacting them and urging them to register. By 2016, it was used by twenty states and helped break a national record for registered voters. But most of these new voters were not in the swing states. Something else was going on in states that determined who controlled the US House, that kept turning purple states red and was pivotal in electing the president. As I was drawn into covering 2016's recounts, I was jarred by a series of deepening reminders that whatever progress had been made was being overshadowed where it mattered the most.\n\nI knew what Friesdat thought in Racine. This was not the democracy we expect, not when citizens cannot verify the count. She was doing math in her head\u2014just like we did in Ohio. If the voting machine error she saw was anything close to routine in a state with 2,976,150 presidential votes, it could mean up to 150,000 ballots were incorrectly tallied. That might negate Trump's 22,000-vote victory. It might not. The only way to find out was not widely happening\u2014observable hand counts. Of course, it is easy to say this is reading too much into one incident. But I, too, kept seeing a system not defaulting on behalf of voters\u2014America's citizens.\n\nMy moment of outrage came a few days before. In Michigan's capital, Chris Thomas, the white-haired, taciturn, veteran state election director was lecturing reporters about his state's recount that had just begun. Thomas has survived decades of working for Democrats and Republicans. This includes Republicans that glibly say that hordes of phantom Democratic voters are stealing elections and not getting caught, which is a knowing lie. I'd met him in Pew's effort and thought he was shrewd enough to avoid partisan ploys. I was wrong.\n\nThomas cultivated a hardheaded persona. But he had also finessed an impressive work-around for people who, at the last minute, show up at polling places and aren't on voter lists. High-stakes races always have people like that. They are eligible voters but never update their registration or bother to register. During 2016's caucuses and primaries, Democratic Party rules and administrative barriers blocked these enthusiasts from voting. That hurt Bernie Sanders in several key states. New York was the worst, where its rules blocked anyone from voting in April's primary that had not registered with a party six months earlier. Thomas devised a slick way to allow these impulsive people to vote while cutting red tape. Instead of giving them a provisional ballot\u2014which isn't counted right away because it has to be verified, he created a form with an oath. They would swear they were state residents, US citizens, eligible voters, and would vote via a regular ballot. That is unadvertised Election Day registration\u2014a pro-voter move that Republicans have shut down in many red states.\n\nBut now, in December 2016, Thomas was on Michigan TV stating nearly 60 percent of Detroit's ballots could not be recounted. The reason given was because poll workers had mislabeled the boxes the paper ballots were put in while cleaning up on Election night. Or the numbers of ballots listed outside the boxes didn't match the ballots inside. Most were off by a few, often less than ten. Or seals on storage boxes did not look perfectly tight. Or the boxes were banged up. To the GOP, this apparent sloppiness was turned into accusations of ballot tampering by Democrats. Imaginary threats to the integrity of the process are what Republicans love to cite and hype\u2014especially if it helps them win. They never put voters first and abide by the results. More than half of Detroit's presidential ballots were deemed ineligible for the recount. Votes that might have tipped the state were taken off the table.\n\nThat conclusion was just what Thomas' political bosses wanted. Michigan Republicans, led by the state Attorney General Bill Schuette, joined Trump's campaign to block a recount. They ignored Trump's rants that the voting was error prone and not-to-be-trusted until he won. They declared there was no reason to recount the state. Such double standards are predictable in this fold. They shrugged at the fact that Michigan had Trump's closest margin nationwide: just eleven thousand votes out of 4.8 million ballots cast. Some seventy-five thousand ballots did not show a vote for president, Michigan's secretary of state office reported on its website.\n\nThat last omission is always suspicious. That's because people tend to vote for the high-profile races if they vote at all. Maybe some of these seventy-five thousand ballots had presidential votes, but they were not properly scanned. If a good number were from around Detroit, which went two-to-one for Clinton, maybe Trump did not win Michigan after all. Local election activists pointed to a 1950s state law that gave discretion to election officials to examine and recount every paper ballot. But Thomas went on TV saying it would not be done. Detroit's election director followed his cue, apologizing for the sorry state of voting in his city.\n\nKnowing someone who could have quickly solved one of 2016's answerable questions rattled me. Thomas and Detroit officials could have sided with that black majority city's voters. Instead, this incident was a brazen reminder of the GOP ethic that one can never over-police the process or do too much to twist the rules in pursuit of victory. Michigan's GOP didn't care if a display of apparent institutional racism and voter suppression were a means to that end. Not surprisingly, they stopped the recount in court days later\u2014much as the Supreme Court had infamously shut down Florida's recount in 2000, making George W. Bush president.\n\nFollowing the recount was mesmerizing and jarring. It wasn't doing what a democratic process is supposed to do\u2014provide assurances that the result is accurate. Instead, it was exposing a starker and more cynical side of elections. The answer to Michigan's seventy-five thousand under-votes was not computer science. It was examining the vote: looking at individual ballots to see if a vote was there or not. That is straightforward. That was not the same as bringing in cyber detectives to hunt for code-altering fingerprints in voting software to see if the count is being tweaked to one side's benefit. That's what the Greens wanted to do in Michigan, Wisconsin, and Pennsylvania. But that wasn't to be, either.\n\nI'd also forgotten how rickety the country's voting machines are. Most are run on operating systems from the early 2000s\u2014before the first iPhones. The software is proprietary, i.e. trade secrets. That is a consequence of Congress' dumb decision to support a privatized voting machine industry after Florida's mess. The election transparency watchdogs I knew were more worried about overly eager Republicans or libertarians with hacking skills than the Russians, despite intelligence agencies saying Russia infiltrated the Democratic Party and several statewide voter registration databases. Officials from Obama to FBI Director James Comey said the vote-counting process was secure. They said the machinery was so old-tech, so decentralized, and not even connected to the Internet that it couldn't be gamed to flip statewide results, or, in turn, seize the presidency.\n\nThose reassurances buckled near Milwaukee. John Brakey, an Arizona voting transparency activist who knows more about the electronic vulnerabilities of voting machines than most local election officials, had gone to Wisconsin to observe the Greens' recount. He spotted a hacking pathway into the vote count tabulators. I am not by nature a conspiracy theorist\u2014especially the possibility that no matter how people vote, that Karl Rove's latest minions or Russians are prefiguring results. Yet in 2016, when almost everyone has been hacked, from the Pentagon to personal emails, and Volkswagen has preprogrammed emission systems to turn themselves off at smog tests, we can't pretend voting machines are magically immune.\n\nBrakey\u2014and a team including Silicon Valley coders and a retired National Security Agency analyst\u2014discovered precincts near Milwaukee used cell-phone modems to send their results to county tabulators. Why was the modem there? Because polls did not always have landline phones, officials said. The NSA analyst sent me links to commercial spyware used by police to intercept this data stream. Their point was Wisconsin's voting machines, like much of the nation, can easily be hacked. These vulnerabilities were identified more than a decade ago.\n\nThe Greens presented testimony from a half-dozen nationally known computer scientists saying these concerns were not a shrill conspiracy theory. Yet Wisconsin officials, echoing a stance taken in Pennsylvania\u2014the third state where the Greens were told they would have to spend millions for a recount\u2014ignored it. This wasn't just Republicans defending Trump. In Pennsylvania, Democrats oversee its elections. Helped by judges, they, too, shut the door on the Greens and blocked its recount. So more questions about the vote went unanswered. State and local officials thwarted inquires and evidence that could have assured the public.\n\nThe more I looked at various phases and features of the 2016 election, the more I saw a spectrum of antidemocratic features. That was also true at the process's start. It looked to me (and also to the Des Moines Register editorial page) like Sanders beat Clinton in Iowa's first-in-the-nation contest. I saw other tactics in Nevada, such as deciding where its caucuses were held and where they weren't, that made it easier for Clinton's supporters to participate. But beyond the anti-Sanders biases from top aides of DNC Chairwoman Debbie Wasserman Schultz, beyond scheduling televised candidate debates on weekends where audiences were smaller, and beyond shutting access to its national voter database as Sanders's grassroots donations soared, was the Democrats' antidemocratic super-delegate system.\n\nThese were 712 federal and state elected officials, party leaders, and lobbyists whose votes count toward the 2,384 needed for the presidential nomination. They preempt millions of citizens voting in caucuses and primaries. Most were with Clinton before Sanders entered the race and 570 stayed with her. Superdelegates are another example of a system filled with rules that privilege the few at the expense of many. This template keeps governing classes in place. But so, too, is who gets to vote, and whether those votes count.\n\nWhat Democrats did to Sanders was nasty. But it pales next to what Republicans have done to Democrats and the vast majority of Americans. That is what was different about 2016 as well as every congressional and state election cycle since President Obama's election in 2008. It is what sets this decade apart, as a party representing a minority has seized power and steadily imposed a regime of oligarchy, plutocracy, and theocracy on its fellow citizens. It is one thing when Democrats run presidential nomination contests like private clubs with varying membership rules. But it is another when Republican strategists plot to win dozens of state legislatures and House seats, and then sabotage their critic's ability to compete in elections. The country is clearly living with the results of a subverted democracy.\n\nPolitics has always had hardball tactics, but as this introduction is being written in late 2017, Trump's national approval ratings were 35 percent. \"Every party, gender, education, age, and racial group disapproves except Republicans,\" an August 23 Quinnipiac University poll reported. A week later, a poll by pro-Republican Fox News said 56 percent of voters \"say Trump [is] tearing the country apart.\"\n\nThink about it, how can 2016's presidential election come down to less than eighty thousand votes in Wisconsin, Pennsylvania, and Michigan, yet two-thirds of the US House seats in those states\u2014and even more seats in their legislatures\u2014end up as red super-majorities? All things are not being equal. This same pattern is also found, with slight variations, in Florida, Georgia, North Carolina, Ohio, Virginia\u2014all states where, if demographics were destiny, the political complexion should be purple turning blue or blue.\n\nWhat's different now is the mechanisms of the Republican's advantage can be quantified. But you have to look carefully. Beyond all of 2016's headline-grabbing features\u2014Trump's ugliness, Clinton's foibles, Sanders's populism, Russian meddling\u2014what unfolded was the Democrats did not know what they were up against. They did not appreciate how rigged the process was\u2014structurally stacked against them. Republicans had built in a starting-line lead of ten points or more. This isn't in every race. But it is enough of the ones that determine statewide rule in otherwise purple states and determines control of Congress. It also is seen in anti-voting laws and barriers in the presidential swing states.\n\nThere is no such thing as 50-percent-plus-one in high-stakes elections. To win in November, the party representing a majority of citizens needs very big margins that cannot be whittled away\u2014as they will be whittled away. There is a wide and deeper dynamic operating here. But to see its entirety, we have to parse the details that comprise the voting process.\n\nWhen I was drawn into the vortex of America's dysfunctional 2016 election via recounts, I did not think I'd emerge with an updated catalog of how the voting process preempts the citizenry and betrays democracy. It's so easy to be diverted into parallel concerns. That's especially true of the hope-deflating swamp of voting machine hacking, which, if it were widespread, renders all other efforts to improve the process futile. There are other big issues in the voting sphere, such as ex-felon disenfranchisement. That is why Florida is politically red, not blue. Florida's lifetime voting ban for its 1.6 million residents with a felony record, most nonviolent drug offenses, is a Jim Crow\u2013era holdover.\n\nNonetheless, I keep coming back to a big picture where Democrats, progressives, and independents do not understand the fundamentals they are up against. In far too many states, Republicans have restructured the voting process in antidemocratic ways. The stakes are bigger for the country than the fate of which party or factions win and lose. The GOP has undermined the process that is the basis for representative government. Unless stakeholders, meaning citizens who put democracy before party, work to restore the voting process, the legitimacy of its outcomes will remain dubious and our political system will unravel.\n\nThis book is divided into four sections. It starts with the Democrats' nominating contests. The goal is not to rehash the Sanders-Clinton split within the party in 2016 and since. What Democratic national and state leadership threw at Sanders is a useful way to start looking at a landscape where insiders game the rules of voting. But even as these features undermine the Democrats' moral authority as the party that defends voting rights, their flaws are of a different and lesser caliber than the Republicans' attacks on voting in America.\n\nThe second section focuses on Republican efforts to disenfranchise opponents and preempt representative government. This section, the book's core, shows readers a range of strategies with their impact. As the nation looks ahead to elections in 2018 and 2020, what's new now is that the most pernicious tactics can be quantified. It's no longer a guessing game about which anti-voter laws and rules are the most harmful to citizens and voting.\n\nThe third section continues with the recount, where both major parties frustrated efforts to verify the vote count, and where the prospect of hacking was raised, insufficiently answered, and dismissed. (By mid-summer 2017 hacking resurfaced, although it appeared the Russian interventions were more impactful with hurting Democrats on social media than with the voting machinery.) The book concludes with an afterword focusing on 2017 and the Trump administration, where a rogues' gallery of Republicans have resurfaced in powerful posts and are poised to recycle their anti-voter playbook.\n\nThere is a way forward for Americans who seek a government that represents the nation's multiracial, multi-class, increasingly diverse citizenry. It starts with seeing what has happened to voting, how the process has been hijacked to subvert democracy, and acting to surmount it. Come take a look.\nSECTION I\n\nTHE DEMOCRATS\n1\n\nBERNIE SANDERS\n\nBERNIE SANDERS HAS ALWAYS BEEN A very able, savvy, and visceral politician. When I briefly had a front-row seat to Sanders, as press secretary on the campaign that first elected him to the House of Representatives in 1990, he was championing many of the same stances that resonated with millions in 2016. By the time he launched his presidential campaign in mid-2015, he was more of a political pro than anything I'd seen in Vermont decades before.\n\nBut neither Bernie nor his earliest supporters\u2014practitioners of a \"now I'm inside, now I'm outside\" Democratic Party\u2014nor his legions of youthful volunteers quite knew what they were getting into. They were taking on the Democratic Party's national establishment, state party leaders, a rickety election system, stone-faced administrators, political rituals manned by bungling locals, and, of course, the Clintons, now in their fourth presidential campaign. Bernie had never run for anything outside of Vermont. He had not faced real competition or dirty electioneering tactics in years. Nor had he interacted with the party insiders and others running caucuses and primaries, nor the vagaries of voting rules and administration.\n\nTop campaign officials I spoke to in spring 2017 agreed that Bernie didn't know what he was getting into. But they said the Democratic Party didn't see him coming either, and as an institution it didn't respect the man, what he stood for, or his strengths as a candidate. The party's establishment had lost touch with its energized grassroots and younger base, which propelled a candidate who came out of nowhere to win 45 percent of the elected delegates to the Democratic National Convention. But it was not the party's blindness, nor its insider-favoring rules alone that prevented Sanders from breaking through. There also were party traditions whose intent wasn't nefarious, but whose effects were antidemocratic.\n\nStarting in 2015, Bernie had spoken at events set up by Progressive Democrats of America and others. Growing crowds clung to his words and embraced a refreshing and more blunt champion than Hillary Clinton. Bernie had a message that people wanted to hear and one he wanted to spread. Iowa's Caucuses, 2016's first testing ground, were in a state that was not that different from Vermont. It was mostly rural, white, and had progressive streaks.\n\nBut just because you want to run as a Democrat for president doesn't make it happen. You have to get on the ballot. The Democratic National Committee (DNC) didn't know what to make of his decision to seek their nomination. They knew of his old speeches condemning them and the GOP in one fell swoop. They also knew he had been a reliable Democratic vote for twenty-five years in the House and then the Senate. Mostly, they didn't take him seriously, just as they misread the loose energy on the land that is a feature of presidential elections. These are people who ditch their day jobs, throw caution to the wind, and join a political road show they deeply believe would change the nation and world.\n\nLong after the party realized that Bernie was in it to win, and not just to hound members about their progressive populist past, I asked Debra Kozikowski, the Massachusetts Party vice-chair and a grassroots activist for three decades, what had her peers been thinking? They backed putting Bernie on the ballot after he promised he would publicly support the 2016 nominee, she said. Their thinking was a calculated afterthought. They knew Bernie would bring along voters not drawn to Hillary. They thought he would boost her by a few points in key states. They never thought he would run neck and neck with the anointed nominee.\n\nThe first sign that Bernie was a very serious contender came in late 2015 when campaign finance reports were filed. Bernie was hardly the first progressive to reject donations from corporate PACs (political action committees) or newer deeper-pocket front groups, Super PACs. In 1992, ex-California Gov. Jerry Brown was mocked for announcing his 1-800 number for donations, but it worked. In 2016, Bernie used the Internet to raise nearly a quarter-billion dollars in funds averaging $27 per donation. As late 2015's fund-raising reports were filed, the Clintons, and party insiders with signed photos of the Clintons on their desks, faced an intra-party split that was only deepening. The pro-Clinton party establishment would have to turn to hard(er) ball tactics to impede his candidacy.\n\nThe tension broke into public view in bizarre but telling ways. It's an example of barriers facing progressives, who wonder why there aren't more candidates like them in the party. The answer is because national leaders and pro-establishment liberals don't want them and haven't for years. This is the Clinton side of the party. These more traditional Democrats can be found in northeastern Virginia, New Jersey, Massachusetts, and even in California. By mid-fall 2015, the disputes between Bernie's surging grassroots campaign and Clinton allies inside the DNC began to break into the open.\n\nThe first frays were inside ball, but they revealed the DNC was privately operating as a subsidiary of the Clinton campaign (even if publicly posing as a neutral referee). In mid-October, Wikileaks\u2014we now know, assisted by Russian intelligence agency hackers\u2014posted an April 2015 memo from Charlie Baker, Clinton campaign chief administrative officer, to the Clinton campaign chairman, John Podesta. It outlined the televised debate schedule that her camp wanted\u2014and got. They were to be minimal, held on weekends when audiences were smaller, and were announced without input from Sanders.\n\nBy mid-December, the disputes reached a new orbit. As part of his candidacy, Sanders agreed to use the DNC's national voter database. Both major parties have these, where, in addition to all the public information from state voter registration files (name, address, party, etc.), the campaigns add tidbits like whether a voter likes their candidate, what issues matter to them, whether they will volunteer, and more. That information comes from phone calls and by going door-to-door by campaign workers. Bernie kept his donor list separate from the DNC, but his campaign had been using the party's voter file as they built a national organization.\n\nBesides mutual feelings of distrust, there were accusations that Clinton's staffers, and then Sanders's staffers, had been spying on each other's voter files. DNC Chairwoman Debbie Wasserman Schultz responded by shutting off the Sanders campaign's access to the database in mid-December, at a time when Sanders was raking in millions from small grassroots donations daily. Groups like MoveOn launched petitions. The Sanders campaign sued the DNC before cooler heads prevailed. The episode, nonetheless, cemented its belief the party was all in for Hillary.\n\nThese brazen attempts to tip the scales were precursors to what unfolded in the first electoral contests involving voters\u2014the party caucuses and primary elections. Even more examples surfaced as the nominating season ended, when Wikileaks released stolen emails from top DNC staff with anti-Bernie plots, including possible lines of attack. Those came as Clinton loyalists fumed that Sanders was not conceding and had the audacity to keep campaigning through the convention. They forget that Clinton kept pushing similarly in 2008 to regain delegates she lost in Florida and Michigan after those states unilaterally moved up their primaries and were penalized by the DNC. Some of these behaviors and reactions were due to the emotional turmoil found on every presidential campaign. But there were other ways state parties and the DNC rigged their voting rules.\n\nSanders was more than aware of the DNC's biggest insider bias: its superdelegate system. This is where 713 of 2,383 national convention voters\u2014to nominate their 2016 presidential candidate\u2014were preselected Democratic elected officials, state party leaders, loyalists, and lobbyists. Superdelegates only exist for one purpose: to take the nomination away from duly cast primary votes and caucuses. While party brass said that has not happened, and in fact would never happen, their existence is not merely an antidemocratic facet of its way of doing business. Superdelegates undermine the Democrats' moral authority to criticize the blatant and often racist efforts by the GOP to disenfranchise voters, especially nonwhites in blue epicenters.\n\nSanders downplayed the DNC's affirmative action for insiders, believing these Democrats would back him if he won the popular vote in the caucuses and primaries. But unlike Bill and Hillary Clinton, his team had yet to compete in these contests, let alone master each state's intricacies where party loyalists intentionally acted and deprived him of victories. That inside edge surfaced in the very first contest, the Iowa caucuses in late January.\n2\n\nWHO WON IOWA?\n\nBERNIE HAD EVERY REASON TO BE upbeat about Iowa. He followed a trail blazed by George McGovern in 1972 and Jimmy Carter in 1976. After Democrats had a disastrous 1968 convention in Chicago, where Democratic Mayor Richard Daley ordered city police to crack down on anti-Vietnam War protesters, the DNC was forced to respond to entirely accurate accusations that their power brokers controlled the nomination process. They responded by moving Iowa's grassroots caucuses to its first-in-the-nation position.\n\nMcGovern, a senator from nearby South Dakota, didn't win Iowa's 1972 caucuses, even though his antiwar platform eventually brought him the nomination. However, Carter, a relatively unknown governor from Georgia, started showing up in Iowa in 1974 and set the precedent followed by most candidates since. He went from town to town, stopping in diners, giving speeches before local civic clubs, recruiting volunteers, appearing on local media, and ignoring the national press, which dismissed anyone they disliked. Two years later, Carter was the nominee and then the next president. After he lost the 1980 election to Ronald Reagan, a DNC commission reinstated superdelegates to make sure an outsider like him was never nominated again.\n\nBernie moved to northeastern Vermont four years before the Chicago debacle. Like other 1960s countercultural types, he was an urban intellectual refugee drawn to a state whose rural lifestyle was punctuated by short summers, long winters, mud seasons, hardscrabble farming, and voting for Republican presidents. He fell in with the proud but tolerant locals, a mix of personal libertarians, neighborly socialists, and antiestablishment types. His tell-it-like-it-is\u2014but never lie\u2014style worked. Old Vermonters touted their self-reliance, but they knew what life was like before Franklin Delano Roosevelt's New Deal brought electricity to their farms. The natives may not have liked Bernie's politics, but they knew where he stood. They respected his consistency and, over time, started voting for him.\n\nNearly a half century later when Bernie began campaigning in Iowa, he was more than at home. He even had an advantage that Clinton did not. He was new; a Socialist novelty who quickly showed he was a dogged campaigner and dead serious on issues. The last socialist candidate who ignited as big a grassroots movement was Eugene Debs, who had long inspired Sanders. Debs ran for president a century ago, including once from a jail cell, after Democratic president Woodrow Wilson had him arrested for demanding peace and social justice.\n\nEverything Sanders had going for him, from his moral compass to his passionate delivery to his long-distance runner stamina from an athletic youth, served him well. As the Iowa's caucuses approached, his campaign estimated that he had held more than a hundred rallies and town meetings and had spoken to seventy thousand people. His staff was filled with supporters who would not be put off by participating in that wintery event, including young techies who devised apps to track voting.\n\nWhat Bernie's campaign didn't have, notwithstanding the loyalty among Iowans that the Clintons had built up over the decades, was the support of the Iowa Democratic Party's top leaders. Nor did his burgeoning campaign have an understanding of how chaotic, opaque, and unaccountable the caucus process could be. Vermonters were familiar with their state's presidential caucuses and town meetings, where, after debate and huddling in the corners of libraries and gymnasiums, the winners and issues are openly decided. But that's not what unfolded in 2016's opening contest where the coverage and messaging would be critical.\n\nWhat unfolded on caucus night was as predictable as it was offensive to participants and observers who expect our democracy to be fair, transparent, and accountable. As the results started pouring in, Sanders and Clinton were running neck and neck. Other Democrats, like Maryland's ex-Gov. Martin O'Malley, were behind. The Sanders team used its app to report local results to headquarters, while the caucus chairs were calling in or using a Microsoft app from the party. As results started adding up, so did complaints about the process.\n\n\"Too many questions have been raised,\" was how the editorial writers at Iowa's foremost newspaper, the Des Moines Register, put it days later. \"Too many accounts have arisen of inconsistent counts, untrained and overwhelmed volunteers, confused voters, cramped precinct locations, a lack of voter registration forms and other problems. Too many of us, including members of the Register editorial board who were observing caucuses, saw opportunities for error amid Monday night's chaos.\"\n\nLet's pause there. Opportunities for error, coming after the DNC colluded with Clinton in the debates and tried to impede Sanders's fund-raising, begin to show how winning is not simply a matter of getting 50 percent plus one. To win in modern America, whether as a Democratic Party presidential contender in a state party caucus or in a general election against Republicans, requires specific margins, given the rules of the contest at hand.\n\nThe Iowa caucuses winner was not who won the popular vote, but who won the most delegates to the next stage in the process, its state party convention. As Jonathan Allen and Amie Parnes wrote in their book, Shattered: Inside Hillary Clinton's Doomed Campaign, Iowa's caucuses were \"a math exercise that didn't necessarily require Clinton to win more total votes but to take a majority at the right set of caucus sites.\"\n\nClinton's campaign knew this. They positioned themselves to quickly declare victory. At 2:30 AM on Tuesday, February 2, Iowa Democratic Party Chairwoman Andy McGuire announced that Clinton had won 699.57 \"state delegate equivalents\" to the next stage in the process, while Sanders had won 695.49 \"state delegate equivalents.\" Only one precinct out of 1,682 across Iowa had not reported. That locale, Des Moines No. 42, where Sanders won, was presided by a voter who only agreed to do so after showing up late and was unaware he had to report the results that night. Five minutes after McGuire's announcement, Matt Paul, Clinton's Iowa director, issued a statement claiming victory and declaring nothing would change it. Her campaign staff was in full giant-killing mode.\n\n\"Hillary Clinton has won the Iowa caucus,\" Paul said. \"After thorough reporting\u2014and analysis\u2014of results, there is no uncertainty and Secretary Clinton has clearly won the most national and state delegates. Statistically, there is no outstanding information that could change the results and no way that Senator Sanders can overcome Secretary Clinton's advantage.\"\n\nThat no outstanding information claim was a lie and Iowa's politicos knew it. The headline of the Register's analysis was, \"Iowa's Nightmare Revisited: Was Correct Winner Called?\" Four years earlier, Mitt Romney was named the GOP winner, but an audit of the counties showed Rick Santorum had more votes\u2014along with finding that some precincts had not reported. However, unlike Iowa Republicans, McGuire and her party rejected an audit and refused to release the raw vote counts. The party wouldn't budge from its opaque delegate-allocation math. Clinton's victory would stand.\n\nThe Register editorialized but got nowhere. \"Something smells in the Democratic Party,\" it wrote, fingering McGuire, who \"dug in her heels.\" A two-tenths of 1 percent difference between candidates would \"trigger automatic recounts in other states,\" it wrote, hoping a reasoned argument might have traction. \"Her path forward is clear: Work with all the campaigns to audit results. Break silly party tradition and release the raw vote totals. Provide a list of each precinct coin flip [when candidates tied] and its outcome, as well as other information sought by the Register. Be transparent.\"\n\nBy the time McGuire made her announcement, both Sanders and Clinton had left the state. Both were headed to New Hampshire for the nation's first primary, where voters cast ballots and every vote is counted equally. \"I can tell you, I've won and I've lost there, and it's a lot better to win,\" Clinton said the next day. Sanders, in his post-2016 book, Our Revolution, didn't say anything about Iowa's antidemocratic tally or its flawed caucus process.\n\n\"Most of the media correctly perceived the night as a victory for us,\" he wrote. \"From the first day of the campaign, we knew that we would have to do well in the early states to establish credibility and let the world know that we were in this for the long haul.\"\n\nBernie's retelling took the high road. His campaign manager, Jeff Weaver, started to say that he was \"the most electable Democrat,\" a line he would repeat throughout the spring. Weaver may have been right. But Iowa's caucus was not an election, per se. It was a high-stakes process where those at the helm in precincts\u2014doing it the way it has always been done\u2014were more likely to be incompetent than corrupt.\n\nAt the state party level, there was a different intransigence: a system of counting that is not the equivalent of one person, one vote. Should anyone have been surprised that Iowa's party hid behind math to help Clinton, not someone who had not been a Democrat until running for president?\n\nThis is hardball, not the fair, rational process the Register's editorial page sought. Anyone doing homework should have seen this coming. In 2008, Clinton's campaign made a similar move when it stretched the truth to say it had won Texas\u2014which national media reported\u2014even though Obama later emerged as the victor. Democrats in Texas cast ballots in a two-step process: a primary during the day and caucus at night. Clinton won the primary. But Obama won the caucuses. As reporters waited for her speech in Columbus, Ohio\u2014also holding its primary that day\u2014her campaign ludicrously said tens of thousands of Texans may have illegally caucused for Obama. That baseless accusation finessed her victory declaration in two states that night, which she needed to try to blunt Obama.\n\nAfter Iowa's caucuses, Clinton's campaign panicked. With the help of allies in the DNC, they added an unscheduled televised debate before New Hampshire's primary. That broke the schedule they helped write and refused to budge on a few months before. But the extra debate didn't help Clinton and probably further legitimized Sanders's stature. He won New Hampshire, Vermont's eastern neighbor. Weaver kept saying that Sanders was the most electable Democrat. But what Sanders's campaign would soon find out in the next state they needed to win was there were other fine-print rules and barriers to a fair contest.\n3\n\nLAS VEGAS RULES\n\nSANDERS HAD ONE BIG CHANCE TO upset the Clinton campaign before the race headed to South Carolina and then to early March's Super Tuesday, where eleven mostly southern and central states were expected to deliver for Clinton. That was Nevada, a western state with a racially diverse population centered in Las Vegas to the south and Reno to the north. Nevada is not a wealthy state. But it is nothing like Iowa and New Hampshire. It is more like California, a mix of rustic downtowns and sprawling suburbs and residential neighborhoods. Las Vegas was home to a very powerful union, the Culinary Workers local, whose fifty-eight thousand members, mostly Latino and black, work in the corporate-run hotel, casino, and hospitality industry.\n\nBernie hadn't spent as much time in Nevada as he had Iowa and New Hampshire. In Our Revolution, he said that was the right choice because he had to make a strong start to be taken seriously as a candidate. By the time Nevada's mid-February caucuses came, he'd held seventeen events in the state, including a large outdoor rally at the University of Nevada in Reno. On Election Day, he stopped by Culinary Workers picket lines in Las Vegas and visited hotels and casinos along the strip, including some where the Nevada Democratic Party had set up caucus sites in rooms normally used for conventions. Bill and Hillary Clinton also made the rounds, visiting employee lounges and urging workers to vote.\n\nLike Iowa, Nevada's Democratic Party ran the caucuses according to their rules. It had an obscure delegate allocation formula and multistep nominating process that began at local precincts, moved to county, and then to state conventions. Sanders ran into obstacles not seen in Iowa and New Hampshire. These were institutional decisions that affected who would likely participate and how easy that would\u2014or not\u2014affect voter turnout. It well may have been that despite growing momentum, he was not going to beat Clinton as he had run out of time to have become known in the state where no candidate goes town to town, or holds forth in diners and civic clubs. But his chances would have been better if the Nevada Party\u2014and its true leader, Senator Harry Reid\u2014had played fairer and squarer.\n\nJon Ralston, the Nevada press corps dean, recounted in a USA Today commentary how the party rescued Clinton. Ralston said Reid, the retiring Senate Minority Leader, claimed he was neutral but was working for a nervous Clinton camp. First, Reid told the Las Vegas Culinary Workers Union president not to endorse a candidate. He agreed and predicted his members would skip the vote. Then, Reid lobbied casinos to let the workers get time off to caucus. He then pushed the union to deploy \"swarms of organizers to turn out mostly Latino workers, who would likely vote for Clinton.\" I was at one of the casino sites and saw waves of employees arrive just before the cut-off time to participate. Most, especially women, said they were for Clinton. When asked why, several replied that if she were in trouble, Hillary's husband would tell her what to do.\n\nRalston said the credibility of \"Prince Harry\" was on the line and he delivered a result \"that could indeed prove to be a contest that saved Clinton's campaign.\" The rescue was a good story and surely contained elements of the truth. But it doesn't note what else happened that added to Clinton's victory on February 20. Out of the eighty thousand or so participants statewide, the party said Clinton had won 52.64 percent of the delegates to the county forums while Sanders won 47.29 percent. One vote doesn't equal one delegate equivalent in the party's math, but there were not more than twenty-five hundred participants at the six Las Vegas casino precincts. What Ralston didn't say, but what was in Reno newspapers, was a laundry list of factors that could tilt close elections.\n\nNevada Party leaders know that whoever won Las Vegas would win the state. So what did they do for which Sanders had no recourse and was beyond Reid's efforts? First, the party's caucuses are closed to voters who are not registered as Democrats. Closed primaries are not unique to Nevada, but they prevent last-minute enthusiasts or independents from participating if they didn't register as Democrats. The party also had no caucus sites at Las Vegas universities, a setting where Sanders was likely to draw supporters. In Reno, there were caucus sites on campuses, which contributed to Sanders winning northern Nevada despite other snafus.\n\nThe Reno Gazette Journal's coverage two days later categorized the problems they has witnessed. These included: \"strange delegate math,\" \"long check-in process,\" \"wrong location,\" \"people who left were improperly counted,\" \"people with disabilities not allowed to exercise rights,\" \"disorganized and unprofessional,\" \"Clinton voters allowed to break rules,\" \"Wi-Fi overload,\" \"text check-in not working,\" \"cards not counted,\" \"voters turned away.\" These complications, which are not unique to Nevada's caucuses, refer to various steps in the process. This list deserves closer scrutiny.\n\nThese factors, some due to incompetence and others due to institutional decisions, make participating easier or harder. Some help or hurt specific constituents, such as location. Then there are other ways of deterring voters and eroding turnout, from poor planning that leads to long check-in lines at precincts to overcrowding and inexperienced staffing. Those are all details the state party should anticipate and address. Not doing so in Nevada in 2016, a year where one-third fewer Democrats caucused compared to the 2008 Clinton-Obama race, suggests the party either intended Reno's voting to be messy or it should not be running caucuses. Many academics drew the latter conclusion.\n\nThe problem with barriers to participation is they stay below the radar until it's too late. Come Election Day, especially in close races, complaining goes nowhere. After Iowa's and Nevada's opening caucuses, Sanders's campaign wised up. They knew the voting and count weren't straightforward or precise. On April 2, when the party's county conventions were held, only two-thirds of the delegates elected in February showed up: 2,964 for Sanders, and 2,386 for Clinton. Sanders ended that day by sending 1,613 delegates to the state convention, while Clinton sent 1,298.\n\nThat turnaround prompted Sanders's campaign to tweet that they had staged a comeback. But they hadn't. By that time, the campaign press wasn't following steps between the first caucuses and national party convention, except for pages dedicated to delegate counts. Sanders's surge didn't make much news. Moreover, his win was among the elected delegates, not the DNC's superdelegates. His Nevada campaign won a key battle but lost the war.\n\nI briefly ran into Weaver the night before opening caucuses. The campaign held a rally at an amphitheater in Henderson, east of Las Vegas. I had not seen him in decades. We agreed no one who knew Bernie back then could have imagined that he would be running for president and become a real contender. Weaver hoped the campaign would squeak out a Nevada win and confided that the next stretch of states\u2014South Carolina and southern states on Super Tuesday\u2014would go for Clinton. But after that, he said there would be a month or more where everything should line up behind Bernie.\n\nThat's what happened. But in key states where primary elections ensued, their campaign ran into other antidemocratic hurdles. Some came from state party policies. Some came from administrative blunders and stonewalling by local election officials. Some were by media coverage and discouraged turnout. All of these hurdles underscore that nobody, especially outsiders like Bernie, can win unless they generate turnout that becomes a wave big enough to breach the barriers. The contest's next phase illustrated this point again and again.\n4\n\nNEW QUESTIONS, NEW BARRIERS\n\nFROM SUPER TUESDAY IN EARLY MARCH until late April's primary in New York, the pendulum swung from Clinton's recovery in the South to a Sanders surge in the Midwest and West, where he won seven primaries and caucuses in a row. Clinton's streak, the first part of that progression, was unsurprising. But her pattern of repeatedly outperforming Election Day media exit polls did not sit well with election integrity activists who saw foul play in the discrepancies between the exit polls and official results.\n\nTheir suspicions kept rising after exit polls forecasting Clinton's victory were off by more than ten points in South Carolina, Alabama, and Georgia; and by slightly less in Tennessee, and in Democratic strongholds in Missouri, Illinois, and Ohio. (When the nominating season ended, these data-crunching activists found other odd patterns, such as Sanders winning in smaller precincts where ballots were counted by hand but losing in larger ones where votes were tallied electronically.) To them, this points to hacking or altering counts by tampering with voting machine software. Party leaders and election officials routinely ignore these discrepancies and complaints, and officials never let anyone examine their voting systems. Candidates also stay away. Sanders avoided this issue entirely in his campaign book.\n\nI'm averse for other reasons. I know hacking is possible, respect academics who have shown how it can be done, and suspect it has happened in races where fewer people are watching, such as in local noncandidate bond issues. But I have not seen sufficient proof in my years of covering federal elections. That doesn't mean it's not out there somewhere. But all we see are patterns that raise questions that are never sufficiently answered. After a point, I return to more tangible and traceable factors that shape elections and voting. Sanders kept facing these more concrete partisan barriers and administrative oopsies that consistently favored Clinton as the race unfolded.\n\nHis book mentions how mistakes by county election officials\u2014beyond antidemocratic state party rules\u2014also affected thousands of voters. These are in a different class than bungling precinct captains, incompetent state party chairs, and antidemocratic party rules. These are less excusable because these are coming from government officials on public payrolls.\n\nOne of the foremost examples was in Arizona's largest county, Maricopa, where Phoenix is located. Its top election official, Helen Purcell, who had held that post for years, made a bad judgment call when she closed 140 out of 200 voting centers for the primary. It is likely she never would have been allowed to do this had the enforcement trigger in the 1965 Voting Rights Act had not been declared unconstitutional by the Supreme Court in 2013. The Justice Department had previously questioned having one poll for every 21,000 voters, compared to an average of 2,500 for the rest of the state. What could go wrong?\n\nPurcell, like many administrators, sought to make voting more efficient. Her mistake was implementing this in a presidential primary and not a lower-turnout race. To say that voters were caught off-guard is an understatement. People didn't know where to vote. Some stood in line for five-plus hours to get into the voting centers. Local activists, national civil rights law groups, Sanders, Clinton, and the DNC all filed lawsuits. One suit sought to throw out the primary results. Another challenged Purcell's rationale but also took aim at other anti-voter tactics, such as a 2016 state law making it a felony to collect another person's ballot and bring it to a poll (making it harder for Native Americans and seniors). Another wanted the county's elections under court supervision. Nothing changed the results in the jurisdiction where 70 percent of Arizona's Latinos live.\n\nSanders mentioned this morass in his book, saying he was stunned by how an administrative decision undermined the finish line of a hard-fought campaign. \"I thought we had a good shot to win there. We didn't and I still don't know why. One thing I do know was that the voting process in Arizona was an absolute disaster and an embarrassment to American democracy,\" he wrote. \"Nobody can ever really know which candidate was hurt more by this travesty, but I don't think it helped us. Many of Clinton's supporters were older and voted by mail. Most of our supporters came out on Election Day with the obvious expectation that they would be allowed to vote in a reasonable period of time.\"\n\nThe Sanders campaign would soon discover that sloppy election administration could be as damaging to his efforts as antidemocratic state party rules. As his April momentum grew, the media's eyes turned to New York and the Big Apple where many news outlets are based. Its April 19 primary had every angle for maximum drama. A Brooklyn-born socialist returned home to face the state's former two-term US senator, who also was the former secretary of state and First Lady. Hovering below this hype was a big blue state's notoriously stubborn Democratic Party and intransigent election officials.\n\nThe primary run by the New York State Democratic Committee is one of two dozen in the United States that are closed, meaning only registered Democrats can participate. Like Nevada, independents or voters who joined another party cannot make a last-minute decision to cast a vote for a Democrat that will count. They might show up and insist on voting, but they will get a provisional ballot after not being listed as a Democrat. Their ballot will not count when counties finalize their tallies. New York's closed primaries meant three million Independents could not cast a vote for Sanders, even though he was elected as one to Congress for more than a quarter century.\n\nThat wasn't the worst of it. New York has the nation's longest party registration deadlines. It's been this way for decades and nobody in power is inclined to change it. To vote in the Democrats' April 2016 primary, you would have had to register as a party member by October 2015. So even as Bernie held rallies before tens of thousands across New York City, many people who heard and liked him could not vote for him. If that wasn't frustrating enough, another barrier emerged in a key borough targeted by his campaign. As April began, news broke that the Board of Elections (BOE) in Kings County, better known as Brooklyn, had purged\u2014or removed\u2014122,000 voters from the official rolls, and had not told those people they could not vote in the primary.\n\nThe rough outline of what happened merits explaining. These nuts and bolts are some of what Republicans seize and distort when they make claims about voter fraud, their made-up menace and excuse for passing anti-voter laws in red states. Basically, Brooklyn's election office failed to keep their voter rolls current, and was slammed in public reports for things that sounded worse than they were\u2014but it still kept messy bookkeeping. Then the county's Board of Elections (BOE) made matters worse by locally purging voters but not telling the state's election agency\u2014whose online records said those people could vote in 2016's primary.\n\nThe problem that kept Brooklyn's voter rolls from being accurate is one faced by every county in America. After registered voters move, marry and change their names, or die, almost nobody notifies local election offices. Thus, voter rolls, which are not magically updated, are in a state of flux because a county's demographics are constantly changing. Some states have reliable data systems that track and update these changes, but not New York State and not Brooklyn. Dead people can remain listed, as well as people who move and register elsewhere. The dead don't vote. Neither do most people who move, although there's always a handful\u2014usually in single digits\u2014who think they can vote in more than one state. Voter roll errors are an easy target for sensation-seeking media that don't think through what they are writing. The GOP also targets outdated lists as it looks to overly police the process, usually by creating impediments to blue voting blocs.\n\nBefore the 2016 primary, a New York City government audit report hit Brooklyn hard for its sloppiness in voter-list maintenance. Then Kings County didn't follow the steps of a 1993 law (the National Voter Registration Act) that laid out how voter purges are to be done. That law said that nobody who is a registered voter can be purged unless they have not voted in two federal cycles\u2014four years\u2014and not until after county offices try to contact them about losing their voting rights. That was where the BOE began to screw up, because they found 122,000 people who had not voted since 2008, purged them from rolls starting in 2015, but did not tell the state's election agency. When New Yorkers went to the state's website, they saw that they could vote. However, they couldn't. All this surfaced on April Fools' Day.\n\nThe result on April 19 was confusion, frustration, and anger as large numbers of people showed up and found they were not listed as voters, or their party affiliation was incorrect\u2014another bureaucratic snafu\u2014and they couldn't vote in the primary they wanted. As is typical in New York, local officials barely budged. A handful of voters went to court, and, helped by election protection attorneys, won the right to cast a regular ballot. Most did not. New York State's Attorney General Eric Schneiderman said his office received over a thousand complaints. The county BOE director, Michael Ryan, declared that, \"no one was disenfranchised.\" That's absurd, not after 120,000 Brooklyn voters cast provisional ballots, of which only 31,000 ended up being counted. The rest were deemed ineligible.\n\nNew York was the latest big race that Sanders needed to win to blunt Clinton's lead. His campaign threw everything they had at it, including most of their media budget. He also flew to the Vatican to see the pope days before the primary. But it did not happen. Even if all of those voided Brooklyn votes had gone for Sanders, he would have fallen short by more than 100,000 votes statewide. Berniecrats correctly pointed out partisan barriers (closed primaries, registration deadline) and administrative snafus (voter roll purges and other errors) were not confined to Kings County. New York being New York, finger pointing ensued. Brooklyn officials were disciplined. Lawsuits ensued\u2014including the Department of Justice intervening in early 2017. But the primary results remained.\n\nMonths later, Sanders wrote, \"We worked very hard there and ran an excellent campaign, but there were just too many obstacles in our way.\" He forged on and looked to the last big blue state, California, to try to make a case to the DNC superdelegates that he was more electable than Clinton. Like every key state before it, California presented yet more obstacles that he didn't anticipate.\n5\n\nTHE NOT-SO-GOLDEN STATE\n\nAFTER LOSING NEW YORK, DELAWARE, PENNSYLVANIA, Maryland, and Connecticut\u2014but winning in Rhode Island, Indiana, West Virginia, and Oregon\u2014Sanders's campaign made a strategic decision. They would shut operations in other states and tell their best organizers to go to California. Its June 7 primary fell on the last day of all state nominating contests. By then, Sanders had won 45 percent of the pledged delegates. He felt that if they won the biggest state, they might sway the superdelegates. National and state polls kept reporting he was doing better than Clinton in theoretical matchups with Trump.\n\nSanders then did what no presidential candidate has done in decades. He spent a month in California, holding forty rallies that brought out nearly a quarter-million people. Like New York, California had different quirky procedures and party rules regarding who could and could not vote in their primary. California's confusion began when voters were asked if they wanted to join a party on the state registration form. There are a series of boxes to check. The first was \"American Independent Party,\" a right-wing party, which is not an independent. Tens of thousands of Californians inadvertently joined the AIP.\n\nThey should have checked the \"No Party Preference\" (NPP) box. But that wasn't clear. The Los Angeles Times estimated three-quarters of AIP's members had zero intention of joining it\u2014half a million people. That error set in motion a series of dominoes that fell as voting began, first by mail-in ballots and then at precincts. The California Democratic Party allows NPP voters, genuine independents, to vote in its primary, unlike New York. But people who thought they were independents and wanted to vote by mail received the AIP primary ballot. A similar situation unfolded at the polls. California's Democratic Party, led by Gov. Jerry Brown who backed Clinton, did not say anything to clear up this confusion.\n\nStill, Sanders was upbeat. The crowds at his rallies were big and enthusiastic, and ignored East Coast media calls for him to drop out. He had run out of money for statewide advertising\u2014which costs millions per week in California. Nonetheless, Clinton took him seriously and returned for a final series of rallies and fund-raisers.\n\nThen one of 2016's most inglorious moments of media bias struck. One day before the primary, the Associated Press published an article saying its team had reached all of the undeclared superdelegates and found Clinton had secured the nomination. To declare the race over less than twenty-four hours before millions of Californians were to vote in the first primary that mattered in their state in decades was the height of arrogance. The AP is not any wire service. It's a national co-op whose reports become the basis for local TV and radio news, and whose articles fill the pages of regional and local newspapers.\n\nThere is no way to quantify the impact of the AP's ambush of 2016's final primaries. It was an unforeseen antidemocratic deed in a season marred by them. Sanders wrote in his book, \"I strongly believe that their action had a negative impact on voter turnout and hurt us. Why vote if the election was over? Our younger voters were more likely to cast their votes on Election Day rather than voting earlier by absentee ballot, as many Clinton voters did.\"\n\nOn Primary Day, California also experienced its share of administrative snafus that never seem to disappear in American elections. Poll workers gave out provisional ballots instead of regular ones. Some polls had incomplete voter rolls. Some ran out of paper ballots. That night, the media announced that Clinton had won by fourteen points, more than exit polls had forecast. She also won in New Jersey, New Mexico, and South Dakota, while Sanders won Montana and North Dakota.\n\nThe last primaries ended Sanders's presidential hopes. Had he won a majority of elected delegates, he planned to file a series of motions as the Democratic National Convention's opening gavel fell to disqualify superdelegates, and then quickly move to a vote based on pledged delegates. At that point, Sanders would have either won the nomination or walked out, a senior campaign aide told me in spring 2017. This strategy was not reported at the time.\n\nWithin a week of the last primaries, however, Sanders and his senior staff began meeting with Clinton to negotiate what the party would do to honor his issues; in the platform, party rules (reforming the superdelegate system) and with campaigning that fall. Those meetings began before California finished its official vote count. That took time because California uses paper ballots counted by scanners, and millions of mailed-in ballots. Once the count was done, he had 46 percent, cutting her supposed primary night lead in half.\n\nBernie's month of campaigning in California and the state's party's rules of letting independents vote in its primary made a difference, though not enough for him to win. That is a much-overlooked capstone to his performance in California. Think about it: starting with Iowa's caucuses, Sanders encountered a series of state party rules and other decisions that favored his opponent and did little to embrace the millions of voters who were drawn to him. The voters who sent 45 percent of elected delegates to Philadelphia's convention were underestimated, underrepresented, and marginalized by their party's leaders.\n\nOn top of that, Sanders faced the institutional clumsiness of the voting process. In caucus states, inept precinct captains and states' chairs were not up to the task of running contests with presidential nominations at stake. In primary states, many key elements overseen by county government officials\u2014registration deadlines, voter list maintenance and purges, precinct locations, voter lists\u2014complicated rather than clarified voting, especially for first-time voters and independents. Together, what Sanders and his base experienced was that winning takes much more than embracing a candidate you believe in.\n6\n\nTHE DEMOCRATIC NATIONAL COMMITTEE\n\nBUT THE BIGGEST BARRIER TO SANDERS breaking through was the party culture and its leaders. These are political pros that know better, know what nefarious behavior is and isn't, and whose biases blinded them to not just Sanders but to the party's antidemocratic rules.\n\nWhen DNC Chairwoman Debbie Wasserman Schultz was interviewed in February 2016 by CNN after Iowa and New Hampshire, she fiercely defended the superdelegate system. They \"exist really to make sure that party leaders and elected officials don't have to be in a position where they are running against grassroots activists,\" she said. Wasserman Schultz did not see superdelegates as affirmative action for insiders, but said they were to ensure that diverse constituencies have a say in the party. Superdelegates have never departed from primary and caucus results, she said, and chosen a different nominee. This breezy explanation ducked the bigger issues at play, including her loyalty to Clinton.\n\nIn politics, real motives are usually not what is said in public. On the weekend before the Democratic National Convention opened in Philadelphia, Wasserman Schultz was forced to resign after WikiLeaks released emails from her top staff showing plots to attack Sanders\u2014such as questioning his religious beliefs. This was the last straw, because, even if Hillary loyalists smoldered that Sanders had refused to warmly endorse her after California, it told Sanders's supporters, nearly half the party, that they were to be stepped on and sidelined.\n\nOf course, the DNC's bias was evident before the first votes were cast. Beyond its superdelegates, most of whom were Clinton supporters before Iowa and stayed that way, were other candidate-centered aggressions and antidemocratic slights. The DNC cut off Sanders's access to its national voter database as his fundraising peaked in late 2015. It scheduled debates on weekends, when fewer people would become familiar with him as a candidate and his policies, as well as his critiques of Clinton. The DNC created a joint fund-raising effort with the Clinton campaign before the 2016 nominee was known. WikiLeaks showed that Donna Brazile, the ex-DNC chair who was then a CNN analyst, shared debate questions with the Clinton campaign before its primary debate. Incredibly, Brazile took the DNC helm after Wasserman Schultz resigned in disgrace.\n\nAnother open admission came in February 2017 when the soon-to-be elected DNC chairman, Obama's former labor secretary Tom Perez, told Kansas lawmakers that 2016 was rigged from the inside. \"We heard loudly and clearly yesterday from Bernie supporters that the process was rigged, and it was,\" Perez said. \"You've got to be honest about it. That's why we need a chair who is transparent.\" Needless to say, Perez immediately got flack for that honesty and later tweeted that he \"misspoke.\"\n\nThe notion that political ends justify the means, and doing whatever it takes is loyal and patriotic, is as old as politics itself. But the modern Democratic Party is supposed to be the party fighting voter suppression, not deploying it. It is the party that keeps going into court to defend voting rights from Republican attacks. To run its presidential nominating contest as inside ball violates what the party says it stands for. Take Wasserman Schultz's defense of superdelegates to ensure minority representation. That definition excluded the grassroots, which, as it turned out, were up for grabs in 2016 and which Trump deftly exploited.\n\nSanders chose his words carefully when he discussed voter suppression in his post-2016 book. As you would expect, he first noted how Supreme Court campaign finance rulings have empowered billionaires to single-handedly finance candidates. But Sanders then turned to the Court's 2013 gutting of the Voting Rights Act. That 1965 law was passed after many states, mostly in the South, adopted rules to keep blacks from voting. However, the Supreme Court's Republican-appointed majority, led by Chief Justice John Roberts, said voter suppression was no longer a problem that required the federal government to sign off on any new election law or procedure in covered states. In response, GOP-run states moved very quickly to pass a series of laws making it harder for Democrats to vote.\n\nSanders is unsparing in his criticism of those who would rig the rules rather than win by the power of their ideas. \"Political cowards are doing everything they can to keep people from voting,\" he wrote. \"They are making it harder for people to register and to participate in the political process.\"\n\nMany of the things that Sanders correctly criticizes the GOP for\u2014creating voting barriers, hiding behind opaque rules and procedures, failing to ensure election administration is fully modernized and professional, and ensuring accurate and verifiable counts\u2014have analogues in the Democrats' nominating process. While the Republicans are far more accomplished in the dark arts of undermining voting and elections, one has to wonder if voters experiencing a more open Democratic contest would bring about fairer and more accountable elections.\n\n\"Democracy should be easy,\" Sanders wrote. \"All adults must have the right to vote. We can and should spend the money necessary to defend democracy by making sure polling places are adequately staffed, that voting machines function well, and that however voting is tabulated, there are paper ballots that can be counted and audited in cases where the voting is close or contested.\"\n\nThat sounds simple, but it isn't simple in a nation where our constitutional system has its roots and retains key characteristics of the late 1700s, when slavery, racism, classism, sexism, and other hard lines enshrined economic elites, while excluding those who challenge that order.\n\nToday, there is a party that believes the country does not benefit when every eligible adult votes and unapologetically keeps passing laws and adopting rules in that antidemocratic spirit. That is the Republicans. Its path to power starts with creating red-state monopolies and extends to dominating the US House and presidential swing states. Until their game plan and tactics are understood, unmasked, and bested, a minority party that consistently elicits fewer popular votes nationally, will not be relegated to history.\nSECTION II\n\nTHE REPUBLICANS\n7\n\nGAMING THE RULES OF POLITICS\n\nIT'S TIME TO SET ASIDE THE notion that American democracy is based on free and fair elections, epitomized by one-person\u2013one-vote. Instead, we need to recognize the frame that the nation's founders created: a system that elevates elites and manipulates the rules of voting to shape outcomes. This legacy endures today, despite twentieth-century US Supreme Court decisions affirming the ideal of one person, one vote.\n\nPerhaps the best-known example of this enduring legacy is the Senate. Wyoming, with less than six hundred thousand residents, has the same representation as California with 37 million residents. Another is the Constitution's Three-Fifths Clause in Article 1, which allocated House seats based on a state's population and counted slaves as three-fifths of a person. Historians have noted that ugly formula gave more clout to slave states via overrepresentation. But its shadow continues today in the Electoral College, where each state's president-selecting vote is based on its congressional delegation's size. Tinkering with underlying rules for elections by insiders acting on \"superior knowledge\" and \"sufficient virtue\" is as old as America. It just takes new forms: some more obvious and some more odious.\n\nIn American politics, there have always been factions who argue that allowing free and fair elections invites mob rule. That's why the state legislatures appointed the US Senate (almost uniformly wealthy white men) until the Seventeenth Amendment was enacted in 1913. There are right-wingers today that want to go back to that way of selecting US senators, just as House members like Florida Republican Ted Yolo believe that only property owners should vote. Pundits and scholars say politics has always been this way: whoever holds power will publicly offer high-minded words about democratic principles, but behind closed doors use every available advantage to game the system's operating rules. To expect anything else is naive.\n\nMost political reporting looks at the results in elections, not the underlying \"porous and partisan\" structures, rules, and administration that feed into the outcomes, as voting rights attorney Michael Waldman described it in his 2016 book, The Fight to Vote. Yet, in order to understand why the nation has had the Republican-majority House and red supermajority state governments this decade, we must examine the less-visible features of voting that do not treat all citizens equally, or technicalities that benefit Republicans or GOP rule. These are especially important in close races when public enthusiasm is waning because they decisively affect voting blocs.\n\nNo matter which verb you pick\u2014biasing, tilting, gaming, rigging\u2014the outcome in many races is unfairly handicapped before the votes are cast. This is why a Democratic president could be reelected in 2012 with 5 million more popular votes, and 1.4 million more Democrats than Republicans voted for House candidates nationwide, yet the GOP emerged with 33 more House seats. That same \"red shift\" persisted in November 2016, when Clinton won 2.9 million more popular votes than Trump, but the GOP won a forty-seven-seat US House majority and ended up with full control of thirty-two state legislatures.\n\nWhat Democrats did to Bernie and progressives was bad. But compared to what the GOP has done to all voters this past decade, it was child's play. There is little on the Democratic side of the aisle that compares to what Republicans have done on a national scale to build a structural advantage into winning key state legislatures and House races. The GOP crafted a pre-Election Day advantage that boosts their starting-line odds of winning. Beyond that baseline, the party has been pushing and passing scores of state laws that further erode Democratic voting blocs at key steps in the voting process. In state after state, they have ruthlessly targeted the same voters Democrats are seeking through its registration drives and get-out-the-vote efforts\u2014nonwhites, students, and the poor. The GOP's motives are clear: hold onto power and protect its shrinking and aging white base. (In 1992, nonwhites were 13 percent of the national electorate. In 2016, it was nearly a third.)\n\nRemarkably, the most impactful Republican action to restructure the electoral landscape flew below the national media's radar and was ignored by Democrats until recently\u2014years after it began. Data analysts, however, say that by 2016, the GOP had created upwards of a 6\u20138 percent built-in lead for many of its candidates for state legislatures and the House. How did they do that? They saw and seized an old but audacious way to resegregate each party's most reliable voters. These are voters who turn out for local, state, and congressional races\u2014not just presidential elections. Republicans identified each party's reliable voters in enough key states, and drew election district lines where more Republicans were likely to show up and vote than Democrats. Because they saturated these districts with their voters and diluted the Democratic base, Republicans could much more easily reach winning majorities.\n\n\"The Democrats fell asleep at the switch,\" said David Daley, author of Ratf**ked: The True Story Behind The Secret Plan to Steal America's Democracy, which details how a small team of Republicans foresaw, plotted, and created that advantage after losing nationally in 2008. \"They [Democrats] did not pay attention to it. They did not pay attention to it even after Karl Rove laid out the playbook in the Wall Street Journal [in March 2010]. They didn't have the strategic imagination to come up with a plan or even the tactical ability to play defense once Karl Rove spelled it out on the op-ed page of the country's largest newspaper. That's bad. And they have paid the consequences for that this entire decade.\"\n\nThe Republicans assembled majorities by concentrating as many Democrats as they could in few urban districts while spreading out the rest in seats where they would not reach winning majorities. They did that by drilling down and literally choosing which neighborhoods were included in state legislative and House districts. These are not statewide races, such as for governor or US senator or president. These are the jigsaw-puzzle shape districts that elect state legislators and House members. A handful of astute, well-organized Republicans sensed this opportunity in 2009. They realized whichever party held majority power in state legislatures after the 2010 election would redraw political boundaries lasting for the next decade, as most redistricting follows the once-a-decade US census (which recurs in 2020).\n\nThe GOP did not have to draw lines that followed logical demarcations like county lines, school districts, or population centers. They forged enough bizarrely shaped districts with a predictably higher Republican turnout to win state legislative majorities and keep the US House red. As Richard Wolf, one of the few national reporters who follows redistricting wrote for USA Today, the results were seen immediately. \"In 2012, Republicans won 53 percent of the vote but 72 percent of the House seats in [twenty] states where they drew the lines.\"\n\nThis is called gerrymandering\u2014or extreme redistricting. It's wonky, little understood, and underappreciated. But everyone has seen its impact. Think of 2016's presidential results. In Pennsylvania, Trump won statewide by forty-four thousand votes. In Wisconsin, it was by twenty-three thousand votes. In Michigan, it was by eleven thousand votes. In these states, the presidential result\u2014the statewide vote\u2014was split nearly 50\u201350. That reveals these states are evenly divided before cracking them into smaller pieces. Yet, in these and other purple states this decade, the GOP has routinely won two-thirds or more of state legislative and House seats. How can that be?\n\nThe answer is their political maps. There are some instances of Democrats doing this to Republicans\u2014such as a 2011 Maryland ploy to oust a ten-term, Republican congressman (which both parties have litigated and the Supreme Court may hear in 2018). But these are not part of a national Democratic strategy. What Democrats and their allies have typically done in cycle after cycle is register voters, do polling that shapes political advertising and get-out-the-vote (GOTV) efforts, and hope this will coalesce. The Republicans are more aggressive and cynical. They register voters, do analytics, messaging and GOTV, but they don't stop there. They want more certainty and a proven baseline. Redistricting gives them that edge.\n8\n\nCREATING PARTISAN GHETTOS\n\nTHIS DECADE'S REPUBLICAN STRATEGY STARTS WITH extreme redistricting: deciding beforehand which electoral races each party's base will vote in. Redistricting is not the elephant in the living room, that is to say, a large and unacknowledged presence we don't want to talk about. Redistricting is the living room. It is the field and stage upon which everything else follows.\n\nRedistricting is the biggest verifiable way Republicans have gamed and added points for their side in state and federal elections before candidates are known and voters chime in. That is not all they have done. They keep hyping virtually nonexistent voter impersonation fraud to justify new state-based anti-voter laws such as stricter ID requirements in order to get a ballot at polling places, which nonpartisan analysts such as the US Government Accountability Office found shaves another 2\u20133 percentage points off the November turnout. Other GOP tactics that reduce turnout include partisan purges of voter lists, limiting early voting, barring same-day registration and voting, and restricting voting by mail. But those suppressive tactics come later, long after extreme redistricting sets the stage.\n\nExtreme is the key word here. Republicans have not abided by what the federal courts have long said should be the standard. It is not even close. A half century ago, at the height of the civil rights struggle in 1962, the Supreme Court issued a ruling, Baker v. Carr, which articulated the ideal of one person one vote when drawing political districts. In 2004, in one of its last decisions before Republicans seized this process in 2011, the court urged partisans to behave. \"The orderly working of our Republic, and the democratic process, depends on a sense of decorum and restraint in all branches of government, and in the citizenry itself,\" wrote Justice Anthony Kennedy for the majority in Vieth v. Jubelirer. What the GOP did seven years later, however\u2014where the cases and maps coming before federal courts today originated\u2014showed no restraint or decorum. It was the rawest of power plays, where they saw and executed a ruthless strategy to win.\n\nDaley's book, Ratf**ked: The True Story Behind The Secret Plan to Steal America's Democracy, describes what unfolded in excruciating and brilliant detail. A small team, starting with a handful of political junkie consultants, envisioned and executed a two-part plan. Much like the way Democrats felt after 2016's pounding, most Republicans concluded they were headed for political exile after Obama's 2008 landslide victory and the Democrats' sweep of Congress. But a respected Republican political consultant who specialized in tracking state-based campaigns, Chris Jankowski, saw a path back into power.\n\nJankowski knew the president's party almost always loses seats in midterm elections, the non-presidential years. Moreover, as you go down the ballot to statehouse races, he knew most people voted on party lines because they do not know the candidates. The GOP team he led, known as REDMAP, for REDistricting MAjority Project, looked at the country. It concluded that if Republican candidates won slightly more than one hundred state legislative races in key states like Ohio, Pennsylvania, Wisconsin, and Michigan, and a few governors' races in 2010, that nothing could stop them from a virtually unassailable national comeback.\n\nFirst, they had to win enough state races to get majorities in the legislative chambers that would redraw maps after the 2010 census. They identified 107 local races in sixteen states that accounted for 190 US House seats and jumped in. REDMAP launched some of the most vicious attacks seen in years in these states. These were radio and TV ads, or mailers landing ten days before 2010's Election Day, when it was too late to respond with media buys. Democratic incumbent women were compared to prostitutes. Longtime incumbents who governed by making compromises were called the worst sellouts. Republicans won twenty-one state legislative chambers in 2010, enough to take over the redistricting process.\n\nRepublicans kept their eye on the prize. What does this process look like? Imagine sitting with consultants in front of a bank of computers in a law office\u2014so all conversations are confidential under attorney-client privileges\u2014or in hotel suites near the Capitol building. You stare at something akin to a Google map you can zoom in and out of. The goal is creating districts with roughly the same number of people when the seats are not statewide, countywide, or citywide. In the same way you drag a mouse to change a driving route following different highways, mapmakers could examine neighborhoods down to every block. Their software had voting histories of residents and other demographic data. Jankowski's team tracked the GOP voters who backed their presidential nominee John McCain in 2008\u2014a year they lost badly\u2014and who voted in state elections. They identified their most reliable voters. In some southern states, they tracked voters by race, assuming most non-whites were Democrats.\n\nBased on that threshold of each party's reliable base, they massaged, moved, or radically reset boundaries for state legislative and House races so turnout by Republicans would almost always defeat Democrats. They packed Democrats into blue epicenters, usually cities. That let the GOP to posture their maps as supporting minority candidates. (The Voting Rights Act of 1965 says that is the only allowable use of race\u2014to create districts where minorities can elect representatives.) Republicans knew what they were doing. They would never win certain seats, but they intentionally concentrated more Democrats so the nearby districts would be more firmly red. In some states, black Democrats even cut deals with the GOP mapmakers to get winning House seats. Their state parties had spurned them.\n\nThe first hint of what was to come for the most of this decade was seen immediately after the 2012 election. Obama was reelected with a nationwide 5 million popular vote majority. That showed that Democrats could win majorities when there are no boundaries inside states. But when it came to the US House, even though Democrats cast 1.4 million more votes nationally, the GOP ended up with thirty-three more seats. Democrats who won House seats often had higher overall percentages than Republican victors, but the GOP took more seats. REDMAP successfully resegregated voters: by race, party, or both. They erased competitive districts. Most of these elections were over before they began.\n\nAs is always the case in politics, there are publicly given explanations to deflect the truth. One academic study in 2013 downplayed the impact of redistricting, saying Democrats got clobbered because they lived in cities\u2014pushing the geography, not partisan intervention, explanation. But later studies by other experts, such as Princeton University mathematics professor Samuel Wang, found Democrats lost twenty-two seats in 2012 due to gerrymandering.\n\nLooking at each election this decade, a pattern has emerged and held firm. Republicans keep winning an outsized share of state legislative and US House seats despite the closeness of statewide contests. This continued in 2016. Republicans emerged with forty-seven more US House seats than Democrats. That's nearly 11 percent more, when, nationwide, 1.1 percent more GOP votes were cast in House elections. In states targeted by REDMAP\u2014where Trump and Clinton ran neck and neck\u2014the results are stunning. Pennsylvania's US House delegation is 13\u20135 red to blue. Ohio's is 12\u20134. Michigan's is 9\u20135. Virginia's is 7\u20134. North Carolina's is 10\u20133. Florida's is 16\u201311. A May 2017 report by the Brennan Center for Justice at NYU Law School concluded, \"extreme partisan bias in congressional maps account for at least 16\u201317 Republican seats in the current Congress\u2014a significant portion of the 24 seats Democrats would need to gain control of the House.\"\n\nAs bad as the congressional picture is, it's worse in the state legislatures. Since World War II, the president's party has typically lost 450 legislative seats as voter sentiment swings in the opposite direction during his term. Under Obama, Democrats have lost nearly one thousand seats. Daley writes the post-2010 redistricting is \"the story of how Republicans turned a looming demographic disaster [its aging shrinking white base] into legislative majorities so unbreakable, so impregnable, that none of the outcomes are in doubt until after the 2020 Census.\" His book came out before the 2016 election.\n\n\"State legislative numbers have not budged over the course of this decade,\" he said in 2017. \"The Ohio House is 66\u201333 Republican. Pennsylvania is 121\u201382 Republican. Michigan is 63\u201347. But what's really interesting to note is that it hasn't budged in those states in 2012, 2014, 2016. It's pretty much [at] the exact same place. They have taken all of the swing districts out of these states and Democrats have not been able to make even the slightest incremental gains in the most important states over this decade\u2014to which I'd add North Carolina, Virginia, Wisconsin, and Florida.\"\n\nHow big a built-in lead do Republicans have? \"It takes 55 to 56 percent of the popular vote of reliable Democrats voting] to have a 50\u201350 chance\" of winning, he said, adding that was an average and in some states it was higher. Other analysts like [FiveThirtyEight.com's Harry Enten said the GOP starting line advantage is closer to 8 percent. \"The mapmakers know the voters who turn out,\" Daley said. \"These districts are drawn by people with all that information preloaded into their software. They draw lines knowing who turns out and knowing what kind of elections they turn out for.\"\n\nOther attempts to measure the impact of redistricting agree with these figures. A major development since Daley's book is that federal courts have been looking at a wave of new redistricting lawsuits for the first time in years. With the exception of a GOP suit over a US House seat in Maryland, all of the cases have been filed against Republicans and date back to 2011 and REDMAP's handiwork. In a Wisconsin lawsuit that will be heard by the Supreme Court in its fall 2017 term, voting rights advocates are using a relatively new measure called wasted votes, which is expressed as a percentage of voter turnout.\n\nIt was developed by Nicholas O. Stephanopoulos, a University of Chicago law professor leading the team that challenged Wisconsin's Republican gerrymander as illegally partisan. It looks at the percentage of one party's votes that aren't needed to win\u2014that are wasted\u2014when voters are packed into districts (or conversely diluted from winning majorities). That gap in Wisconsin was 13.3 percent in 2012 and 9.6 percent in 2014. Stanford University political scientist Simon Jackson issued a 2015 report based on the 2012 and 2014 cycles that found similar gaps of more than 10 percent in Florida, Indiana, Kansas, Michigan, Missouri, North Carolina, New York, Ohio, Rhode Island, Virginia, Wisconsin, and Wyoming. Only two of these states are blue\u2014New York and Rhode Island.\n\nBy mid-2017, federal court rulings over 2011's redistricting seemed to be articulating new rules for when gerrymandering was and wasn't unconstitutional. There are two kinds of redistricting cases. The first involves partisan greed: grabbing as many seats as possible via jigsaw puzzle\u2013like districts. That's where the GOP lost in federal court in Wisconsin, the first ruling in a decade to throw out maps for partisan excess. The GOP's appeal of that ruling was accepted by Supreme Court, which has never thrown out political maps citing excessive partisanship.\n\nThe second category of redistricting cases involves the illegal use of race when creating districts. Republicans lost on that issue in lower federal court in Texas. In Virginia and Alabama, the Supreme Court told lower courts to go back and use different standards to examine how the GOP targeted voters by race. In May 2017, the Court\u2014voting 5\u20133 without its newest member, conservative Neil Gorsuch, ruled that North Carolina had illegally used race to create two House districts. They added tens of thousands of black voters to seats where black congressmen were winning\u2014a tactic to bleach surrounding districts and make them more solidly Republican.\n\nBefore North Carolina's ruling, the Court was hesitant to separate illegal racial factors from legal partisan actions. The majority rejected the Republican's rhetoric that they were merely looking for McCain voters and were trying to comply with the Voting Rights Act (VRA) to help create black seats. The majority concluded that North Carolina's GOP subverted the VRA because it first sorted voters by race, not party, and packed thousands into House districts where blacks were already winning. Dissenting justices said North Carolina's GOP didn't use race that way, accused the majority of looking for a \"smoking gun,\" and declared they had no problem with partisan redistricting \"while some might find it distasteful.\"\n\nRemarkably, the majority's opinion also quantified the advantage that extreme redistricting brought North Carolina's Republicans. It noted the GOP has consistently won House races with 56 percent of the vote this decade. In contrast, North Carolina's three Democratic House members were reelected in 2016 with 67 to 69 percent of the vote. This disparity comes from \"cracking\" and \"packing\" voters, the slang used by political consultants. In June 2017, the Supreme Court affirmed a lower court ruling that found North Carolina's GOP also had racially gerrymandered twenty-eight state legislative districts. What happens next in the state, as well as in other states where federal courts have ruled against Republicans, is an open question. That is because GOP-run legislatures that have redrawn maps have been accused by Democrats of doing too little to restore balances. That's been the case in Georgia, Alabama, and Texas.\n\nThe REDMAP plan didn't just seal a House majority for this decade. It also locked up state legislatures. This structural advantage has fueled ensuing red state rollbacks against voting rights, collective bargaining and union dues, women's reproductive rights, and LGBT rights. It's what has enabled red states' lawsuits against Obamacare, affirmative action in higher education, and challenges to Obama's climate change\u2013related actions. Democrats in 2017 are saying they will never allow a REDMAP-like project to recur. They say the courts are siding with them and ordering red-run states to revise state and congressional districts before 2018. They are focusing on key legislative and gubernatorial races in 2017 and 2018. They say they will have a place at 2021's redistricting and mapmaking tables\u2014with Democratic governors to veto bad maps. We will see.\n\nBeyond the audaciousness and longtime thinking that Republicans saw and imposed on a nation with more Democratic voters, REDMAP underscores how the GOP view elections. They see a system and rules that can be bent to serve their party before all other concerns. April 2017's vote by the Senate's GOP majority to revoke the filibuster rule for Supreme Court confirmations is another example of this pattern. Until then, sixty votes\u2014a bipartisan coalition\u2014were needed to end debate and move to a Senate vote. Majority Leader Mitch McConnell did everything to game the rules to get a right-wing court majority. The GOP blocked Obama's nominee and changed the filibuster rule to seat Gorsuch. They put their party before country, played dirty, and won.\n\nDemocrats typically take more fair-minded approaches, such as expanding the electorate. They are generally more naive and less cynical, less ruthless. They express those values with voter registration drives in underrepresented communities, by filing lawsuits to uphold voting rights, and by trying to prompt new voters to turn out\u2014especially the young and nonwhites. That's wide-eyed compared to the surgical thinking executed by Jankowski and REDMAP. Targeting reliable proven voters is more concrete. Republicans do not solely rely on enlarging their base and getting out that vote.\n\nIt's taken me a while to see this. I first heard the left's registration boasts in Ohio in 2004, when EMILY's List, the PAC for women candidates, said they helped register hundreds of thousands of voters. They did not know or appreciate that Ohio's hyper-partisan secretary of state, Republican J. Kenneth Blackwell, had overseen purges of large numbers of voters in the bluest cities. (From 2011\u20132016 under his GOP successor, Ohio has purged 2 million voters, more than any other state.) That tactic was one of many that Blackwell used to undermine Democrats. EMILY's List had perfectly good intentions. But they did not realize how far behind they were to start. They were optimistic they could get new voters to cast ballots. The Republican calculus was more precise and cutthroat. Jankowski's tactics were an expression of his party's do-everything-to-win philosophy.\n\nNo other voter suppression tactic scales with the impact of extreme redistricting. The GOP wants their voters to turn out and they work at it. They don't mind if Democrats turn out as long as that's in districts Democrats were going to win anyway\u2014and the GOP controls the levers of state power and the US House. What's revealing about the 2011 gerrymander is it's the one strategy they are least likely to brag about\u2014notwithstanding Rove's Wall Street Journal op-ed in March 2010. Compared to their rants about voter fraud and Democrats cheating, it is all but invisible. That affirms how significant it is.\n\nIn politics, insiders who know what's going on have an unspoken code of silence. The rarest commodity in politics is trust. It's only recently\u2014mostly since Daley's book in mid-2016\u2014that REDMAP's architects are telling their story. Other GOP mapmakers have kept mum. Some have been sued in states like Florida, where they recruited people to submit maps to supposedly nonpartisan state redistricting commissions. A quarter of the states have these panels. These consultants are playing dumb in court by discounting the power and impact of gerrymandering. (They may not have to do that after the Supreme Court's North Carolina ruling, where its conservatives said there was nothing illegal about partisan power grabs\u2014the issue in the Wisconsin case.)\n\nWhile extreme redistricting is emerging from the shadows as 2018 and 2020 loom, we need to turn to other antidemocratic features the GOP has placed atop this foundation. But it is critical to hold onto the view that the playing field is not level at its starting line. It is not even close. The goddess of liberty is not holding one scale in each hand, awaiting voters to weigh in equally. Aided by no serious opposition from Democrats in 2010, Republicans segregated voters to regain power. They built a 6 to 8 percent starting-line head start for scores of state and federal contests. However, even that advantage wasn't enough.\n\n\"The pattern is really, really clear. The first thing gerrymandered legislatures do in most of these states is try and limit voting rights. The second thing they do is go after labor rights. The third thing is women's rights,\" Daley said. \"There is a deep connection. It makes it much easier to win closer statewide races when you can limit the Democratic turnout and minority turnout with new voter ID laws, by making it harder to register, by conducting purges of voting rolls, by eliminating early voting days, all that.\"\n9\n\nLOOKING FORWARD, NOT BACKWARD\n\nSOME OF THE MOST CONSEQUENTIAL TRENDS in elections arrive with little fanfare. That was the case with the GOP's extreme 2011 gerrymander. The same is true with their next serious attack: suppressing major Democratic voting blocs. Unlike the twentieth-century's civil rights movement where the most visible bad guys were local police using clubs and fire hoses, this century's bad guys are a Brooks Brothers Riot of briefcase-carrying lawyers armed with shrill testimony and partisan propaganda. But they seek the same result: to keep the electorate from growing and diversifying by complicating the process and restricting who gets to vote.\n\nA dozen years ago, I witnessed a moment marking the arrival of Republican's fabricated crusade against what they call voter fraud. I remember thinking this cannot be serious. It was after 2004's presidential election in Ohio, where Republicans vehemently went after Democratic voters. A little-known lawyer showed up at a poorly attended congressional field hearing to offer the GOP's counter-narrative: cheating Democrats were the problem that needed fixing. Partisan accusations have been around forever. But his tirade had no basis. Hordes of imposters inundating the polls? Voting for Democrats but not getting caught? And all of this in red-run states such as Ohio?\n\nThis scenario was right out of nineteenth-century America. Then, urban machines and rural bosses paid so-called \"repeaters\" to do just this, as the University of Kentucky's Tracy Campbell describes in his 2005 book, Deliver The Vote: A History of Election Fraud as an American Political Tradition\u20141742\u20132004. A century ago, little was akin to today's statewide registration rolls, polling place voter lists, and check-in protocols. If you sign in to get a ballot with a fake name today, you are signing a perjury confession. That's a serious crime, ticket to a large fine, and possibly jail.\n\nThis is not to say that individuals do not make mistakes or try to game the system. Every election sees a handful of ex-felons who have not regained their voting rights trying to cast a ballot. There are a handful of people who think they can vote in more than one state as long as they own property there. There also are a few people who try to impersonate another voter\u2014the textbook GOP definition of voter fraud. They typically are a candidate's relative or a local official seeking reelection. Those are real-world examples. But these incidents literally are once-in-a-million voter occurrences, infinitesimal exceptions in a nation where multimillions routinely vote. They are so rare, so inconsequential, compared to real margins swinging elections, that something else is going on. And indeed there is.\n\nRepublicans know they cannot hold onto power when their base is aging and shrinking compared to the rest of the country's demographics. So they have seized upon a prejudice-fueled strategy as an excuse to overly police the process in pursuit of a phantom menace: Illegal Democratic voters. Despite producing no evidence that this is a real problem of any magnitude, combating voter fraud has become the GOP's go-to excuse this century to pass scores of state laws that complicate voting for entire populations trending blue. Those laws peel off slices of Democratic voters\u2014percent by percent. Every year, hundreds of thousands of ballots are spoiled, misread, or disqualified across the country. Yet Republicans don't ever complain about that\u2014or make federal cases over it.\n\nNonetheless, because the voter fraud meme draws on partisan vitriol, serves a partisan purpose, and crucially\u2014there is no federal agency that definitively tracks it\u2014the GOP has been using it to pass a spectrum of state laws and rules to intentionally disrupt voting. The GOP persists despite investigations by the Department of Justice, federal courts, Congress, academics, and journalists, some of whom have noted real-world instances of illegal voting are literally rarer than getting hit by lightning. Another confirmation came in April 2017 in a federal court ruling that threw out a Texas 2011 voter ID law as racially discriminatory against blacks and Latinos. The ruling noted that Texas had \"two convictions for in-person voter impersonation fraud out of 20 million votes in the decade leading up to SB 14's [the law's] passage.\"\n\nNumbers like that are not unusual. After the incumbent GOP governor of North Carolina, Pat McCrory, narrowly lost his 2016 reelection bid, his campaign consultants challenged the results by saying six hundred people fraudulently voted. They hastily fabricated and filed legal papers that named innocent people who were smeared in the local press. It was a ruse to try to throw the governor's race to its legislature, where Republicans held the majority. The State Board of Elections' April 2017 report found only two people impersonated another voter in 2016\u2014\"family members voting in place of a recently deceased loved one.\" That was out of 4.6 million votes cast.\n\nThe right-wing Heritage Foundation's post-2016 report of all national examples of false registrations, ineligible voting, fraudulent use of absentee ballots, and duplicate voting cites 492 cases and 733 convictions from 1982 through 2016. That is one case for every 2 million presidential election voters from 1984 to 2016 (approximately 980 million votes were cast). If you count by convictions, where some people pleaded to more than one charge, that total was still less than one in a million voters. Although there were more elections in America and multimillions more people voting in this period, which further dilutes their figures, Heritage nonetheless uses its report to further police the polls.\n\nBut made-up menaces serve a purpose. Republicans know it doesn't matter if they are pilloried in the press as long as they manipulate the rules of voting and elections. They don't care if political gain means disadvantaging opponents, including racially hostile tactics and openly discriminatory results. Looking back, I was present when a messenger arrived to herald the return of this phantom menace and its sleazy rationale. It was more layered and consequential than I realized at the time. But not taking this ruse seriously is exactly how Republicans finesse restrictions on voting into law and elections.\n\nThe confrontation pitted some of the civil rights movement's heroines\u2014women born under segregation who became members of Congress\u2014against this century's voting suppressors. Like Karl Rove's Wall Street Journal op-ed on REDMAP's strategy, it heralded a national wave of anti-voting tactics that continues today. Trump's Commission on Election Integrity, announced in spring 2017, is being cochaired by, and has seated people, whose attacks on nonwhite voters began during this time. Since he took office, more states have passed anti-voter laws than during the Obama Administration's final two years. The rationale was countering the made-up menace of voter fraud.\n\nThe setting in 2004 was a city auditorium in Columbus, Ohio. The Committee on House Administration, which oversees elections, was holding a hearing blandly entitled, 2004 Election and the Implementation of The Help America Vote Act. That law provided several billion dollars to states to buy new electronic voting machines as the response to Florida's problem-plagued presidential election in 2000. The hearing's title papered over two competing agendas. The committee's Republican chairman, Ohio's Bob Ney, wanted to place a whitewashed version of what happened in Ohio into the record. Weeks before, the House Judiciary Committee's Democrats had other ideas and took testimony detailing partisan barriers that delayed and disenfranchised tens of thousands of voters.\n\nSeated besides Ney were two now-deceased black Democratic congresswomen. Stephanie Tubbs-Jones, Ohio's first black congresswoman, was an ex-judge and local prosecutor from Cleveland. Six weeks before she stood in the US Capitol building with California Democratic Sen. Barbara Boxer to challenge the ratification of the 2004 Electoral College vote. They forced the House and Senate to return to their chambers and hear complaints about GOP-led voter suppression for two hours. Juanita Millender-McDonald was a PhD educator representing Los Angeles who grew up in Alabama. They came to question Republicans such as Ohio Secretary of State Kenneth Blackwell, who refused to shake Tubbs-Jones' hand and then bragged about that as a witness. Bleackwell is at the time of this writing on Trump's election integrity commission.\n\nState Senate President Jeff Jacobson testified he didn't understand why Ohio had become a poster child of problem voting. Without blinking, Millender-McDonald asked about one notorious tactic. Why had Blackwell's office told county officials to reject registrations that weren't printed on 80-pound paper\u2014a heavier stock than copied forms used by grassroots groups? Jacobson's surreal reply hinted at the GOP's arguments to come, and not just in Ohio\u2014but on Fox News for years.\n\n\"I would just say that fraudulent registrations are not part of the American way and groups that are paid to come in and end up registering Mickey Mouse and some of the other people that were registered in return for crack cocaine, the millions of dollars that poured in, in an attempt to influence Ohio, I think, is not normal and I would just state that it did make all of our jobs quite a bit more difficult,\" Jacobson said, justifying efforts to reject registrations turned in by civil rights groups like the NAACP or antipoverty groups like ACORN, the Association of Community Organizers for Reform Now.\n\nMillender-McDonald was as dignified as the state senator was unbound.\n\n\"Mr. Jacobson, I would hate for you to characterize voters as crack cocaine,\" she began, carefully choosing her words. \"Please do not characterize those who are doing the registration of people\u2014\"\n\n\"You may not be aware of what specifically happened,\" he interrupted. \"A gentleman was arrested and I think pled guilty that he was paid in crack cocaine for submitting registration cards and the registration cards that he submitted included Donald Duck, Mickey Mouse, Mary Poppins, and all kinds of others.\"\n\n\"Of course we have heard that, but is that not one incident of the many?\" she said. \"We have to be very careful that we do not show any semblance of arrogance on the parts of those who wish to vote and let their vote be counted. There are many minorities in this state and in every state who have been at the throws of not letting their votes be counted.\"\n\nThat exchange showed one party giving voters the benefit of the doubt and the other party doubting voters. It pitted the lessons from last century's voting rights movement against this century's vote suppressors. Despite his bluntness, Jacobson signaled the start of the GOP argument and strategy: seize on singular incidents of malfeasance\u2014such as a registration group's bad hire who forged voter applications, which is easily caught and prosecuted. Then hype that lone example into a menace to justify new laws that intentionally complicate voting for entire statewide populations, but especially targeting Democrats.\n\nMillender-McDonald urged him not to respond disproportionately\u2014but that was what the Republicans in Ohio and nationwide intended to do. Later that afternoon as she left for the airport, the most important witness testified and submitted a written report. Mark 'Thor' Hearne cut his remarks short because of the hour. He was a Missouri lawyer who worked with Rove on Bush's 2004 campaign. No Democrat there knew him. Nor had anyone heard of the American Center for Voting Rights he represented. Its website was launched only a few days before. As the Democrats focused on the election that barely ended, Hearne and the GOP looked to the future. He had come to launch a new wave of voter suppression.\n\n\"Reprehensible acts\" marred the 2004 election, Hearne said, such as the submission of fake registration forms by left-leaning groups associated with the Democrats. He spoke of \"an onslaught made against the Ohio election system,\" including lawsuits from civil rights groups that \"caused great chaos and confusion and difficulty for the election officials seeking to implement Ohio law and made it more difficult to have a fair and honest election.\" His prepared remarks, which he submitted after condensing them, were more pointed. \"Voter fraud was reported in every corner of the state and fraudulent voter registrations totaled in the thousands,\" his script said. \"Ohio citizens deserve the confidence that they\u2014the voters\u2014not trial lawyers, activist judges and special interest groups soliciting voters with crack cocaine determine the result of Ohio elections.\"\n\nHearne also submitted a report for the Congressional Record, which Ney accepted. It was a declaration of political war. Hearne and its coauthors, all Republican lawyers, went after every major Democratic political committee, voting rights organization, labor union, and antipoverty group. It cites \"MoveOn operatives,\" polls \"stalked by ACLU and 'voter rights' operatives,\" and even a \"table operated by MoveOn that promised 'Free Coffee.'\" These examples were beyond shrill, but they served a partisan point. Everyone Hearne cited not only supported expanding voting rights, but was a coconspirator in a vast conspiracy\u2014a characterization that endures today as Republicans keep pounding the voter fraud drum.\n\nHearne's report did cite one-off mistakes by people working for registration groups like ACORN\u2014who, as noted in the Heritage Foundation's 2016 report, were quickly caught by authorities. Every election has human errors, hubris, and passionate excess. But in a state with 5.7 million presidential voters in 2004, there is a false equivalency between isolated events where no illegal votes resulted, and statewide remedies complicating the process or preventing tens of thousands of voters. As Millender-McDonald warned, what matters is separating the one from the many.\n\nNeither Hearne nor any Republicans at that hearing cared about twenty-six thousand Ohioans whose regular ballots were not counted, nor about thirty thousand Ohioans given provisional ballots that were later disqualified, nor about tens of thousands who waited hours in the rain because polls in Democratic cities didn't have enough voting machines. Instead, they signaled the GOP would soon revive its pursuit of voter fraud with the White House fully behind it.\n\nAs Michael Waldman put it in his book, The Fight To Vote, Republicans were chasing the one problem that did not exist. But they chased it. Two years later, after three hundred investigations by the Justice Department, 119 people had been charged with voter fraud nationwide, with eighty-six convictions. One was a Wisconsin ex-felon who voted before his voting rights had been restored. In West Virginia, a sheriff, police chief, and a court clerk were caught paying residents $20 to vote for a slate of local candidates. This is truly small-time stuff. That record, nonetheless, prompted Bush's Attorney General, Alberto Gonzales, to fire seven federal prosecutors\u2014which caused a scandal eventually leading to resignations by Gonzales and Rove, then a senior White House strategist.\n\nThe voter fraud meme did not stop there. It solidified into a doctrinaire Republican belief that Democrats steal elections\u2014when, if anything, it is the Republicans who are guilty of this crime in the early twenty-first century. As the 2008 presidential race closed, GOP nominee John McCain accused ACORN of \"perpetuating massive voter fraud,\" even though there was no proof. After Democrats won the White House and Congress that year, REDMAP emerged. The GOP won key statehouse majorities in November 2010, but Republicans did not only set their sights on redrawing districts. Many believed McCain's finger-pointing. Fox News and Breitbart.com kept the hysteria alive. They now had majorities to solidify Republican advantages by rewriting more rules for voting.\n\nThe first thing Republicans did after their extreme redistricting was go after voting rights. In legislative sessions starting in January 2011, GOP voter suppression moved forward big time. Backed by front groups and right-wing legislation mills like the American Legislative Exchange Council, GOP legislators introduced more than 180 bills in forty-one states to toughen voter ID laws, cut back early voting, and penalize registration drive errors. The argument\u2014made by some of the same people now on Trump's \"election integrity\" commission\u2014was to combat voter fraud. Rhetoric aside, it was a war over access to the ballot.\n10\n\nTHE VOTING WARS\n\nIT'S AN OPEN QUESTION IF REPUBLICANS outside of elite circles like REDMAP knew how effective their 2011 gerrymander would be\u2014locking up state legislatures and House majorities for every election cycle this decade. But party leaders knew measures like tougher photo ID requirements to get a ballot at precincts would cut into Democratic turnout. Royal Masset, the Texas Republican Party political director, told the Houston Chronicle in 2007 that stricter ID laws could \"add 3 percent to the Republican vote.\"\n\nA 2014 study by the congressional nonpartisan Government Accountability Office on tougher voter ID laws in Kansas and Tennessee found that Masset was not far off. Those laws reduced turnout in Kansas by 1.9 percent and in Tennessee 2.2 percent, the GAO said. A May 2017 study by Demos, a think tank headed by former Connecticut secretary of state, Miles Rapoport, a Democrat, found that stricter ID laws in Wisconsin, Virginia, and Mississippi caused turnout to drop by an average of 1.7 percent statewide. However, \"turnout in counties with a high black population\" was 2.8 percent lower.\n\nOther research parses the impact of strict voter ID laws on non-whites even further. Three academics analyzing 2006\u20132014 data from the Cooperative Congressional Election Study, which sampled over a third of a million voters, found starker numbers. Compared to less restrictive states, stricter ID states\u2014such as Texas, Florida, Arizona\u2014saw November black turnout fall by 2.1 percent, Asian turnout fall by 5.0 percent, and Latino turnout fall by 8.3 percent. The drop-off was even larger in primary elections. \"By instituting strict voter ID laws, states can alter the electorate and shift outcomes toward those on the right,\" coauthors Zoltan L. Hajnal of the University of California San Diego, Nazita Lajevardi of Michigan State University, and Lindsay Neilson of Bucknell University wrote for the Washington Post in February 2017.\n\nWhat's happened to the electoral landscape for the past decade has been nothing less than an ongoing war of attrition. The pendulum first swings to the right after red states pass laws to make it harder for Democratic voting blocs to participate. That's been followed by a push back in court by civil rights groups seeking to restore procedural and voting options, where some of the partisan advantages favoring Republicans are undone. In states where voting wars have been most intense, voters, poll workers, and election officials have been left confused by last-minute court rulings and administrative directives. That's not an accident. Complicating the process also discourages participation, and lower-turnout elections favor the GOP's base. This dynamic is not just seen in presidential swing states. States that could put the House in play\u2014or should be a swing state like Georgia\u2014have watched Republicans postpone what should be their relegation to history books\u2014if demography was destiny.\n\nMore than one voting rights attorney has referred to the teens as a lost decade. What followed 2010's electoral GOP victories and redistricting was a pattern that has been repeated in each two-year cycle since. Soon after new legislatures convene\u2014typically January in odd-numbered years\u2014numerous red states saw dozens of bills introduced to make voting more onerous for those states' Democrats. Republicans almost never propose anything that affects voting by mail\u2014absentee ballots\u2014which their wealthier suburban base prefers. It is almost all aimed at polling places and precinct-based voting, which many urban voters prefer and where first-time voters are required to cast their ballot.\n\nSome of the most aggressive voter suppression bills were rushed through legislatures without hearings and were quickly signed into law. This started in 2011, when scores of bills were introduced in forty-one states. In nineteen states, twenty-five bills became law along with gubernatorial executive orders. These included tougher voter ID requirements, such as Texas barring state university photo IDs but allowing state handgun licenses. Wisconsin also created a photo ID requirement that excluded university IDs. Florida adopted big fines for voter registration drive errors, which stopped the League of Women Voters' effort for the 2012 presidential election. Ohio ended early voting on the last Sunday before Election Day, when clergy led congregations in a \"souls to the polls\" push. States scheduled voter purges in Democratic strongholds shortly before the presidential election. Kansas added a proof of citizenship requirement to its state registration form and passed a tougher ID law. Both measures discouraged nonwhites and youths from voting in state elections. All were aimed at Democrats.\n\nAgainst this array of barriers came a rapid response by civil rights groups and Obama's Department of Justice. Many laws were challenged in state and federal courts. The DOJ came in under the Voting Rights Act of 1965. That law, a high watermark of the Civil Rights Movement, requires the Justice Department to approve any change in voting laws or rules in states and counties with histories of racial discrimination in elections. It applied almost completely to the Old South, however, and not states such as Wisconsin, which have led Republican voter suppression efforts this decade. Nonetheless, under its so-called preclearance provision, the DOJ could veto anything it found would discriminate against minorities, compared to that jurisdiction's current rules.\n\nThe ensuing court fights did not correspond neatly to the election cycles. Many started in one and continued into the next. In some cases, new laws were put on hold while lawsuits unfolded. In other instances they were not. When the spate of new anti-voter bills was first introduced in 2011, the Brennan Center estimated the new laws could make it harder for five million people to vote nationally. That became a front page New York Times story. But after many legal challenges were filed, the Center reported anti-voter laws in fourteen states \"have been blocked, blunted, repealed or postponed\" just before the 2012 election.\n\nThat red-to-blue swing of the pendulum faced an abrupt reversal in 2013 that endures today. In a lawsuit and ruling that voting rights activists long feared, a right-wing Supreme Court majority ruled that preclearance\u2014VRA's Section 5\u2014was unconstitutional. Chief Justice John Roberts wrote race relations had evolved so the DOJ's oversight was obsolete. In a flash, states and counties with histories of discrimination in voting were deregulated. States under Section 5 jurisdiction in 2013 were Alabama, Alaska, Arizona, Georgia, Louisiana, Mississippi, South Carolina, Texas, and Virginia. Preclearance also covered counties and cities in California, Florida, New York, North Carolina, South Dakota, and Michigan.\n\nThe lawsuit prompting the ruling came from Shelby County, Alabama. It was designed by right-wing lawyers specifically to challenge the VRA's toughest provision. These advocates also drafted the federal suit challenging affirmative action in state university admissions that made it to the Supreme Court\u2014but lost. These are white supremacists that say civil rights groups are unduly negative and backward by continuing to focus on racial identity, racial differences, and discrimination. Publicly, they feign the nation has moved beyond racial divides. While this was all before Trump's overtly racist 2016 campaign, they were aware what their GOP allies would do\u2014restore race-based voter suppression measures in states where white governing classes are threatened by diversity.\n\nWithin hours of the Supreme Court's ruling, states demonstrated the absurdity of Justice John Roberts's opinion. Texas, North Carolina, and Alabama immediately acted to pass or announce they would begin enforcing anti-voter laws that had been stuck in court. Overnight, the voting rights pendulum swung hard to the right. Voter suppression had been legalized anew for the 2014 midterms, 2016 presidential election, and local races\u2014and was embraced in red states.\n\nWhat's often overlooked in analyses of Shelby County v. Holder is how many voting law changes came before the DOJ for their preapproval\u2014and involved local elections, such as those for city councils and university boards. The DOJ website said for the decade before Shelby it had \"received between 4,500 and 5,500 Section 5 submissions, and had reviewed between 14,000 and 20,000 changes in voting laws and procedures annually. A June 2016 report by the NAACP Legal Defense Fund noted, \"More than 85 percent of pre-clearance work previously done under Section 5 was at the local level.\"\n\nThe NAACP report highlighted the most common discriminatory tactics that ensued after the VRA was gutted, all of which dilute turnout or tilt the outcome in close races. These \"include reducing the number of polling places; changing or eliminating early voting days and\/or hours; replacing district voting with at-large elections [undercutting strong local candidates]; implementing onerous registration requirements like proof of citizenship; and removing qualified voters from registration lists.\" What the NAACP didn't quantify was what the GAO and Demos documented, and what Texas GOP Chair Royal Masset had said: that tactics like these shave a few points off of voter turnout, especially among Democrats.\n\nBy 2017, many voting rights groups have been saying that the pendulum was starting to swing back their way, as growing numbers of state and federal judges had been offended by the Republican's antidemocratic playbook. They note that courts are ordering states to redraw political districts marred by racially motivated gerrymandering, to stop enforcing new voter ID laws, and to reverse other anti-voter measures like cutting early voting. They cite progress that Democrats and their allies have made, such as states instituting more registration options and Oregon's model of universal automatic registration.\n\nThat view is accurate but incomplete. What it omits is that in the states that determine federal power\u2014House races and presidential swing states\u2014and states where the political complexion is purple or trending blue, that advocates of expanding voting rights are on the defensive. Also, lower-court victories have been hamstrung by GOP-filed appeals\u2014or by legislators dragging their feet to fix laws, or unfairly doing so, despite losing in court. Meanwhile, voting rights groups face new barriers in these states, as Republicans seeking more restrictive voting keep launching new offensives. The result is an ongoing spectrum of rules and requirements that makes voting anything but easy. The NAACP's post-Shelby report described the pinpoint targeting of attacks that led up to and fed into 2016. Each attack seeks to peel away blue voters: A few percent here, a few percent there.\n\nLAYERED AND NUANCED ATTACKS\n\nTAKE THE PURPLE STATE OF VIRGINIA. Consider what's unfolded since Shelby. In 2013, after the Supreme Court threw out the DOJ's veto power over local changes, Virginia\u2014which had been under the VRA\u2014did what a lot of red states did. It wasn't the most egregious. Its GOP governor and legislature quickly passed a tougher voter ID requirement. It was to take effect the following year in June 2014. Before Shelby, the DOJ would likely have blocked this law. Civil rights groups sued and noted 197,000 Virginians didn't have a state driver's license\u2014an acceptable form of ID. Many were nonwhites and lower-income.\n\nShrewdly, the ID law's authors added a provision so Virginians could get a free state ID. That was what they cited in court. Virginians had to make an effort, they said. Democracy should not be taken for granted, they argued, and the new law provided such a path. Yet a year later, fewer than five thousand Virginians had gotten that free state ID in order to vote. That result was what Republicans expected. That's how voter suppression works. The federal district judge who tossed Texas's ID law also noted 608,000 residents lacked ID, whereas the state issued 279 substitute IDs\u2014showing the same voter suppression template. Virginia's voter ID law remained in effect for 2016's election and was upheld in court that December.\n\nVirginia's voter ID fight wasn't in a vacuum. It had other voter suppression battles. After the 2010 election, Republicans aggressively packed black voters into twelve legislative districts in the Richmond-Hampton Roads region along its southeastern border. That gerrymander was challenged in court as racially discriminatory, prompting a federal court to order the state legislature to redraw its maps in early 2016. New maps were created, but that legal fight has continued with the US Supreme Court telling a lower court to use different standards in assessing whether the gerrymander was partisan, which is legal, while racial sorting is not. Both Democrats and Republicans claimed victory, but contested lines remain in effect.\n\nThere's more. In 2013, Virginia's political landscape shifted. That November, Democrat Terry McAuliffe was elected governor, starting a four-year term in 2014. Two years later, he signed an executive order restoring voting rights to two hundred thousand Virginians convicted of nonviolent felonies who served their sentences and completed probation or parole. Many were youths caught up in the war on drugs. The state was one of four in the country that banned ex-felons from voting for life. Virginia Republicans sued McAuliffe, arguing his clemency was an abuse of power because it did not look at individual cases. State courts put the rights restoration on hold. In July 2016, the Virginia Supreme Court killed it and ordered 11,662 ex-felons who had registered to vote to be removed from voter rolls. McAuliffe responded that he would act to individually restore their rights. In early 2017, McAuliffe reported that he had restored the voting rights of 140,000 people. Virginia's Republicans have yet to respond, but may be waiting until McAuliffe's term ends in January 2018.\n\nThese pendulum swings show how layered and deliberate Republicans are with trying to nullify slivers of the Democrats' base. It is maddening. State courts uphold a stricter voter ID law that adds complexities to the process for all of a state's voters, but block a governor who restores voting rights to a disadvantaged class. Republicans in Virginia and elsewhere don't care about ex-felons' due process rights. They care about preserving political power. As Marty Connors, then chairman of the Alabama Republican Party, told the Washington Post in 2004 when asked about felon re-enfranchisement, \"As frank as I can be, we're opposed to [it] because felons don't tend to vote Republican.\"\n\nNeedless to say, the states drawing the hardest lines against felon voting are states where restoring that right would threaten Republican rule. Nationally in 2016, six million ex-felons had lost their voting rights, according to the Brennan Center. A third were black\u2014one-in-thirteen voting-age blacks across the United States, Brennan reported. Half a million are in Texas. A quarter million are in Georgia. But Florida is the worst. There, 1.6 million\u2014one in ten voting-age adults\u2014are disenfranchised for life (although civil rights groups are hoping to put a state constitutional amendment changing that before voters in 2018).\n\nVirginia is a typical state in the voting war of attrition. It's purple and a presidential swing state. It's in the GOP mainstream and has close ties to the national party in Washington. Its tug-of-war on voting rights are indicative of this past decade, where political and legal fights have unfurled in state-by-state silos\u2014even as they use a similar playbook. Unless you live in these states, you are unlikely to hear about these power plays, ongoing court fights, their complexities and often-unresolved nature. Yet, these battles are all preambles to the stage upon which the 2016 election was set, including purple-trending states that saw aggressive anti-voter attacks\u2014such as Georgia and North Carolina.\n\nWe'll soon delve into what went on there\u2014because it is part of a landscape that is setting the stage for 2018 and 2020. But first we need to step back to get a better foundation in order to understand what the GOP's hit men are pursuing in the fine print of election law. They are burrowing into deeper parts of the process where federal and state laws can be ambiguous and have not been subject to enough lawsuits for clear national standards to emerge. So far, we have touched on big frames\u2014redistricting, changing the rules to one side's benefit, propaganda about fake threats to facilitate restrictive laws. Now we have to step back.\n\nThere's another lens that can clarify where other serious anti-voter attacks are occurring. If we think about voting as a narrative with a beginning, middle, and end, we can pinpoint why what's described as a technicality is anything but. Otherwise, it is easy to get lost in details of state-by-state fights, or get distracted by meaningless digressions from election officials or aggressive partisans.\n\nOne irrelevant detail that repeatedly surfaces is the claim that voter rolls are a gold mine for fraudulent behavior, when, in reality, they are not. Think about how many times you have heard a TV news reporter blare that voter rolls include dead people. \"People that have died 10 years ago are still voting,\" Trump said at an October 2016 rally in Green Bay, Wisconsin. Well, dead people do not notify election offices that they have died and ask to be removed. They also don't vote\u2014unless they voted early and passed away before Election Day, which is legal. It's an empty concern.\n\nThat illegal voter meme points to a corner of election law that has come under increasing GOP attack: voter registration rolls and how they are kept current. There are protocols about how county officials clean up their rolls, and how they face the same problem with people who die as with people who move\u2014and they're not notified of the change. These steps are in the legal nuts and bolts of the voter registration process. If we want to see how the GOP is playing dirty, and how the same people who have been behind the worst suppression gambits are resurfacing under Trump, we must examine what voter registration is and isn't, as well as voter purges.\n11\n\nTHE STARTING LINE\n\nTO BEGIN TO SEE HOW A straightforward process such as voter registration can be needlessly twisted, we first turn to Alabama. Its political culture repeatedly elected Jefferson Beauregard Sessions III to positions of power, first as state attorney general and US senator, before being picked by Trump as US Attorney General, where he now sits atop national voting rights enforcement. Sessions' namesake are his father and grandfather, both named after Jefferson Davis, the Confederate's president, and Pierre Gustave Toutant-Beauregard, the general who ordered the attack on Fort Sumter that started the Civil War. As a federal prosecutor in 1985, he tried but failed to convict three Alabama activists for voter fraud after they helped with absentee ballots. One was a former aide to Dr. Martin Luther King Jr. That sounds like long ago, but the past is never far behind the fine print of American elections.\n\nAlabama didn't just pass a stricter voter ID law within days of the Shelby decision with redistricting and announcements of intention to enforce it. As 2016's cycle began, the state tried to close thirty-one state motor vehicles offices across Alabama's \"Black Belt\" counties, where residents could get an approved ID. It backtracked by reopening those offices for one day each month, prompting civil rights groups to sue. Alabama's GOP-led redistricting also has been in federal court for most of this decade, where civil rights groups repeatedly won rulings that its state Senate maps were unconstitutional racial gerrymanders. Instead of appealing for a second time to the US Supreme Court, its Republican-majority legislature adopted new maps in May 2017\u2014on a party-line vote. That prompted black Democrats to cry foul and predict they would be back in court. That fight may leave Alabama's maps in place for 2018, however.\n\nBut that's not all. Alabama is among a handful of states where Republicans want to impose a secondary voter registration system to winnow who gets to vote in its state elections. This would not affect voting for Congress and president, but for local and state contests on the same day. This is another strategy for resegregating who gets to vote\u2014atop gerrymandering\u2014and is more than reminiscent of Jim Crow days. As Ari Berman has noted in The Nation, the impact of these anti-voter laws in the South is turning Democrats into a party of nonwhites and Republicans into a party of whites.\n\nThese states seeking a separate, unequal, and secondary state registration process\u2014Alabama, Arizona, Kansas, and Georgia\u2014updated the voter fraud meme in a way that anticipated Trump's big lie that millions of illegal immigrants voted in 2016, keeping him from a popular victory. These states want people registering to vote to provide paper proof of American citizenship. That's usually a birth certificate, passport, or naturalization papers, which not everyone has and takes time and money to get. As with voter fraud, there's no federal agency that maintains an authoritative US citizenship database.\n\nA handful of powerful Republicans in these states unapologetically keep pushing for this secondary standard despite seeing variations of this ploy lose in federal court. They know their allies in high statewide office can keep throwing the first punch by implementing new voter requirements and then fighting via ensuing litigation. That's because voter registration is done at the state level. The fact there is no authoritative federal database of all citizens creates an ambiguity they knowingly exploit. To understand why this is not just odious but could open a major new frontier for GOP-led voter suppression, we have to parse the starting line of voting in America, the registration process.\n\nUnlike most western countries and Canada, voter registration is not automatic for all adult citizens. Most people know they have to register, and if the state doesn't offer it online then federal registration forms are available at the post office. The exceptions are North Dakota, the only state that doesn't require any voter registration, and Oregon, which in 2016 became the first state to automatically enroll everyone who was getting a driver's license. (A half dozen other blue states have taken steps to follow Oregon's example.)\n\nAt first glance, registration seems simple. To be an eligible voter, you must be a citizen, a state resident, and at least eighteen years old. That's cited in most state Constitutions. Additional requirements often include not being a convicted felon (most states), not judged mentally incompetent, and swearing to uphold the state and federal constitutions. The federal voter application form seeks this basic information, which is then copied by election office workers from paper forms into computerized voter databases. The data fields or categories are clear: one's first, middle, and last name; street address, city, county, and state; political party if you're joining one; and an identifying ID\u2014usually a state driver's license or the federal Social Security number. Registrants also sign their name, which provides a signature (and signature image) and crucially constitutes a legal oath under penalty of perjury\u2014a major offense.\n\nFrom this baseline, officials who run elections\u2014usually at the county level\u2014will ensure you get the right ballot with party primaries and correct local and regional races. They will add you to polling place lists (if you are not voting by mail). This is straightforward until snafus emerge with partisan overtones\u2014such as removing infrequent voters from rolls, or moving local polling places without telling voters. But those come later. States also ask people that apply for a driver's license if they want to register, under the National Voter Registration Act of 1993 (also called the motor voter law). The NVRA requires military recruiters and other state agencies to offer registration. But many states have not done so at their welfare offices\u2014prompting lawsuits. (The NVRA also specifies how state officials are to conduct voter list purges, to protect voters but to remove people who move or die.)\n\nSome constraints appear at this electoral starting line. The first, imposed by states and parties, are registration deadlines. These recur before every election, including possibly the upcoming one if you are a new voter. In 2016, only thirteen states and Washington, DC, had Election Day registration, which allowed any eligible person to show up, register, and vote. Two more states, Ohio and Maryland, allow people who come to early-voting centers to do all this. The rest have deadlines lasting from several days to a month before. Some states, like New York, have longer deadlines for party primaries. No one who had not registered six months prior could vote in its 2016's presidential primary.\n\nThen there are requirements that have nothing to do with one's legal eligibility to vote. This is where a simple process gets murky. The most obvious is requiring a specific state-issued photo ID to get a ballot. No state constitution says a plastic card is a requirement to be an eligible voter. But come Election Day, there it is in seventeen states, especially in red states. One exception had been Arkansas, where its Supreme Court in 2014 overturned a GOP-sponsored ID law. The Court said that was an unconstitutional \"additional qualification\" and compared it to the state's segregation-era poll tax. In March 2017, however, the Republican Gov. Asa Hutchinson signed a bill reinstating the photo ID law. He claimed that new legislation was different and would withstand legal scrutiny.\n\nBut the constraint that is likely to assume a higher profile under Trump is the paper proof of citizenship to register. That's because it surfaced in a state where his attorney general hails from and is the brainchild of the man Trump appointed in May 2017 to be vice-chair of his Advisory Commission on Election Integrity, Kansas Secretary of State, Kris Kobach. Vice President Mike Pence is the panel's titular chairman. Kobach built his career by targeting the nonwhite populations that Republicans want to suppress in elections and now has a White House\u2013sanctioned podium.\n\nPaper proof of citizenship emerged a dozen years ago in Arizona. Kobach, then a lawyer and anti-immigrant activist in his late thirties, saw that adding it to the Republican catalog of voting \"reforms\" could help launch a political career. And it did. He helped draft a ballot measure in 2004, Arizona's Proposition 200, which included requiring citizenship proof to register for state elections. Around this time, Kobach and his allies sued three states for not charging higher out-of-state tuition rates to the undocumented students at state colleges and universities (they lost in court). He later befriended Maricopa County's anti-immigrant Sheriff Joe Arpaio and helped draft a state law letting police demand proof of immigration status to anyone pulled over in a traffic stop. Arizona voters passed Prop. 200 in 2004, starting a legal fight that most people thought ended in 2013 when the Supreme Court threw out most of the law. But Kobach\u2014this decade's Thor Hearne\u2014doesn't take no for an answer.\n\nDuring the race for Kansas secretary of state, his campaign website claimed that voter fraud had \"burrowed into every corner of our country\" and \"illegal registration of alien voters has become pervasive\" in his state. Weeks after taking office in 2011, Kobach convinced his legislature to adopt a tougher voter ID law and to require registrants to provide paper proof of citizenship. (They later gave him power to prosecute voter fraud, unlike any other secretary of state.) Kobach found like-minded Republicans in Alabama, Georgia, and Arizona. They did not need to be told the benefits of raising the bar to participate in state elections. In 2009, Georgia had passed its proof of citizenship law. In 2011, Alabama passed its law. Arizona Republicans, meanwhile, did not let the Supreme Court's 2013 ruling stop them from instituting the proof in its elections.\n\nTwo years later, this cadre launched a sly maneuver. The secretaries of state in Arizona and Kansas followed a suggestion in the majority opinion written by Justice Antonin Scalia that there might be another way to achieve their goal. The states sued an obscure federal agency, the US Election Assistance Commission (EAC), whose purview included the federal voter registration form. These Republicans did not just want stricter voter requirements and applications for their states. They wanted the federal agency to modify its registration form to help them impose the requirement\u2014and give its imprimatur so other states might follow. It sounds like splitting hairs, but state and federal elections are different legal silos. Kobach and his allies were retreating to state's rights terrain over how they ran their state elections. But they also sought a stamp of approval and help from a federal agency that oversees federal elections. In 2016, a Kobach ally who became EAC executive director, Kansan Brian Newby, unilaterally changed the instructions for the federal form to add citizenship demands by Alabama, Georgia, and Kansas. That was met with more lawsuits\u2014where the Brennan Center, representing the League of Women Voters in these states, submitted filings saying 7 percent of voters do not readily have paper proof. Their filing said Kobach's hurdles had blocked thousands of Kansans from voting.\n\nThat 7 percent figure is worth repeating. That is what's at stake in this arcane fight. Preempting another 7 percent of eligible voters who the GOP believes tilt to the left comes after redistricting's built-in advantage of 6 or more percentage points. It is on top of stricter voter ID laws that add additional 2\u20133 percent of November election advantage\u2014with a higher percentage advantage against non-white ethnicities. It is apart from the millions of ex-felons whose voting rights have not been restored. These are all ways that the GOP cumulatively targets the opposition. Many Democrats do not realize how these factors add up. The proof of citizenship fight hasn't gone away because the stakes are huge if the GOP can find a way to nullify signing one's name as a legal oath to register to vote.\n\nIn September 2016, a federal court blocked Newby's action. In early 2017, a court told the EAC to assess whether Newby had the authority to do what he did. That means this fight is not over. There's more evidence of that conclusion back in Kansas. The ACLU has sued Kobach for a new program called Birth Link, which seeks to automate the proof of citizenship process by comparing voter applications to state birth records. That discriminates against otherwise eligible voters born outside Kansas, one more example of his nativism.\n\nLike Trump, Kobach has said millions of noncitizens voted in 2016, but he can't offer any proof\u2014as there is none. But in the GOP's prejudice-filled world of voting, we have seen facts don't matter\u2014even if research repeatedly proves it. In May 2017, the Brennan Center released a report to rebut Trump's claims of immigrants illegally voting. It looked in forty-two jurisdictions in twelve states with the country's largest noncitizen populations. Out of 23.5 million votes cast in those jurisdictions in 2016, election offices could only cite thirty instances of \"suspected noncitizen voting,\" noting that figure was \"0.0001 percent of the votes cast.\"\n\n\"In California, Virginia, and New Hampshire\u2014the states where Trump claimed the problem of noncitizen voting was especially acute\u2014no [election] official we spoke with identified any incidents of noncitizen voting in 2016,\" the Center said. But during the 2016 campaign and afterward, Kobach stood by Trump's big lie. You can be sure the proof of citizenship charade will resurface as he cochairs the White House's \"election integrity\" commission. While the panel's work is likely to be ignored like past presidential panels, Kobach has continued to promote outright falsehoods as partisan gospel.\n\nDays after Trump named Kobach to his commission, The New York Times reported that he \"declared victory in the fraud wars\" by claiming he had unearthed 125 non-citizens out of 1.8 million registered voters in Kansas. That's 0.007 percent of its voters. Many of the noncitizens never voted, the Times said, but reported their mistake to local officials. \"Most of Mr. Kobach's nine fraud convictions involve people who voted in two states,\" it said. \"Neither a citizenship requirement nor an ID would have prevented those offenses.\"\n\nBut Kobach did accomplish partisan goals. When a federal court struck down some of his restrictions in 2016, it said these laws \"denied more than 18,000 Kansans their constitutional right to cast ballots,\" according to the Times. \"Whenever I hear Kris Kobach use the words 'voter fraud,' what that means in English for regular folks is voter suppression,\" Rep. Jim Ward, Kansas House Minority Leader, told the newspaper. \"Most secretaries of state see their job to be a fair arbiter of elections. Kris has believed that the secretary of state is a partisan tool to affect the results of elections.\"\n\nANOTHER HIJACKED STATE: GEORGIA\n\nIN THE 2016 CYCLE, THERE WERE other examples of Republican secretaries of state targeting the registration process to suppress votes. Georgia's Republicans showcased bad behavior in 2016\u2014just as it did in the 2014\u201315 cycle. Georgia is another state whose complexion would be purple were it not for aggressive efforts by Republicans to preempt voters and stymie voting to keep its rule red. This pattern can be seen locally and at the highest reaches of state government.\n\nOne eyebrow-raising local example came in 2015, when white Republicans in rural Hancock County, who are a majority on its county board of elections, sent sheriff deputies knocking on the front doors of more than 180 black voters in the town of Sparta. They were not concerned about real crime; they were questioning voter registrations. Their goal in dogging one-fifth of the town's voters was to help a white mayoral candidate, a lawsuit that followed said. The county attorney, a Republican state legislator, defended the police action by alleging sloppy voter rolls. This was intimidation by whites to keep blacks from voting.\n\nWhile Sparta was a local drama, a fight with statewide consequences involving registrations broke out that same year between Georgia's Republican Secretary of State, Brian Kemp, and its House Minority Leader, Democrat Stacey Abrams. Seen from afar, it looked like Kemp was modernizing the process by instituting online voter registration that year. But Abrams was leading an effort called the New Georgia Project that was running traditional voter drives, where canvassers went door-to-door and registrants filled out paper forms. In October 2014, it emerged that more than forty-one thousand of the eighty-seven thousand forms filed by the Project were not going to be processed by Election Day.\n\nThis disclosure was not accompanied by the usual gripe from local officials about voter drives\u2014that activists dump mountains of paper on their desks at the last minute. Instead, Kemp announced he was launching a major voter fraud investigation as one of his state's biggest drives in years was cresting. That accusation and the bureaucratic stonewalling that ensued was completely disingenuous. His investigation, completed months after Election Day, found problems with twenty-five of the eighty-seven thousand voter forms submitted\u20140.03 percent. But there's far more that Kemp did to block voters in the 2015\u20132016 cycle.\n\nRemember the data fields that comprise what information is required to register\u2014one's name, address, birth date, and driver's license or Social Security number? Kemp deliberately politicized a process that is similar to what is used in many states. When people do not register online, local election offices end up typing the information from paper forms into computers and the statewide voter file. Beyond illegible handwriting and typos that occur\u2014and cause some people to not be registered\u2014local officials also do electronic checks to verify the information. States validate a would-be voter's identity by pinging their driver's license database and the federal Social Security database using the last four digits of an applicant's Social Security number. Some also ping prison records to screen for felons. Under Kemp, any unconfirmed match rejected the registrants.\n\nWhere this becomes insidious is every one of those databases has had known accuracy problems or shortcomings that disqualify eligible voters. In 2009, the Social Security Administration (SSA) Inspector General's office assessed how reliable its voter verification was. Compared to \"other [Social Security number-based verification] programs used by the states and employers... (the voter registration) no-match rate was two-to-five times higher,\" rejecting an additional 16 percent of registrants, it found. Why? That bigger gap was because other SSA programs used the full nine-digit Social Security number, while the voter registration program only used the last four digits. Less precise matching yields more errors.\n\nIn 2005, Congress's GAO issued a report finding state drivers license databases confuse names, including full names, names with or without middle initials, aliases, etc. \"Even a 1 percent error rate on a match validating names, driver license numbers, etc., could generate tens of thousands of bad matches,\" the nonpartisan congressional analysts reported.\n\nThe Atlanta-Journal Constitution noted on how this process unfolded in their state in late 2014. \"If everything goes right, a match comes back and the voter's name is sent to the Secretary of State's Office. But it sometimes does not go right,\" they wrote. They cited the problems with Social Security and drivers' license data, and noted why state prison records also were unreliable. \"County officials must also ping the state Department of Corrections database, which may lag by up to several months in its information,\" the paper wrote. \"The Social Security Administration database only comes back as a 'yes' or 'no' match, giving county officials no help in determining what further information may need to be provided by the applicant. A hyphenated last name can cause hiccups with the state's identification database, as can unintentional data entry errors by county clerks, said former North Carolina [state] election director Gary Bartlett, who submitted an affidavit on behalf of the voter groups [who sued over the botched processing of registrations].\"\n\nIn other words, Kemp did not just masquerade behind the badge of fighting nonexistent voter fraud to thwart a registration drive led by a powerful elected Democrat. He also was winnowing voter rolls to impact the finish line\u2014who can vote on Election Day. And there's more. Between October 2012 and November 2014, he purged more than 370,000 inactive voters, which exceeded the number of newly registered voters. But at the starting line, voter registration, he was knowingly using a shoddy verification process to disqualify or hold registrants in limbo. Using porous data and imprecise analytics produces a politically expedient result: false positives. A lawsuit over Kemp's matching that was settled in early 2017 found that of the nearly thirty-five thousand registrants whose registrations he canceled between July 2013 and July 15, 2016, 64 percent were black, 8 percent were Latino, 5 percent were Asian, and 14 percent were white. Meanwhile, of those forty-one thousand registrants that Kemp kept from voting in 2014, eighteen thousand were approved after the election in which they registered to vote. This is how nuanced and targeted voter suppression works.\n\nKemp's tactics, especially using error-prone data mining, are not unique. One of its biggest proponents is Kobach, who, as Kansas secretary of state, oversees the Interstate Voter Registration Crosscheck Program. That project was created in 2005 by state election directors with the goal of identifying voters registered in more than one state and instances of illegal double voting. It does so by matching first and last names, birth dates, and state turnout data. It was used by twenty-five states in 2016. You would be correct if you concluded that it surfaces hardly any illegal activity\u2014single-digit instances from states each with tens of millions of voters. But its imprecise methodology generates hundreds of thousands of false positives.\n\nIn 2015, Crosscheck helped Kemp identify some 540,000 Georgia voters who were in danger of being purged before the 2016 election. That's one-eighth of the Georgians who voted for president that fall. This points to another fight in the lead up to the election\u2014mass purges by GOP secretaries of state of tens of thousands of infrequent but legally registered voters in Democratic strongholds. In early 2017, Georgia and more than a dozen red states filed a brief at the Supreme Court urging it to hear a case in its fall term over Ohio's purge of 144,000 infrequent voters in blue epicenters before the 2016 election. These red states sided with Ohio's partisan mass purge. The Supreme Court will hear the case, where Ohio acted after exploiting another ambiguity in federal voting laws.\n\nMeanwhile, the battle between Georgia Secretary of State Brian Kemp and House Minority Leader Stacey Abrams is not over. Both are running for governor in 2018 (as is Kobach in Kansas). But the tactic that Kemp and Kobach engaged in\u2014targeting and freezing out new voters, and partisan purges of infrequent voters\u2014is the next stage in the process where the GOP has an indisputable record of cynically exploiting a simple bookkeeping process.\n12\n\nTHE FINISH LINE\n\nTHE DEEPER WE DELVE INTO DETAILS, the more we see the GOP hijack technicalities. One of the most persistent is selectively purging voters, which is when registered voters are removed from voter rolls under the guise of keeping the lists current. When these voters show up to vote in states without Election Day registration, they are told they can reregister for the next election\u2014but they can not vote that day. Many protest and poll workers end up giving them provisional ballots. But those ballots are usually disqualified and don't count.\n\nThere are many legitimate reasons to keep rolls up-to-date by removing names, especially in our mobile society. List maintenance, as it's called, streamlines the process at the polls, saving counties and campaigns money on printing ballots and mailings, and helps with planning for turnout at precincts. But what should be routine bookkeeping can also be darkly politicized.\n\nThere are other recent examples where Republicans combined their propagandistic voter fraud meme and sought to use, or did use, sloppy analytics to target legal voters for mass removals. An especially eyebrow-raising example came from Florida in 2014, when GOP Secretary of State Ken Detzner scrapped a statewide purge after county officials openly rejected his attempt to impose mistake-ridden purge lists on them.\n\nThat rebellion was a remarkable development in the usually compliant world of election officials. But Florida is not just any state when it comes to purges. It got a black eye after the 2000 election for a catalog of dysfunction. One of its worst episodes involved Florida Secretary of State (and George W. Bush campaign cochair) Katherine Harris. She hired a contractor, Database Technologies, to winnow the state's voter rolls. The National Voter Registration Act lays out what counties are supposed to do, which, needless to say, Florida didn't do. The NVRA says no registered voter is to be removed unless they have not voted in four years\u2014two full federal cycles\u2014and only after postcards were mailed to them advising them they were being placed on inactive lists and could be purged if they do not resume voting. The 1993 law also says infrequent voting is not a sufficient reason to be removed from official lists, creating an ambiguity that's come under GOP attack\u2014but more on that later.\n\nWhat Harris did was piggyback on another racist legacy, Florida's lifetime voting ban for felons. While denying the vote to felons has constitutional precedents, it has been twisted in a handful of states to disenfranchise an outsized percent of nonwhites\u2014especially non-violent drug offenders in the war on drugs. In 2000, more than eight hundred thousand Floridians, about 7.5 percent of its voting-age population, had permanently lost their voting rights this way. (The figure in 2016 is double that: 1.6 million citizens, including 21 percent of voting age blacks.) Harris hired Database Technologies to create a list of \"possible\" and \"probable felons\" who were on Florida's rolls. It presented her office with 82,389 names, which was 1 percent of its electorate. In turn, she instructed county election officials to remove the voters who Database Technologies identified.\n\n\"The tallies were wildly inaccurate,\" wrote Michael Waldman in his book, The Fight to Vote. \"The search used primitive methods, such as identifying voters who share a common name and birth date with someone who had a criminal record.\" Thousands of legitimate voters were wrongly purged in the race where the Supreme Court stopped a recount and George W. Bush \"won\" by 537 votes. In his book, Give Us The Ballot, Ari Berman said that \"no one could ever determine precisely how many voters who were incorrectly labeled felons were turned away from the polls.\" He notes a legal settlement between the NAACP and Florida re-ran Harris's purge list with \"stricter criteria\" and found twelve thousand errors. More than poorly designed ballots\u2014where hundreds of Palm Beach County Democrats mistakenly voted for Pat Buchanan\u2014the purge was a bigger factor in tilting the odds to Bush via a hidden administrative process.\n\nFast forward to 2012 when Detzner revived Harris's script with a twist. He swapped her fear of felons with his fear of noncitizens. Two years before, Detzner, whom Republican Gov. Rick Scott appointed, claimed that there were 180,000 possible noncitizen voters based on state driver's license records. After pushback by voting rights advocates that number shrank to 2,600, and then to 200, and finally 85 people were removed from Florida's voter rolls. In 2014, Detzner tried again. This time, he wanted to use a federal Department of Homeland Security database to identify noncitizens, even though the agency said it was incomplete. While Detzner and DHS argued in court, his county election supervisors refused to accept purge lists from Florida's state election director. Their routine list maintenance process worked, they said. They didn't need a solution for a nonexistent problem.\n\nYou would think in the relatively small world of election administration that such brazen moves would be known and widely rejected. But across the country, local election officials rarely are that bold\u2014or some agree with the GOP's goals. In mid-2016, Kim Strach, the North Carolina Board of Elections executive director, also appointed by Republicans, used a PowerPoint demonstration to augment her testimony before her GOP-run legislature. It said 35,750 voters were \"registered in North Carolina and another state and voted in both [states for the] 2012 general election,\" Greg Palast noted in August 2016's Rolling Stone.\n\nWhere did her number originate? In part, the Interstate Crosscheck program run by Kobach. Under media scrutiny, Strach backtracked. \"We were not suggesting that 35,750 voters had committed any type of fraud,\" she told Palast. \"There were 35,750 people who voted in North Carolina whose first and last names and dates of birth matched persons who voted in the same election in another state.\" Strach was using porous data and knew it. \"Despite hiring an ex-FBI agent to lead the hunt, the state has charged exactly zero double voters from the Crosscheck list,\" the Rolling Stone article said.\n\nPalast then quoted Mark Swedlund, a \"database expert whose clients include eBay and American Express,\" who was stunned after looking at Crosscheck's Georgia and Virginia lists. \"God forbid your name is Garcia, of which there are 858,000 in the US, and your first name is Joseph or Jose. You're probably suspected of voting in 27 states,\" he said. Their point is partisan purges based on flimsy data are another large-scale means to take out eligible Democrats when the number of purged voters can exceed margins of victory.\n\nWhat's emerged since Palast's report is that North Carolina, through Crosscheck, had identified 455,891 suspicious voters as of February 2016. In Ohio, it identified 386,092 suspicious registrations. In Michigan, it was 406,268. In Arizona, it was 240,277. In the twenty-five participating states analyzed in 2016, more than 5.9 million registrants were flagged as suspicious by this Republican-led fishing expedition that pretends to protect the integrity of the vote as a supposedly effective administrative bookkeeping tool. These figures are from a March 2017 brief filed at the Supreme Court by Republican former Justice Department officials, including those who led the Bush administration's voter fraud witch hunts. They urged the Court to hear a case over Ohio Republicans' mass purge of hundreds of thousands of infrequent voters between 2012 and 2016.\n\nIn June 2017, the Court took the case, which will be heard in their fall 2017 term. In A. Phillip Randolph Institute v. Husted, Ohio's Republican Secretary of State Jon Husted is appealing a September 2016 federal appeals court ruling that found the purges he ordered violated the NVRA. Husted created a \"so-called 'Supplemental Process,'\" the ruling said, to identify all voters who had not voted in two years and put them on a track for removal. The Appeals Court said Ohio should not have purged infrequent voters\u2014as many people only turn out in presidential elections every four years. A dissenting opinion said the NVRA was \"ambiguous,\" because, on one hand it laid out a process to remove voters who did not show up for four years\u2014two federal cycles, but on the other hand it also said voters could not be removed \"by reason of the person's failure to vote.\"\n\nThis ambiguity was seized and exploited by Husted\u2014just as Republicans in other states seized upon the lack of a definitive federal citizenship database to demand paper proof of citizenship. Under Husted, Ohio has a notorious record for purging voters. As Ari Berman noted in The Nation after the Supreme Court said it would hear the NVRA case, \"From 2011 to 2016, Ohio purged 2 million voters from the rolls\u20141.2 million for infrequent voting\u2014more than any other state.\" The purges have unabashedly favored Republicans. In June 2016, Reuters reported that twice as many voters were purged in \"Democratic-leaning neighborhoods\" compared to \"Republican-leaning\" neighborhoods in the counties where Ohio's three biggest cities are located\u2014Cleveland, Cincinnati, and Columbus.\n\n\"The disparity is especially stark in Hamilton County, where affluent Republican suburbs ring Cincinnati, which has one of the highest child-poverty rates in the country,\" Reuters reported. \"In the heavily African-American neighborhoods near downtown, more than 10 percent of registered voters have been removed due to inactivity since 2012. In suburban Indian Hill, only 4 percent have been purged due to inactivity.\"\n\nThis disparity is another facet of what twenty-first-century voter suppression looks like. Democrats are targeted and peeled away, a few percentage points at a time. After the Sixth Circuit Court of Appeals ruled against Husted in September 2016, 7,500 of the purged Ohio voters were restored. (That's because the NVRA prohibits removals within ninety days of a federal election.) But that victory only restored a fraction of the otherwise eligible voters who were removed. It underscores how the GOP benefits by throwing the most destructive first punches in a voting rights war of attrition.\n\nWhen thinking about Ohio's purges, Crosscheck's use in North Carolina, Georgia's use of imprecise voter verification databases, or Florida's bid to seize any government file to find noncitizens, one must not only ask why are Republicans doing everything within reach to obstruct voting. One also must ask why are voter file procedures and systems so prone to partisan abuses? Why are they so outdated and unsophisticated compared to private-sector data mining, which tracks people's identity, address, and contact information, and lets no online search go unrewarded without deluging the searcher with banner ads?\n\nPart of the answer is politicians don't want to change a system that they mastered\u2014and that serves them. That is also true with the federal campaign finance system. As a result, elections often are kept as a backwater in state and federal budgets. They are underfunded and have aging voting machines that are hardly ever replaced. That takes on added significance when we look at the computer systems used\u2014some of which were targeted by Russians in the run-up to 2016's Election Day.\n\nBut mostly, Republicans don't want high turnout elections\u2014and that starts with registration. In 2010, I briefly worked with the Pew Center on the States as part of its Voter Registration Modernization Design Working Group. It was clear that states could use better data to verify registrations\u2014such as changes of address from credit rating agencies and the post office. It was also clear states could find and register all eligible voters if there was the political will\u2014which Oregon later did. Pew devised a plan for a far better data-mining operation than Crosscheck that top state election officials could convince legislatures to use. It was a way to make rolls more accurate, efficiently managed, and cost-effective. But it wouldn't automatically register voters, even though it identified all of the eligible but unregistered voters in participating states.\n\nRed state election directors said they could only take eligible voters up to the finish line, but those people would have to cross it on their own. They talked about citizens taking the final affirmative step. That's as far as sympathetic Republicans went. Under Pew's blueprint, the participating states would identify all their eligible voters and contact them to urge them to register. It also used better data to maintain voter rolls under the NVRA's timetable. The system it created, called ERIC\u2014Electronic Registration Information System\u2014was used by twenty states in 2016. Some of these states are also in Crosscheck's network, but do not use Crosscheck's data due to its shortcomings. ERIC was a reason why 200 million people were registered to vote in 2016. (Notably, no ERIC states are represented on Trump's election integrity commission led by Kobach.) But registration alone is not the key to winning. REDMAP showed it's not just about candidates and registrations, but getting a reliable base to the polls in sufficient numbers. Clinton's presidential campaign showed you could not alienate a sizable slice of your party's base and get all of them to turn out.\n\nPredictably, as the 2016 election traversed from summer into fall and Clinton led in the polls, Trump began making repeated claims that the election would be stolen by voter fraud. \"More than 1.8 million deceased individuals, right now, are listed as voters,\" he said, citing Pew's work. Of course, that Pew study never made claims about the dead voting; it was intended to show dead weight that needed to be removed by modernizing how voter rolls are maintained\u2014part of making a case for ERIC. But, as the right-wing blogosphere is wont to do, that statistic was distorted and exaggerated. By January 2017, Trump's Press Secretary Sean Spicer said, \"There's one report that came out of Pew in 2008 that showed 14 percent of people who voted were non-citizens.\" That Pew issue brief, also created to promote the need for ERIC, did not mention noncitizens either.\n\nBut facts don't matter when it comes to the voter fraud meme: double voters, dead voters, noncitizen voters, whatever. As a former Florida Republican Party chairman, Jim Greer, told the Palm Beach Post in 2012, voter ID laws and cutbacks in early voting were \"done for one reason and one reason only,\" to suppress the Democrats. Greer remarked Florida's Republican political consultants \"never came in to see me and tell me we had a fraud issue. It's all a marketing ploy.\"\n\nAmerica is nothing if not a marketing society. It's unclear how much Trump's revival of the voter fraud meme motivated his voters, apart from inflaming anti-Clinton sentiments among swing voters or preaching to the GOP's converted. A poll in mid-September 2016 from the Washington Post\u2013ABC News found that nearly half of registered voters across the country believed voter fraud occurs \"somewhat\" or \"very\" often. As The New York Times editorialized about that finding, \"Another 26 percent of American voters said that fraud 'rarely' occurs, but even that characterization is off the mark. Just 1 percent of respondents gave the answer that comes closest to reality: Never.'\"\n\nBy October, Trump wasn't just tweeting, \"Stop Crooked Hillary From Rigging This Election!\" He was trying to recruit \"Trump Election Observers\" to police the polls. As in past presidential years, Trump was sending out a call for GOP loyalists to go to precincts in blue strongholds and challenge the credentials of anyone who didn't appear to them to be a legitimate voter. There's a long history of Jim Crow-era incidents of this ilk. But beyond the GOP putting up billboards in recent cycles in minority neighborhoods in battleground states threatening arrest should someone not present the proper voter ID\u2014another lie\u2014this is more bark than bite. Still, there were a few overt voter vigilantes in 2016.\n\nVOTING VIGILANTES\n\nMOST POLICE-THE-POLLS THREATS REMAIN JUST THAT\u2014THREATS to discourage voting. But North Carolina saw a notable vigilante effort as the 2016 Election Day approached. A longtime right-wing activist, Jay DeLancy, who runs a group called Voter Integrity Project of North Carolina, targeted thousands of voters to be purged from three counties with large black populations. What he did is called caging\u2014a tactic the Republican National Committee was caught doing in the 1990s, and signed a federal consent decree agreeing to stop. In 2013, DeLancy told a local weekly newspaper, \"I can't be a Republican and do what I'm doing.\" But, of course, he is a Republican\u2014and an activist in a state where party and race overlap. DeLancy said TrueTheVote, a Tea Party-based group that started in Houston, Texas, and sees itself as a polling place posse, had inspired him. TrueTheVote subsequently failed to go national in 2012.\n\nDelancy's group specialized in a tactic that's an electoral mugging. They sent postcards to addresses of registered voters in three key counties\u2014people who had not voted recently. The postcards said, \"do not forward\" and thousands were therefore returned. They took those cards to county election offices as \"evidence\" to challenge those voters' registrations. The NAACP heard about this and sued. One week before the 2016 election, a federal judge ordered the counties to restore 6,700 voters.\n\nUS District Judge Loretta Biggs ordered the State Board of Elections to restore voters who had been purged within ninety days of the fall election. Needless to say, county election boards knew the NVRA's procedures\u2014it's been law since 1993. They should have ignored DeLancy's vigilantism. However, the voter suppressor, in remarks that echoed Kobach's ongoing proof-of-citizenship gambit, hewed to a state's rights line. \"The state did nothing wrong with following the law. We did nothing wrong in following the law,\" DeLancy told the Associated Press. The wire service wrote, \"he hopes the federal voting rules that Biggs cited in her ruling are ultimately challenged.\" DeLancy had plenty of company among North Carolina's voter suppressors. At the GOP-run State Board of Elections, staff attorney Josh Lawson said it would comply with the court order but might \"appeal facets of the ruling, potentially after the election.\"\n\nSloppy partisan purges, using bad voter data in accusatory ways, taking preemptive steps to deprive others of voting rights\u2014these tactics are taken for granted by fervent Republicans in North Carolina and elsewhere. They persist despite forceful court rulings in the past two years, starting with rejecting race-based gerrymanders in North Carolina, Texas, Alabama, and Virginia. Federal Courts have overturned other anti-voter laws in states like Wisconsin, Pennsylvania, and Kansas. But North Carolina stood apart in 2016\u2014even surpassing Texas, where courts have struck down laws as racially discriminatory and Republicans keep trying to reinstate them.\n\nIn July 2016, a federal appeals court threw out five North Carolina anti-voter laws that its opinion said \"target African-Americans with almost surgical precision.\" These included a more restrictive voter ID requirement, reduced early-voting options, and fewer registration opportunities, from eliminating same-day registration to blocking teenagers from registering before they turn eighteen. All were adopted after the Supreme Court gutted the Voting Rights Act in 2013. Less than a year later in spring 2017, the Supreme Court refused to hear the state's appeal\u2014saying it would reserve the right to return to those issues another time, but not in a state with such messy politics. Later that term, the Court ruled in two cases that North Carolina's redistricting was illegally racist\u2014for two House seats and twenty-eight state legislative seats.\n\nWhat unfolded in North Carolina as the 2016 election crested was as brazen as any of the Republican actions to seize or hold power this decade. It showed that as the pendulum swung back toward voting rights, the GOP had no reluctance to impede nonwhites and Democrats from voting, to disqualify their votes, and to overturn results. North Carolina's story is even more remarkable because before REDMAP targeted it in 2010, it had the most progressive and inclusive voting laws in the South. It was politically purple, but the Republicans viciously rolled back the clock to turn its complexion red.\n\nAll of that was apart from what the Russians may also have been doing in North Carolina by targeting its voting systems. As of late spring 2017, there were questions if hacking could have scrambled voter registration data and e-poll books in Durham County, which may have led to computer breakdowns, delays for voters, and the necessity for poll workers to turn to paper ballots. When asked what caused that those snafus, Durham County BOE staff quickly rejected hacking and blamed \"user error.\" That's voters and poll workers\u2014anyone but those actually responsible for managing the voting systems.\n13\n\nTHE PLAYBOOK: DO EVERYTHING\n\nAS THE ELECTION CAME TO A close, other factors outside of the fine print of voting undoubtedly helped Trump and the GOP. These included former FBI Director James Comey announcing in late October that the bureau was reopening its investigation into Clinton's handling of her classified email as US secretary of state, her campaign's decision not to step up Clinton's visits in Michigan and Wisconsin, the still-simmering rifts between her supporters and the Sanders camp, and the continuing leaks of Clinton campaign emails by Wikileaks\u2014which federal intelligence agencies blamed on Russian hacking. Meanwhile, beyond Trump's taunts to \"Lock her up\" and rants that Democrats always steal millions of votes, the GOP in a handful of states kept returning to the voter fraud meme to tilt the results their way.\n\nStates we haven't discussed much, like Arizona, enacted new barriers to complicate the process. In mid-2016, Arizona made it a felony to collect and turn in another person's absentee ballot. That targeted seniors, especially in Latino communities. Indiana deputized party officials to patrol polling places and demand voters to show their IDs, another intimidation tactic. In Marion County, where Indianapolis is its largest city and many blacks live, GOP election officials only opened one early voting site, compared to more than twenty in adjacent white suburban counties. In all, fourteen states had new restrictions on the process as Election Day approached. In North Carolina, where a federal court in July 2016 threw out anti-voter laws passed after the Supreme Court gutted the Voting Rights Act, the state's GOP drew on its playbook and showed no tactic was too shameless when power was at stake.\n\nAs the November vote approached, North Carolina's Republican-dominated government and election bureaucracy were already ignoring the federal court ruling that its post-VRA laws were illegally racist. It revived Jim Crow era tactics simply making it harder to vote. Black turnout was down 16 percent during the first week of early voting because \"in heavily black counties, there were 158 fewer early polling places,\" The Nation's Joan Walsh reported. On Monday before Election Day, the state party issued a release bragging about this result. \"African-American early voting is down 8.5 percent from this time in 2012. Caucasian voters early voting is up 22.5 percent from this time in 2012.\" The GOP spun this as a sign that Obama's coalition was \"crumbing.\"\n\nIt's worth revisiting fine print in that July 2016 court ruling because it is microcosm of the motives and strategies embodied in the GOP's war on voters during this decade. Despite its extreme redistricting in 2011, US Court of Appeals Judge Diana Gribbon Motz wrote that North Carolina was trending purple until the Supreme Court gutted the VRA in 2013. \"African American registration and turnout rates had finally reached near-parity with white registration and turnout rates. African Americans were poised to act as a major electoral force,\" she wrote for the three-judge appeals panel.\n\nNonetheless, the state's Legislature and governor moved \"within days\" to pass laws that \"target African-Americans with almost surgical precision... and, in fact, impose cures for [alleged voter fraud] problems that did not exist,\" the ruling said, recounting the history. It then cited another federal ruling about similar illegal tactics in Texas\u2014that state's post-VRA voter ID law\u2014where \"the state took away [minority voters'] opportunity because [they] were about to exercise it.\" That's as clear an affirmation of institutional racism in elections as any federal court is apt to make.\n\nJudge Motz didn't stop there. She named the Republicans' motives: To roll back the clock to segregationist days. North Carolina's GOP had been waiting years to do this. Between 1980 and 2013, the DOJ had rejected \"Fifty proposed election law changes,\" the ruling noted, while voting rights groups had won fifty-five cases under the Voting Rights Act. \"The legislation came into being within days of North Carolina's release from the pre-clearance requirements of the Voting Rights Act. That long-ago history bears more heavily here than it might otherwise,\" she wrote for the majority.\n\nMotz's opinion would not call the GOP outright racists: \"Our conclusion does not mean, and we do not suggest, that any member of the General Assembly harbored racial hatred or animosity toward any minority group.\" But it did say that the GOP laws targeted \"voters who, based on race, were unlikely to vote for the majority party.\" That contradiction is imbued in southern politics\u2014taking a genteel stance that a person is not racist even if their actions have racist consequences. But the ruling's footnotes were not blind to that reality and showed some Republicans were unabashedly power hungry, racist, or both.\n\nJudge Motz's appeals court ruling came in response to a lower federal district court ruling that upheld the anti-voter laws passed immediately after the Supreme Court gutted the VRA. The appellate court chastised the district court for acting as if it was still 1955 and segregation reigned.\n\n\"The district court took no issue with one of the Legislature's stated purposes... to 'move the law back to the way it was,'\" Motz wrote. The accompanying footnote was stunning. It quoted a \"Republican precinct chairman who testified before the House Rules Committee that the photo ID requirement would 'disenfranchise some special voting blocks [sic]' and that 'that in itself is the reason to vote for the photo ID, period, end of discussion.'\" That GOP precinct chairman was Don Yelton, an AM talk radio host who called himself the Rush Limbaugh of western North Carolina. He told Comedy Central's The Daily Show in late 2013, \"If [the law] hurts the whites so be it. If it hurts a bunch of lazy blacks that want government to give them everything, so be it.\"\n\n\"The sheer outrageousness of these public statements by a party leader does provide some evidence of the racial and partisan political environment in which the General Assembly enacted the law,\" that footnote ended, returning to the cautious wording of federal court decisions. As expected, North Carolina's GOP vehemently denounced the court and ruling, saying Democrats had appointed Motz and her colleagues.\n\nBut they were not the only red state getting slapped for anti-voter actions in 2016. The same day of the North Carolina ruling, federal District Court Judge James Peterson struck down some Wisconsin laws, writing \"a preoccupation with mostly phantom election fraud leads to real incidents of disenfranchisement.\" A week earlier, Texas reached a settlement with the DOJ and civil rights groups to loosen its new voter ID rules after being sued for racial discrimination. For a moment, it seemed like the pendulum in this decade's voting rights war was swinging back in the direction of civil rights.\n\nThese rulings should have been the end of the story, but they were not. As Election Day neared, North Carolina was like many states. Despite a close governor's race, Clinton and its Democratic US Senate candidate, Deborah Ross, a former state ACLU director and state representative, were ahead in polls. Reporters like The Nation's Walsh were upbeat about \"a new South led by North Carolina.\" She mentioned repeated visits by President Obama and the First Lady, by Vice President Joe Biden, and by Bill Clinton to fight \"powerful headwinds, given the state's herculean efforts at voter suppression.\"\n\nWe know what happened. On November 8, Clinton lost the state by nearly 4 points. Ross, like the five other Senate candidates who were key to regaining a majority in that body, also lost. In the governor's race, the incumbent Republican, Pat McCrory, was trailing the Democratic state Attorney General Roy Cooper. As results trickled in\u2014each county had to count its regular ballots, provisional ballots, absentee, and overseas votes\u2014McCrory's campaign consultants began accusing Democrats of voter fraud. Hundreds were named as illegal voters in vote-contesting filings and were smeared in the press. The state party filed an emergency motion with the US Supreme Court, under a complex theory it might get a ruling throwing the selection of the next governor to the GOP-led legislature.\n\nThese moves showed that just when you thought the election might be over, it's not. The same was also true for Democrats across the nation who did not understand why Clinton lost\u2014especially after polls showed her leading despite Comey's resurrection of the FBI's investigation. In Michigan, more than seventy-five thousand ballots were initially reported on the state's website as missing presidential votes. That state had the nation's tightest finish, with eleven thousand votes separating Clinton and Trump. In Wisconsin, where a tough voter ID law was in effect for April's primaries but thrown out by a federal judge in July, Trump won by nearly twenty-three thousand votes. Milwaukee County, with its largest population of blacks and tens of thousands of students, cast sixty thousand fewer votes than in 2012. (Only two years before, another court noted that an estimated three hundred thousand Wisconsin residents, mostly poor and students, lacked required IDs to get a ballot.) And in Pennsylvania, voting machines in dozens of counties that went for Trump were entirely paperless, which meant that there was no way to audit or verify the reported result.\n\nAs the campaign peaked, Clinton was crestfallen by Comey's investigation, preoccupied with Russian interference, and believed that she was ahead in key Midwestern states\u2014Michigan and Wisconsin. Her campaign sent Clinton and its resources elsewhere. These observations and miscalculations are detailed in the 2017 book Shattered: Inside Hillary Clinton's Doomed Campaign by Jonathan Allen and Amie Parnes.\n\nWhat's not in that book and most post-2016 accounts is how Democrats, including many Berniecrats, voted for Clinton\u2014even if it was more out of opposition to Trump than affirmation of her. But she was an insider in a season favoring outsiders. The Democratic establishment anointed Clinton expecting she would be facing ex-Florida Gov. Jeb Bush. Nonetheless, many coastal Democrats rallied and believed she was more progressive than her husband. Wall Street, Silicon Valley, and even northeastern Virginia, where Pentagon contractors and federal employees live, saw her as a technocrat who understood them. But across interior America, Clinton was unconvincing for many reasons. Those ranged from her cautiousness on the stump to decades of being pilloried as an elitist, liberal feminist by right-wing media in those regions. Too many women, let alone men threatened by strong women, rejected her.\n\nAll these factors, which politically attentive Americans are aware of, played into Trump's favor. But what is overlooked in those cultural and political analyses is how the stubborn Republican game plan of doing everything to rig the rules and tilt the field comes into play when contests are tight and the turnout is less than a landslide. The 2016 election didn't see the turnout waves and popular vote margins of victory that Obama inspired in 2008 (9.5 million votes) and 2012 (5 million votes). While her 2.9 million popular vote victory was bigger numerically and percentage-wise than John Kennedy's in 1960 and Richard Nixon's in 1968, it wasn't sufficient to translate into an Electoral College victory in 2016.\n\nRepublicans turn to a do-everything strategy because our elections are so decentralized. There are ten thousand jurisdictions overseeing one hundred thousand Election Day precincts nationwide\u2014to say nothing of states that vote early and vote by mail. It is virtually impossible to predict the final swing precincts, in the swing counties, in the swing states. No pundit, pollster, or data analytic expert accurately foresaw where she would fall short by tens of thousands of votes compared to Obama in 2012.\n\nThat's why Republicans deploy such a vast playbook of starting line advantages, such as extreme redistricting giving them upwards of a 6\u20138 point lead, and other barriers such as strict voter ID laws that also shave another 2\u20133 percent off of Democratic voting blocs in November (and even more in the primaries). Nobody precisely knows where the fault lines will appear in a country with 137 million presidential voters, especially in a year like 2016 with no final battleground state\u2014unlike Ohio in 2004 and Florida in 2000.\n\nWhile Democrats were mortified with the outcome\u2014Trump's victory and the Democrats' loss of the Senate\u2014the travesties didn't end on Election Night. They continued when Americans wanted to know what happened, including who voted for Trump and why. They found elections are filled with additional opacities, namely, a non-transparent vote verification process, and partisan and institutional reluctance to recount close elections.\n\nThat absence, after a race that never seemed to end, was another antidemocratic feature of our system of voting. When citizens are left grasping for facts and officials don't or won't provide them, then the most cynical, biased, and conspiratorial forces win. Voters deserved to know who voted for Trump and trust that counts were accurate\u2014or realize why voting machinery is unreliable. The Green Party's recounts sought those answers, and they were blocked on almost every front.\nSECTION III\n\nTHE RECOUNTS\n14\n\nWHAT HAPPENED?\n\nNO DEMOCRAT OR PROGRESSIVE NEEDS TO be reminded how they felt on 2016's Election Night. During the day, precinct exit polls by a national media consortium of the print and broadcast outlets reported a tight race but likely a Clinton victory. But as results trickled in and states assumed to be reliably blue lined up behind Trump, and the GOP took control of Congress, disbelief and shock set in. At about 11:15 PM Eastern, the AP called North Carolina for Trump. A half hour later Fox News called Wisconsin for Trump. Clinton was ahead in Pennsylvania, but her strongholds had all reported and Trump was gaining. At midnight, David Simas, Obama's White House political director, called Clinton Campaign Manager Robby Mook and urged them to not \"drag this out.\"\n\nClinton wouldn't concede until the next morning. Immediately across America, voters, campaign volunteers, and others who did not imagine a Trump victory and GOP sweep, did not accept this result. The public, especially Democrats, wanted to know what happened, what went wrong and hoped there might be some way to prevent Trump and GOP from taking power in January 2017. An angst-filled campaign ended with even more angst.\n\nJust when explanations and understanding were needed, mainstream media failed to deliver. The biggest outlets, which for months hawked their latest polls as gospel, did not discuss what they had missed or misreported. Nor did they consider pausing to verify the results in swing states, where, only days before, polls had projected Clinton as winning. Instead, the outlets did what they always do in the aftermath of close and controversial elections. They seized upon an easy narrative to explain away the results, while ignoring a genuine desire for actual proof. The most seized-upon narrative was low-income, working-class whites came out for Trump in a spasm to preserve their place in an increasingly diverse nation. (By mid-2017, academics found Trump's base was white but was not low-income; most lacked college degrees but two-thirds made more than $50,000 (just below America's median household income), with a third making more than $100,000.)\n\nWhat ensued after November 8 was schizophrenic. Beyond the GOP hyping Trump's win as a mandate and Trump blaring he would have won the popular vote if millions of immigrants had not voted illegally\u2014a lie\u2014state election directors saw November 8 as a wide success. Compared to 2012, the top officials said there were fewer delays and polling place problems. Across the country, voting in communities of color and white areas was more similar than not, they added, applauding themselves. These same officials spent the fall downplaying reports that Russians had breached their computerized voting systems. By spring 2017, it emerged that thirty-nine states were targeted and one had been penetrated\u2014more than anyone reported during the election. But Election Day did not produce obvious telltale signs of interference, such as outsized requests for provisional ballots by voters not on precinct lists. The officials overseeing elections wanted to slam the door on 2016, sweeping under the rug the less visible, antidemocratic, poorly performing, or possibly malicious features of the election. They turned to certifying the results and taking vacations.\n\nThat landscape left a handful of voting rights activists, constitutional lawyers, computer science experts, and journalists looking for answers and explanations\u2014or offering their own. Chris Thomas, who retired in early 2017 after thirty-six years as Michigan's election director, called these people \"wonks,\" \"crusaders,\" and \"professional irritants\" in an exit interview with a trade publication, reflecting how many in his profession see these advocates.\n\nNonetheless, one of the first analyses to emerge came from a cadre of election integrity activists who, since 2004's presidential election, have been tracking discrepancies between media exit polls and the Election Night results. (These counts are not the official totals that come days or weeks later, but unofficial ones used by the media to project winners and by the candidates to declare victory or concede.) These activists emphasized that international election observers use exit polls as a big clue for detecting fraudulent results. They were astonished. As Election Defense Alliance cofounder Jonathan Simon wrote, \"By early morning [November 9] I had begun circulating tables documenting the most egregious 'red shift' exit poll to vote count disparities ever recorded in the computerized voting era. Even for those accustomed to the mysterious and pervasive rightward shifts between exit poll and vote count results, the results were eye-popping.\"\n\nBy \"red shift,\" Simons meant exit polls that predicted Democratic victories in contrast to results adding up to a Republican sweep. He wrote, \"Ohio had shifted from an exit poll dead heat to an 8.1 percent Trump win; North Carolina from a 2.1 percent Clinton win to a 3.6 percent Trump win; Pennsylvania from 4.4 percent Clinton to 0.7 percent Trump; Wisconsin from 3.9 percent Clinton to 0.7 percent Trump; Florida from 1.3 percent Clinton to 1.2 percent Trump; and Michigan from a dead heat to 0.3 percent Trump... If the exit polls rather than the vote counts were accurately capturing voter intent, the Electoral College majority would have gone to Clinton and it would not have been close.\"\n\nSimons and his colleagues were grabbing at the only available information to question whether the counts had been hijacked. Needless to say, they believed that it had. Election officials ignored them, as did the media consortium and its pollsters. As has been the case for years, the consortium refused to release any raw precinct-level data. (During the day, they revise projections to reflect a better picture of electorate.) As number crunchers on list-serves screamed this was proof of yet another stolen election, exit pollsters countered these activists were gadflies reading too much into imperfect exit polls. (One of their team who conducted exit polls in Berkeley, California, the publisher of Ballot Access News, Richard Winger, told me that no voter he interviewed said that they had voted for Trump\u2014but there were several who did in the official count at his precinct.) The exit poll discrepancy stalemate left a trail of dots that could not be connected without more proof, which was not forthcoming.\n\nAnother early line of inquiry came from computer scientists and voting rights lawyers who were not convinced that foreign or domestic hackers\u2014Russians or Republicans\u2014had not targeted and planted malware that fractionally readjusted the electronically tabulated counts. Barbara Simons, a retired IBM programmer, past president of the Association for Computing Machinery, advisor to the US Election Assistance Commission since 2008, and board chair of the advocacy group Verified Voting, suspected there was more going on than what the White House and federal intelligence agencies said. James Clapper, then National Director of Intelligence, said the spy agencies had concluded that the Russian government was behind the hacks and thefts of emails from the Democratic National Committee and Clinton campaign that were posted on Wikileaks. While that theft and publication served its purpose as anti-Clinton propaganda, FBI Director Comey and President Obama said the voting machinery was separate, secure, and that the vote count could be trusted. Simons knew the machinery was not secure. For more than a decade, her colleagues had routinely used voting machines in university classrooms to demonstrate hacking. Some reproduced those same experiments in congressional hearings as far back as 2006, where she was among those testifying to the House Committee on Administration.\n\nBarbara Simons talked to associates in North Carolina. In blue epicenters like Raleigh-Durham, there had been problems with electronic poll books producing precinct voter lists. She knew that the private contractor that had maintained that database and other state voter registration databases\u2014Florida-based VR Systems\u2014had been hacked by Russians. Simons wondered if Russian hackers had scrambled North Carolina's poll books. As Steve Friess wrote in a lengthy piece in February 2017 for the New Republic, five days after November 8, Simons, Amy Rao\u2014a Silicon Valley CEO and Democratic fund-raiser, and David Jefferson\u2014a computer scientist at Lawrence Livermore National Laboratory\u2014got John Podesta on a conference call to urge he \"call for an official review of the election results.\"\n\nThe scientists made the case that has been the baseline for more than a decade in activist circles seeking transparent and verifiable elections. The voting machines and tabulators installed after Florida's 2000 disaster with computer punch cards were aging, prone to inaccuracies, easily breached by hackers, maintained by private companies with partisan leanings, and, in a quarter of the country where voting systems were paperless\u2014like most of Pennsylvania\u2014impossible to recount. Podesta asked if they had proof the vote had been stolen. They replied no, but said a targeted audit of key counties and states would affirm who had won. Podesta demurred.\n\nMeanwhile, another line of inquiry was developing. John Bonifaz, a voting rights attorney who won a MacArthur \"genius\" award in the 1990s, also saw anomalies. Bonifaz was not another attorney frustrated with the prospect of a Trump presidency. He was the Green Party's lead attorney in their 2004 presidential recount in Ohio. (He also represented the Libertarian party.) When he realized that the Greens' 2016 presidential candidate, Jill Stein, could file for recounts if she and her party acted quickly, Bonifaz took on convincing her and trying to raise the legal fees. I've known John for two decades and wasn't surprised when he called a week after Election Day and asked whether I could write an article that suggested a recount could address serious questions.\n\n\"A lot is happening. We're looking at anomalies beyond the exit poll questions. We're looking at potential recounts in several states\u2014not audits, but recounts,\" he said. \"Jill Stein, on Sunday, said she's prepared to take this on. Only candidates can ask for recounts. She's willing to do it.\"\n\nBonifaz explained the filing deadlines and trail they hoped to follow. In Wisconsin, he said paperless machines had markedly different victory margins than scanned paper ballots. In Michigan, the secretary of state's office reported eighty-seven thousand paper ballots lacked presidential votes\u2014almost double the number from 2012. (That would soon be adjusted to 75,335 undervotes.) In Pennsylvania, sixteen counties were using the same vote-count tabulators that computer scientists had shown a decade before were hackable. Bonifaz cited VR Systems' work in North Carolina and said there were questions about whether volumes of absentee ballots could be fabricated in Michigan, Wisconsin, and Pennsylvania. He wondered if the VR Systems hack could provide a pathway where election workers would unwittingly transport malware from voter registration databases to separate vote-counting computers. \"That's the theory. It's strong enough for us to use as a basis for a recount,\" he said.\n\nOn November 18, AlterNet published my story suggesting that lingering questions were raising the prospect of filing for recounts in several states. While Bonifaz tried to raise money for the filings, which were estimated to cost several million dollars, I reached out to state officials and others I knew from my work with Pew to see what they were hearing and knew. They frowned on Bonifaz's theories as dots that would not connect, even as they said they also were trying to figure out what had happened. \"People are flailing and looking for something,\" a former official who ran elections in blue and red states said. He was skeptical that recounts would change anything, but suggested the Greens look at the machines' \"logic and accuracy\" tests. (Activists counter that's insufficient, as it doesn't reflect what happened on Election Day.)\n\n\"If they don't have them, that gives you something to poke at then,\" he said, before making what turned out to be the most prescient point. \"What judges are looking for when you file suits is real evidence that results should be overturned. It has to be so screwed up that a judge will overturn it. That's not the same as a recount... recounts don't prove much.\"\n\nLater that week, in the small world of election activists, it was inevitable that Bonifaz and Simons would join forces to again try to persuade Podesta to help\u2014privately if not publicly. The Clinton campaign, like John Kerry in Ohio in 2004, met their voters' desire for answers with silence, not even suggesting they were evaluating the outcome. Alex Halderman, the director of the Center for Computer Security and Society at the University of Michigan, had listened to the first conference call with Podesta while aboard a flight. Afterward, he recruited thirty nationally known computer security experts and statisticians to join him to discuss possibly auditing the results. Some wanted to hand count seven hundred thousand ballots in twenty-nine states, saying such selective audits would provide certainty about the election. Others said auditing ten states could suffice. As the New Republic noted, the prospect of audits was more of an academic opportunity to examine the machinery than a legal procedure that could overturn results. Nonetheless, Halderman and the experts added credibility. A second conference call was held with Podesta, longtime Clinton adviser Jake Sullivan, DNC general counsel Marc Elias, and others. By this point, Bonifaz was hoping the Clinton campaign\u2014which had taken no public position\u2014would quietly point major donors his way.\n\nHowever, that didn't happen either. Instead, news of the conference call slipped out and was reported by New York magazine, which named Bonifaz and Halderman as coleaders of a likely recount effort. It wrote they \"found persuasive evidence that results in Wisconsin, Michigan, and Pennsylvania may have been manipulated or hacked.\" Needless to say, that breathy assertion was quickly seized upon by many\u2014from Democrats in Congress to journalists to voters everywhere\u2014as an offer of hope that somehow the election results could be challenged or even reversed. Bonifaz asked me to hold off on writing more about formally filing in those states, as he was trying to raise millions for filing and legal fees.\n\nAs the Thanksgiving holiday neared, he wasn't succeeding. Stein and the Greens decided to try to crowdfund the effort by turning to their grassroots base and frustrated voters. To everybody's surprise, starting midday Wednesday and continuing over the weekend, the Greens raised $7.3 million in donations from 161,000 online donors, and began recruiting what became ten thousand volunteers. That unprecedented response for an unprecedented multistate presidential recount was treated with scorn and suspicion by mainstream media and the Republicans' top propaganda outlets, Fox News and Breitbart.com.\n\nEven though it was starting late\u2014two weeks after Election Day\u2014the recounts held a faint prospect of getting some answers. Namely, were the votes cast accurately and legitimately counted? That would begin to reveal which communities had surprisingly supported Trump and the GOP, abandoned the Democrats, or had not voted at all. What ensued frustrated all those expectations. As Stein wrote in an email blast on the eve of Trump's inauguration, \"We had little idea how hard the political establishment would fight against transparent, accountable elections.\"\n15\n\nTHE FINAL SWING STATES\n\nTHE BIGGEST PLAYERS IN OUR ELECTIONS\u2014STATE administrators, political parties, the press, voting technology companies\u2014all take it as gospel that the system is as democratic, above board, and as free from malfeasance and error as any major endeavor can be. That credo becomes an excuse to defend a status quo that serves them differently, even as power shifts. Nowhere is this more forcefully seen than with the subject of hacking or electronically stealing votes.\n\nTo dismiss hacking as trivial or conspiratorial in 2017 is naive. Yet that's been the line for years\u2014continuing even as Russia got caught stealing DNC and Clinton campaign email, fabricating emails from contractors sent to election officials, and getting inside the Illinois voter registration database. As ex-CIA and NSA Director Michael Hayden told a Heritage Foundation audience in October 2016, hacking adversaries' electronic data is old hat. \"I have to admit my definition of what the Russians did is, unfortunately, honorable state espionage,\" he said. \"A foreign intelligence service getting the internal emails of a major political party in a major foreign adversary? Game on. That's what we do. By the way, I would not want to be in an America court of law and be forced to deny I never did anything like that as director of the NSA.\"\n\nCloser to home, police regularly use hacking technology to intercept and capture online, Wi-Fi, and cell phone-band communications, including devices where the FBI has gone to court to try to keep a lid on sales to departments. These are the same data pathways that voting machines use to report precinct results to central tabulators. More concretely, former state election directors have told me that state motor vehicles databases, which are linked to voter registration databases, are hit routinely\u2014\"on a weekly basis\"\u2014by overseas hackers, especially from China. There is a false equivalency in the media, which covers GOP claims about voter fraud extensively\u2014even though it barely exists, while ignoring hacking of government systems including those tied to voting and elections. There is an establishment bias that does not want to admit that shortcomings can undermine public trust in elections, instead of proactively addressing them and showing the results can be trusted.\n\nThe central problem in this sphere is that there haven't been any serious public examinations, ever, of looking for hacking in voting systems in the immediate aftermath of nationwide elections for the purpose of possibly overturning results. Some academics say it would be impossible to trace malware that fractionally adjusted vote totals at key central tabulation nodes, because these bits of code can erase their electronic footprints. Others say it's still worth looking at. Either way, the voting machinery, with few exceptions, is privatized, protected from scrutiny as trade secrets, and maintained in most counties by contractors. All that feeds the accurate characterizations that it is an unnecessary black box. None of this is reassuring.\n\nStill, I do not first plant myself in the company of those who say nothing else in elections matters if votes can be electronically stolen, because that is ignoring too many other antidemocratic features. It's not an either-or choice. The GOP's attacks on voting are politically shrewd but administratively crude. They take aim at broad swaths of the electorate in a do-everything strategy, because elections are so full of unanticipated twists and turns.\n\nJust as crucially, in more than a dozen years of covering electoral nuts and bolts, I have never seen sufficient proof of major federal elections being stolen electronically. By proof, I mean beyond inferring conclusions from suspicious data points such as exit poll discrepancies, or documenting dubious behavior like Ohio's GOP secretary of state posting the 2004 results on a state website where the RNC had backdoor access to its server; and thus access to what was posted for the media and public. (I helped break that story. George W. Bush came from behind to defeat John Kerry after Ohio's results were frozen for ninety minutes.) To know that the rest of the voting process has been micro-targeted by the GOP for partisan gain but to think that voting machines and tabulators are immune or sacred would be illogical. There may be proof of hacking that fractionally adjusts vote counts, but we just haven't seen it. This fraught terrain set the stage for 2016's recounts.\n\nWhat the Greens wanted was to verify the vote by having the country's most respected computer scientists look inside the black boxes and share what they found. \"History came knocking\u2014who was I to say no to this effort,\" Stein said. Also, her party had a track record, unlike the Democrats. In Ohio in 2004, the Greens' involvement in the recount led to some structural reforms, at least by then Democratic secretaries of state in Ohio and California. Those states started a national wave of slowly replacing entirely paperless voting systems with hand-marked paper ballots that were electronically scanned, but could be manually examined in recounts and audits. Michigan used paper-based systems in 2016. That's also true for Wisconsin, except for touch screens with printouts for voters with disabilities. Pennsylvania mostly used the oldest entirely paperless systems.\n\nThe Greens never came close to looking inside. As Stein summarized in her preinaugural email to donors,\n\nIn Michigan, Trump's GOP cronies stopped the recount [after one-third of the state reported] despite\u2014or maybe because of\u2014revelations about major problems with the vote count, particularly in under-resourced black and brown communities. In Wisconsin, although the law was on our side, many low-income communities of color most vulnerable to tampering never got the hand-count needed to verify the vote. But the worst mess was in Pennsylvania, where over 80 percent of voters use touch-screen machines with no tangible paper backup. These paperless voting machines are so vulnerable to hacking and so difficult to verify that they have been banned in most states.\"\n\nWhat emerged from the recount were more partisan attacks that were on par with the most antidemocratic features of the primary season and General Election. In Michigan, Bill Schuette, the Republican attorney general, led the legal charge on Trump's behalf to convince its courts to stop the recount, because, among other arguments, Stein had no real chance of winning. What was overlooked in that argument was the most easily answered question surrounding the state with the closest margin between Clinton and Trump\u2014why did seventy-five thousand ballots lack presidential votes? Most explanations were deeply unsatisfying, amounting to, as Philip Bump wrote in the Washington Post in December, \"It happens.\" He reported that nationally, 1.7 million people in thirty-three states cast no presidential vote. Michigan was in the bottom third of states, percentage-wise. What Bump and many mainstream media reporters never argued was that a hand recount of paper ballots, not just running them through high-speed scanners, would not just reveal whether those ballots missed votes. It would reveal where Trump's base was, where people who skipped voting lived, and other important facts that would have helped the country move forward.\n\nThe highest-profile snafus that emerged in Michigan's brief recount were in Detroit. There, it became apparent that the city's elections were a mess. More than half the precincts were barred from the recount because the number of ballots inside tabulator bins did not match the number of voters written on the outside of those bins by poll workers. As Jan BenDor, the Michigan Election Reform Alliance's statewide coordinator, noted at the time, the state has a 1954 law giving full power to local officials to look at all the ballots in recounts. But state election officials, pressured by Republicans in constitutional offices, countered that a 1979 administrative rule prevented them from doing so. Many progressives and civil rights attorneys saw that decision as institutional racism, treating Detroit separately and unequally. When Secretary of State Ruth Johnson published a report on Detroit's election woes in 2017, the incongruities were mostly blamed on poll workers. While their incompetence may have been real, just as local Iowans and Nevadans running Democratic Party presidential caucuses botched those contests, that explanation\u2014beyond reassigning blame\u2014underscores the system is not up to its assigned task.\n\nOther postelection analyses noted additional institutional barriers in Michigan. The affidavit developed by state Election Director Chris Thomas\u2014where people lacking ID or not listed in precinct poll books swore they were registered voters and received a regular ballot\u2014was not promoted by poll workers, according to a summer 2017 report by Miles Rapoport, the former Connecticut secretary of state and Democrat, who was president of Demos and Common Cause. He also wrote, \"Michigan is another Crosscheck state, where five hundred thousand 'potential duplicates' were found, and at least fifty-five thousand voters were taken off the rolls as a result.\"\n\nIn neighboring Wisconsin, different obstacles emerged in a state that until recently had the best nonpartisan election overseers in the country, a bipartisan panel of retired state judges known as the Government Accountability Board. However, Republican Gov. Scott Walker and a GOP-majority legislature retired that panel, among other anti-voter laws they passed, and raised the statewide recount filing fee from about $18,000 to $3.5 million. After Stein's campaign paid up, the newly created Wisconsin Elections Commission said counties could run the paper ballots through high-speed scanners instead of recounting them by hand.\n\nThat decision flustered the Greens' intent of a verifiable recount and led them to sue. Their court affidavits from computer scientists, such as from University of California Berkeley computing and mathematics professor Philip B. Stark, noted the 22,525-vote difference between Clinton and Trump was 0.776 percent of the 2.9 million votes cast. That margin, percentage-wise, is less than the standard error rates of high-speed scanners, he said. Dane County Circuit Judge Valerie Bailey-Rihn ruled that even though a hand recount was preferable, she could not force counties to do that. As a result, Wisconsin's most populous counties\u2014Milwaukee (440,000 presidential votes), Waukesha (237,000 votes), Brown (129,000 votes), and Racine (94,000 votes)\u2014all rescanned their ballots, flouting best practices to verify the vote.\n\nActivists watching the state's hand recounts tallied nearly eighteen thousand votes that were incorrectly scanned on November 8. But when it was all over, officials said that Trump had increased his victory margin by about four hundred votes\u2014prompting media yawns and ridicule of the Greens. Meanwhile, Arizona-based activist John Brakey noticed and confirmed, contrary to many statements by officials, that one of the most commonly used voting machines in Wisconsin and nationally, ES&S model DS200, used cell phone modems to transmit precinct results to central tabulators. That potential hacking pathway cast doubts on the results, Brakey said. Legal discovery by the Greens also revealed the state did not require the contractors who maintained and programmed the voting machines to take extra security steps. Despite these windows into Wisconsin's less-than-ideal infrastructure, it was the only state to complete a recount, and no new numbers emerged to change the result.\n\nThe state that received the least attention in the recount, Pennsylvania, was the most antidemocratic. \"The primary reason Pennsylvania was recounted was to take a close forensic look at the electronic voting machines, since the state is not designed to be recounted,\" wrote Bob Fitrakis, an Ohio-based attorney and longtime legal advisor to the Greens on elections, in the Columbus Free Press, where he is editor and publisher. Local election integrity activists like Dr. Michelle Zuckerman-Parker, a medical researcher in Allegheny County, where Pittsburgh is located, typified the volunteers who tried to help the Greens. She documented inconsistencies among \"data values\" in the county and state systems managing voter rolls and tabulating votes, and was hoping a recount might spotlight these lapses. While that was the agenda of computer scientists and election integrity activists, the Greens' grassroots supporters faced an insurmountable task on the ground: submitting recount petitions and notarized affidavits from three voters in each of the state's 9,158 precincts and laying out $450,000 in filing fees.\n\nAs Jonathan S. Abady, the lead lawyer for the Greens' legal team wrote to Attorney General Loretta Lynch on December 23, 2016, about Pennsylvania, \"This year, voters who sought to confirm that their votes had been accurately recorded by requesting recounts were thwarted by a byzantine, unworkable, legal regime.\" County boards conducting official vote counts don't disclose when they've finished, Abady noted, which is supposed to trigger a narrow recount filing window. \"Two of the state's largest counties (Allegheny and Delaware) admitted in court that their boards of elections do not comply with the legally mandated process for completing the count. Not even the state's top election officials knew when the various counties had finished counting the vote...\"\n\nAbady didn't stop there. He noted how one state court demanded a $1 million bond from ordinary voters filing recount petitions. He said that the need for recounts was prompted by \"at least two types of vulnerabilities: susceptibility to malicious interference and poor performance.\" Turning to other recount states, Abady cited eighty-seven broken machines \"in Detroit alone.\" He noted Halderman \"has personally hacked into several voting machines as part of a research study, including the optical scan model used in Michigan.\" Abady wrote, \"FBI Director James Comey's sworn testimony to Congress that no election machines are ever connected to the internet... is inaccurate,\" citing the new \"ES&S DS200 model is in broad use in approximately twenty-five states.\" He concluded, \"The United States should investigate these irregularities and vulnerabilities,\" adding the country needed \"paper ballot-based systems,\" \"automatic audits, for every election,\" and \"adequate funding of our election system to maintain voting machines and train election staff.\"\n\nNeedless to say, neither the federal government, nor the Clinton campaign, nor Pennsylvania's Democratic executive branch leadership including Secretary of the Commonwealth Pedro Cortes, nor its state judiciary bothered to reply to the substance of the Greens' concerns. As Fitrakis wrote in early January, \"U.S. District Judge Paul Diamond states that suspicion of a hacked election in the Quaker State 'borders on the irrational.' In prohibiting a presidential recount, Diamond held that there was 'no credible evidence that any \"hack\" occurred and compelling evidence that Pennsylvania's voting system was not in any way compromised.' Diamond's ruling came a day after the Washington Post claimed that a secret CIA report found the Russians interfering in the US election electronically in order to hurt Clinton and aid Trump.\"\n\nStill, many voting rights attorneys, even those who have led the fights against GOP anti-voter laws, are more than reluctant to say administrative mistakes and misdeeds should be subject to criminal penalties. They fall under the \"bad policy\" umbrella, which lawsuits can challenge and legislatures can rectify, these lawyers explain, rather than statutory crimes such as explicit voter intimidation and specific Voting Rights Act violations. (While I've known and respected these lawyers for many years, I can only wonder aloud if the betrayal by our voting system of American democracy requires its custodians to be held to higher standards and penalties\u2014even if they will be deeply opposed by most public officials.)\n\nBy early fall 2017, when this book is being finished, we know a lot more about what Russia did and didn't do. When Trump finally met Vladimir Putin at a July G-20 meeting in Germany, both perfectly played the part of any longtime politico who never would admit to anything that would undermine their power. They agreed that Russia didn't interfere in the election or collude with the Trump campaign. (The Justice Department's Special Counsel, former FBI director Robert Mueller, may have more to say about that narrative.) Russia clearly interfered and had influence, said Timothy Snyder, a Yale historian of early twentieth-century authoritarian regimes\u2014including the Soviet Union. In a mid-June speech, Snyder explained the Russians didn't need to tinker with voting machinery, even though they were probing those systems, because they found a much more effective way to influence voters: inundate them with propaganda on social media.\n\nSnyder said we know that Russians hacked the DNC and Clinton campaign emails before spring and summer 2016. Trump also invited them to do more of it while campaigning, he noted. (That was before disclosures in 2017 congressional testimony that dozens of states were targeted, and the early June publication by TheIntercept.com of a National Security Agency document with more details, such as fake emails sent by Russia to election vendors and local offices. Notably, most actions cited by the NSA were after October 27, 2016, long after voter registration had closed in Florida, Georgia, Michigan, Nevada, North Carolina, Ohio, Pennsylvania, and Virginia.)\n\nSnyder explained that Russia didn't have to focus on voting machinery because they had a bigger, more effective strategy. As Bloomberg's Joshua Green and Sasha Issenberg reported in late October 2016, the campaign's social media operation targeted Clinton supporters on Facebook and Twitter with negative messaging to discourage them from voting for her. These weren't just traditional political ads that had been diverted online, but as Facebook said in its April 2017 report, \"Social media accounts and pages were created to amplify news accounts of and direct people to the stolen data. From there, organic proliferation of the messaging and data through authentic peer groups and networks was inevitable.\" In other words, Trump's campaign bombarded swing-state Democrats with political ads and fake news, some of which were based on information stolen in Russian hacks of the DNC and Clinton campaign.\n\n\"The main element of the Russian intervention in the election had to do with gathering political data about tens of millions of Americans, and then using Facebook and other platforms to target fake news to people who were regarded as susceptible, especially in the last weeks before the election, which probably determined the outcome in critical states that were targeted,\" Snyder said. \"This isn't to say that there aren't other reasons, and in my view, good reasons, why people could vote for Mr. Trump. It isn't to say that Hillary Clinton ran a wonderful campaign either. It's just to say that in a very close election, that probably made the difference.\"\n\nIf Snyder is correct, Russia more than learned the lessons of the Republican playbook. They looked at 2016's political landscape and found vulnerabilities where targeted interventions could scale. Why mess with Michigan's 1,600 local election jurisdictions or Wisconsin's 1,850 election jurisdictions, and their respective computer systems, when they could use social media to target those states' voters with provocative propagandistic messaging? That tactic was akin to the \"honorable state espionage\" cited by former CIA and NSA Director Hayden in his October 2016 speech at the Heritage Foundation. Strategically, this is no different than any other negative political advertising campaign seeking to discourage the opposing side's voters. Instead of deluging voters with attack-dog mailings, they used social media platforms like Facebook and promoted propagandistic stories in American right-wing media that drew on information obtained by Russia's hacking of Democrats.\n\nBut let's step back and grasp the big picture. Whether extreme partisan redistricting, newly restrictive voter ID requirements, paper proof of citizenship requirements, limiting early voting opportunities, moving polling places, repealing Election Day registration, there is a long list of Republican strategies that shave points off of likely turnout by Democratic voting blocs. These tactics are all politically shrewd because they are administratively crude. They take aim widely at multitudes and await the consequences. It appears that Russia followed this same template.\n\nMeanwhile, research since the election addressed some of the anomalies cited by Bonifaz on the eve of the recounts. A study by the University of Michigan's Walter Mebane Jr., a professor of political science and statistics, and Matthew Bernhard, a PhD computer science scholar, examined the recount data in Michigan and Wisconsin to see if the different voting technologies\u2014ink-marked paper ballots versus electronic touch screens\u2014treated Clinton and Trump differently. \"Presumably, if there had been a hack to benefit or harm one candidate, the voting machines would have systematically under- or over-counted one candidate's ballots more than the other,\" they wrote for the Washington Post's \"Monkey Cage\" blog in early June 2017. \"That didn't happen.\"\n\n\"Our analysis assumes that the recount, when done by hand, accurately counted the votes,\" they said, agreeing with the Greens' methodology. \"In brief, we find no evidence that the voting technology favored one candidate more than the other... The tests uncovered nothing suspicious. That supports a conclusion that voting machines themselves were not hacked.\"\n\nThat analysis put to rest one concern that emerged immediately after the election. However, it didn't suggest that electronic voting machines weren't vulnerable. Hackers attending the DEF CON25 conference in July 2017 in Las Vegas \"claimed to have found major vulnerabilities or claimed to have breached every voting machine and system present,\" wrote Lulu Freisdat, who videotaped their effort. She reported the breached systems included \"the Sequoia AVC Edge, currently in use in thirteen states and the AccuVote TSX, in use in nineteen states.\" Hackers also took \"complete control of an e-poll book\" and scrambled their files, she said, noting these hacks took from a few minutes to a few hours\u2014another metric about our elections.\n\nBut Mebane and Bernhard's analysis of the Greens' recount had a bombshell\u2014though not enough to unseat Trump. They didn't find evidence of hacking, but they did find something else that could have changed the result. \"We ran into a different problem in Michigan: only a subset of precincts were recounted,\" they said. \"Serious problems have been documented in Detroit that potentially could have changed the outcome in the state.\"\n\nRapoport, too, believed that the aborted recount, reports of spoiled ballots in Detroit, and partisan voter purges facilitated by Crosscheck's lousy data all \"had a cumulative impact that likely exceeded Trump's 11,000-vote margin.\"\n\nAnother postelection study, led by Professor Kenneth R. Mayer of the University of Wisconsin-Madison, found that Wisconsin's strict voter ID law discouraged 16,800 voters from casting ballots in the state's two most populous counties. Those suppressed votes were equal to 11.2 percent of those counties' electorate, leading many observers to conclude that, statewide, the voter ID law could have led to Trump's nearly 23,000 vote victory.\n\nSo Clinton should have won Michigan and Wisconsin, but that still wasn't enough to win the presidency.\n16\n\nTHE ELECTORAL COLLEGE\n\nTHE RECOUNTS EXPOSED MORE UNSAVORY FEATURES of American elections. Yet, as has been the case for years, election officials grudgingly accepted these slights as an inevitable part of the process. The recount's shortcomings joined the list of Election Day impediments compiled by groups like the League of Women Voters, whose post-2016 Wisconsin report cited many incidents of voters being shuffled between polls, confusion over voter ID requirements, and provisional ballots. In a state with millions voting however, hundreds of random victims were seen as an unfortunate but not outcome-changing cost to be tolerated.\n\nNonetheless, the cumulative toll of a presidential campaign that was widely called one of the \"worst ever\" kept prompting people to look for a final way to stop Trump. As the recounts churned along, the Republican Party's members of their state Electoral College prepared to meet in their Capitol buildings to ratify the presidential results. That led to an effort by a group called the Hamilton Electors, named after Alexander Hamilton, who argued that the GOP's almost-final gatekeepers owed it to their country to reject Trump and pick someone else.\n\n\"I am asked to cast a vote on Dec. 19 for someone who shows daily he is not qualified for the office,\" wrote Christopher Suprun, a Texas paramedic and Republican elector, for The New York Times' opinion page. \"The election of the next president is not yet a done deal. Presidential electors have the legal right and constitutional duty to vote their conscience. I believe electors should unify behind a Republican alternative.\"\n\nSuprum's Electoral College brethren did not respond to his call to arms. In fact, the only electors to reject their party's nominee were a few Democrats who abandoned Clinton. That left one final possibility, urging members of Congress to reject the Electoral College's vote for ratification in its joint session in early January. The last time members of both chambers rose to block the ratification was in 2005. That January, Ohio's Rep. Stephanie Tubbs-Jones and California's Sen. Barbara Boxer forced the House and the Senate to return to their chambers for a two-hour debate, where they heard now-familiar complaints about voter suppression. As in 2005, civil rights activists urged the Congressional Black Caucus to take the lead.\n\n\"We are writing to you as a member of the Black Caucus to ask you to file an objection before January 6, 2017, when you and your colleagues reconvene to count and authenticate the ill-gotten votes that some electors cast for Donald Trump,\" Ruby Sales, Amy Billingsley, and Ann Massaro wrote. \"Whether you realize it or not, there is another powder keg that is percolating in Black, Brown, and progressive communities who voted in large numbers for Secretary Clinton only to have their votes stolen and disregarded not only by Republicans but seemingly by Democrats.\"\n\nTheir letter cited the seventy-five thousand Michigan ballots lacking votes for president, the \"massive non-count of ballots in Milwaukee [that] doubtless contain the tally that defeated Mr. Trump,\" and \"missing tally from communities of color\" in Pennsylvania, among anti-voter actions. Their conviction and determination was laudable, but some facts were not on their side such as how some sixty thousand fewer voters turned out in Milwaukee compared to 2012, and how it was virtually impossible to know what was or wasn't missing in Pennsylvania because its voting machinery was almost all paperless, including in Philadelphia and Pittsburgh.\n\nJust days before Congress was to convene, emergency petitions were filed at the Supreme Court. Jerroll M. Sanders, a Missouri activist, argued the election should be voided because federal officeholders were elected by a foreign power\u2014Russia. That went nowhere. In the meantime, as a handful of House Democrats said they would contest the ratification, Ryan Clayton, with the group Americans Take Action, released a thousand-page report from a team of lawyers and law students that claimed fifty Republican members of the Electoral College were illegally seated.\n\nThis should have been a bombshell. The \"Electoral Vote Objection Package\" said sixteen GOP electors lived outside of the congressional districts they purportedly represented, and thirty-four more Republican electors were dual officeholders in states that barred elected officials from holding more than one office. Being an Electoral College member, however brief the term, was an elected office. \"Ironically, Florida Attorney General Pam Bondi, who has issued a number of Advisory Legal Opinions on dual-officeholding, was a presidential elector,\" their executive summary said. \"Joe Negron, who also cast an electoral vote, is currently President of the Florida Senate.\"\n\nThe findings by Clayton's team were a stunning final example of a party that is willing to ignore or revise the rules of voting in an endless pursuit of power. When a few progressive House members rose to challenge the Electoral College certification, they barely got to say more than a sound bite beyond the formalities. (Had their challenge succeeded, the selection of the president would have gone to the House. It didn't get that far.)\n\nWith Vice President Joe Biden presiding, Massachusetts Rep. James McGovern rose. \"I object to the certificate from the state of Alabama on the grounds that the electoral votes were not under all of the known circumstances regularly given and the electors were not lawfully certified, especially given the confirmed and illegal activities engaged in by the Government of Russia that were designed to interfere with our election and widespread violations of the Voting Rights Act that unlawfully suppressed thousands of votes in the state of Alabama.\"\n\n\"Is the objection in writing and signed not only by a Member of the House but also by a Senator?\" Biden replied.\n\n\"The objection is in writing and is signed by a Member of the House of Representatives but not yet by a Member of the United States Senate,\" McGovern countered.\n\n\"In that case, the objection cannot be entertained,\" Biden said.\n\nAnd so it went. Maryland's Jamie Raskin, Washington's Pramila Jayapal, California's Barbara Lee, and Texas' Shiela Jackson Lee all rose to file objections and were gaveled down. California's Maxine Waters rose, saying, \"I wish to ask: is there one United States Senator who will join me in this letter of objection?\"\n\nNone replied.\n\nThe certification continued with one more attempted objection by Arizona's Raoul Grijalva.\n\nBut Biden cut him off, saying, \"There is no debate in the joint session... Is there a signature from a Senator?\" He replied no, prompting Jackson Lee to rise again, saying she objected to North Carolina's certification, \"because of the massive voter suppression and the closing of voting, massive suppression that occurred from African-American\u2014\" Tersely, Biden gaveled her down. \"The tellers will continue the count.\"\n\nIn her extension of remarks that were added to the Congressional Record, Jackson Lee recounted the election's arc. Historians will \"surely record that the 2016 campaign was one for the ages,\" she wrote. Clinton was widely considered \"the most qualified person ever nominated for the office.\" Trump \"is the first person to gain an Electoral College majority with no experience whatsoever in elective or appointive governmental office or public service.\" It was the \"first time in history that a Director of the Federal Bureau of Investigation had ever injected himself and his agency\" into the race. It was \"the first American presidential election that the Intelligence Community has confirmed was the subject of cyberattacks and other subversive activities of entities allied with the Government of Russia.\" She tried to strike a bright note, saying, \"Hillary Clinton received more votes for president than any person in history not named Barack Obama, which means the two greatest vote getters in American political history are an African-American male and white female, which is a testament to how far America has traveled on the path to equality.\"\n\nBut then Jackson Lee returned to brass tacks. \"A switch of less than eighty thousand votes in just three states\u2014Pennsylvania, Michigan, and Wisconsin\u2014would have secured an Electoral College majority for Hillary Clinton,\" she said, but \"that fact is of little consolation and practical consequence to the situation and task now before us.\" She ended, \"There is compelling evidence that activities engaged in by state officials violated the Voting Rights Act of 1965 and disenfranchised thousands of voters and resulted in the unlawful certification of electors... The fate of our democracy is at stake.\"\n\nBrave words, but not one senator would sign her letters of objection to voting in North Carolina, Pennsylvania, and Wisconsin. Democracy Spring activists shouted protests from the gallery and were arrested. The presidential election was officially over. The following week, Bret Sablosky, a paralegal who moved to Washington to block Trump, filed a federal court motion to postpone the inauguration. \"The popular vote plurality is the only Constitutional method of electing the President and Vice President,\" he argued, saying the Electoral College was an unconstitutional under slavery-ending amendments.\n\nHis lawsuit was dismissed.\nAFTERWORD\n\nAS I WRITE ALMOST A YEAR into Trump's presidency, the voting rights pendulum is heading back to the dark side. The forces in American politics that want to stop all eligible citizens from voting are newly empowered. States with supermajority red legislatures are not just trying to suppress Democratic voters, but preempting the very issues that blue voters and municipalities can vote on. These threats to American democracy and representative government pose profound questions that transcend not only the future of the parties, but what it means to be citizens now and in the future.\n\nThe Republican war on voting has continued, taking its cues from President Trump. Even before his swearing in, he told congressional leaders that between three and five million illegal votes were cast for Clinton to deny him a popular vote majority. Never mind that the AP found Gregg Phillips, the source for Trump's false claim, \"was registered to vote in multiple states in the 2016 election.\" That underscores, yet again, that all the GOP noise about policing the vote is as hypocritical as it is fake. The press may expose Trump's tweets as lies, but that has not stopped top White House aides from repeating them.\n\nNational Policy Director Stephen Miller echoed Trump by telling a national TV audience that he only lost New Hampshire because illegal voters were bused in from Massachusetts. Rep. Steve, R-Iowa, told another national audience it was plausible. Bill Gardner, who has been New Hampshire's secretary of state since 1976, said there was no proof. Tom Rath, the state's former attorney general and a Republican said that it was \"shameful to spread these fantasies.\" Even Trump's first campaign manager, Corey Lewandoski, who lives near the border, said he didn't see any buses.\n\nBut that propaganda circus is typical of Trump and obscures more serious GOP actions. As Washington Post political reporter Dan Balz noted about Trump's presidential-size illegal voter claim, \"There is no benign explanation... It is either a deliberate attempt to undermine faith in the democratic process, an exhortation to those who favor new restrictions on access to the ballot box, or the worrisome trait of someone with immense power willing to make wild statements without any credible evidence.\"\n\nAll of the above are true. But what has emerged since Trump took office is especially Balz's second prospect, which does not bode well for American democracy. The larger stakes are what is happening under this administration to further unwind representative government.\n\nTrump has empowered vote suppressors to transform the Department of Justice, whose Civil Rights Division just celebrated its sixtieth anniversary. The Attorney General, Jeff Sessions, began his career prosecuting one of Dr. Martin Luther King's associates for voter fraud. By late February, the Department said it would not pursue its lawsuit against Texas's voter ID law, which it had been fighting for the past several years. As a Huffington Post headline succinctly observed, \"Obama's DOJ Fought Texas Voter ID Law. Trump's New Civil Rights Chief Offered Tips On Writing It.\"\n\nThe voter suppressors were just getting going. At the end of March, two dozen conservative activists and lawyers, including men who oversaw the DOJ's voter fraud investigations during George W. Bush's administration, wrote an unsolicited letter to Sessions. They presented a classic white supremacist agenda: back off laws protective of minority voters and start pressuring states to aggressively purge their voter lists.\n\n\"We have witnessed longstanding conventions held from the mid-20th century prove outmoded in recent years and discovered new fronts in need of protection where civil rights are concerned\u2014with particular respect to voting,\" said the letter signed by Kansas Secretary of State Kris Kobach, The Heritage Foundation's Hans von Spakovsky, and the Public Interest Legal Foundation's J. Christian Adams, among others.\n\nIn early May, Trump appointed these three men to a Presidential Advisory Commission on Election Integrity, where Kobach is vice-chair. (Vice President Mike Pence is the titular chair.) While their initial meetings were held in secret, the March letter to Sessions laid out their agenda. They seek a \"return to race-neutral Voting Rights Act enforcement,\" \"an end to politically driven pursuits against state photo voter identification requirements, citizenship verification in voter registration, and common-sense adjustments to early voting periods,\" and \"return to enforcing Section 8\" of the NVRA, which governs voter purges. They ended, \"The American electorate is crying out to see protections against political enforcement of the law.\"\n\nThese missives and presidential missions are the opening moves in the next round of an ugly fight over who gets to vote in America and how arduous that process will be. This especially is the case in states that have been most heavily gerrymandered and are key to control of the House (and include presidential swing states). These are states like Michigan, Wisconsin, Pennsylvania, North Carolina, Ohio, Virginia, Florida, Georgia, and Texas.\n\nOne of Kobach's first tactics has been to demand all fifty states turn over their voter files, including lists of inactive voters and confidential data like Social Security numbers. While progressives, Democrats, civil rights groups, and others have loudly protested and sought to defy Kobach's commission, some election experts who have watched him for years believe that this data grab is a giant bait and switch. (Kobach has not said why the commission is seeking the statewide voter data or what it will do with it.)\n\nKobach knows most states can't meet his deadline even if they wanted to, Michael McDonald, a University of Florida associate professor specializing in American elections, tweeted after Kobach's letter became public. That's because, among other things, many states are not authorized by legislatures to share everything he wants. Knowing that is likely to be the case positions him and Trump to say that states are hiding evidence of voter fraud. That's exactly what Trump tweeted on July 1, \"Numerous states are refusing to give information to the very distinguished VOTER FRAUD PANEL. What are they trying to hide?\"\n\nBy late July, seven lawsuits had been filed against the commission. Dozens of states were saying they would only partly comply and some top state election officials openly said they didn't trust Kobach. At the panel's first public meeting, Trump stopped by and said, \"one has to wonder what they are worried about,\" referring to recalcitrant states. And then the panel's discussion, chaired by Kobach and dominated by von Spakovsky, suggested using known problematic government databases to vet (and erroneously reject) voter applications.\n\nThose records included the US Department of Homeland Security's incomplete citizenship database\u2014which Florida Republicans previously sought to use. Von Spakovsky also wanted jury duty records, to see if people who said they were noncitizens to avoid jury duty were registered voters. He also wanted state tax records to see if someone had registered to vote at a business address instead of a home address. The panel also wanted a list of elections from across the country where the winner was decided by single-digit margins.\n\nThese demands, especially from von Spakovsky, did not surprise election lawyers who favor expanding the franchise. Von Spakovsky's role during the Bush administration of undermining the Justice Department's pro-voting rights mission was widely documented, starting with his false and outsized claims about voter fraud. As Rick Hasen, an election law expert at the University of California at Irvine blogged in late June 2017, \"The appointment is a big middle finger from the president to those who are serious about fixing problems with our elections.\"\n\nAll of this points to fabricating a pretext for newly restrictive federal anti-voter legislation\u2014or red states with super majorities to further complicate the process of voting. Recall there is no nationwide federal database for voter fraud, or for citizenship\u2014a useful absence that allows Republicans to say the process must be better policed; and then using imprecise methodologies that disproportionately target blue voting blocs.\n\nMore potentially insidious, on the same day that Kobach sent his letter seeking state voter files and related data, the DOJ's Voting Section also sent the same states a demand letter on a parallel fishing expedition. It asked all the states to provide everything they use\u2014\"all statutes, regulations, written guidance, internal policies, or database use manuals that set out the procedures\"\u2014relating to the \"list maintenance provisions\" of the National Voter Registration Act (NVRA) and Help America Vote Act (HAVA). Their letter said nothing about expanding registration, only about removing previously registered voters who may have moved, died, been convicted of a felony, or were on inactive voter lists. It doesn't take much to see where this is going. Trump's commission and DOJ's Voting Section are following the template laid out in the late March 2017 letter from longtime GOP vote suppressors, including those who led the Bush administration's voter fraud inquests.\n\nWhat may happen is that Kobach and company will hold off issuing any report until 2018, at the earliest. That's because the Supreme Court's fall 2017 term contains two big voting cases, where its right-wing majority might issue rulings enshrining key pieces of the GOP's anti-voter playbook. (Early in 2018 is also when state legislatures reconvene, and where red states could fast-track any recommendations before the 2018 elections.)\n\nThe first case coming before the Court, from Ohio, concerns how soon infrequent voters can be purged under the NVRA. The law has an ambiguity that Republicans identified and are trying to exploit. It lays out a four-year timetable for voter purges, but also says that no one can be removed for being an infrequent voter, an apparent contradiction. Georgia's GOP Secretary of State Brain Kemp, now running for governor\u2014like Kansas' Kobach\u2014filed a brief on behalf of a dozen red states urging the court take the case. In August 2017, Trump's Justice Department filed a brief supporting Ohio's GOP-led purge, reversing the DOJ's stance taken during the Obama administration.\n\nThe second case is from Wisconsin on whether extreme partisan redistricting is legal. More than a decade ago, Justice Anthony Kennedy wrote an opinion suggesting the Court would like to see a standard for assessing when partisan gerrymanders are so extreme they become unconstitutionally antidemocratic. Law professors have created a metric based on \"wasted votes\" using Wisconsin as an example. But in the Supreme Court's late May ruling about illegal racial gerrymandering in North Carolina, Kennedy signed a dissent suggesting he may have changed his mind about extreme redistricting. The dissenters said such hyper-partisanship was distasteful but part of human nature and the political process. The ruling in these cases will determine the electoral playing field for years to come.\n\nMeanwhile, despite populist expectations among Democrats that they can retake the House in 2018, there are other signs the national landscape is tightening in antidemocratic ways. Looking at 2016's results, REDMAP states continued to see red supermajorities elected to their House delegations and state legislatures. Extreme redistricting is to blame for the noncompetitive races that produced the political landscapes in these states and Congress, Daley reiterated in a post-2016 epilogue to his book, Ratf**ked. After the 2016 election, the GOP held sixty-nine of the niety-nine state legislative chambers. There is little reason to think Republican supermajority states will see Democratic takeovers in 2018 or 2020\u2014the year before their legislatures draw political maps lasting for the 2020s, he said. That likelihood makes 2018's governor races in states where Republican incumbents face term limits\u2014such as Florida, Michigan, Ohio, and Georgia\u2014the most important races if the GOP's lock is to be broken. Only Democratic governors can veto the bad maps that likely will be drawn in 2021.\n\nMeanwhile, when it comes to retaking the US House, 2016's results showed the 2011 gerrymander continues to hold. Even though Republican candidates received 1.2 percent more votes nationwide, they ended up with 10.8 percent more seats. What most Democrats don't grasp looking at 2018 is how steep the climb back into power is. As Daley said in July 2017, Democrats have only flipped one seat from red to blue in five key states since 2011's maps\u2014in Michigan, Ohio, Pennsylvania, Wisconsin, and Virginia. That seat was in Virginia, after a court ruled it had been an illegal race-based gerrymander. When Democrats say they are looking at taking the twenty-three GOP seats in House districts where Clinton defeated Trump\u2014they need twenty-four to regain a majority\u2014Daley countered that's not a reassuring strategy. They need to compete in fifty competitive races, he said, and he doesn't see that taking shape.\n\nThere are other ominous trends. One of the biggest trends following red super majorities is not merely restricting the process of voting, but going further to restrict what citizens locally vote on. This trend has the clunky name of preemption, but is very serious. In the most gerrymandered states, there are schisms between red-run legislatures and blue-run cities, which have passed all kinds of progressive laws\u2014from higher minimum wages, to banning natural gas hydraulic fracturing, to being sanctuaries for immigrants, to taxing soda and tobacco, to banning guns in public, and more. Red legislatures have responded with new statewide laws to reverse and ban local initiatives. The GOP, which once venerated local control, are now more than willing to preempt what entire populations can vote on. Segregating voters via extreme redistricting and blocking access to the polls is antidemocratic enough. But censoring what citizens can vote on is astounding.\n\nWhat Democrats, progressives, and independents must realize, whether they are focusing on the House in 2018 or state legislatures, are the structural disadvantages they face. The GOP, in the most heavily gerrymandered districts, has a 6\u20138 percentage point starting line advantage. These are states where red legislatures have kept passing stricter voter ID laws\u2014which shave off another 2\u20133 points in the November election, and much more in primaries among blacks, Asians, and Latinos. These states may follow Kobach's lead and impose new documented proof of citizenship requirements for registrants, which 7 percent of the public lacks. These tactics are followed by other anti-voter moves, from limiting early voting options to ending Election Day registration.\n\nBut by summer 2017, it was not clear the GOP's structural advantages were understood. Few apart from the Democratic Governors Association talk about the importance of upcoming governors' races, compared to retaking the House. Meanwhile, other Republican gambits in order to preserve their power are surfacing and have to be taken seriously.\n\nWhen ALEC, the right-wing American Legislative Exchange Council, met in July 2017, it discussed two structural reforms to create a permanent Republican majority. The first was restoring the appointment of US Senators by state legislatures by repealing the Seventeenth Amendment. (After November 2016, the GOP held both chambers in thirty-two states, which, if such a repeal were law, would mean they would have sixty-four senators.) As far-fetched as that sounds, ALEC has been pushing red states to pass legislation calling for a federal Article Five constitutional convention, where such a proposal could be raised. So far, twenty-seven states out of the thirty-four states needed have passed resolutions for that convention.\n\nA more likely near-term threat is gerrymandering the Electoral College. Republicans could do that by replacing the statewide winner-take-all system with awarding delegates by House district. If that were law in Virginia, where a bill was introduced but did not pass in 2017, the state's eleven Electoral College votes would not have gone to Clinton; Trump carried six of its House districts. Legislation to do that was proposed in a few states in early 2017, but did not go anywhere.\n\nToo few people pay attention to these potential power grabs. But they would continue what Republicans have done this decade. They have transformed voting to such a degree that it has subverted our democracy. There's no indication they intend to reverse course and put representative government before their partisan power. If anything, the opposite is true. They are betraying the citizenry by hijacking voting at key nodes across the process.\n\nEvery month into Trump's presidency underscores the antidemocratic result: a political party representing a minority keeps pushing an authoritarian regime of oligarchy, plutocracy, and theocracy. The vast majority of citizens don't support Trump or the GOP's agenda, yet they are witnessing their government being taken from them. The stakes for the country are bigger than which party, or party factions, win or lose. If voting is rendered meaningless, representative government will disintegrate and the American experiment will end.\n\nThe only way to stop this trajectory is to understand what's taking place and dismantle the structural advantages the Republicans built into the voting process. This book has tried to describe the antidemocratic features of the 2016 election, which is the latest manifestation of that landscape. It's possible to quantify these features with metrics that didn't exist a decade ago. Understanding what is impeding an increasingly diverse citizenry from voting to elect a representative government is the first step in countering the assault on our democracy.\n\nOn the Democratic side of the aisle, the party that is the champion of voting rights must step up to the moral challenge to representative government. The party loses its moral authority when it continues to have features like its superdelegate system, a national apparatus that takes sides in presidential nominating contests, registration deadlines and rules that prevent independents from voting in its primaries, as well as presidential caucuses that are manned by inept volunteers. Party loyalists need to stop defending the indefensible and side with all voters.\n\nBut the hurdles posed by Republicans are far more serious, insidious, and embedded in law. We have seen enough of the GOP playbook to do more than object. We know how restrictive voter ID laws impact voter turnout, especially along racial lines. We know about the extreme gerrymandering. We know about possibly replacing a legal oath by signature on voter registration forms with paper proof of citizenship. We also know vast numbers of Republicans believe the myth of voter fraud and support policing the vote. We also know half of Republicans would accept a political coup\u2014postponing the 2020 election if Trump and Congress called for that, according to an August poll by the Washington Post. (Even if the US Constitution's Twentieth Amendment sets terms of office with expiration dates.)\n\nBut we can quantify how much of the electorate the GOP seeks to preempt. This matters because when people start arguing about elections, the first thing that is lost is magnitude and scale: what is the possible or real impact on voters? Arguing about elections is filled with false equivalencies, hyped impacts, distortions, and prejudice. The Republicans cite one-in-a-million incidents as reasons to pass state laws affecting millions of voters. Citizens who want fair and accountable elections, and representative government, need to push back wisely, and know what matters most.\n\nThey need to realize the stakes and yet not give up hope. Nobody knows where the Trump-Russia probes will go. Nor if disgusted voters will turn out in such large numbers in 2018 that they will swamp the GOP's ten-point starting line advantage in a wave sweeping them from office. Meanwhile, voters should look at all possible ways to take back this process. Extreme redistricting is increasingly in the news, with maverick Republicans like former California governor Arnold Schwarzenegger saying backroom mapmakers need to be replaced by citizens' commissions. The political class needs to know the public has their number.\n\nOne such glimmer of hope emerged from the Greens' recount. Unlike most of this book, it concerned voting's finish line: verifying the count. John Brakey, the election transparency activist, pointed to a development that can lead to citizen-led verifiable vote counts, even with electronic voting machines. It's a humble prospect, but in the right direction.\n\n\"In an 'ah-ha' moment, Election Integrity investigators realized that 40 percent of the voting machines presently in use have the built-in ability to transparently verify reported election results RIGHT NOW,\" he wrote for a presentation at the Left Forum in New York City in June 2017. \"The machines that are used to count hand-marked paper ballots are rapidly being replaced with technology that uses digital imagery to count scanned ballot images. At the current rate, by the next presidential election, approximately 85 percent of the country will generate scanned ballot images.\"\n\nBrakey envisions crowd-sourcing recounts and audits using images of marked ballots. It's not entirely magical thinking, when you consider NASA puts online telescope images of distant galaxies so thousands of volunteers parse them to find new stars. Of course, it would be better if American elections were held using ink-marked paper ballots that were openly counted and rigorously audited. Brakey believes in the better angels of democracy. But he also knows who and what he's fighting. Many local election officials do not want citizens to pore over ballots to see if their votes have been accurately counted.\n\nSuch is the state of our elections in 2018. The closer you look, the more antidemocratic features emerge. Right now, American democracy is in deep trouble. Anything that restores rights of voters and engages citizens cannot be overlooked. Nothing less than the nation's future, the majority voices of citizens, and our future, is at stake.\nACKNOWLEDGMENTS\n\nTHE OBSERVATIONS AND CONCLUSIONS PRESENTED IN this book span more than a decade of being a journalist who was given the opportunity to interview, observe, collaborate with, and learn from scores of people with different roles in the voting sphere. These include journalists, lawyers, activists, scholars, foundation officials, candidates, campaign workers, and many authors I have not met but whose work I've read for years. As Bill Moyers said, journalists are beachcombers on the shores of other people's wisdom. That is the case here.\n\nThis book began with an unexpected call from John Bonifaz, a voting-rights attorney I met in the 1990s, whose work has crossed my path over the decades. My coverage at AlterNet of the 2016 recounts caught Skyhorse's attention and launched this journey. But this book also reaches back to Vermont in the 1980s, when I was a reporter and met Bernie Sanders and became press secretary on the first campaign electing him to the House. Bill Stetson led me to many people in Bernie's campaign and in Democratic Party circles, including some who freely gave insights but asked not to be named, as they still work in politics.\n\nThe recounts put me back in touch with a community of activists, attorneys, and journalists I met after Ohio's 2004 election. They include Mark Crispin Miller, Bob Fitrakis, Jonathan Simon, Harvey Wasserman, John Brakey, and others. Special thanks go to Don Hazen who has supported my work for many years at AlterNet starting with covering voting rights in 2008, and Bill Moyers and the Schumann Center for Media and Democracy.\n\nI also must thank David Becker and John Lindback, who let me assist Pew's effort in 2010 to modernize voter registration and in so doing introduced me to many dedicated civil servants, academics, and technologists who want to improve elections and expand the vote. They gave me a greater appreciation of the complexity and layers involved in the process and still do.\n\nThen there are the election law groups and affiliated scholars whose work has helped to document and quantify how voters are marginalized on partisan lines. These include the Brennan Center for Justice at New York University School of Law, Advancement Project, NAACP Legal Defense Fund, Demos and others cited in the endnotes. Their reports, legal briefs, and resulting court rulings from their work all contributed\u2014as did legal documents, reports, and other records from those who wanted to tilt the field.\n\nThere are also many analysts and journalists who I have not met but whose work informs these pages. These include the Cook Political Report's David Wasserman, FiveThirtyEight's Harry Enten, and many excellent reporters at the Washington Post, The New York Times, Bloomberg, Politico, USA Today, The Atlanta Journal-Constitution, Reuters, The Chicago Tribune, The Intercept, Associated Press, and local TV.\n\nNumerous books also contributed. Especially noteworthy are those by Michael Waldman and David Daley. David was very generous with his time and encouraged this effort to peel back the layers of how elections are structured. Other big contributions include drawing on voluminous reporting from The Nation's Ari Berman, efforts pressing public officials for answers by Greg Palast, and videos by filmmaker Lulu Friesdat. Pro-democracy activists also played a part, such as early postelection analyses by Election Justice USA, recount organizers like Warren Linney, and Verified Voting's Barbara Simon. Thanks to Keith Conkin for proofreading drafts and suggesting what needed clarification.\n\nI must also thank my family for their support and encouragement. My parents, Ronnie and Sheldon Rosenfeld, continue to encourage me and take pride in my work; my siblings for their feedback; Beth and Danya Sauerhaft who marveled that I made and met my deadlines.\n\nBut most of all, I want to thank my partner, Lynda Beth Unkeless, who encouraged this project from the start, was endlessly patient as it took up our spare time, and still found surprises and contributions in libraries and articles that sharpened its content and focus.\n\nFinally, I thank my editor, Jonathan McCullough, for his incisive eye and comments taking the manuscript to a higher level.\nENDNOTES\n\nINTRODUCTION\n\n. \"It's time to bust the myth: Most Trump voters were not working class,\" by Nicholas Carnes and Noam Lupu, Washington Post, June 5, 2017. This report said two-thirds of Trump's voters earned more than the annual national median household income. https:\/\/www.washingtonpost.com\/news\/monkey-cage\/wp\/2017\/06\/05\/its-time-to-bust-the-myth-most-trump-voters-were-not-working-class\/?utm_term=.fc6044e39471&tid=a_inl.\n\n. \"New Video: Watch Wisconsin Election Officials Reject Hand Counts After Electronic Scanners Make Big Mistake: 'I am just stunned at how inaccurate the whole thing is,'\" by Lulu Friesdat, AlterNet.org, December 16, 2016, <http:\/\/www.alternet.org\/new-video-watch-wisconsin-election-officials-reject-hand-counts-after-electronic-scanners-make-big>.\n\n. \"Trump Is Dividing The Country, US Voters Say 2\u20131, Quinnipiac University National Poll Finds; Most Trust Media More Than President,\" Quinnipiac University, August 23, 2017, <https:\/\/poll.qu.edu\/national\/release-detail?ReleaseID=2482>.\n\nAlso see: \"Fox News Poll: Voters' mood sours, 56 percent say Trump tearing country apart,\" by Dana Blanton, FoxNews.com, August 30, 2017, \"The number of voters happy with how things are going in the country is down 10 percentage points since April and stands at just 35 percent.\"\n\n<http:\/\/www.foxnews.com\/politics\/2017\/08\/30\/fox-news-poll-voters-mood-sours-56-percent-say-trump-tearing-country-apart.html>.\n\nSECTION I: THE DEMOCRATS\n\nCHAPTER 1: BERNIE SANDERS\n\n. \"Democratic National Convention, 2016,\" BallotPedia.org. Sanders had 1,832 pledged delegates from the caucuses and primaries, 45 percent, compared to Clinton's 2,219 pledged delegates. There were 713 superdelegates in addition. <https:\/\/ballotpedia.org\/Democratic_National_Convention,_2016>.\n\n. \"Bernie 2016,\" Presidential Principle Campaign Committee, US Federal Election Commission. This is the FEC summary page, showing Sanders's campaign raised $237.6 million. <https:\/\/www.fec.gov\/data\/committee\/C00577130\/?cycle=2016>.\n\n. \"Revised Debate Memo,\" April 27, 2015, wikileaks.org. This email contains the Clinton campaign memo describing how the DNC coordinated with her campaign to schedule the televised debates in late 2015 and early 2016. After getting the schedule it sought, her campaign pushed the DNC to add another debate before the New Hampshire primary.\n\n<https:\/\/www.wikileaks.org\/podesta-emails\/emailid\/5688#searchresult>.\n\nAlso: \"Leaked Email Proves DNC Rigged Primary Debates,\" by Dan Wright, shadowproof.com, October 12, 2016, <https:\/\/shadowproof.com\/2016\/10\/12\/leaked-email-shows-dnc-rigged-primary-debates\/>.\n\n. \"Sanders campaign sues DNC after database breach,\" by Catherine Treyz, Dan Merica, Jeremy Diamond, Jeff Zeleny, CNN.com, December 21, 2015, <http:\/\/www.cnn.com\/2015\/12\/18\/politics\/bernie-sanders-campaign-dnc-suspension\/index.html>.\n\nAlso: \"Debbie Wasserman Schultz: Give the voter file back to Bernie Sanders' campaign,\" Petition by Carl Gibson, MoveOn.org, <https:\/\/pac.petitions.moveon.org\/sign\/debbie-wasserman-schultz-2>.\n\nCHAPTER 2: WHO WON IOWA?\n\n. \"The Iowa caucuses: An accident of history,\" by David Jackson, USA Today, January 29, 2016, <http:\/\/www.desmoinesregister.com\/story\/news\/politics\/elections\/2016\/01\/29\/iowa-caucuses-history-jimmy-carter-julian-zelizer\/79426692\/>.\n\n. \"Hunt Commission,\" wikipedia.org. The DNC convened the Hunt Commission in 1981 after a tumultuous 1980 convention, which, in turn, created its super-delegate system. <https:\/\/en.wikipedia.org\/wiki\/Hunt_Commission>.\n\n. Eugene V. Debs, wikipedia.org. Excerpt: \"Debs ran as a Socialist candidate for President of the United States five times, including 1900 (earning 0.63% of the popular vote), 1904 (2.98%), 1908 (2.83%), 1912 (5.99%), and 1920 (3.41%), the last time from a prison cell.\" <https:\/\/en.wikipedia.org\/wiki\/Eugene_V._Debs>.\n\n. Our Revolution: A Future to Believe In, by Bernie Sanders, Thomas Dunne Books, 2016. p. 170\u2013171. <https:\/\/us.macmillan.com\/ourrevolution\/berniesanders\/9781250132925>.\n\n. \"Editorial: Something smells in the Democratic Party,\" The Des Moines Register editorial, February 3, 2016, questioning whether Clinton had won and noting the ineptly run contest. <http:\/\/www.desmoinesregister.com\/story\/opinion\/editorials\/caucus\/2016\/02\/03\/editorial-something-smells-democratic-party\/79777580\/>.\n\n. Shattered: Inside Hillary Clinton's Doomed Campaign, by Jonathan Allen and Amie Parnes, Crown, 2016, page 111, <http:\/\/www.penguinrandomhouse.com\/search\/shattered?q=shattered>.\n\n. \"Statement on Hillary Clinton's Victory in the Iowa Caucus,\" by Hillary for America's Iowa State Director Matt Paul, February 2, 2016, <https:\/\/www.hillaryclinton.com\/briefing\/statements\/2016\/02\/02\/iowa-caucus-victory\/>.\n\n. Iowa's nightmare revisited: Was correct winner called? Report by Jennifer Jacobs, Feb. 2, 2016, questioning the accuracy of Iowa's Democratic Caucus count and winner, <http:\/\/www.desmoinesregister.com\/story\/news\/elections\/presidential\/caucus\/2016\/02\/02\/iowas-nightmare-revisited-correct-winner-called-caucus-night\/79702010\/>.\n\n. Same as 12: \"Editorial: Something smells in the Democratic Party,\" The Des Moines Register editorial, February 3, 2016.\n\n. Same as 11: Our Revolution: A Future to Believe In, by Bernie Sanders. p. 167.\n\n. \"Texas Democratic primary and caucuses, 2008,\" wikipedia.org. This entry recounts how Clinton's 2008 presidential campaign filed challenges of the Texas caucus results, seeking to nullify the votes of tens of thousands of attendees. <https:\/\/en.wikipedia.org\/wiki\/Texas_Democratic_primary_and_caucuses,_2008>.\n\nCHAPTER 3: LAS VEGAS RULES\n\n. \"Harry Reid delivers for Hillary Clinton: Nevada's 'neutral' power players may have saved a campaign and changed history.\" by Jon Ralston, USA Today, February 20, 2016, <http:\/\/www.usatoday.com\/story\/opinion\/2016\/02\/20\/hillary-clinton-wins-nevada-caucus-harry-reid-culinary-union-jon-ralston\/80688750\/>.\n\n. \"Results from the 2016 Nevada GOP caucuses,\" by Lily Mihalik, Anthony Pesce and Ben Walsh, Los Angeles Times, February 23, 2016. This map shows Sanders winning Nevada's northern counties, including Reno and Carson City. <http:\/\/graphics.latimes.com\/election-2016-nevada-results\/>.\n\n. \"Update: Suspicious votes, long times at Dem caucus,\" by Mark Robison, Reno Gazette-Journal, Feb. 20, 2016. This report categorizes more than a dozen problems associated with inexperience and mismanagement in running the Democratic caucuses in Reno. <http:\/\/www.rgj.com\/story\/news\/politics\/2016\/02\/20\/long-lines-too-few-ballots-dem-caucus-locations\/80666630\/>.\n\n. \"Don't Call It a Bern-Back: According to social media rumors, Bernie Sanders claimed a win in Nevada in a surprise post-caucus upset, but the reality was far more complex and less definitive,\" by Kim LaCapria, snopes.com, April 9, 2016. This report describes the next stage in Nevada's delegate selection process, county conventions, and statements by the campaigns about who was winning. <http:\/\/www.snopes.com\/bernie-sanders-won-nevada\/>.\n\nCHAPTER 4: NEW QUESTIONS, NEW BARRIERS\n\n. \"Democracy Lost: A Report on the Fatally Flawed 2016 Democratic Primaries,\" by Election Justice USA, August 2016, ElectionJusticeUSA.org. See pages 36\u201348, for media exit poll discrepancies with official results. <https:\/\/drive.google.com\/file\/d\/0B5O9I4XJdSISNzJyaWIxaWpZWnM\/view>.\n\n. \"There Were 5-Hour Lines to Vote in Arizona Because the Supreme Court Gutted the Voting Rights Act,\" by Ari Berman, The Nation, March 23, 2016, <https:\/\/www.thenation.com\/article\/there-were-five-hour-lines-to-vote-in-arizona-because-the-supreme-court-gutted-the-voting-rights-act\/>.\n\nAlso see: Democracy Diminished: State and Local Threats to Voting Post-Shelby County, Alabama v. Holder, by NAACP Legal Defense Fund, 2016, pages 9\u201312. These reports describe how hours-long delays in Maricopa County, due to switching from precinct voting to voting centers, had not been approved by the Justice Department before the US Supreme Court invalidated DOJ oversight in its 2013 ruling on the Voting Rights Act. <http:\/\/www.naacpldf.org\/files\/publications\/Democracy%20Diminished-State%20and%20Local%20Voting%20Changes%20Post-Shelby%20v.%20Holder_4.pdf>.\n\n. Same as 23: See pages 13\u201317. \"Democracy Lost: A Report on the Fatally Flawed 2016 Democratic Primaries,\" by Election Justice USA, August 2016.\n\n. Same as 11: Our Revolution: A Future to Believe In, by Bernie Sanders. p. 174.\n\n. \"Closed primary,\" ballotpedia.org. This lists all states with closed presidential primaries. <https:\/\/ballotpedia.org\/Closed_primary>.\n\n. \"It's Far Harder To Change Parties In New York Than In Any Other State,\" by Leah Lebresco, FiveThirtyEight.com, April 19, 2016, <https:\/\/fivethirtyeight.com\/features\/its-far-harder-to-change-parties-in-new-york-than-in-any-other-state\/>.\n\n. Same as 11: Our Revolution: A Future to Believe In, by Bernie Sanders. p. 175.\n\n. Justice Department Suing NYC Board Of Elections Over Last Year's Pre-Primary Mass Voter Purge, by Nathan Tempey, gothamist.com, January 13, 2017. This report recounts the sloppy voter purge by Kings County (Brooklyn) before the 2016 Democratic presidential primary that led to 122,000 infrequent voters being removed. <http:\/\/gothamist.com\/2017\/01\/13\/doj_sues_boe_brooklyn_voter_rolls.php>.\n\n. Same as 30: Justice Department Suing NYC Board Of Elections Over Last Year's Pre-Primary Mass Voter Purge, by Nathan Tempey, gothamist.com.\n\n. \"Statement From A. G. Schneiderman On Voting Issues During New York's Primary Election,\" The New York State Office of the Attorney General, April 20, 2016. \"I am deeply troubled by the volume and consistency of voting irregularities, both in public reports and direct complaints to my office's voter hotline, which received more than one thousand complaints in the course of the day yesterday.\" <https:\/\/ag.ny.gov\/press-release\/statement-ag-schneiderman-voting-issues-during-new-york%E2%80%99s-primary-election>.\n\n. Same as 11: Our Revolution: A Future to Believe In, by Bernie Sanders. p. 178.\n\nCHAPTER 5: THE NOT-SO-GOLDEN STATE\n\n. Same as 11: Our Revolution: A Future to Believe In, by Bernie Sanders. p. 179.\n\n. \"Would-be independents joining the American Independent Party could blame California's voter registration card,\" by Jon Myers, Los Angeles Times, April 19, 2016. Excerpt: \"A Times investigation found widespread confusion among California voters who choose the American Independent Party, an ultra-conservative organization that's been largely invisible from most campaigns. A poll of AIP voters found 73% mistakenly thought they were 'independent' of all parties. Those voters should have chosen the 'no party preference' option.\" <http:\/\/www.latimes.com\/politics\/la-pol-ca-american-independent-party-california-voter-registration-card-20160419-story.html>.\n\n. \"Primary Elections in California,\" by Alex Padilla, California Secretary of State.\n\nThis page from the Secretary of State's website explains how the Democratic presidential primary allows independent voters to cast ballots in the party's contest. \"The Democratic and American Independent parties notified the Secretary of State that they will allow voters who did not state a political party preference to vote the presidential ballot of their parties in the upcoming June 7, 2016, Presidential Primary Election. Their notifications can be found in CC\/ROV Memorandum #16036 (PDF).\" <http:\/\/www.sos.ca.gov\/elections\/primary-elections-california\/>.\n\n. \"Snagged Votes In Los Angeles,\" by Judy Frankel, Huffington Post, July 5, 2016. This report notes certain counties opted for special crossover ballots for independent, or nonparty preference (NNP), voters, and the confusion in Los Angeles over handling these ballots. <http:\/\/www.huffingtonpost.com\/judy-frankel\/snagged-votes-in-los-angeles_b_10794718.html>.\n\n. \"AP count: Clinton has delegates to win Democratic nomination,\" by Hope Yen, Stephen Ohlemacher, Lisa Lerer and Catherine Lucey, The Associated Press, June 7, 2016. This report, dateline Los Angeles, announced the Democratic nomination was locked up by Clinton one day before California held its presidential primary. <https:\/\/apnews.com\/4c9c850385c84b12ad5b85fda49743f9\/after-weekend-wins-clinton-cusp-democratic-nomination>.\n\n. Same as 11: Our Revolution: A Future to Believe In, by Bernie Sanders. p. 179.\n\n. \"California Democratic primary, 2016,\" wikipedia.org. This report of the final vote count shows Hillary Clinton's election night margin of 14 percent shrinking to 7 percent after all California primary ballots were counted. <https:\/\/en.wikipedia.org\/wiki\/California_Democratic_primary,_2016>.\n\nCHAPTER 6: THE DEMOCRATIC NATIONAL COMMITTEE\n\n. \"DNC Chair Says Superdelegates Exist to Protect Party Leaders,\" by Independent Voter, YouTube.com, February 12, 2016. This video clip shows a Feb. 12, 2016, CNN interview where DNC Chairwoman Debbie Wasserman Schultz said that superdelegates ensure inclusivity but also ward against grassroots activists. <https:\/\/www.youtube.com\/watch?v=w5llLIKM9Yc>.\n\n. \"The 4 Most Damaging Emails From the DNC WikiLeaks Dump,\" by Alana Abrasion and Shushannah Walsh, ABC News, July 25, 2016, <http:\/\/abcnews.go.com\/Politics\/damaging-emails-dnc-wikileaks-dump\/story?id=40852448>.\n\n. \"Carol Wilding et al vs. DNC Services Corporation and Deborah Wasserman Schultz,\" United States District Court, Southern District of Florida, Case 0:16-cv-61511-XXXX, entered June 28, 2016. This class-action lawsuit, filed by Sanders supporters against the DNC before the Democratic Convention, accuses the party of violating its charter to be a neutral party in nominating contests and cites Wikileaks' email hacks as evidence. <http:\/\/jampac.us\/wp-content\/uploads\/2016\/06\/1-CLASS-ACTION-COMPLAINT-6-28-16.pdf>.\n\n. \"Will Keith Ellison Move the Democrats Left? By running for the D.N.C. chair, the Minnesota congressman hopes to lead a populist opposition against Trump,\" by Vinson Cunningham, The New Yorker, February 27, 2017. This profile of a bid by Rep. Keith Ellison, D-MN, to be DNC chair gives two more examples of the DNC colluding with Clinton's campaign in the primary season. First is establishing a joint fund-raising operation before the national party convention. Second is ex-chairwoman Donna Brazile, then a CNN commentator, alerting the campaign to questions for an upcoming televised debate. <http:\/\/www.newyorker.com\/magazine\/2017\/02\/27\/will-keith-ellison-move-the-democrats-left>.\n\nAlso see: \"Donna Brazile,\" wikipedia.org. Brazile's biograhy, including stints as Al Gore's presidential campaign manager in 2000 and interim DNC chair in 2011. <https:\/\/en.wikipedia.org\/wiki\/Donna_Brazile>.\n\n. \"Top Democrat backpedals after saying primaries were 'rigged,'\" by Yaron Steinbuch, The New York Post, Feb. 10. 2017. This report notes Tom Perez, then a candidate for DNC chair, told Kansas lawmakers that the 2016 nominating contest was rigged against Sanders, and then backtracked. <http:\/\/nypost.com\/2017\/02\/10\/ex-labor-secretary-backpedals-on-rigged-democratic-primaries\/>.\n\n. Same as 11: Our Revolution: A Future to Believe In, by Bernie Sanders. p. 195.\n\n. Same as 11: Our Revolution: A Future to Believe In, by Bernie Sanders. p. 185, 197.\n\nSECTION II: THE REPUBLICANS\n\nCHAPTER 7. GAMING THE RULES OF Politics\n\n. \"Why Did the Founding Fathers Choose a Republic?\" by Thomas DeMichele, May 12, 2016, FactMyth.com. This article concisely describes why the United States is a republic with democratic features, but not a democracy where popular votes determine governing institutions. <http:\/\/factmyth.com\/why-did-the-founding-fathers-choose-a-republic\/>.\n\nAlso see: \"Child of the Enlightenment,\" by Garry Wills, The New York Times Magazine supplement, July 2, 2017. This describes the founder's competing considerations, including the distrust of popular factions and the belief the high reaches of government should be run by people of \"sufficient virtue and intelligence.\" <https:\/\/usedbooksinclass.com\/tag\/garry-wills-child-of-enlightenment\/>.\n\n. The American Slave Coast: A History of the Slave-Breeding Industry, by Ned and Constance Sublette, Lawrence Hill Books, 2016. This authoritative economic history of the Atlantic slave trade describes its impact on the colonial and pre-Civil War eras. http:\/\/www.chicagoreviewpress.com\/american-slave-coast\u2014the-products-9781613748206.php.\n\n. The Great Suppression: Voting Rights, Corporate Cash, and the Conservative Assault on Democracy, by Zachary Roth, Crown, 2016. See Introduction and Chapter 4. Also same as 47: Garry Wills' New York Times Magazine special section.\n\n. \"ALEC Is Talking About Changing the Way Senators Are Elected and Taking Away Your Vote: A proposed resolution advocates for overturning the 17th Amendment so Republican-controlled state legislatures could pick senators,\" by John Nichols, The Nation, July 18, 2017, <https:\/\/www.thenation.com\/article\/alec-wants-to-to-change-the-way-senators-are-elected-and-take-away-your-vote\/>.\n\n. The Fight to Vote, by Michael Waldman, Simon and Schuster, 2016,\n\n<http:\/\/www.simonandschuster.com\/books\/The-Fight-to-Vote\/Michael-Waldman\/9781501116483>.\n\n. \"United States House of Representatives elections, 2012,\" wikipedia.org. Excerpt: \"Although Democratic candidates received a nationwide plurality of more than 1.4 million votes (1.2%) in all House elections, the Republican Party won a 33-seat advantage in the state-apportioned totals, thus retaining its House majority by 17 seats.\" <https:\/\/en.wikipedia.org\/wiki\/United_States_House_of_Representatives_elections,_2012>.\n\n. \"United States House of Representatives elections, 2016,\" wikipedia.org, <https:\/\/en.wikipedia.org\/wiki\/United_States_House_of_Representatives_elections,_2016>.\n\n. 2016 Pre- and Post-Election State Legislative Control, by National Conference of State Legislatures, November 9, 2016. Excerpt: \"This map shows post-election legislative control as of Wednesday morning. Republicans control both chambers of the legislature in 32 states, Democrats control both chambers of the legislature in 13 states and three states split control.\" <http:\/\/www.ncsl.org\/research\/elections-and-campaigns\/2016-pre-and-post-election-state-legislative-control.aspx>.\n\n. \"October House Overview: GOP Risk Factors,\" by David Wasserman, The Cook Political Report, October 17, 2013. Redistricting expert Wasserman writes the GOP has a 6.8 percent head start in the competitive congressional races that will determine the House majority in 2014's elections. \"By our calculations, House Democrats would need to win the national popular vote for the House by between six and seven points in order to win the barest possible majority of 218 seats. How did we calculate this? In 2012, Democratic candidates won 49.16 percent of the national popular House vote to 48.03 for Republicans, but only won 201 of 435 seats. If every seat shifted uniformly in Democrats' direction, Democrats' lead nationally would have needed to expand to roughly 6.8 percent to drag 17 more GOP-won seats to Democrats' side.\" <http:\/\/cookpolitical.com\/story\/6330>.\n\nAlso see: \"A Very Early Look At The Battle For The House In 2018: Donald Trump is unpopular enough that Republicans could lose the House, but there's a lot of uncertainty.\" by Harry Enten, FiveThirtyEight.com, February 15, 2017. Enten writes the GOP has a 5.5 to 8.0 percent starting line advantage due to 2011's extreme redistricting in the races that will control the House in 2018's election. \"Democrats are essentially spotting the GOP 5.5 points in the battle for control of the House. And even that may be underestimating Republicans ability to win a majority of seats without a majority of the vote. Since 2012 (or when most states instituted the current House district lines), Republicans have won, on average, 51 percent of the two-party House vote and 55 percent of House seats. If that difference holds for 2018, Democrats would need to win the House popular vote by about 8 percentage points to win half the House seats.\" <https:\/\/fivethirtyeight.com\/features\/a-very-early-look-at-the-battle-for-the-house-in-2018\/>.\n\n. Ratf**ked: The True Story Behind the Secret Plan to Steal America's Democracy, by David Daley, Norton, 2016, <http:\/\/books.wwnorton.com\/books\/Ratfked\/>.\n\n. \"The GOP Targets State Legislatures: He who controls redistricting can control Congress,\" by Karl Rove, The Wall Street Journal, March 4, 2010. Rove's commentary laying out the RedMap plan began, \"The political world is fixated on whether this year's elections will deliver an epic rebuke of President Barack Obama and his party. If that happens, it could end up costing Democrats congressional seats for a decade to come.\"\n\n<https:\/\/www.wsj.com\/articles\/SB10001424052748703862704575099670689398044>.\n\n. \"2011 State and Legislative Partisan Composition,\" National Conference of State Legislatures. For 2011, when stage legislatures or commissions redrew legislative and US House districts, Democrats fully controlled eleven states: AR, CA, CN, DE, HI, IL, MD, MA, VT, VA, and WV. Republicans controlled twenty states: AL, AZ, FL, GA, ID, IN, KS, ME, MI, ND, OH, OK, PA, SC, SD, TN, TX, UT, WS, and WY. <http:\/\/www.ncsl.org\/documents\/statevote\/LegisControl_2011.pdf>.\n\n. \"Supreme Court forced to confront the 'unsavory' politics of district lines,\" by Richard Wolf, USA Today, May 25, 2017. Wolf notes the blue states where extreme redistricting is most prevalent are MD, MA, IL. \"In 2012, Republicans won 53% of the vote but 72% of the House seats in states where they drew the lines. Democrats won 56% of the vote but 71% of the seats where they controlled the process.\" <https:\/\/www.usatoday.com\/story\/news\/politics\/2017\/05\/25\/supreme-court-faces-decision-politics-election-districts\/102020210\/>.\n\nAlso see: \"The State of Redistricting Litigation (March 2017 edition),\" Brennan Center for Justice at New York University Law School, <https:\/\/www.brennancenter.org\/blog\/state-redistricting-litigation-march-2017-edition>.\n\nCHAPTER 8. CREATING PARTISAN GHETTOS\n\n. Issues Related to State Voter Identification Laws, US Government Accountability Office, September 2014. This report analyzes how stricter voter ID laws diminished turnout in Kansas and Tennessee. \"GAO found that turnout among eligible and registered voters declined more in Kansas and Tennessee than it declined in comparison states\u2014by an estimated 1.9 to 2.2 percentage points more in Kansas and 2.2 to 3.2 percentage points more in Tennessee\u2014and the results were consistent across the different data sources and voter populations used in the analysis.\" <https:\/\/www.gao.gov\/products\/GAO-14-634>.\n\n. Issues Related to Registering Voters and Administering Elections, US Government Accountability Office, June 2016. This report has a chart on page 40 that shows same-day registration and voting by mail have the largest measurable effects of increasing turnout. <http:\/\/www.gao.gov\/assets\/680\/678131.pdf>.\n\n. Baker v. Carr, Wikipedia. This is the US Supreme Court's 1962 decision stating the goal of redistricting was to create political districts based on the principle of one-person, one-vote. <https:\/\/en.wikipedia.org\/wiki\/Baker_v._Carr>.\n\n. Vieth v. Jubelirer 02\u20131580], [FindLaw.com. The Supreme Court's 2004 ruling, Vieth v. Jubelirer where Justice Anthony Kennedy wrote said political \"decorum\" should guide mapmakers. <http:\/\/caselaw.findlaw.com\/us-supreme-court\/541\/267.html>.\n\n. Same as 57: Ratf**ked: The True Story Behind the Secret Plan to Steal America's Democracy, by David Daley, Norton, 2016.\n\n. Same as 60: \"The State of Redistricting Litigation (March 2017 edition).\" This Brennan Center summary tracks federal redistricting litigation. The only challenge filed by Republicans against Democrats is in Maryland. It notes that redistricting cases fall under two categories, those challenging the illegal use of race and those asserting excessive partisanship.\n\nAlso see: \"Republican redistricting is taking a beating in the courts, right now,\" by Amber Phillips, The Washington Post, January 28, 2017. This report summarizes recent rulings in federal court that have gone against the GOP. In Wisconsin, a district court found partisan advantage. In Texas, Alabama, Virginia, and North Carolina, the courts found racial discrimination. <https:\/\/www.washingtonpost.com\/news\/the-fix\/wp\/2017\/01\/28\/republican-redistricting-is-taking-a-beating-in-the-courts-right-now\/?utm_term=.0b9cfe552d74>.\n\n. \"Three Tests for Practical Evaluation of Partisan Gerrymandering,\" by Samuel S.-H. Wang, Stanford Law Review, June 2016. Princeton professor Samuel S. H. Wang's study concluding that gerrymanders had cost Democrats as many as twenty-two House seats in the 2012 election\u2014nearly enough to flip the chamber's control. <http:\/\/www.stanfordlawreview.org\/wp-content\/uploads\/sites\/3\/2016\/06\/3_-_Wang_-_Stan._L._Rev.pdf>.\n\n. \"New Report: Extreme Partisan Maps Account for 16\u201317 Republican Seats in Congress,\" Brennan Center for Justice, May 16, 2017. This study finds sixteen to seventeen GOP House seats are affected by extreme redistricting. <https:\/\/www.brennancenter.org\/press-release\/new-report-extreme-partisan-maps-account-16-17-republican-seats-congress>.\n\n. Same as 55. NCSL report. 2016 Pre- and Post-Election State Legislative Control, by National Conference of State Legislatures, November 9, 2016.\n\n. Same as 56: \"A Very Early Look At The Battle For The House In 2018: Donald Trump is unpopular enough that Republicans could lose the House, but there's a lot of uncertainty,\" by Harry Enten, FiveThirtyEight.com, February 15, 2017.\n\n. Same as 66: \"The State of Redistricting Litigation (March 2017 edition),\" by the Brennan Center For Justice, and \"Republican redistricting is taking a beating in the courts, right now,\" by Amber Phillips, The Washington Post, January 28, 2017.\n\n. See 59 for GOP-held states in 2011. Also: \"When Does Political Gerrymandering Cross a Constitutional Line?\" by Adam Liptak, The New York Times, May 15, 2017. This update on gerrymandering says the Supreme Court has never struck down a map based on outsized partisan gains, and describes the analysis that is being used in the Wisconsin partisan-based challenge that calculates what are called wasted votes. It lists the states with biggest \"wasted vote\" gaps. Republican-held: FL, IN, KS, MI, MO, NC, OH, VA, WS, WY; Democrat-held: NY, RI. https:\/\/www.nytimes.com\/2017\/05\/15\/us\/politics\/when-does-political-gerrymandering-cross-a-constitutional-line.html?rref=collection%2Fbyline%2Fadam-liptak&action=click&contentCollection=undefined&region=stream&module=stream_unit&version=latest&contentPlacement=3&pgtype=collection.\n\n. Same as 56: \"A Very Early Look At The Battle For The House In 2018: Donald Trump is unpopular enough that Republicans could lose the House, but there's a lot of uncertainty.\" by Harry Enten, FiveThirtyEight.com, February 15, 2017.\n\n. \"Republican redistricting is taking a beating in the courts (again),\" by Amber Phillips, The Washington Post, May 22, 2017. This article reviews recent federal court redistricting decisions (NC, VA, TX, AL, WS) and notes that they stemmed from Republicans winning twenty-one state chambers in 2010. It also notes that the November 2016 ruling against Wisconsin's GOP was the first federal court ruling in a decade that voided a map on excessive partisan grounds. The Supreme Court will hear an appeal in its fall 2017 term. <https:\/\/www.washingtonpost.com\/news\/the-fix\/wp\/2017\/05\/22\/republican-redistricting-is-taking-a-beating-in-the-courts-again\/?utm_term=.499e563318b8>.\n\n. \"Key Question for Supreme Court: Will It Let Gerrymanders Stand?\" by Michael Wines, The New York Times, April 21, 2017. This article previews the Wisconsin gerrymandering case that's headed to the Supreme Court in the fall 2018 term. It notes how a lower federal court agreed with Democrats that it was an unconstitutional partisan power grab, but also that the US Supreme Court has never thrown out maps on partisan greed. (It doesn't discuss the other gerrymandering cases that are illegal, which is why they are based on overt and provable racial discrimination.) It also references the Court's 2004 ruling, in Vieth v. Jubelirer, where Justice Kennedy wrote said political \"decorum\" should guide mapmakers. https:\/\/www.nytimes.com\/2017\/04\/21\/us\/democrats-gerrymander-supreme-court.html?ref=todayspaper&_r=0.\n\n. Same as 60: \"The State of Redistricting Litigation (March 2017 edition),\" by the Brennan Center For Justice.\n\n. Cooper v. Harris, The US Supreme Court's May 2017 ruling 5 to 3 that North Carolina's 2011 gerrymander of two US House districts was racially discriminatory and unconstitutional. The dissent, starting on page 46, by Justice Samuel Alito and agreed to by Chief Justice John Roberts and Associate Justice Anthony Kennedy said the majority went looking for race, adding that extreme partisan redistricting was constitutional. <https:\/\/www.supremecourt.gov\/opinions\/16pdf\/15-1262_db8e.pdf>.\n\n. \"For Voting Rights Advocates, Court Decision Is 'Temporary Victory,'\" by Michael Wines, The New York Times, May 16, 2017. This analysis, coming after the Supreme Court's North Carolina ruling that concluded that the state's 2011 gerrymander was racially discriminatory, suggested that ruling was a temporary victory in a longer war. It discussed how the Court sees the two categories of gerrymandering cases, extreme partisanship and racial discrimination, differently, even as it struggles to separate them when those factors overlap. <https:\/\/www.nytimes.com\/2017\/05\/16\/us\/for-voting-rights-advocates-court-decision-is-temporary-victory.html?_r=0>.\n\nAlso see tweets by nationally known redistricting expert, University of Florida's Michael McDonald, who notes the North Carolina dissent by conservatives doesn't bode well for Wisconsin Democrat's excessive partisanship argument: <https:\/\/twitter.com\/ElectProject\/status\/866675389778624512>.\n\n. Same as 75: \"Key Question for Supreme Court: Will It Let Gerrymanders Stand?\" by Michael Wines, The New York Times, April 21, 2017. Also see 77, Cooper v. Harris, with dissent by Justice Samuel Alito starting on page 46.\n\n. \"Revised Redistricting Plans Face Strong Public Opposition: GOP lawmakers on a key Senate committee approved a revision Friday of controversial redistricting plans, but Democrats and others accused Republicans of lacking transparency,\" by Ezra Kaplan, Associated Press, March 24, 2017. This report about revising Georgia's state legislative districts had Democrats accusing majority Republicans of steamrolling the process by acting secretly and still maintaining an advantage. \"Jerry Gonzalez, the CEO of the Georgia Association of Latino Elected Officials, is concerned that the lack of transparency could have more nefarious results. 'Elected officials are rigging the system to racially pick their voters,' Gonzalez said to the committee. 'This process occurred under the cover of darkness and without public input.'\" <https:\/\/www.usnews.com\/news\/best-states\/georgia\/articles\/2017-03-24\/revised-redistricting-plans-face-strong-public-opposition>.\n\n. \"Senate Approves Redistricting Plan; Black Democrats Object: The Alabama Senate approved new legislative districts over the objections of black Democrats,\" by Kim Chandler, The Associated Press, May 4, 2017. This report says the GOP's redrawn state legislative districts in Alabama still are not fair. \"A contentious point in the debate was control of Jefferson County, home to the state's largest city, Birmingham, which is majority black and often the focus of partisan disputes over legislation impacting local governments. Democrats want an even number of Democrats and Republicans in the Jefferson County delegation. The Republican map, however, creates districts that would most likely elect four Republicans and three Democrats. The current split is five Republicans and three Democrats. Sen. Rodger Smitherman, D-Birmingham, said the GOP-drawn map stretches suburban districts into the county to maintain GOP control. 'It has smidgens of Jim Crow,' Smitherman said.\" <https:\/\/www.usnews.com\/news\/best-states\/alabama\/articles\/2017-05-04\/senate-approves-redistricting-plan>.\n\n. \"With Supreme Court appeal, Texas wants to keep congressional map intact: Attorney General Ken Paxton revealed that Texas has no plans to ask lawmakers to redraw the state's Congressional map in a fresh round of legislative overtime. Instead, Paxton is appealing the ruling to the US Supreme Court,\" by Jim Malewitz, The Texas Tribune, August 18, 2017, <https:\/\/www.texastribune.org\/2017\/08\/18\/paxton-redistricting-filing\/>.\n\n. \"A Harsh Dose of Electoral Reality: Democrats Have Uphill Battle in '18 and Need to Elect Governors to Fight Gerrymandered GOP Monopoly,\" by Steven Rosenfeld, AlterNet.org, April 7, 2017. Excerpt: \"'There is no question that Republican-led gerrymandering in 2011 has rigged the system against Democrats... What we need to do is level the playing field, unrig that system, so that Democrats can compete and we can translate those campaign tactics into electoral success that actually reflects the will of the voters, which is not happening right now,' Kelly Ward, executive director of the newly formed National Democratic Redistricting Committee said. 'That is what the National Democratic Redistricting Committee is trying to lead.'\" <http:\/\/www.alternet.org\/activism\/democratic-partys-future-hinges-more-winning-2018-governors-races-taking-back-house>.\n\n. \"The Supreme Court Could Make It Easier for States to Purge Voters; The Court has decided to hear a new case from Ohio, where Democrats were twice as likely as Republicans to be purged from the rolls in the state's largest counties.\" by Ari Berman, TheNation.com, May 30, 2017. This article reports on Ohio's voter purges of infrequent voters that are at the center of a case to be heard by the Supreme Court in its fall 2017 term. \"From 2011 to 2016, Ohio purged 2 million voters from the rolls\u20141.2 million for infrequent voting\u2014more than any other state.\" <https:\/\/www.thenation.com\/article\/supreme-court-make-easier-states-purge-voters\/>.\n\n. Judges Find Wisconsin Redistricting Unfairly Favored Republicans, by Michael Wines, November 21, 2017. This report discusses why the WI case is the first one in three decades to bring the issue of an unconstitutional partisan gerrymander to the Supreme Court. (It doesn't say that lower federal courts have been making these rulings.) It also notes that WI's partisan tilting of the playing field is of the same magnitude as VA, NC, and MI. https:\/\/www.nytimes.com\/2016\/11\/21\/us\/wisconsin-redistricting-found-to-unfairly-favor-republicans.html?mcubz=2&_r=0.\n\nAlso see 72: \"When Does Political Gerrymandering Cross a Constitutional Line?\" by Adam Liptak, The New York Times, May 15, 2017. This analysis discusses why the WI case, even with its new math showing outsize partisan impacts benefiting GOP, faces an uphill fight at the Supreme Court. \"The Supreme Court has never struck down an election map on the ground that it was drawn to make sure one political party would win an outsize number of seats. But it has left open the possibility that some kinds of political gamesmanship in redistricting may be too extreme.\"\n\nCHAPTER 9. LOOKING FORWARD, NOT BACKWARD\n\n. Deliver The Vote: A History of Election Fraud, An American Political Tradition\u20141742\u20132004, by Tracy Campbell, Carroll & Graf Publishers, 2005. <https:\/\/www.amazon.com\/Deliver-Vote-Election-Political-Tradition-1742-2004\/dp\/078671591X>.\n\n. Same as 52: The Fight to Vote, by Michael Waldman. For voter fraud investigation statistics, see pages 186, 187. For GAO study on impact of stricter voter ID requirements, see page 209. For Texas Republican Royal Masset's estimate of 3 percent reduction in Democratic turnout but to stricter voter ID laws, see page 190.\n\n. \"Over 220,000 Ballots Didn't Count In The Presidential Primaries; Try to cast a regular ballot, rather than a provisional one, when voting this fall,\" by Samantha Lachman, Huffington Post, June 16, 2016, \"Over 220,000 provisional ballots have already been discounted in the presidential primaries this year, according to a survey by The Huffington Post, with many thousands more likely to be tossed as states finish certifying their primaries.\" <http:\/\/www.huffingtonpost.com\/entry\/provisional-ballots-not-counted_us_5761bb92e4b0df4d586f15fe>.\n\n. Same as 61: Issues Related to State Voter Identification Laws, US Government Accountability Office, September 2014. This report discusses the most authoritative research on voter fraud, says its occurrence is exceptionally rare\u2014single-digit instances in state where million vote\u2014and that no federal agency tracks it. That absence helps to perpetuate fictitious claims about its presence. See numbered pages 62\u201374. \"Based on our own review of federal and state information sources, we identified challenges such as there is no single source of information on possible instances of in-person voter fraud and variation exists among federal and state sources in the extent to which they collect information on election fraud.\"\n\n. \"New Voting Restrictions in America,\" by the Brennan Center for Justice, 2016. This is an interactive webpage and map that details restrictive voting laws passed by states since the 2011 gerrymander. Of the twenty states listed, only two\u2014Rhode Island and Illinois\u2014are reliably blue. \"Overall, 20 states have new restrictions in effect since then\u201410 states have more restrictive voter ID laws in place (and six states have strict photo ID requirements), seven have laws making it harder for citizens to register, six cut back on early voting days and hours, and three made it harder to restore voting rights for people with past criminal convictions.\" <https:\/\/www.brennancenter.org\/new-voting-restrictions-america>.\n\n. \"Debunking the Voter Fraud Myth,\" by the Brennan Center for Justice, January 31, 2017. This is the center's most up-to-date report on this issue. It found most purported incidents of voter fraud are traceable to other causes, such as clerical errors or bad data matching. The report reviewed elections that had been meticulously studied for voter fraud, and found incident rates between 0.0003 percent and 0.0025 percent. It is more likely that an American \"will be struck by lightning than that he will impersonate another voter at the polls.\" <https:\/\/www.brennancenter.org\/analysis\/debunking-voter-fraud-myth>.\n\n. Texas NAACP v. Steen (consolidated with Veasey v. Abbott), the Brennan Center for Justice, July 6, 2017. This webpage gives the history of the cases, NAACP v. Steen: <https:\/\/www.brennancenter.org\/legal-work\/naacp-v-steen> This is the US District Court ruling issued on April 10, 2017, <https:\/\/www.brennancenter.org\/sites\/default\/files\/legal-work\/2017-04-10_Order_Intent.pdf>.\n\n. The Deceit of Voter Fraud, by Bob Hall and Isela Gutierrez, Democracy North Carolina, April 2017. This report details how Gov. Pat McCrory's reelection campaign fabricated voter fraud charges against six hundred people in a failed attempt to overturn the results of his narrow loss in November 2016. It includes the sloppy analysis used, interviews with victims, and even comments by Republican activists who dutifully filed charges contesting these voters registrations but who were left empty-handed by McCrory's campaign during subsequent hearings by county boards of election. <http:\/\/nc-democracy.org\/wp-content\/uploads\/2017\/04\/FraudReport.pdf>.\n\n. A Sampling of Election Fraud Cases From Across The Country, The Heritage Foundation, 2017. This report lists 492 cases and 733 convictions of false voter registrations, noncitizen voting, fraudulent use of absentee ballots, and duplicate voting from 1984 through 2017. <http:\/\/thf-legal.s3.amazonaws.com\/VoterFraudCases.pdf>.\n\n. \"Voter Turnout in Presidential Elections: 1828\u20132012,\" The American Presidency Project, University of California Santa Barbara. This chart lists the presidential election turnout through 2012. <http:\/\/www.presidency.ucsb.edu\/data\/turnout.php>. Also see, \"2016 November General Election Turnout Rates,\" The United States Election Project, the University of Florida. This chart lists the 2016 turnout. Together, 980 million people voted for president between 1984 and 2016. <http:\/\/www.electproject.org\/2016g>.\n\n. \"Voting Laws Roundup 2017,\" by the Brennan Center for Justice, May 10, 2017. Excerpt: \"At this point in the year, every state's legislature is either in session or has completed its 2017 calendar. As has been the case all decade, legislators across the country are trying to reshape state voting laws. In several places, this means it will soon be harder to vote: Five states have already enacted bills to cut back on voting access, and one more is on the verge of doing so. By comparison, three states enacted voting restrictions in 2015 and 2016 combined.\" <https:\/\/www.brennancenter.org\/analysis\/voting-laws-roundup-2017>.\n\n. \"Hearing Before The Committee on House Administration, Columbus Ohio, March 21, 2005.\" This is the transcript and submissions, including testimony by Mark \"Thor\" Hearne. https:\/\/babel.hathitrust.org\/cgi\/pt?id=pur1.32754078873928;view=1up;seq=31.\n\n. Same as 97: All the quotations in the paragraphs that follow are from this source.\n\n. Same as 97: Hearne testimony and submitted report begin on page 273.\n\n. Same as 97.\n\n. What Happened in Ohio? A Documentary Record of Theft and Fraud in the 2004 Election, by Robert J. Fitrakis, Steven Rosenfeld and Harvey Wasserman, The New Press, 2006. Pages 306\u2013313 summarize impact of anti-voter actions taken by Secretary of State Kenneth Blackwell. <http:\/\/thenewpress.com\/books\/what-happened-ohio>.\n\n. Same as 52: The Fight to Vote, by Michael Waldman. See pages 186\u20137 for examples of voter fraud that federal investigations found\u2014all were errors or small-time ploys unworthy of Justice Department scrutiny. <http:\/\/www.simonandschuster.com\/books\/The-Fight-to-Vote\/Michael-Waldman\/9781501116483>.\n\n. \"Why Republicans Can't Find the Big Voter Fraud Conspiracy: If the last federal investigation is any guide, the answer is simple: It probably doesn't exist,\" by Lisa Rab, Politico Magazine, April 2, 2017. This feature reviews the history of the GOP's pursuit of voter fraud, including the George W. Bush administration's firing on federal prosecutors for failing to deliver voter fraud convictions, the DOJ's investigation results and prosecutions, and other evidence pointing to its real-life status as a political fiction. <http:\/\/www.politico.com\/magazine\/story\/2017\/04\/why-republicans-cant-find-the-big-voter-fraud-conspiracy-214972>.\n\n. \"ACORN Accusations: McCain makes exaggerated claims of 'voter fraud.' Obama soft-pedals his connections,\" by Jess Henig, October 21, 2008, <http:\/\/www.factcheck.org\/2008\/10\/acorn-accusations\/>.\n\n. \"Election 2012: Voting Laws Roundup,\" by the Brennan Center for Justice, October 11, 2012, <http:\/\/www.brennancenter.org\/analysis\/election-2012-voting-laws-roundup>.\n\nCHAPTER 10: THE VOTING WARS\n\n. Same as 52: The Fight to Vote, by Michael Waldman. See page 186\u20137 for federal voter fraud investigations, page 190 for estimate by Texas Republican Party's Royal Masset that stricter voter ID laws would undercut Democratic turnout by 3 percent, and page 209 for GAO studies. <http:\/\/www.simonandschuster.com\/books\/The-Fight-to-Vote\/Michael-Waldman\/9781501116483>.\n\n. Same as 61: Issues Related to State Voter Identification Laws, US Government Accountability Office, September 2014.\n\n. \"Voter Suppression Works,\" by Sean McElwee, demos.org, May 9, 2017. This study on the impact of stricter voter ID compared laws between red and blue states. It found stricter ID laws in Mississippi, Virginia, and Wisconsin caused turnout to drop by 1.7 percent. When Wisconsin was compared to Minnesota, turnout fell by 2.2 percent in high black population counties. <http:\/\/www.demos.org\/blog\/5\/9\/17\/voter-suppression-works>.\n\n. \"Do voter identification laws suppress minority voting? Yes. We did the research.\" by Zoltan Hajnal, Nazita Lajevardi, Lindsay Nielson, The Washington Post, February 15, 2017. This analysis suggests that stricter voter ID laws undercut progressive candidates in primaries and pushes the Democratic Party toward the center. \"All else equal, when strict ID laws are instituted, the turnout gap between Republicans and Democrats in primary contests more than doubles from 4.3 points to 9.8 points. Likewise, the turnout gap between conservative and liberal voters more than doubles from 7.7 to 20.4 points... Unsurprisingly, these strict ID laws are passed almost exclusively by Republican legislatures.\" <https:\/\/www.washingtonpost.com\/news\/monkey-cage\/wp\/2017\/02\/15\/do-voter-identification-laws-suppress-minority-voting-yes-we-did-the-research\/?utm_term=.a248da5d08ce>.\n\n. \"Voting Law Changes: Election Update,\" by Wendy Weiser and Diana Kasdan, The Brennan Center for Justice, October 28, 2012. <http:\/\/www.brennancenter.org\/publication\/voting-law-changes-election-update>.\n\n. \"The Voting Rights Act of 1965,\" The US Department of Justice. This is the DOJ's online description of the history of the Voting Rights Act and subsequent amendments in 1970, 1975, 1982, and 2006. <https:\/\/www.justice.gov\/crt\/history-federal-voting-rights-laws>.\n\n. Same as 110: \"Voting Law Changes: Election Update,\" by Wendy Weiser and Diana Kasdan, The Brennan Center for Justice, October 28, 2012.\n\n. Shelby County v. Holder, The Supreme Court's 2013 Shelby County v. Holder decision striking down the enforcement formula in Section 5 as unconstitutional, <https:\/\/www.supremecourt.gov\/opinions\/12pdf\/12-96_6k47.pdf>.\n\n. \"About Section 5 of the Voting Rights Act,\" The US Department of Justice. This is the DOJ's summary of states and counties previously covered under Section 5 of the Voting Rights Act. <https:\/\/www.justice.gov\/crt\/about-section-5-voting-rights-act>.\n\n. \"Special Report: Behind US race cases, a little-known recruiter,\" by Joan Biskupic, Reuters, December 4, 2012. Excerpt: \"Over the past 20 years, Blum has similarly launched at least a dozen lawsuits attacking race-based protections. In addition to the Fisher [affirmative action in university admissions] and Shelby County [Voting Rights Act] cases, two other Blum-backed cases reached the Supreme Court. One struck down majority-black and majority-Latino voting districts in Texas. The other prompted the court to suggest it might eliminate a major portion of the Voting Rights Act of 1965, which the conservative-majority bench may now be poised to do.\" <http:\/\/www.reuters.com\/article\/us-usa-court-casemaker-idUSBRE8B30V220121204>.\n\nAlso see: \"Meet the Brains Behind the Effort to Get the Supreme Court to Rethink Civil Rights: Edward Blum has orchestrated a string of big cases targeting voting rights and affirmative action,\" by Stephanie Mencimer, Mother Jones, March\/April 2016. <http:\/\/www.motherjones.com\/politics\/2016\/04\/edward-blum-supreme-court-affirmative-action-civil-rights>.\n\n. \"The GOP is Dying Off. Literally,\" by Daniel J. McGraw, May 17, 2015, Politico Magazine. <http:\/\/www.politico.com\/magazine\/story\/2015\/05\/the-gop-is-dying-off-literally-118035>.\n\n. \"Voting Law Roundup 2013,\" by the Brennan Center for Justice, Dec. 19, 2013. This summary lists the 2013 legislation and resulting laws that would restrict and also expand voting rights following the Supreme Court's Shelby County ruling. Typically, red states, including some presidential swing states, moved to restrict rights, while blue states went in the opposite direction. <http:\/\/www.brennancenter.org\/analysis\/election-2013-voting-laws-roundup> Also see: \"Voting Laws Roundup 2014,\" by the Brennan Center for Justice, December 18, 2014. <http:\/\/www.brennancenter.org\/analysis\/voting-laws-roundup-2014>.\n\n. Same as 114. \"About Section 5 of the Voting Rights Act,\" Department of Justice.\n\n. Same as 24: Democracy Diminished: State and Local Threats to Voting Post-Shelby County, Alabama v. Holder, by NAACP Legal Defense Fund, June 9, 2016. This report details post-Shelby voter suppression tactics and legal fights that continued into the 2016 election. It has sections on Alabama, Arizona, Arkansas, Florida, Georgia, North Carolina, South Carolina, and also describes the fight over requiring documented proof of citizenship on voter registration forms at US Election Assistance Commission.\n\n. Same as 96: \"Voting Laws Roundup 2017,\" by the Brennan Center for Justice, May 10, 2017. This update notes that more anti-voter laws were passed in 2017 than in the two prior years, but notes that blue states have been looking at expanding voting rights. Those blue states are not considered swing states for the House majority or presidential elections.\n\n. \"Supreme Court: Virginia redistricting must be reexamined for racial bias,\" by Robert Barnes, Chicago Tribune, March 1, 2017. This report discusses the Supreme Court's ruling that sent the Virginia redistricting case back to lower courts to assess the latest maps. While it is nominally a victory for minority voters and Democrats, the political boundaries that have been challenged as racially discriminatory remain. <http:\/\/www.chicagotribune.com\/news\/nationworld\/ct-supreme-court-virginia-redistricting-racial-bias-20170301-story.html>.\n\nAlso see: \"Will third time be a charm for those who want NC Supreme Court to invalidate election maps?\" by Anne Blythe, The News and Observer, May 30, 2017. This report recounts how the North Carolina GOP has thwarted repeated federal court rulings to redraw state and congressional district maps. <http:\/\/www.newsobserver.com\/news\/politics-government\/politics-columns-blogs\/under-the-dome\/article153347589.html>.\n\nAlso see: \"Revised Redistricting Plans Face Strong Public Opposition: GOP lawmakers on a key Senate committee approved a revision Friday of controversial redistricting plans, but Democrats and others accused Republicans of lacking transparency,\" by Ezra Kaplan, Associated Press, March 24, 2017. This AP report about revising Georgia state legislative districts also had Democrats accusing the majority GOP of acting secretly to retain their advantage. <https:\/\/www.usnews.com\/news\/best-states\/georgia\/articles\/2017-03-24\/revised-redistricting-plans-face-strong-public-opposition>.\n\n. \"Alabama's GOP approves new maps; Dems vow repeat court fight: Alabama's overwhelmingly Republican legislature has sent a new legislative map to the governor after a rancorous session featuring a racially charged email comparing lawmakers to monkeys.\" by Kim Chandler, Associated Press, May 20, 2017. What's unfolded in Alabama is typical of GOP efforts to keep their structural advantage despite losing in court. \"'It seems like we are going to end up in court again,' said Legislative Black Caucus Chairman John Knight, D-Montgomery. 'It's clear. You can look at the [newly redrawn] map. There is racial gerrymandering.'\" <https:\/\/www.apnews.com\/526ad1f61e32407ab28c40616efbbcd7>.\n\n. Same as 24: Democracy Diminished: State and Local Threats to Voting Post-Shelby County, Alabama v. Holder, by NAACP Legal Defense Fund, June 9, 2016.\n\n. \"Plaintiffs Claim Bias During Closing Argument Against Texas Voter ID Law,\" by Manny Fernandez, The New York Times, Sept. 22, 2014. Excerpt: \"Richard Dellheim, a lawyer for the Justice Department's Voting Section, said in his closing argument that more than 608,000 registered voters in Texas lack the types of ID required by the law and that the percentage of Hispanics and blacks without the necessary identification is substantially higher than the percentage of whites without that ID... Texas has issued only 279 election ID cards. A similar program in Georgia has issued 2,200 cards, Mr. Dellheim said.\" <https:\/\/www.nytimes.com\/2014\/09\/23\/us\/plaintiffs-assert-bias-during-closing-argument-against-texas-voter-id-law.html?mcubz=2>.\n\n. Same as the first item in 121: \"Supreme Court: Virginia redistricting must be reexamined for racial bias,\" by Robert Barnes, Chicago Tribune, March 1, 2017.\n\n. \"Va. Supreme Court strikes down McAuliffe's order on felon voting rights,\" by Graham Moomaw, Richmond Times-Dispatch, July 23, 2016. This report recounts the state Supreme Court ruling that threw out Gov. McAuliffe's first executive order restoring ex-felon voting rights, which resulted in 11,662 ex-felons to register. Virginia's Supreme Court ordered those ex-felons to be removed from voter rolls. McAuliffe replied he'd take steps to reinstate their voting rights. <http:\/\/www.richmond.com\/news\/virginia\/government-politics\/va-supreme-court-strikes-down-mcauliffe-s-order-on-felon\/article_718d04d8-70b2-5bfb-aa8c-0ff1ca108b8d.html>.\n\n. \"Governor McAuliffe Provides Update on Restoration of Rights Numbers,\" by WWIR, NBC29.com, January 18, 2018. This report describes McAuliffe's second clemency effort, which restored 140,000 ex-felon's voting rights. <http:\/\/www.nbc29.com\/story\/34203534\/governor-mcauliffe-provides-update-on-restoration-of-rights-numbers>.\n\nAlso see: \"Virginia's breakthrough on rectifying an enormous injustice,\" Editorial Board, The Washington Post, June 11, 2017. This editorial reports the number of ex-felons whose voting rights were restored as 156,000, and that 26,000 of those individuals had voted in November 2016. <https:\/\/www.washingtonpost.com\/opinions\/virginias-breakthrough-on-rectifying-an-enormous-injustice\/2017\/06\/11\/c90e3808-4d52-11e7-9669-250d0b15f83b_story.html?utm_term=.b0528b5bfd05>.\n\n. \"Are Felons More Likely to Vote for Democrats over Republicans?\" by ProCon.org, March 7, 2014. This website contains statements by Republicans and Democrats about the partisan implications of restoring ex-felon voting rights, including the 2004 statement by Marty Connor to the Washington Post that ex-felons are likely to support Democrats. <http:\/\/felonvoting.procon.org\/view.answers.php?questionID=000662>.\n\n. \"Florida: An Outlier in Denying Voting Rights,\" by Erkia Wood, The Brennan Center for Justice, December 16, 2016. This report has national and state-based statistics on felon disenfranchisement. \"Florida is one of only three states with a lifetime voting ban for people with felony convictions. The strict law disenfranchises 1.6 million citizens, including 21 percent of the state's voting-age African Americans.\" <https:\/\/www.brennancenter.org\/publication\/florida-outlier-denying-voting-rights>.\n\nAlso see: \"Florida's felons inch closer to regaining right to vote,\" by Dan Sweeney, Sun-Sentinel, December 31, 2016. Ten percent of Florida's voting age population can't vote because of criminal records. <http:\/\/www.sun-sentinel.com\/news\/florida\/fl-voting-rights-felons-20161231-story.html>.\n\nAlso see: \"Epic Fight Over Florida's Political Future Looms As Constitutional Amendment to Restore Felon Voting Rights Clears Hurdle,\" by Lulu Friesdat, AlterNet.org, April 27, 2017. This is an update on a prospective state ballot measure to amend the Florida Constitution to restore felon voting. <http:\/\/www.alternet.org\/activism\/epic-fight-over-floridas-political-future-looms-constitutional-amendment-restore-felon>.\n\n. \"No, 1.8 million dead people aren't going to vote in November,\" by Phillip Bump, The Washington Post, October 18, 2016. This piece recounts Trump's oft-recreated claims that more than 1 million dead people will be voting\u2014but not likely for him. The Post's traces the source of that claim, notes how Trump manipulates a nonissue into a fake threat, and shows how dead people do not vote. The real-life challenge for election officials is keeping their voters lists accurate after people die or move, in both cases not being informed of the change. <https:\/\/www.washingtonpost.com\/news\/the-fix\/wp\/2016\/10\/18\/no-1-8-million-dead-people-arent-going-to-vote-in-november\/?utm_term=.16e1f2247798>.\n\nCHAPTER 11: THE STARTING LINE\n\n. \"Jeff Sessions,\" wikipedia.org. Jeff Sessions' biography. <https:\/\/en.wikipedia.org\/wiki\/Jeff_Sessions>.\n\nSee 103, for 1985 voter fraud prosecutions as US Attorney in Alabama: \"Why Republicans Can't Find the Big Voter Fraud Conspiracy: If the last federal investigation is any guide, the answer is simple: It probably doesn't exist,\" by Lisa Rab, Politico Magazine, April 2, 2017. See: \"Names of the American Civil War,\" wikipedia.org, for Confederate names of American Civil War: <https:\/\/en.wikipedia.org\/wiki\/Names_of_the_American_Civil_War>.\n\n. Same as 24: Democracy Diminished: State and Local Threats to Voting Post-Shelby County, Alabama v. Holder, by NAACP Legal Defense Fund, June 9, 2016.\n\n. Same as 81: \"Alabama House Approves Redistricting Bill Over Objections,\" by Kim Chandler, Associated Press, May 18, 2017.\n\nAlso see 66: \"The State of Redistricting Litigation (March 2017 edition),\" by the Brennan Center For Justice.\n\n. \"How the GOP Is Re-segregating the South: Republicans are using the redistricting process to undermine minority voting power and ensure their party's dominance,\" by Ari Berman, The Nation, February 20, 2012. This report describes GOP redistricting is an attempt to turn the GOP into a party of white voters in the South, while making Democrats a bastion of nonwhites. <https:\/\/www.thenation.com\/article\/how-gop-resegregating-south\/>.\n\nAlso see: \"Separate and Unequal Voting in Arizona and Kansas: Two states want to create a two-tiered voting system reminiscent of the Jim Crow South,\" by Ari Berman, The Nation, October 15, 2013. This report described how Arizona and Kansas \"want to create a two-tiered voting system reminiscent of the Jim Crow South.\" <https:\/\/www.thenation.com\/article\/separate-and-unequal-voting-arizona-and-kansas\/>.\n\n. Same as 108: \"How Do Proof-of-Citizenship Laws Block Legitimate Voters?\" by Stuart Naifeh, demos.org, August 25, 2014. This report summarizes the various state and federal laws and lawsuits surrounding the attempt to add documented proof of citizenship to the voter registration process.\n\nAlso see 24: Democracy Diminished: State and Local Threats to Voting Post-Shelby County, Alabama v. Holder, by NAACP Legal Defense Fund, June 9, 2016.\n\nAlso see: \"Voter Identification, Proof of Citizenship for Voting and Provisional Ballots: A National Overview,\" by Wendy Underhill, National Conference of State Legislatures, August 25, 2016. This paper presented at NCSL's 2016 conference is an overview of proof of citizenship laws, including how Tennessee's law differs from Arizona, Kansas, Alabama, and Georgia. \"The Tennessee approach is different than the Kansas approach because it only comes into play when there is reason to doubt citizenship, rather than creating a requirement for all voters to show proof of citizenship.\" http:\/\/www.legis.nd.gov\/files\/committees\/64\u20132014%20appendices\/17_5143_03000appendixf.pdf.\n\n. \"Additional Data Could Help State and Local Elections Officials Maintain Accurate Voter Registration Lists,\" US Government Accountability Office, June 2005. The report notes there is no single federal citizenship database, which has not changed since then. See numbered pages 14\u201317. \"Two types of standard sources of identification, such as a driver's license, state identification or social security numbers are not useful because neither are evidence of citizenship.\" <http:\/\/www.gao.gov\/new.items\/d05478.pdf>.\n\n. \"Register to Vote and Confirm or Change Registration,\" USA.gov. This is the US government's webpage with voter registration basics and links to state deadlines and voter ID requirements. <https:\/\/www.usa.gov\/register-to-vote>.\n\n. \"Register To Vote In Your State by Using This Postcard Form and Guide,\" vote.gov. This is the federal voter registration form and accompanying state registration details. Under Kansas, it described the documented proof of citizenship requirements and filing deadlines. <https:\/\/vote.gov\/files\/federal-voter-registration_1-25-16_english.pdf>.\n\n. \"Same-day voter registration,\" ballotpedia.org. This lists the states with same-day or Election Day registration. Thirteen states and Washington, DC, had this option for the 2016 election. Three more states have passed it but it had not taken effect. <https:\/\/ballotpedia.org\/Same-day_voter_registration>.\n\n. \"Voter Identification Requirements | Voter ID Laws,\" by Wendy Underhill, National Conference of State Legislatures, June 5, 2017. This 2017 summary of state voter ID requirements says seventeen states in 2017 require state-issued photo IDs. \"A total of 34 states have laws requesting or requiring voters to show some form of identification at the polls. 32 of these voter identification laws are in force in 2017. West Virginia's law, signed on April 1, 2016, goes into effect in 2018 and Iowa's law, signed on May 5, 2017, will not be fully implemented until 2019... The remaining 18 states use other methods to verify the identity of voters. Most frequently, other identifying information provided at the polling place, such as a signature, is checked against information on file.\" <http:\/\/www.ncsl.org\/research\/elections-and-campaigns\/voter-id.aspx#Details>.\n\n. Mark Martin, Secretary of State, et al v. Freedom Kohls, Toylanda Smith, Joe Flakes and Barry Haas, Supreme Court of Arkansas, No. CV-14-462, October 15, 2014.\n\nThis Arkansas Supreme Court ruling threw out the state's 2013 voter ID law as an unconstitutional \"additional qualification\" for eligibility. <https:\/\/posting.arktimes.com\/media\/pdf\/voter_id_opinion.pdf>.\n\nAlso see: \"Arkansas governor signs bill to reinstate voter ID law,\" by Andrew Demillo, Associated Press, March 24, 2017. The state's Republican Governor, Asa Hutchinson, signed a new voter ID law in early 2017 reinstating the requirement. <http:\/\/www.pbs.org\/newshour\/rundown\/arkansas-governor-signs-bill-reinstate-voter-id-law\/>.\n\n. \"Kris Kobach,\" by wikipedia.org. This biography article cites Kansas Secretary of State's history, including anti-immigrant activism. <https:\/\/en.wikipedia.org\/wiki\/Kris_Kobach>.\n\nAlso see: \"Advocate for Tough Voting Rules to Steer Trump's Elections Commission,\" by Michael Wines and Julie Bosman, The New York Times, May 14, 2017. This profile follows his appointment as vice-chair by Trump to his Advisory Commission on Election Integrity. https:\/\/www.nytimes.com\/2017\/05\/14\/us\/kris-kobach-voter-fraud.html?rref=collection%2Fsectioncollection%2Fus&action=click&contentCollection=us&region=rank&module=package&version=highlights&contentPlacement=6&pgtype=sectionfront&_r=0.\n\n. \"Arizona Taxpayer and Citizen Protection, Proposition 200 (2004),\" ballotpedia.org. The text and litigation summary of Arizona's 2004 state ballot measure, Proposition 200. <https:\/\/ballotpedia.org\/Arizona_Taxpayer_and_Citizen_Protection,_Proposition_200_(2004)>.\n\nAlso see: \"Uncovering Kris Kobach's Anti-Voting History,\" by Tomas Lopez and Jennifer L. Clark, The Brennan Center for Justice at New York University School of Law, May 11, 2017. <https:\/\/www.brennancenter.org\/blog\/uncovering-kris-kobach%E2%80%99s-anti-voting-history>.\n\n. Same as 142, \"Kris Kobach,\" Wikipedia section on \"Immigration lawsuits.\"\n\nAlso see: \"The Man Behind Trump's Voter Fraud Obsession,\" by Ari Berman, The New York Times Magazine, June 18, 2017. This further details Kobach's anti-immigrant work. https:\/\/www.nytimes.com\/2017\/06\/13\/magazine\/the-man-behind-trumps-voter-fraud-obsession.html?mcubz=2&_r=0.\n\n. Same as 134: \"Separate and Unequal Voting in Arizona and Kansas: Two states want to create a two-tiered voting system reminiscent of the Jim Crow South,\" by Ari Berman, The Nation, October 15, 2013. This report recounts the efforts by Kansas and Arizona to add documented proof of citizenship to their state voter registration forms, notes how Arizona law blocked more than thirty thousand people from registering, notes its Kansas corollary stopped over seventeen thousand people from registering, and reports the 2013 US Supreme Court ruling against Arizona.\n\nAlso see: Arizona v. Inter Tribal Council of Arizona, Inc. (Amicus Brief), by the Brennan Center for Justice at New York University School of Law, June 17, 2013. This webpage summarized the Supreme Court case and has links to all the briefs filed by voting rights advocates. <http:\/\/www.brennancenter.org\/legal-work\/arizona-v-inter-tribal-council-arizona-inc-amicus-brief>.\n\n. Same as 103: \"Why Republicans Can't Find the Big Voter Fraud Conspiracy: If the last federal investigation is any guide, the answer is simple: It probably doesn't exist,\" by Lisa Rab, Politico Magazine, April 2, 2017. This feature includes Kobach's 2010 comments on voter fraud burrowing into every corner of United States and illegal aliens voting in Kansas.\n\n. \"Proof of US Citizenship for Voter Registration: NOTE: This requirement is effective January 1, 2013.\" GotVoterId.com. This is Kansas's website page describing its proof of citizenship requirement for voter registration. <http:\/\/www.gotvoterid.com\/proof-of-citizenship.html>.\n\nAlso see: \"Secretary of State Kris Kobach announces first conviction of noncitizen voting in Kansas,\" by Hunter Woodall, KansasCity.com, April 12, 2017. This report notes that Kobach is the nation's only secretary with power to prosecute voter fraud. <http:\/\/www.kansascity.com\/news\/politics-government\/article144256424.html>.\n\n. \"Appeals court nixes proof of citizenship for voting in Georgia,\" by Kristina Torres, Atlanta Journal-Constitution, September 11, 2016. This report recounts the history of the state's proof of citizenship laws for voter registration and efforts by Republicans to get the US Election Assistance Commission to include that requirement in its voter registration materials. http:\/\/www.ajc.com\/news\/state--regional-govt\u2014politics\/appeals-court-nixes-proof-citizenship-for-voting-georgia\/wq0WTs7FtiqWeQQNHldaPP\/.\n\n. Kobach et al. v. The United States Election Assistance Commission, by the Brennan Center for Justice at New York University School of Law, June 29, 2015. This webpage summarizes the 2015 suit by Arizona and Kansas against the US Election Assistance Commission over adding the requirement of citizenship to state voter registration forms. In 2013, the US Supreme Court overturned a 2004 Arizona law requiring citizenship proof for the federal voter registration form. <http:\/\/www.brennancenter.org\/legal-work\/kobach-et-al-v-united-states-election-assistance-commission>.\n\nAlso same as 144: \"The Man Behind Trump's Voter Fraud Obsession,\" by Ari Berman, The New York Times Magazine, June 18, 2017.\n\n. Same as 149. Kobach et al. v. The United States Election Assistance Commission, by the Brennan Center for Justice at New York University School of Law, June 29, 2015.\n\n. League of Women Voters v. Newby, by the Brennan Center for Justice at New York University School of Law, June 1, 2017. This page summarizes the litigation surrounding the unilateral decision by US Election Assistance Commission executive director Brian Newby to include proof of citizenship instructions from three states on the EAC's federal voter registration materials. It also notes how prior academic research has found that 7 percent of voters could be disenfranchised by this requirement. <https:\/\/www.brennancenter.org\/legal-work\/league-women-voters-v-newby>.\n\n. Same as 151. League of Women Voters v. Newby, by the Brennan Center for Justice at New York University School of Law, June 1, 2017.\n\n. Same as 144: \"The Man Behind Trump's Voter Fraud Obsession,\" by Ari Berman, The New York Times Magazine, June 18, 2017.\n\n. \"Kris Kobach supports Donald Trump's assertion millions voted illegally,\" by Tim Carpenter, The Topeka Capital-Journal, November 30, 2016. Kansas Secretary of State Kris Kobach agrees with Trump that as many as 3 million noncitizens voted illegally. <http:\/\/cjonline.com\/news\/2016-11-30\/kris-kobach-supports-donald-trump-s-assertion-millions-voted-illegally>.\n\n. \"Noncitizen Voting: The Missing Millions,\" by Christopher Famighetti, Douglas Keith, Myrna Perez, by the Brennan Canter for Justice at New York University School of Law, May 5, 2017. This report surveyed the ten counties with the largest noncitizen populations in United States to see the extent of noncitizen voting in 2016\u2014following Trump's claims that upwards of 3 million immigrant illegally voted, thus preventing him from a national popular vote majority. The Brennan Center found only thirty cases of noncitizen voting out of 23.5 million votes cast in 2016 from these counties. <https:\/\/www.brennancenter.org\/publication\/noncitizen-voting-missing-millions>.\n\n. Same as 142: \"Advocate for Tough Voting Rules to Steer Trump's Elections Commission,\" by Michael Wines and Julie Bosman, The New York Times, May 14, 2017. This profile of Kris Kobach's career notes his anti-immigrant advocacy, exaggerated claims of noncitizen voting and voter fraud, and how his policies have disenfranchised thousands while only one noncitizen voter was found\u2014only nine people were successfully prosecuted for voter fraud.\n\n. \"Critics See Efforts by Counties and Towns to Purge Minority Voters From Rolls,\" by Michael Wines, The New York Times, July 31, 2016. This report recounts how in Sparta, Georgia, local Republicans deployed sheriff deputies to question the voting credentials of blacks in 2015 before a mayoral election. https:\/\/www.nytimes.com\/2016\/08\/01\/us\/critics-see-efforts-to-purge-minorities-from-voter-rolls-in-new-elections-rules.html?mcubz=2&_r=0.\n\n. \"New Georgia Project releases financial records, renews voter push,\" by Kristina Torres, The Atlanta Journal-Constitution, February 11, 2016. This report recounts how the New Georgia Project submitted forty-one thousand voter registrations before the 2014 November election that were not processed\u2014denying those registrants the vote that fall. After Election Day, eighteen thousand of those registrants were added to voter rolls. http:\/\/www.myajc.com\/news\/state\u2014regional-govt\u2014politics\/new-georgia-project-releases-financial-records-renews-voter-push\/JoatzFC6XHZQ8GcACOsRN\/.\n\n. \"Quick Response Evaluation: Accuracy of the Help America Vote Verification Program Responses,\" June 22, 2009 memo from Inspector General Patrick P. O'Carroll Jr., Social Security Administration, memo to SSA Commissioner. This internal agency report notes how using the last four digits of Social Security numbers, verses the full nine-digit numbers, creates twice as many false positives. \"Our comparison of the FY 2008 verification responses for HAVV [SSA's system] and three other verification programs used by the States and employers \u2014SSNVS, E-Verify, and SSOLV\u2014showed that HAVV had a significantly higher no-match response rate. The three verification systems were developed for different purposes and [they] use varying tolerances because they require the full nine-digit SSN for verification. HAVV, on the other hand, does not use these same tolerances because it requires use of the last four digits of the SSN as mandated by the legislation. The no-match response rate for the three other programs ranged from 6 to 15 percent, whereas HAVV's no-match response rate was 31 percent. Therefore, HAVV's no-match response rate was about two to five times higher than the no-match response rate for the other three programs.\" <https:\/\/oig.ssa.gov\/sites\/default\/files\/audit\/full\/html\/A-03-09-29115_7.html>.\n\n. Same as 136: \"Additional Data Could Help State and Local Elections Officials Maintain Accurate Voter Registration Lists,\" US Government Accountability Office, June 2005. This report notes matching used by states to verify voter registrations is error prone. It cites a Wisconsin study of its motor vehicle database, saying, \"The matching and validation of names are complex and made more so when considering aliases and name changes, as are matches such as 'Margie L. Smith' with 'Margaret Smith' according to a Wisconsin study. The study estimated that even a 1 percent error rate on a match validating names, driver license numbers, etc. could generate tens of thousands of bad matches.\"\n\n. \"Georgia judge sides with state in 'missing' voters case,\" by Kristina Torres, The Atlanta Journal-Constitution, October 29, 2014. This report describes the electronic matching process used by county election offices to verify the identities of new voters under direction of the Georgia Secretary of State. In addition to \"pinging\" the state driver's license database and federal Social Security database, it also pings the state Department of Corrections, whose information is not up-to-date. http:\/\/www.myajc.com\/news\/state\u2014regional-govt\u2014politics\/georgia-judge-sides-with-state-missing-voters-case\/AkvPr6mP9FiMXglh5NUlHJ\/.\n\n. Same as 161: \"Georgia judge sides with state in 'missing' voters case,\" by Kristina Torres, The Atlanta Journal-Constitution, October 29, 2014.\n\n. \"From a Contentious Election to a Stronger Democracy: Strengthening democracy is the key to all other reforms,\" by Miles Rapoport, The American Prospect, October 6, 2016. <http:\/\/prospect.org\/article\/contentious-election-stronger-democracy>.\n\n. \"Voting Advocates Announce a Settlement of 'Exact Match' Lawsuit in Georgia: Minor Typos and Data Entry Errors will No Longer Deny Eligible Georgians the Right to Register and Vote,\" press release from Lawyers Committee for Civil Rights Under Law and others, February 10, 2017. <https:\/\/lawyerscommittee.org\/press-release\/voting-advocates-announce-settlement-exact-match-lawsuit-georgia\/>.\n\n. Same as 62: Issues Related to Registering Voters and Administering Elections, US Government Accountability Office, June 2016. This report describes how The Interstate Crosscheck program works on page 30 in footnote 45. \"Participating states agree to provide information such as full name and date of birth for those registered as well as turnout data, to Interstate Voter Registration Crosscheck program administrators in January following a general election. Using the information provided by member states, the Interstate Voter Registration Crosscheck program then provides states with data on potential duplicate registrations and potential double voters.\"\n\n. BRIEF OF FORMER ATTORNEYS OF THE CIVIL RIGHTS DIVISION OF THE UNITED STATES DEPARTMENT OF JUSTICE AS AMICI CURIAE IN SUPPORT OF PETITIONER, filed March 9, 2017, at the US Supreme Court. This brief filed by former Republican appointees at the Justice Department under President George W. Bush urges the Supreme Court to hear an appeal of a lower-court ruling that found Ohio had illegally purged inactive voters under the National Voter Registration Act of 1993. On pages 16\u201319, it describes the 2015 to 2016 activity of the Interstate Voter Registration Crosscheck Program, which tries to identify possible duplicate registrations and combat double voting. It notes that thirty states are participants, but only twenty-five states used it as of February 2016. It reported that Georgia had 540,245 \"potential voter registration duplications.\" In these states, including Ohio, North Carolina, Michigan, Virginia, Nevada, and Colorado, more than 5,935,000 voters were flagged as suspicious based on Crosscheck's imprecise matching criteria. <http:\/\/www.scotusblog.com\/wp-content\/uploads\/2017\/03\/16-980-cert-amicus-former-doj-attorneys.pdf>.\n\n. BRIEF OF GEORGIA AND 14 OTHER STATES AS AMICI CURIAE IN SUPPORT OF PETITIONER, filed March 10, at the US Supreme Court. This is the brief filed by Georgia and fourteen other states urging the Court to hear Ohio's appeal of a decision that found it had unconstitutionally purged 144,000 voters before the 2016 election. That ruling found the state did not follow the protocols in the National Voter Registration Act of 1993. Georgia and other red states said the law was ambiguous and they should be allowed to purge inactive voters. The Supreme Court will hear the case in its fall term. <http:\/\/www.scotusblog.com\/wp-content\/uploads\/2017\/03\/16-980-cert-amicus-Georgia.pdf>.\n\nCHAPTER 12: THE FINISH LINE\n\n. \"Florida halts purge of noncitizens from voter rolls,\" by Steve Bousquet and Amy Sherman, Tampa Bay Times, March 27, 2014. Florida secretary of state stops a 2014 mass purge after local election officials revolted about his inaccurate voter purge lists. <http:\/\/www.tampabay.com\/news\/politics\/elections\/florida-halts-purge-of-noncitizens-from-voter-rolls\/2172206>.\n\n. Same as 52: The Fight to Vote, by Michael Waldman, See pages 174\u20138 for Florida's massive 2000 voter purge.\n\n. \"The National Voter Registration Act Of 1993 (NVRA): Questions and Answers,\" The United States Department of Justice. <https:\/\/www.justice.gov\/crt\/national-voter-registration-act-1993-nvra>.\n\n. Same as 129: \"Florida: An Outlier in Denying Voting Rights,\" by Erika Wood, The Brennan Center for Justice at New York University School of Law, December 16, 2016.\n\n. Same as 52: The Fight to Vote, by Michael Waldman, See pages 174\u20138 for Florida'a massive 2000 voter purge.\n\n. \"This Election Is Being Rigged\u2014but Not by Democrats: The GOP's voter-suppression efforts are the real voter fraud,\" by Ari Berman, The Nation, October 17, 2016. This report cites of the results of a legal settlement in NAACP v. Florida where rerunning then secretary of state Katherine Harris's felon list with tighter criteria found twelve thousand voters were mistakenly purged in 2000. It also notes Trump's cry to \"Stop Crooked Hillary From Rigging This Election,\" the crowd's \"Lock her up\" chants, and Trump's baseless accusations that Democrats steal millions of votes. <https:\/\/www.thenation.com\/article\/this-election-is-being-rigged\/>.\n\n. Same as 168: \"Florida halts purge of noncitizens from voter rolls,\" by Steve Bousquet and Amy Sherman, Tampa Bay Times, March 27, 2014.\n\n. \"The GOP's Stealth War Against Voters: Will an anti-voter-fraud program designed by one of Trump's advisers deny tens of thousands their right to vote in November?\" by Greg Palast, Rolling Stone, August 24, 2016. <http:\/\/www.rollingstone.com\/politics\/features\/the-gops-stealth-war-against-voters-w435890>.\n\n. Same as 166: BRIEF OF FORMER ATTORNEYS OF THE CIVIL RIGHTS DIVISION OF THE UNITED STATES DEPARTMENT OF JUSTICE AS AMICI CURIAE IN SUPPORT OF PETITIONER, filed March 9, 2017, at the US Supreme Court. This brief was filed by six former Justice Department officials, including four who were appointed by George W. Bush and oversaw the DOJ's unsuccessful voter fraud prosecutions. They include: Colonel Karl S. \"Butch\" Bowers, Robert D. Popper, Bradley J. Schlozman, and Hans A. von Spakovsky. Pages 16\u201319 details the Interstate Voter Registration Crosscheck Program's 2016 activities, including twenty-five states' numbers of suspicious registrations totaling 5.9 million voters.\n\n. A Philip Randolph Institute, Northeast Ohio Coalition for the Homeless, Larry Harmon v. Jon Husted, Secretary of State, September 23, 2016 decision by the US Court of Appeals for the Sixth Circuit, File Name 16a0241p.06. This is the ruling that found Ohio had violated the National Voter Registration Act by purging inactive voters. The state appealed it to the US Supreme Court, which will hear it in its fall 2017 term.\n\n<http:\/\/www.scotusblog.com\/wp-content\/uploads\/2017\/02\/16-980-op-bel-6th-cir.pdf>.\n\n. Same as 84: \"The Supreme Court Could Make It Easier for States to Purge Voters: The Court has decided to hear a new case from Ohio, where Democrats were twice as likely as Republicans to be purged from the rolls in the state's largest counties,\" by Ari Berman, The Nation, May 30, 2017. This update on the Supreme Court agreeing to hear Ohio's appeal notes that the GOP-led purges were disproportionately in black and Democratic neighborhoods in its three largest cities: Cleveland, Columbus, and Cincinnati. \"At least 144,000 voters in Ohio's three largest counties, home to Cleveland, Columbus, and Cincinnati, were purged since the 2012 election, with voters in Democratic-leaning neighborhoods twice as likely to be removed as those in Republican-leaning ones, according to a Reuters analysis. These cities are heavily Democratic, with large minority populations, and there was a clear partisan and racial disparity to the state's voter purge.\"\n\n. \"Use it or lose it: Occasional Ohio voters may be shut out in November,\" by Andy Sullivan and Grant Smith, Reuters, June 2, 2016. The investigative report found the GOP secretary of state had disproportionately purged inactive voters in blue epicenters compared to nearby red regions. \"In Cleveland's Cuyahoga County, 5 percent of voters in neighborhoods that backed Obama by more than 60 percent in 2012 were purged last year due to inactivity, according to the Reuters analysis of the voter lists. In neighborhoods where Obama got less than 40 percent of the vote, 2.5 percent of registered voters were removed for that reason. In Franklin County, home to the state capital Columbus, 11 percent of voters in Democratic-leaning neighborhoods have been purged since 2012 due to inactivity. Only 6 percent of voters in Republican-leaning neighborhoods have been purged. The disparity is especially stark in Hamilton County, where affluent Republican suburbs ring Cincinnati, which has one of the highest child-poverty rates in the country. In the heavily African-American neighborhoods near downtown, more than 10 percent of registered voters have been removed due to inactivity since 2012. In suburban Indian Hill, only 4 percent have been purged due to inactivity.\" <http:\/\/www.reuters.com\/article\/us-usa-votingrights-ohio-insight-idUSKCN0YO19D>.\n\n. Same as 84: \"The Supreme Court Could Make It Easier for States to Purge Voters: The Court has decided to hear a new case from Ohio, where Democrats were twice as likely as Republicans to be purged from the rolls in the state's largest counties,\" by Ari Berman, The Nation, May 30, 2017. This report notes that 7,500 voters were put back on Ohio rolls after the Sixth Circuit ruled in September 2016.\n\n. \"Upgrading Democracy: Improving America's Elections by Modernizing States' Voter Registration Systems,\" by The Pew Center on The States\/Election Initiatives, November 2010. This is the report of Pew's Voter Registration Modernization Design Working Group. <http:\/\/www.pewtrusts.org\/~\/media\/legacy\/uploadedfiles\/pcs_assets\/2010\/upgradingdemocracyreportpdf.pdf>.\n\n. \"Who We Are,\" by EricStates.org\/. About the Electronic Registration Information Center, which in 2016 had twenty states participating. <http:\/\/ericstates.org\/whoweare>.\n\n. Same as 103: \"Why Republicans Can't Find the Big Voter Fraud Conspiracy: If the last federal investigation is any guide, the answer is simple: It probably doesn't exist,\" by Lisa Rab, Politico Magazine, April 2, 2017. This report notes Trump incorrectly citing a Pew study at an October 2016 rally to make accusations of hordes of dead voters casting ballots. It includes responses by David Becker, who led Pew's Voter Registration Modernization Design Working Group and oversaw a series of studies and reports to make the case for better registration data systems.\n\n. \"Sean Spicer wrongly uses Pew study to bolster claim that noncitizens vote in large numbers,\" by Lauren Carroll, politifact.com, January 25, 2017. This report notes Spicer's January 2017 comment mistakenly cites Pew's data to make false claims about noncitizen voting. <http:\/\/www.politifact.com\/truth-o-meter\/statements\/2017\/jan\/25\/sean-spicer\/sean-spicer-wrongly-uses-pew-study-bolster-claim-n\/>.\n\nAlso see: \"Inaccurate, Costly, and Inefficient: Evidence That America's Voter Registration System Needs an Upgrade,\" by The Pew Center on The States\/Election Initiatives, 2012. This is the Pew report that Spicer referenced: <http:\/\/www.pewtrusts.org\/~\/media\/legacy\/uploadedfiles\/pcs_assets\/2012\/pewupgradingvoterregistrationpdf.pdf>.\n\n. \"The Success of the Voter Fraud Myth,\" by The Editorial Board, The New York Times, September 20, 2016. This editorial contains the 2102 quote by Jim Greer on voter fraud being a marketing ploy. This editorial cited a Washington Post\/ABC News poll on voter fraud where only 1 percent of respondents correctly answered how often it occurs. <https:\/\/www.nytimes.com\/2016\/09\/20\/opinion\/the-success-of-the-voter-fraud-myth.html?mcubz=2>.\n\nAlso see: \"Poll: Nearly half of Americans say voter fraud occurs often,\" by Emily Guskin and Scott Clement, The Washington Post, September 15, 2016. <https:\/\/www.washingtonpost.com\/news\/the-fix\/wp\/2016\/09\/15\/poll-nearly-half-of-americans-say-voter-fraud-occurs-often\/?utm_term=.cd1ceb14af2b>.\n\n. Same as 185: \"The Success of the Voter Fraud Myth,\" by The Editorial Board, The New York Times, September 20, 2016.\n\n. \"Ohio Voter Fraud Billboards Target Minorities,\" by Lauren Feeney, BillMoyers.com, October 11, 2012. This report described voter intimidation in three swing state cities with large black populations: \"Billboards warning that voter fraud is a felony punishable with up to 3-and-a-half years and a $10,000 fine have been popping up in predominantly black neighborhoods in Cleveland, Cincinnati, and Milwaukee. The ads are a blatant attempt at voter intimidation, say community leaders in Cleveland, where the signs were first noticed.\" <http:\/\/billmoyers.com\/2012\/10\/11\/ohio-voter-fraud-billboards-target-minorities\/>.\n\n. \"Voter Integrity Project boot camp teaches how to challenge voters,\" by Jane Porter and Billy Ball, IndyWeek.com, September 18, 2013. This report contains an interview with DeLancy where he said he could not be affiliated with the Republican Party and work the voter caging operations that his organization was doing. The report also says DeLancy took his inspiration from True The Vote. <http:\/\/www.indyweek.com\/indyweek\/voter-integrity-project-boot-camp-teaches-how-to-challenge-voters\/Content?oid=3721401>.\n\n. \"Judge: North Carolina counties must restore voters removed in 'insane' process,\" by Tribune News Services\/Associated Press, November 4, 2016. This AP report recounts North Carolina restoring 6,700 voter registrations after a NAACP suit, where the \"Voter Integrity Project\" tried to purge thousands of voters. <http:\/\/www.chicagotribune.com\/news\/nationworld\/ct-north-carolina-voter-purge-ruling-20161104-story.html>.\n\n. North Carolina State Conference of the NAACP et al. v. North Carolina State Board of Elections et al, November 4, 2016 ruling by US District Court Judge Loretta C. Biggs ordering North Carolina counties to restore any voters purged within ninety days of the presidential election, noting that this violated the National Voter Registration Act of 1993.\n\nhttps:\/\/assets.documentcloud.org\/documents\/3214462\/Order.pdf.\n\n. \"Abbott signs voter ID, end of straight-party voting into law,\" by Chuck Lindell, Austin American-Statesman, June 1, 2017. This report notes how GOP Texas Gov. Greg Abbott signed into law a new slightly less restrictive voter ID bill and also eliminated the option of voting a straight party ticket starting in 2020. <http:\/\/www.statesman.com\/news\/abbott-signs-voter-end-straight-party-voting-into-law\/5vb95W3p0406a5mRgysEyL\/>.\n\nAlso see: \"Redistricting,\" TexasTribune.org. This index of ongoing articles includes a May report that \"Several Texas Republicans in Congress told the Tribune they want Gov. Greg Abbott to call a special session to redraw the state's congressional map. Yet a court filing from Texas Attorney General Ken Paxton Thursday made clear Abbott isn't interested.\" <https:\/\/www.texastribune.org\/tribpedia\/redistricting\/>.\n\n. North Carolina State Conference of the NAACP et al. v. Pat McCrory et al. This is the July 29, 2016 ruling by the United States Court of Appeals for the Fourth Circuit, No. 16\u20131468, written by Judge Diana Gribbon Motz. The ruling struck down five anti-voter laws in North Carolina, saying they targeted black voters with \"almost surgical\" precision. <https:\/\/www.ca4.uscourts.gov\/Opinions\/Published\/161468.P.pdf>.\n\nAlso see: \"Strict North Carolina Voter ID Law Thwarted After Supreme Court Rejects Case,\" by Adam Liptak and Michael Wines, The New York Times, May 15, 2017. This report notes the US Supreme Court rejected the state's appeal of the Fourth Circuit's ruling, saying it would like to rule on voting issues raised but the state's politics were too messy to take the case. https:\/\/www.nytimes.com\/2017\/05\/15\/us\/politics\/voter-id-laws-supreme-court-north-carolina.html?rref=collection%2Fbyline%2Fadam-liptak&action=click&contentCollection=undefined&region=stream&module=stream_unit&version=latest&contentPlacement=2&pgtype=collection.\n\n. Same as 192: North Carolina State Conference of the NAACP et al. v. Pat McCrory et al. This is the July 29, 2016 ruling by the United States Court of Appeals for the Fourth Circuit, No. 16\u20131468, written by Judge Diana Gribbon Motz.\n\n. \"Some States Review Election Systems for Signs of Intrusion,\" by Emery P. Dalesio and Geoff Mulvihill, Associated Press, June 8, 2017. The report notes a Florida company, VR Systems, whose voter registration database, which created precinct e-poll books in a handful of counties in eight states, was hacked by Russia. The AP wrote, \"North Carolina state elections board director Kim Westbrook Strach said her office had not been contacted by any federal officials about whether any of the 21 county election offices that use VR software were targeted. Still, her office was contacting county boards about potential breaches. The news of a reported Russian hacking attempt surprised Bill Brian, elections board chairman in Durham County, which experienced problems with VR Systems' electronic poll books on Election Day. The issue forced officials to abandon the system, issue paper ballots and extend voting hours, but officials there said that trouble did not appear to have been caused by hacking.\" https:\/\/www.apnews.com\/2dfaa4ca0053447d868eeb0dd4153c2a?utm_campaign=SocialFlow&utm_source=Twitter&utm_medium=AP_Politics.\n\n. \"Top-Secret NSA Report Details Russian Hacking Effort Days Before 2016 Election,\" by Mathew Cole, Richard Esposito, Sam Biddle and Ryan Grimm, TheIntercept.com, June 5, 2017. The report presents a classified May 2017 National Security Agency document that recounts how Russia intelligence agencies targeted private contractors servicing voter registration systems before the 2016 election. It notes how in Durham County, North Carolina, e-poll books failed at precincts \"causing chaos and long lines, which triggered election officials to switch to paper ballots and extend voting later into the evening.\" George McCue, BOE deputy director, said, \"There was some investigation there, essentially no evidence came out of it indicating there was any problem with the product,\" he said. \"It appears to be user errors at different points in the process, between the setup of the computers and the poll workers using them.\" <https:\/\/theintercept.com\/2017\/06\/05\/top-secret-nsa-report-details-russian-hacking-effort-days-before-2016-election\/>.\n\nCHAPTER 13: THE PLAYBOOK: DO EVERYTHING\n\n. \"Donald Trump is wrong. Rigging an election is almost impossible: This is how hard it would be to pull off widespread voter fraud,\" by Ari Berman, The Washington Post, August 5, 2016. This report notes when Trump started raising voter-fraud claims in the campaign. <https:\/\/www.washingtonpost.com\/posteverything\/wp\/2016\/08\/05\/donald-trump-is-wrong-rigging-an-election-is-almost-impossible\/?utm_term=.053d77ddbd53>.\n\n. \"Republicans limit early voting in Marion County, letting it bloom in suburbs,\" by Fatima Hussein, IndyStar.com, August 10, 2017. Excerpt: \"Most telling, Hamilton County saw a 63 percent increase in absentee voting from 2008 to 2016, while Marion County saw a 26 percent decline. Absentee ballots are used at early voting stations.\" <http:\/\/www.indystar.com\/story\/news\/2017\/08\/10\/silencing-vote-data-shows-unequal-barrier-indiana-polls\/435450001\/>.\n\n. \"New Voting Restrictions in Place for the 2016 Election,\" by the Brennan Center for Justice at New York University School of Law, November 2, 2016. This report notes new restrictions were in place in fourteen states, including: Alabama, Arizona, New Hampshire, Ohio, South Carolina, Texas, Virginia, and Wisconsin. <https:\/\/www.brennancenter.org\/sites\/default\/files\/analysis\/New_Restrictions_2016.pdf>.\n\n. \"Will North Carolina Lead the Way to a New South? Progressives like Reverend Barber and candidates like Deborah Ross are hoping to assemble a multiracial coalition to defeat a conservative backlash,\" by Joan Walsh, The Nation, November 7, 2016. This report notes black turnout in early voting had fallen by 16 percent because there were 158 fewer polling places in forty counties. <https:\/\/www.thenation.com\/article\/will-north-carolina-lead-the-way-to-a-new-south\/>.\n\n. \"NCGOP Sees Encouraging Early Voting, Obama\/Clinton Coalition Tired, Fail to Resonate in North Carolina,\" North Carolina Republican Party, November 6, 2016. This press release said black voting was down and white voting was up, compared to 2012. http:\/\/us2.campaign-archive1.com\/?u=f3100bc5464cbba2f472ddf2c&id=e4b9a8fb19.\n\n. Same as 192: North Carolina State Conference of the NAACP et al. v. Pat McCrory et al. This is the July 29, 2016 ruling by the United States Court of Appeals for the Fourth Circuit, No. 16\u20131468, written by Judge Diana Gribbon Motz, striking down North Carolina's omnibus anti-voter legislation following the US Supreme Court's 2013 gutting of the Voting Rights Act of 1965.\n\n. League of United Latin American Citizens v. Perry 548 US 399 (2006). This is the US District Court ruling cited by the Fourth Circuit opinion over North Carolina's 2013 anti-voter laws. The US Supreme Court affirmed the District Court's conclusions. <https:\/\/supreme.justia.com\/cases\/federal\/us\/548\/399\/opinion.html>.\n\n. Same as 192: North Carolina State Conference of the NAACP et al. v. Pat McCrory et al.\n\n. Same as 192: North Carolina State Conference of the NAACP et al. v. Pat McCrory et al. See footnote 7, page 47.\n\n. \"Don Yelton, GOP Precinct Chair, Delivers Most Baldly Racist Daily Show Interview of All Time,\" by Joe Coscarelli, New York magazine, October 24, 2013. This is the article cited in the footnote in the Fourth Circuit's ruling against North Carolina with Don Yelton's statement on \"lazy blacks\" and North Carolina's 2013 voter ID law. <http:\/\/nymag.com\/daily\/intelligencer\/2013\/10\/don-yelton-racist-daily-show-interview.html>.\n\n. Same as 192: North Carolina State Conference of the NAACP et al. v. Pat McCrory et al. See footnote 7, page 47.\n\n. Same as 196: \"Donald Trump is wrong. Rigging an election is almost impossible: This is how hard it would be to pull off widespread voter fraud,\" by Ari Berman, The Washington Post, August 5, 2016. The report quotes US District Judge James Peterson's Jury 2016 ruling that threw out Wisconsin's stricter voter ID law. \"A preoccupation with mostly phantom election fraud leads to real incidents of disenfranchisement, which undermine rather than enhance confidence in elections.\"\n\n. \"Texas Agrees to Soften Voter ID Law After Court Order,\" by Michael Wines, The New York Times, August 4, 2016. https:\/\/www.nytimes.com\/2016\/08\/04\/us\/texas-agrees-to-soften-voter-id-law-after-court-order.html?mcubz=2&_r=0.\n\n. Same as 199: \"Will North Carolina Lead the Way to a New South? Progressives like Reverend Barber and candidates like Deborah Ross are hoping to assemble a multiracial coalition to defeat a conservative backlash,\" by Joan Walsh, The Nation, November 7, 2016. This report cites the final 2016 campaign efforts by Democrats.\n\n. Same as 93: The Deceit of Voter Fraud, by Bob Hall and Isela Gutierrez, Democracy North Carolina, April 2017. This report goes into great detail how Gov. Pat McCrory's reelection campaign fabricated voter fraud charges against six hundred people in a failed attempt to overturn the results of his narrow loss in November 2016.\n\n. Ruthelle Frank, et al., v. Scott Walker, Case no. 11-CV-01128, League of United Latin American Citizens (LULAC) of Wisconsin, et al., v. Judge David Deininger, et al, Case no. 12-CV-00185, United States District Court Eastern District of Wisconsin, decision issued April 29, 2014. This ruling noted that Wisconsin's restrictive voter ID law prevented three hundred thousand people from voting. <https:\/\/www.scribd.com\/document\/221004483\/WiscVoterID-195-Decision>.\n\nAlso see: \"Texas, Wisconsin Photo ID Wins Could Help Nearly 1 Million Registered Voters,\" by the Brennan Center for Justice at New York University School of Law, October 10, 2014. <http:\/\/www.brennancenter.org\/press-release\/texas-wisconsin-photo-id-wins-could-help-nearly-1-million-registered-voters>. But by 2016, Wisconsin had adopted a new voter ID law. See: \"Voting Laws Roundup 2016,\" by the Brennan Center, April 18, 2016. <http:\/\/www.brennancenter.org\/analysis\/voting-laws-roundup-2016>.\n\n. Same as 13: Shattered: Inside Hillary Clinton's Doomed Campaign, by Jonathan Allen and Amie Parnes, Crown, 2016. <http:\/\/www.penguinrandomhouse.com\/books\/247611\/shattered-by-jonathan-allen-and-amie-parnes\/9780553447088\/> See pages 396\u201398.\n\nSECTION III: THE RECOUNTS\n\nCHAPTER 14: WHAT HAPPENED?\n\n. Same as 13: Shattered: Inside Hillary Clinton's Doomed Campaign, by Jonathan Allen and Amie Parnes, Crown, 2016. See pages 382\u2013394.\n\n. Same as 1: \"It's time to bust the myth: Most Trump voters were not working class,\" by Nicholas Carnes and Noam Lupu, Washington Post, June 5, 2017. Excerpt: \"Among people who said they voted for Trump in the general election, 35 percent had household incomes under $50,000 per year (the figure was also 35 percent among non-Hispanic whites), almost exactly the percentage in NBC's March 2016 survey. Trump's voters weren't overwhelmingly poor. In the general election, like the primary, about two thirds of Trump supporters came from the better-off half of the economy.\"\n\n. \"Elections: State Progress, Federal Train Wreck; State secretaries bask in smooth Election Day, joust in Washington's battles,\" by Miles Rapoport, The American Prospect, March 2, 2017. Rapoport is the former secretary of state in Connecticut, and president of both D\u0113mos and Common Cause. \"The chaos and conflict at the polls that was feared by many did not materialize... Two issues, however, were too fraught with partisan conflict to achieve any consensus on the part of the assembled secretaries of state: Russian hacking and calculated interference in the election, and the president's claim of massive voter fraud.\" <http:\/\/prospect.org\/article\/elections-state-progress-federal-train-wreck>.\n\n. \"Russian Cyber Hacks on US Electoral System Far Wider Than Previously Known,\" by Michael Riley and Jordan Robertson, Bloomberg.com, June 13, 2017. Excerpt: \"Russia's cyberattack on the US electoral system before Donald Trump's election was far more widespread than has been publicly revealed, including incursions into voter databases and software systems in almost twice as many states as previously reported... In all, the Russian hackers hit systems in a total of 39 states, one of them said.\" <https:\/\/www.bloomberg.com\/news\/articles\/2017-06-13\/russian-breach-of-39-states-threatens-future-u-s-elections>.\n\n. \"Exit Interview: Michigan Elections Director Christopher Thomas: Thomas looking forward to having time to actually think about elections,\" ElectionLine Weekly, June 29, 2017, <http:\/\/www.electionline.org\/index.php\/2017\/2233-electionlineweekly-june-29-2017>.\n\n. \"Donald Trump Warned Of A 'Rigged' Election, Was He Right?\" by Jonathan D. Simon, MintPress News, January 24, 2017, <https:\/\/www.mintpressnews.com\/donald-trump-warned-of-a-rigged-election-was-he-right\/224326\/>.\n\n. Same as 218: \"Donald Trump Warned Of A 'Rigged' Election, Was He Right?\" by Jonathan D. Simon, MintPress News, January 24, 2017.\n\n. \"Hacking The Diebold Machine,\" YouTube.com, August 2, 2007. In 2006, Princeton University's Ed Felten testified before Congress and demonstrated the hack, saying it was possible to \"silently transfer votes from one candidate to another,\" and that \"launching it requires access to a single voting machine for as little as one minute.\" <https:\/\/www.youtube.com\/watch?v=HBqGzgxcfAk>.\n\nFor full hearing, with Barbara Simons testimony, see: \"Electronic Voting Machines: Verification, Security, and Paper Trails,\" by Committee on House Administration, September 28, 2006, <https:\/\/archive.org\/details\/gov.house.cha.2006.09.28>.\n\n. \"Feds believe Russians hacked Florida election-systems vendor,\" by Evan Perez, Shimon Prokupecz, and Wesley Bruer, CNN.com, October 10, 2016. <http:\/\/www.cnn.com\/2016\/10\/12\/politics\/florida-election-hack>.\n\n. \"Inside The Recount: Jill Stein and a ragtag team of computer experts decided to take America's elections to court. Here's how it all went wrong,\" by Steve Friess, NewRepublic.com, Feb. 15, 2017, <https:\/\/newrepublic.com\/article\/140254\/inside-story-trump-clinton-stein-presidential-election-recount>.\n\n. Same as 222: \"Inside The Recount: Jill Stein and a ragtag team of computer experts decided to take America's elections to court. Here's how it all went wrong,\" by Steve Friess, NewRepublic.com, Feb. 15, 2017.\n\n. \"A Fair Election? Serious, Hard-to-Explain Questions Arise About Trump Vote Totals in 3 Key States: Voting rights advocates are scrambling to see if recounts are feasible in Wisconsin, Michigan and Pennsylvania,\" by Steven Rosenfeld, AlterNet.org, November 18, 2016, <http:\/\/www.alternet.org\/election-2016\/fair-election-serious-hard-explain-questions-arise-about-trump-vote-totals-3-key>.\n\n. Same as 222: \"Inside The Recount: Jill Stein and a ragtag team of computer experts decided to take America's elections to court. Here's how it all went wrong,\" by Steve Friess, NewRepublic.com, Feb. 15, 2017.\n\n. \"Experts Urge Clinton Campaign to Challenge Election Results in 3 Swing States,\" by Gabriel Sherman, New York magazine, November 22, 2016, <http:\/\/nymag.com\/daily\/intelligencer\/2016\/11\/activists-urge-hillary-clinton-to-challenge-election-results.html>.\n\n. January 26, 2017, email from Jill Stein, HQ@Jill2016.com, summarizing the recount process and next steps on the eve of Trump's inauguration.\n\n. Same as 227: January 26, 2017 Email from Jill Stein.\n\nCHAPTER 15: THE FINAL SWING STATES\n\n. \"What We Know and Don't Know about Election 'Hacks' in 2016,\" by David Becker, Center for Election Innovation and Research, June 6, 2016, <https:\/\/www.electioninnovation.org\/news\/2017\/6\/6\/what-we-know-and-dont-know-about-election-hacks-in-2016>.\n\n. \"Ex-NSA head suggests US also hacks political parties,\" by Joe Uchill, The Hill, October 18, 2016, <http:\/\/thehill.com\/policy\/cybersecurity\/301667-former-cia-and-nsa-head-suggests-us-also-hacks-political-parties>.\n\n. \"Cyberwar for Sale: After a maker of surveillance software was hacked, its leaked documents shed light on a shadowy global industry that has turned email theft into a terrifying\u2014and lucrative\u2014political weapon.\" by Mattathias Schwartz, The New York Times Magazine, January 8, 2017. <https:\/\/www.nytimes.com\/2017\/01\/04\/magazine\/cyberwar-for-sale.html?_r=0>.\n\n. \"The GOP's cyber election hit squad,\" by Steven Rosenfeld, April 22, 2007, FreePress.org, <http:\/\/freepress.org\/article\/gops-cyber-election-hit-squad>.\n\n. \"The Verifier - Polling Place Equipment - Current,\" by VerifiedVoting.org. This interactive page by Verified Voting allows viewers to zoom into states and see the voting machinery in every county in America. <https:\/\/www.verifiedvoting.org\/verifier\/#>.\n\n. Same as 227: January 26, 2017, email from Jill Stein.\n\n. \"1.7 million people in 33 states and D.C. cast a ballot without voting in the presidential race,\" by Philip Bump, Washington Post, December 14, 2016, <https:\/\/www.washingtonpost.com\/news\/the-fix\/wp\/2016\/12\/14\/1-7-million-people-in-33-states-and-dc-cast-a-ballot-without-voting-in-the-presidential-race\/?utm_term=.ad146cd2a5db>.\n\n. December 21, 2026, email from Jan BenDor, Michigan Election Reform Alliance's statewide coordinator, to author. \"As we explained in our press conference on Monday December 19], Michigan election officials sabotaged the Stein recount by lying about what the 1954 law actually says, and what the 1979 promulgated recount rules actually say.\" See: \"Michigan Election Reform Alliance News Conference,\" by Walter Sorg, [YouTube.com, December 20, 2016. <https:\/\/www.youtube.com\/watch?v=7qSmPymx9G0>.\n\n. \"Voter Suppression in the Mirror and Looking Forward: How much damage occurred in 2016, and what's in store for 2018 and beyond?\" by Miles Rapoport, The American Prospect, August 7, 2017. <http:\/\/prospect.org\/article\/voter-suppression-mirror-and-looking-forward>.\n\n. \"What 6 Top Election Experts Are Saying About the Next Big Step in the 2016 Recount,\" by Steven Rosenfeld, Alternet.org, November 29, 2016, <http:\/\/www.alternet.org\/election-2016\/what-6-top-election-experts-are-saying-about-next-big-step-2016-recount>.\n\n. \"The Real Story of the Recount,\" by Bob Fitrakis, Columbus Free Press, January 4, 2017, <http:\/\/columbusfreepress.com\/article\/real-story-recount>.\n\n. \"Don't Tread on My Vote,\" by Michelle Parker and Sharon Purvis, <http:\/\/mzpiterations.com>, July 2017. This website launches a campaign in Allegheny County to \"push a referendum for the November 7, 2017, ballot that would enact an ordinance that creates a Voting Process Review Commission to evaluate Allegheny County's voting system.\" <http:\/\/mzpiterations.com\/>.\n\n. December 23, 2016, letter from Jonathan S. Abady, lead attorney representing the Greens, to US Attorney General Loretta Lynch. <https:\/\/d3n8a8pro7vhmx.cloudfront.net\/jillstein\/pages\/27984\/attachments\/original\/1482528938\/FINAL_ECBA_letter_to_DOJ_re_election_investigation_(00270677x9CCC2).pdf?1482528938>.\n\n. Same as 241: Letter from Jonathan S. Abady, see page 7.\n\n. Same as 239: \"The Real Story of the Recount,\" by Bob Fitrakis, Columbus Free Press, January 4, 2017.\n\n. October 23, 2017, email from Justin Levitt, Professor of Law and Associate Dean for Research at Loyola Law School, Los Angeles, CA. See: \"Federal Prosecution of Election Offenses, Seventh Edition,\" Department of Justice, May 2007. This is the DOJ's 344-page manual for prosecutors in the Civil Rights Division's Voting Section.\n\n<https:\/\/www.justice.gov\/sites\/default\/files\/criminal\/legacy\/2013\/09\/30\/electbook-0507.pdf>.\n\n. \"On Tyranny: Lessons From The Twentieth Century,\" June 2017 lecture by Timothy Snyder at International Festival of Arts and Ideas, New Haven, Connecticut, posted on Youtube, June 17, 2017, <https:\/\/www.youtube.com\/watch?v=_4T9aGYoONU>.\n\n. Same as 229: \"What We Know and Don't Know about Election 'Hacks' in 2016,\" by David Becker, Center for Election Innovation and Research, June 6, 2016.\n\n. \"Inside the Trump Bunker, With Days to Go: Win or lose, the Republican candidate and his inner circle have built a direct marketing operation that could power a TV network\u2014or finish off the GOP,\" by Joshua Green and Sasha Issenberg, bloomberg.com, October 26, 2016, <https:\/\/www.bloomberg.com\/news\/articles\/2016-10-27\/inside-the-trump-bunker-with-12-days-to-go>.\n\n. \"Information Operations and Facebook,\" by Jen Weedon, William Nuland and Alex Stamos, Facebook, Inc., April 27, 2017. See page 11, section entitled, \"A Deeper Look: A Case Study of a Recent Election.\" <https:\/\/fbnewsroomus.files.wordpress.com\/2017\/04\/facebook-and-information-operations-v1.pdf>.\n\n. \"Connecting The Dots: Political Microtargeting and the Russia Investigation,\" by Kate Brannen, JustSecurity.org, May 18, 2017, <https:\/\/www.justsecurity.org\/41199\/connecting-dots-political-microtargeting-russia-investigation-cambridge-analytica\/> This report includes excerpts from an interview with Sen. Mark Warner, D-VA, the Senate Intelligence Committee vice-chair, where he said the committee was looking at whether Russians were targeting Clinton voters in swing states with fake news via their social media accounts.\n\nAlso see: \"Who Hacked the Election? Ad Tech Did. Through 'Fake News,' Identity Resolution and Hyper-Personalization,\" by Jonathan Albright, Medium.com, July 30, 2017, <https:\/\/medium.com\/@d1gi\/who-hacked-the-election-43d4019f705f>.\n\n. Same as 245: \"On Tyranny: Lessons From The Twentieth Century,\" June 2017 lecture by Timothy Snyder at International Festival of Arts and Ideas, New Haven, Connecticut, posted on Youtube, June 17, 2017.\n\n. Same as 230: \"Ex-NSA head suggests US also hacks political parties,\" by Joe Uchill, The Hill, October 18, 2016.\n\n. \"Were 2016 vote counts in Michigan and Wisconsin hacked? We double-checked,\" by Walter R. Mebane Jr. and Matthew Bernhard, June 6, 2017, <https:\/\/www.washingtonpost.com\/news\/monkey-cage\/wp\/2017\/06\/06\/were-2016-vote-counts-in-michigan-and-wisconsin-hacked-we-double-checked\/?utm_term=.4e769fa2eb9e>.\n\n. \"Watch: Hackers Demonstrate How to Crack Into Electronic Voting Machine in Minutes: Disturbing footage from the DEF CON 25 hacker convention,\" by Lulu Friesdat, AlterNet.org, July 29, 2017, <http:\/\/www.alternet.org\/investigations\/def-con-25-hackers-get-electronic-voting-machines-and-e-poll-books-minutes>.\n\n. Same as 252: \"Were 2016 vote counts in Michigan and Wisconsin hacked? We double-checked,\" by Walter R. Mebane Jr. and Matthew Bernhard, June 6, 2017.\n\n. Same as 237: \"Voter Suppression in the Mirror and Looking Forward: How much damage occurred in 2016, and what's in store for 2018 and beyond?\" by Miles Rapoport, The American Prospect, August 7, 2017.\n\n. \"Wisconsin Strict ID Law Discouraged Voters, Study Finds,\" by Michael Wines, New York Times, September 25, 2017.\n\n<https:\/\/www.nytimes.com\/2017\/09\/25\/us\/wisconsin-voters.html>.\n\nCHAPTER 16: THE ELECTORAL COLLEGE\n\n. \"Why I Will Not Cast My Electoral Vote for Donald Trump,\" by Christopher Suprun, New York Times, Dec. 6, 2016, https:\/\/www.nytimes.com\/2016\/12\/05\/opinion\/why-i-will-not-cast-my-electoral-vote-for-donald-trump.html?mcubz=2&_r=0.\n\n. December 29, 2017, letter to Congressional Black Caucus by Ruby Sales and Ann Billingsley (www.spirithouseproject.org) and Ann Massaro (Women and Allies).\n\n. \"At Least 50 Trump Electors Were Illegitimately Seated as Electoral College Members: More evidence surfaces as calls mount to challenge congressional ratification of Electoral College vote,\" by Steven Rosenfeld, AlterNet.org, January 4, 2017, <http:\/\/www.alternet.org\/election-2016\/least-50-trump-electors-were-illegitimately-seated-electoral-college-members>.\n\n. COUNTING ELECTORAL VOTES\u2014JOINT SESSION OF THE HOUSE AND SENATE HELD PURSUANT TO THE PROVISIONS OF SENATE CONCURRENT RESOLUTION 2, House of Representatives, January 6, 2017. This is the Congressional Record and transcript of the Electoral College ratification debate on January 6, 2017. <https:\/\/www.congress.gov\/congressional-record\/2017\/01\/06\/house-section\/article\/H185-8>.\n\n. PERMISSION TO PLACE IN THE RECORD A STATEMENT REGARDING THE JOINT SESSION OF ELECTION, House of Representatives, January 6, 2017. These are the extended remarks by Rep. Shiela Jackson Lee, D-TX, filed in the Congressional Record. <https:\/\/www.congress.gov\/congressional-record\/2017\/01\/06\/house-section\/article\/H190-3>.\n\nAFTERWORD\n\n. \"The Other Right-Wing Tidal Wave Sweeping America: Federal and State Preemption of Local Progressive Laws: Preemption allows corporations to boost their profits by suppressing local government power, community groups and citizens,\" by Don Hazen and Steven Rosenfeld, AlterNet.org, February 25, 2017, <http:\/\/www.alternet.org\/activism\/other-right-wing-tidal-wave-sweeping-america-federal-and-state-preemption-local-progressive>.\n\n. \"Recidivism Watch: Trump's claim that millions of people voted illegally,\" by Glenn Kessler, Washington Post, January 24, 2017, <https:\/\/www.washingtonpost.com\/news\/fact-checker\/wp\/2017\/01\/24\/recidivism-watch-trumps-claim-that-3-5-million-people-voted-illegally-in-the-election\/?utm_term=.bd3f388d4c36>.\n\n. \"Trump's voter fraud experts registered in 3 states,\" by Garance Burke, Associated Press, January 31, 2017, <https:\/\/apnews.com\/80497cfb5f054c9b8c9e0f8f5ca30a62>.\n\n. \"The Latest Voter Fraud Lie,\" by The Editorial Board, The New York Times, Feb. 13, 2017, <https:\/\/www.nytimes.com\/2017\/02\/13\/opinion\/the-latest-voter-fraud-lie.html?_r=0>.\n\n. \"Trump's Former Aide Concedes There Was No Voter Fraud In New Hampshire: Corey Lewandowski is not reading from the same script,\" by Sam Stein, The Huffington Post, February 20, 2017, <http:\/\/www.huffingtonpost.com\/entry\/trump-voter-fraud-corey-lewandowski_us_58ab2331e4b07602ad56d2f4>.\n\n. \"Trump's voter fraud claims undermine the democratic process and his presidency,\" by Dan Balz, The Washington Post, January 25, 2017, <https:\/\/www.washingtonpost.com\/politics\/trumps-voter-fraud-claims-undermine-the-voting-system-and-his-presidency\/2017\/01\/24\/a71d58ee-e288-11e6-a453-19ec4b3d09ba_story.html?utm_term=.2e1c1a730607>.\n\n. Jeff Sessions, Wikipedia biography. Excerpt: \"In 1985, Sessions prosecuted three African American community organizers in the Black Belt of Alabama, including Martin Luther King Jr.'s former aide Albert Turner, for voter fraud, alleging tampering with 14 absentee ballots. The prosecution stirred charges of selective prosecution of black voter registration. The defendants, known as the Marion Three, were acquitted of all charges by a jury after three hours of deliberation.\" <https:\/\/en.wikipedia.org\/wiki\/Jeff_Sessions#U.S._Attorney>.\n\n. \"Trump administration confirms plans to drop key claim against Texas Voter ID: The US Department of Justice confirmed Monday it plans to ditch its longstanding position that Texas lawmakers purposefully discriminated against minority voters by passing the nation's strictest voter identification law in 2011,\" by Jim Malewitz, The Texas Tribune, Feb. 27, 2107, <https:\/\/www.texastribune.org\/2017\/02\/27\/voter-id-case-trumps-team-drops-argument-against-texas\/>.\n\n. \"Obama's DOJ Fought Texas Voter ID Law. Trump's New Civil Rights Chief Offered Tips On Writing It: That's not a great sign for voting rights,\" by Ryan Reilly, The Huffington Post, January 25, 2017, <http:\/\/www.huffingtonpost.com\/entry\/civil-rights-division-trump-texas-voter-id_us_5887a75fe4b098c0bba6f72d>.\n\n. \"Restore lawful enforcement policies in the Civil Rights Division,\" March 28, 2017 letter to US Attorney General Jeff Sessions, from two-dozen Republicans including former George W. Bush administration DOJ staffers. <https:\/\/www.scribd.com\/document\/343306400\/Letter-to-AG-Sessions-on-Civil-Rights-Division> Also see: \"A Department of Justice, But For Whom?\" by Joshua Matz and Leah Litman, April 7, 2017, TakeCare.com, <https:\/\/takecareblog.com\/blog\/a-department-of-justice-but-for-whom>.\n\n. Same as 166: former Republican DOJ appointees' letter to Sessions. Also see: \"The Justice Department's Voter Fraud Scandal: Lessons,\" by Adam Gitlin and Wendy Weiser, The Brennan Center for Justice at New York University Law School, January 6, 2017. <https:\/\/www.brennancenter.org\/publication\/justice-departments-voter-fraud-scandal-lessons>.\n\n. \"Presidential Executive Order on the Establishment of Presidential Advisory Commission on Election Integrity,\" The White House, May 11, 2017, <https:\/\/www.whitehouse.gov\/the-press-office\/2017\/05\/11\/presidential-executive-order-establishment-presidential-advisory>.\n\n. Same as 166: former Republican DOJ appointees' letter to Sessions.\n\n. June 28, 2017, letter from Kris Kobach, Vice-Chair, Presidential Commission on Election Integrity, to state election directors. <https:\/\/twitter.com\/vanitaguptaCR\/status\/880479649817649152>.\n\n. Tweets by Michael McDonald, University of Florida US election turnout expert, on July 1, 2017. \"Make no mistake, this is a cynical, calculated ploy engineered by Kobach who knew some states could never respond... So when Kobach says states are 'hiding' he knew in advance some states couldn't share data. His request set states up so he can accuse them.\" <https:\/\/twitter.com\/ElectProject\/status\/881139025440124928>.\n\n. Tweet by Donald J. Trump, July 1, 2017, 6:07 AM. <https:\/\/twitter.com\/realDonaldTrump\/status\/881137079958241280>.\n\n. \"The Bogus Voter-Fraud Commission,\" by the Editorial Board, The New York Times, July 22, 2017, <https:\/\/www.nytimes.com\/2017\/07\/22\/opinion\/sunday\/the-bogus-voter-fraud-commission.html?_r=0>.\n\n. \"Presidential Advisory Commission on Election Integrity, Mission, Procedures and Topics for Consideration,\" July 19, 2017, meeting, broadcast by C-SPAN.org, https:\/\/www.c-span.org\/video\/?431521\u20133\/presidential-advisory-commission-election-integrity-mission-procedures-topics-consideration.\n\n. \"President Trump Appoints Country's Worst Vote Suppressor to His 'Election Integrity' Commission,\" by Rick Hasen, Election Law Blog, June 29, 2017. <http:\/\/electionlawblog.org\/?p=93444>.\n\n. June 28, 2017, letter from T. Christian Herren Jr., Chief, Voting Section, US Department of Justice, to state election directors to review \"voter list maintenance procedures in each state.\" <https:\/\/twitter.com\/JessicaHuseman\/status\/880499901179809792>.\n\n. Same as 272: \"The Justice Department's Voter Fraud Scandal: Lessons,\" by Adam Gitlin and Wendy Weiser, The Brennan Center for Justice at New York University Law School, January 6, 2017.\n\n. Husted v. A. Philip Randolph Institute, Supreme Court of the United States blog. \"Issue(s): Whether 52 U.S.C. \u00a7 20507 permits Ohio's list-maintenance process, which uses a registered voter's voter inactivity as a reason to send a confirmation notice to that voter under the National oter Registration Act of 1993 and the Help America Vote Act of 2002.\" <http:\/\/www.scotusblog.com\/case-files\/cases\/husted-v-philip-randolph-institute\/>.\n\n. \"Justice Dept. sides with Ohio's purge of inactive voters in case headed to Supreme Court,\" by Sari Horwitz, The Washington Post, August 8, 2017. Excerpt: \"The move is part of a broader campaign by the Trump administration to support restrictions on who is eligible to vote, a radical change in philosophy from the previous Justice Department, which sued a number of states over voting laws that it deemed discriminatory against minorities.\"\n\n\"<https:\/\/www.washingtonpost.com\/world\/national-security\/justice-department-reverses-position-to-allow-ohio-to-purge-inactive-voters-from-rolls\/2017\/08\/08\/e93c5116-7c35-11e7-9d08-b79f191668ed_story.html?utm_term=.ce60184c4704>.\n\n. Gill v. Whitford, Supreme Court of the United States blog. \"Issue(s): (1) Whether the district court violated Vieth v. Jubelirer when it held that it had the authority to entertain a statewide challenge to Wisconsin's redistricting plan, instead of requiring a district-by-district analysis; (2) whether the district court violated Vieth when it held that Wisconsin's redistricting plan was an impermissible partisan gerrymander, even though it was undisputed that the plan complies with traditional redistricting principles; (3) whether the district court violated Vieth by adopting a watered-down version of the partisan-gerrymandering test employed by the plurality in Davis v. Bandemer; (4) whether the defendants are entitled, at a minimum, to present additional evidence showing that they would have prevailed under the district court's test, which the court announced only after the record had closed; and (5) whether partisan-gerrymandering claims are justiciable.\" <http:\/\/www.scotusblog.com\/case-files\/cases\/gill-v-whitford\/>.\n\n. \"Ratf**ked: The True Story behind The Secret Plan to Steal America's Democracy,\" by David Daley. Printing with post-2016 epilogue available at: https:\/\/www.amazon.com\/gp\/aw\/d\/1631493213\/ref=mp_s_a_1_1?ie=UTF8&qid=1499989700&sr=8\u20131&pi=AC_SX236_SY340_QL65&keywords=ratfuked&dpPl=1&dpID=41vTVqsxuXL&ref=plSrch#featureBulletsAndDetailBullets_secondary_view_div_1500080267818.\n\n. \"Democrats Better Focus on the Races That Matter in '18 or We Will Have Another Decade of Right-Wing Extremists in Charge: Retaking the House in 2018 slows the GOP. Only governors can stop GOP gerrymanders and political monopolies through 2031,\" by Steven Rosenfeld, AlterNet.org, July 14, 2017, <http:\/\/www.alternet.org\/election-03918\/democrats-who-dont-want-live-republican-extremists-another-decade-need-focus-2018>.\n\n. Same as 262: \"The Other Right-Wing Tidal Wave Sweeping America: Federal and State Preemption of Local Progressive Laws: Preemption allows corporations to boost their profits by suppressing local government power, community groups and citizens,\" by Don Hazen and Steven Rosenfeld, AlterNet.org, February 25, 2017.\n\n. \"ALEC's Scary Plan For Electing Your Senators: The radical right is intent on destroying democracy as we know it,\" by David Daley, Alternet.org, July 22, 2017, <http:\/\/www.alternet.org\/alecs-scary-plan-electing-your-senators>.\n\n. \"We're Supposed to Be a Democracy, But Half the GOP Is OK with Postponing 2020's Election: A party full of delusions exposes the dark underbelly of America,\" by Steven Rosenfeld, AlterNet.org, August 10, 2017.\n\n. \"Voting Machine Digital Ballot Images Could Let Public Recount Elections, But Many Locales Aren't Saving or Sharing This Data: Publicly verified elections are key to de-corrupting our democracy,\" by John Brakey, AlterNet, June 8, 2017, <http:\/\/www.alternet.org\/activism\/voting-machine-digital-ballot-images-could-let-public-recount-elections>.\n\n. \"Citizen Scientists Comb Images To Find An 'Overexcited Planet,'\" by Joe Palca, National Public Radio, June 1, 2017, <http:\/\/www.npr.org\/2017\/06\/01\/530766624\/citizen-scientists-comb-images-to-find-an-overexcited-planet>.\nINDEX\n\nA\n\nAbady, Jonathan S., \u2013, n\n\nAbbott, Greg, \u2013n\n\n\"Abbott signs voter ID, end of straight-party voting into law\" (Lindell), \u2013n\n\nABC News, \"Abbott signs voter ID, end of straight-party voting into law\" (Lindell) June 1, 2017, n\n\n\"About Section 5 of the Voting Rights Act,\" n\n\nAbrams, Stacey, \u2013\n\nAbrasion, Alana, n\n\nAccuVote TSX,\n\nACORN (Association of Community Organizers for Reform Now), ,\n\n\"ACORN Accusations: McCain makes exaggerated claims of 'voter fraud.' Obama soft-pedals his connections\" (Henig), n\n\n\"Additional Data Could Help State and Local Elections Officials Maintain Accurate Voter Registration Lists,\" n\n\nAdvisory Commission on Election Integrity (Trump), , , \u2013, , n, n\n\nAdvisory Legal Opinions,\n\n\"Advocate for Tough Voting Rules to Steer Trump's Elections Commission\" (Wines and Bosman), n\n\naffirmative action and gerrymandering, \u2013\n\nAlabama district maps, \u2013\n\nAlabama Republican Party,\n\nAlabama voter ID laws, \u2013\n\n\"Alabama's GOP approves new maps; Dems vow repeat court fight: Alabama's overwhelmingly Republican legislature has sent a new legislative map to the governor after a rancorous session featuring a racially charged email comparing lawmakers to monkeys.\" (Chandler), n\n\nAlbright, Jonathan, n\n\nALEC (American Legislative Exchange Council),\n\n\"ALEC Is Talking About Changing the Way Senators Are Elected and Taking Away Your Vote: A proposed resolution advocates for overturning the 17th Amendment so Republican-controlled state legislatures could pick senators\" (Nichols), n\n\n\"ALEC's Scary Plan For Electing Your Senators: The radical right is intent on destroying democracy as we know it\" (Daley), n\n\nAllen, Jonathan, , , n\n\nAlterNet.org, xiv\n\n\"ALEC's Scary Plan For Electing Your Senators: The radical right is intent on destroying democracy as we know it\" (Daley) July 22, 2017, n\n\n\"Democrats Better Focus on the Races That Matter in '18 or We Will Have Another Decade of Right-Wing Extremists in Charge: Retaking the House in 2018 slows the GOP. Only governors can stop GOP gerrymanders and political monopolies through 2031\" (Rosenfeld) July 14, 2017, n\n\n\"Epic Fight Over Florida's Political Future Looms As Constitutional Amendment to Restore Felon Voting Rights Clears Hurdle\" (Friesdat) April 27, 2017, \u2013n\n\n\"A Fair Election? Serious, Hard-to-Explain Questions Arise About Trump Vote Totals in 3 Key States: Voting rights advocates are scrambling to see if recounts are feasible in Wisconsin, Michigan and Pennsylvania\" (Rosenfeld) Nov. 18, 2016, n\n\n\"A Harsh Dose of Electoral Reality: Democrats Have Uphill Battle in '18 and Need to Elect Governors to Fight Gerrymandered GOP Monopoly\" (Rosenfeld) April 7, 2017, n\n\n\"At Least 50 Trump Electors Were Illegitimately Seated as Electoral College Members: More evidence surfaces as calls mount to challenge congressional ratification of Electoral College vote\" (Rosenfeld) Jan. 4, 2017, \u2013n\n\n\"New Video: Watch Wisconsin Election Officials Reject Hand Counts After Electronic Scanners Make Big Mistake: 'I am just stunned at how inaccurate the whole thing is,'\" (Friesdat) Dec. 16, 2016, n\n\n\"The Other Right-Wing Tidal Wave Sweeping America: Federal and State Preemption of Local Progressive Laws: Preemption allows corporations to boost their profits by suppressing local government power, community groups and citizens\" (Hazen and Rosenfeld) Feb. 25, 2017, n\n\nvote discrepencies, post 2016 election,\n\n\"Voting Machine Digital Ballot Images Could Let Public Recount Elections, But Many Locales Aren't Saving or Sharing This Data: Publicly verified elections are key to de-corrupting our democracy\" (Brakey) June 8, 2017, n\n\n\"Watch: Hackers Demonstrate How to Crack Into Electronic Voting Machine in Minutes: Disturbing footage from the DEF CON 25 hacker convention\" (Friesdat) July 29, 2017, n\n\n\"We're Supposed to Be a Democracy, But Half the GOP Is OK with Postponing 2020's Election: A party full of delusions exposes the dark underbelly of America\" (Rosenfeld) Aug. 10, 2017, n\n\n\"What 6 Top Election Experts Are Saying About the Next Big Step in the 2016 Recount\" (Rosenfeld) Nov. 29, 2016, n\n\nAltlanta Journal-Constitution, \"Appeals court nixes proof of citizenship for voting in Georgia\" (Torres) Sept. 11, 2016, n\n\nAmerican Center for Voting Rights, \u2013\n\nAmerican democracy, threats to,\n\nAmerican Independent Party (AIP), \u2013, n\n\nAmerican Legislative Exchange Council, \u2013\n\nThe American Presidency Project, \"Voter Turnout in Presidential Elections: 1828\u20132012\" (American Presidency Project), n\n\nThe American Prospect\n\n\"From a Contentious Election to a Stronger Democracy: Strengthening democracy is the key to all other reforms\" (Rapoport) Oct. 6, 2016, n\n\n\"Elections: State Progress, Federal Train Wreck; State secretaries bask in smooth Election Day, joust in Washington's battles\" (Rapoport) March 2, 2017, n\n\n\"Voter Suppression in the Mirror and Looking Forward: How much damage occurred in 2016, and what's in store for 2018 and beyond?\" (Rapoport) Aug. 7, 2017, n\n\nThe American Slave Coast: A History of the Slave-Breeding Industry (Sublette), n\n\nAmericans Take Action,\n\nantidemocratic state party rules, \u2013\n\nanti-Sanders biases, xxii. see also Clinton, Hillary; Democratic Party; Sanders, Bernie\n\nanti-voter barriers, xvii, xxiv, , ,\n\nanti-voter laws, , , ,\n\n\"AP count: Clinton has delegates to win Democratic nomination\" (Yen, Ohlemacher, Lerer, and Lucey), n\n\n\"Appeals court nixes proof of citizenship for voting in Georgia\" (Torres), n\n\n\"Are Felons More Likely to Vote for Democrats over Republicans?\" (ProCon.org), n\n\n\"Arizona Taxpayer and Citizen Protection, Proposition 200 (2004),\" n\n\nArizona v. Inter Tribal Council of Arizona, Inc. (Amicus Brief), \u2013n\n\nArizona voting barriers,\n\n\"Arkansas governor signs bill to reinstate voter ID law\" (Demillo), \u2013n\n\nArpaio, Joe,\n\nArticle Five constitutional convention,\n\nAssociated Press (AP), \u2013, ,\n\n\"Alabama's GOP approves new maps; Dems vow repeat court fight: Alabama's overwhelmingly Republican legislature has sent a new legislative map to the governor after a rancorous session featuring a racially charged email comparing lawmakers to monkeys.\" (Chandler) May 20, 2017, n\n\n\"AP count: Clinton has delegates to win Democratic nomination\" (Yen, Ohlemacher, Lerer, and Lucey) June 7, 2016, n\n\n\"Arkansas governor signs bill to reinstate voter ID law\" (Demillo) March 24, 2017, \u2013n\n\n\"Revised Redistricting Plans Face Strong Public Opposition: GOP lawmakers on a key Senate committee approved a revision Friday of controversial redistricting plans, but Democrats and others accused Republicans of lacking transparency\" (Kaplan) March 24, 2017, n\n\n\"Revised Redistricting Plans Face Strong Public Opposition: GOP lawmakers on a key Senate committee approved a revision Friday of controversial redistricting plans, but Democrats and others accused Republicans of lacking transparency\" (Kaplan) March 24, 2017, \u2013n\n\n\"Senate Approves Redistricting Plan; Black Democrats Object: The Alabama Senate approved new legislative districts over the objections of black Democrats\" (Chandler) May 4, 2017, \u2013n\n\n\"Some States Review Election Systems for Signs of Intrusion\" (Dalesio and Mulvihill) June 8, 2017, \u2013n\n\n\"Trump's voter fraud experts registered in 3 states\" (Burke) Jan. 31, 2017, n\n\nAssociation for Computing Machinery,\n\nThe Atlanta Journal-Constitution,\n\n\"Appeals court nixes proof of citizenship for voting in Georgia\" (Torres) Sept. 11, 2016, n\n\n\"Georgia judge sides with state in 'missing' voters case\" (Torres) Oct. 29, 2014, n\n\n\"New Georgia Project releases financial records, renews voter push\" (Torres) Feb. 11, 2016, n\n\nThe Atlanta-Journal Constitution,\n\nAustin American-Statesman, \"Abbott signs voter ID, end of straight-party voting into law\" (Lindell) June 1, 2017, \u2013n\n\nB\n\nbad policy umbrella,\n\nBailey-Rihn, Valerie,\n\nBaker, Charlie,\n\nBaker v. Carr, , n\n\nBall, Billy, n\n\nBallot Access News,\n\nballotpedia.org, n, n, n\n\nBalz, Dan, , n\n\nBarber, William, II, n\n\nBarnes, Robert, \u2013n\n\nbarriers to voting. See anti-voter barriers\n\nBartlett, Gary,\n\nBecker, David, n\n\nBenDor, Jan, , n\n\nBerman, Ari, \u2013, \u2013, , n, n, \u2013n, \u2013n, n\n\nBernhard, Matthew, \u2013\n\n\"Bernie 2016\" US Federal Election Commisssion, n\n\nBerniecrats, ,\n\nBiddle, Sam, n\n\nBiden, Joe, , \u2013\n\nBiggs, Loretta C., , n\n\nBillingsley, Ann, n\n\nBillMoyers.com, \"Ohio Voter Fraud Billboards Target Minorities\" (Feeney) Oct. 11, 2012, \u2013n\n\nBiskupic, Joan, n\n\nBlack Belt counties, Alabama, , n\n\nBlack Caucus,\n\nBlack Caucus, letter to, n\n\nBlackwell, J. Kenneth, ,\n\nBlanton, Dana, n\n\nBloomberg.com\n\n\"Inside the Trump Bunker, With Days to Go: Win or lose, the Republican candidate and his inner circle have built a direct marketing operation that could power a TV network\u2014or finish off the GOP\" (Green and Issenberg) April 27, 2017, n\n\n\"Russian Cyber Hacks on US Electoral System Far Wider Than Previously Known\" (Riley and Robertson) June 13, 2017, n\n\nBlythe, Anne, \u2013n\n\nBoard of Elections (BOE), King's County, \u2013\n\nBondi, Pam,\n\nBonifaz, John, \u2013,\n\nBosman, Julie, n\n\nBousquet, Steve, n\n\nBoxer, Barbara, ,\n\nBrakey, John, xxi\u2013xxii, , , n\n\nBrannen, Kate, n\n\nBrazile, Donna, , n\n\nBreitbart.com, , \u2013\n\nBrennan Center for Justice, \u2013\n\nArizona v. Inter Tribal Council of Arizona, Inc. (Amicus Brief) June 17, 2013, \u2013n\n\n\"Debunking the Voter Fraud Myth,\" n\n\n\"Election 2012: Voting Laws Roundup,\" n\n\nex-felons, voting rights,\n\n\"Florida: An Outlier in Denying Voting Rights\" (Wood) Dec. 16, 2017, \u2013n\n\nKobach et al. v. The United States Election Assistance Commission June 29, 2015, \u2013n\n\nLeague of Women Voters v. Newby June 1, 2017, n\n\n\"New Voting Restrictions in America,\" n\n\n\"New Voting Restrictions in Place for the 2016 Election\" Nov. 2, 2016, \u2013n\n\n\"Noncitizen Voting: The Missing Millions\" (Famighetti, Keith, Perez) May 5, 2017, n\n\nsuppressive voter laws,\n\n\"Texas, Wisconsin Photo ID Wins Could Help Nearly 1 Million Registered Voters\" Oct. 10, 2014, \u2013n\n\nTexas NAACP v. Steen (consolidate with Veasey v. Abbot), \u2013n\n\n\"Uncovering Kris Kobach's Anti-Voting History\" (Lopez and Clark) May 11, 2017, n\n\n\"Voting Law Changes: Election Update\" (Weiser and Kasden), Oct. 28, 2012, n\n\n\"Voting Laws Roundup 2013,\" n\n\nBrennan Center for Justice, May 16, 2017, n\n\nBrennan Law Review, \"Voting Laws Roundup 2017,\" n\n\nBrian, Bill, \u2013n\n\nBRIEF OF FORMER ATTORNEYS OF THE CIVIL RIGHTS DIVISION OF THE UNITED STATES DEPARTMENT OF JUSTICE AS AMICI CURIAE IN SUPPORT OF PETITIONER, n\n\nBRIEF OF GEORGIA AND 14 OTHER STATES AS AMICI CURIAE IN SUPPORT OF PETITIONER, \u2013n\n\nBrown, Jerry, ,\n\nBruer, Wesley, n\n\nBuchanan, Pat,\n\nBump, Phillip, , n, \u2013n\n\nBurke, Garance, n\n\nBush, George W., xiii\u2013xiv, xx, , \u2013, , n, n\n\nBush, Jeb,\n\nC\n\ncaging, \u2013\n\nCalifornia 2016 primaries, \u2013\n\nCalifornia Democratic Party,\n\n\"California Democratic primary, 2016,\" n\n\nCampbell, Tracy, , n\n\nCarnes, Nicholas, n\n\n\"Carol Wilding et al vs. DNC Services Corporation and Deborah Wasserman Schultz,\" \u2013n\n\nCarpenter, Tim, n\n\nCarroll, Lauren, n\n\nCarter, Jimmy,\n\nCBS-TV, xiv\n\nCenter for Computer Security and Society, University of Michigan,\n\nCenter for Election Innovation and Research, \"What We Know and Don't Know about Election 'Hacks' in 2016\" (Becker) June 6, 2016, n\n\nChandler, Kim, \u2013n, n\n\nThe Chicago Tribune,\n\n\"Supreme Court: Virginia redistricting must be reexamined for racial bias\" (Barnes) March 1, 2017, \u2013n\n\n\"Child of the Enlightenment\" (Wills), n\n\nChinese hackers,\n\n\"Citizen Scientists Comb Images To Find An 'Overexcited Planet\" (Palca), n\n\nClapper, James,\n\nClark, Jennifer L., n\n\nClayton, Ryan,\n\nclimate-change and gerrymandering, \u2013\n\nClinton, Bill, ,\n\nClinton, Hillary,\n\nand 2016 election, xiv, xvi, xx, xxii, xxiii, \u2013\n\nCalifornia 2016 primaries, \u2013\n\nchallenges of election results, n\n\nemails of and Comey investigation, \u2013\n\nIowa Caucuses, 2016 election, \u2013\n\nMarch-April primaries, \u2013\n\nmore votes than Trump, 2016,\n\nNevada Caucuses, 2016, \u2013, n\n\nand political microtargeting, n\n\nrecieves more popular votes,\n\nand REDMAP targeting, \u2013\n\nresponse to Trump win,\n\nand Russian hacking, \u2013\n\nTexas Caucuses, 2008, \u2013\n\nvictory in Iowa caucus, n\n\nClinton-Obama Race, 2008, \u2013\n\n\"Closed primary,\" n\n\nCNN.com, \"Feds believe Russians hacked Florida election-systems vendor\" (Perez, Prokupecz, and Bruer) Oct. 10, 2016, n\n\nCole, Matthew, n\n\nColumbus Free Press, \u2013\n\n\"The Real Story of the Recount\" (Fitrakis) Jan. 4, 2017, n\n\nComedy Central,\n\nComey, James, xxi, , , , \u2013\n\nCommittee on House Administration, 2004, \u2013\n\n\"Connecting The Dots: Political Microtargeting and the Russia Investigation\" (Brannen), n\n\nConnors, Marty, , n\n\n\"From a Contentious Election to a Stronger Democracy: Strengthening democracy is the key to all other reforms\" (Rapoport), n\n\nThe Cook Political Report, October 17, 2013, \"October House Overview: GOP Risk Factors\" (Wasserman), \u2013n\n\nCooper v. Harris, n\n\nCooperative Congressional Election Study, \u2013\n\nCortes, Pedro,\n\nCoscarelli, Joe, \u2013n\n\nCOUNTING ELECTORAL VOTES\u2014JOINT SESSION OF THE HOUSE AND SENATE HELD PURSUANT TO THE PROVISIONS OF SENATE CONCURRENT RESOLUTION 2, Jan 6, 2017 House of Representatives, n\n\n\"cracking\" and \"packing\" voters,\n\n\"Critics See Efforts by Counties and Towns to Purge Minority Voters From Rolls\" (Wines), n\n\nC-SPAN.org, \"Presidential Advisory Commission on Election Integrity, Mission, Procedures and Topics for Consideration\" July 19, 2017, n\n\nCulinary Workers Union, \u2013\n\nCunningham, Vinson, n\n\n\"Cyberwar for Sale: After a maker of surveillance software was hacked, its leaked documents shed light on a shadowy global industry that has turned email theft into a terrifying\u2014and lucrative\u2014political weapon.\" (Schwartz), n\n\nD\n\nThe Daily Show,\n\nDalesio, Emery P, \u2013n\n\nDaley, David, , , \u2013, \u2013, , n, n, n\n\nDaley, Richard,\n\nDatabase Technologies, \u2013\n\ndatabases and voter purging, xvii\u2013xviii\n\nDavis, Jefferson,\n\nDavis v. Bandemer, \u2013n\n\n\"Debbie Wasserman Schultz: Give the voter file back to Bernie Sanders' campaign\" (Car), n\n\nDebs, Eugene,\n\n\"Debunking the Voter Fraud Myth,\" n\n\nThe Deceit of Voter Fraud (Hall and Gutierrez), n\n\nDEF CON25,\n\nDeLancy, Jay, \u2013\n\nDeLaney, John, n\n\nDeliver The Vote: A History of Election Fraud, An American Political Tradition\u20141742\u20132004 (Campbell), n\n\nDeliver The Vote: A History of Election Fraud as an American Political Tradition\u20141742\u20132004 (Campbell),\n\nDeMichele, Thomas, n\n\nDemillow, Andrew, \u2013n\n\nDemocracy Diminished: State and Local Threats to Voting Post-Shelby County, Alabama v. Holder, n\n\n\"Democracy Lost: A Report on the Fatally Flawed 2016 Democratic Primaries,\" \u2013n\n\nDemocracy Springs,\n\nDemocratic Governors Association,\n\nDemocratic National Committee, email thefts,\n\nDemocratic National Convention,\n\n\"Democratic National Convention, 2016\" BallotPedia.org., n\n\nDemocratic National Convention (DNC), xxii, \u2013, , n, n\n\nand Bernie Sanders, \u2013\n\nIowa Caucuses, 2016 election, \u2013\n\nDemocratic Party, xix, xxi, xxii,\n\nand Bernie Sanders, \u2013\n\nand civil rights, \u2013\n\nintention to retake the House, \u2013\n\nIowa Caucuses, 2016 election, \u2013\n\nneed to step up,\n\nNew York, April 2016 primary, \u2013\n\n\"now I'm inside, now I'm outside,\"\n\nresponse to Trump win, ,\n\nvote rigging, xxiii\u2013xxiv\n\n\"Democrats Better Focus on the Races That Matter in '18 or We Will Have Another Decade of Right-Wing Extremists in Charge: Retaking the House in 2018 slows the GOP. Only governors can stop GOP gerrymanders and political monopolies through 2031\" (Rosenfeld), n\n\nDemos and Common Cause,\n\ndemos.org, \"Voter Suppression Works\" (McElwee), n\n\nDepartment of Corrections,\n\nDepartment of Homeland Security (DHS),\n\nDepartment of Justice, , n\n\nchallenge of voter ID laws, \u2013\n\nDes Moines Register, xxii, \u2013\n\nDetroit voter discrepencies, \u2013\n\nDetzner, Ken, \u2013\n\nDiamond, Jeremy, , n\n\ndisenfranchisement of voters, xxv\n\n\"DNC Chair Says Superdelegates Exist to Protect Party Leaders,\" n\n\n\"Do voter identification laws suppress minority voting? Yes. We did the research.\" (Hajnal, Lajevardi, Nielson), n\n\nDOJ Voting Section,\n\n\"Don Yelton, GOP Precinct Chair, Delivers Most Baldly Racist Daily Show Interview of All Time\" (Coscarelli), \u2013n\n\n\"Donald Trump is wrong. Rigging an election is almost impossible: This is how hard it would be to pull off widespread voter fraud\" (Berman), n\n\n\"Donald Trump Warned Of A 'Rigged' Election, Was He Right?\" (Simon), n\n\n\"Don't Call It a Bern-Back:.\" (LaCapria), n\n\n\"Don't Tread on My Vote\" (Parker and Purvis), n\n\nE\n\nearly voting days, xvii, , n\n\nearly voting, Election day, xvii\n\n\"Editorial: Something smells in the Democratic Party\" The Des Moines Register, n\n\n\"Election 2012: Voting Laws Roundup\" (Brennan Center for Justice), n\n\nElection Day registration, xix, n\n\nelection district lines, \u2013\n\nelection impediments,\n\nElection Justice USA, Aug. 2016, \u2013n\n\nElection Law Blog, \"President Trump Appoints Country's Worst Vote Suppressor to His 'Election Integrity' Commission\" (Hasen) June 29, 2017, n\n\nelection voiding, call for, \u2013\n\nElectionLine Weekly, \"Exit Interview: Michigan Elections Director Christopher Thomas: Thomas looking forward to having time to actually think about elections,\" \u2013n\n\n\"Elections: State Progress, Federal Train Wreck; State secretaries bask in smooth Election Day, joust in Washington's battles\" (Rapoport), n\n\nElectoral College\n\nand congressional delegation size,\n\ncongressional ratification of, n\n\nand efforts to stop Trump, \u2013\n\nElectoral vote vs. exit polls,\n\nand gerrymandering,\n\ngerrymandering threat to,\n\nillegimately seat electorates, \u2013n\n\nand ratification of the 2004 vote,\n\nSuprun won't vote for Trump, n\n\nand Trump victory, xiv,\n\nElectoral College convention, xiv\n\nelectoral mugging,\n\nThe Electoral Vote Objection Package,\n\nelectoral war of attrition, \u2013\n\n\"Electronic Voting Machines: Verification, Security, and Paper Trails,\" n\n\nEllison, Keith, n\n\ne-mail theft, n\n\nemail theft, ,\n\nEMILY's list, \u2013\n\nEnten, Harry, , \u2013n\n\n\"Epic Fight Over Florida's Political Future Looms As Constitutional Amendment to Restore Felon Voting Rights Clears Hurdle\" (Friesdat), \u2013n\n\ne-poll books, , \u2013n, n\n\nERIC (Electronic Registration Information System),\n\nEricStates.org, n\n\nES&S model DS200, ,\n\nEsposito, Richard, n\n\nEugene V. Debs, n\n\nE-Verify, \u2013n\n\nex-felons, voting rights, , \u2013, , , \u2013, \u2013n, n, n\n\n\"Exit Interview: Michigan Elections Director Christopher Thomas: Thomas looking forward to having time to actually think about elections,\" \u2013n\n\nexit poll\/voter discrepancies, \u2013, \u2013\n\n\"Ex-NSA head suggests US also hacks political parties\" (Uchill), n\n\n\"Experts Urge Clinton Campaign to Challenge Election Results in 3 Swing States\" (Sherman), n\n\nF\n\nFacebook, Inc., \u2013\n\n\"Information Operations and Facebook\" (Weedon, Nuland and Stamos) April 27, 2017, n\n\nFacebook posts, targeted, \u2013\n\nFactMyth.com, May 12, 2016, n\n\n\"A Fair Election? Serious, Hard-to-Explain Questions Arise About Trump Vote Totals in 3 Key States: Voting rights advocates are scrambling to see if recounts are feasible in Wisconsin, Michigan and Pennsylvania\" (Rosenfeld), n\n\nFamighetti, Christopher, n\n\n\"Feds believe Russians hacked Florida election-systems vendor\" (Perez, Prokupecz, and Bruer), n\n\nFeeney, Lauren, \u2013n\n\nFelten, Ed, n\n\nFernandez, Manny, n\n\nThe Fight to Vote (Waldman), , , \u2013, n\n\nFindLaw.com, n\n\nFitrakis, Robert J., \u2013, n, n\n\nFiveThirtyEight.com,\n\n\"A Very Early Look At The Battle For The House In 2018: Donald Trump is unpopular enough that Republicans could lose the House, but there's a lot of uncertainty.\" (Enten), \u2013n\n\nFlorida, 2014 voter purge, \u2013,\n\n\"Florida: An Outlier in Denying Voting Rights\" (Wood), \u2013n\n\n\"Florida halts purge of noncitizens from voter rolls\" (Bousquet and Sherman), n\n\n\"The 4 Most Damaging Emails From the DNC WikiLeaks Dump\" (Abrasion and Walsh), n\n\nFox News, xxiii, , , \u2013\n\nFox News poll, n\n\nFrankel, Judy, \u2013n\n\nFreePress.org, \"The GOP's cyber election hit squad\" (Rosenfeld) April 22, 2007, n\n\nFreisdat, Lulu,\n\nFriesdat, Lulu, xiv, xvi\u2013xvii, xviii, n, \u2013n, n\n\nFriess, Steve, , n\n\nG\n\nGardner, Bill,\n\n\"Georgia judge sides with state in 'missing' voters case\" (Torres), n\n\nGeorgia voter laws, \u2013\n\ngerrymandering, xviii, \u2013, , , \u2013, , , \u2013, n, n, n, \u2013n, n, n, n, \u2013n. see also redistricting, extreme; REDMAP (REDistricting MAjority Project)\n\nget-out-the-votes (GOTV), ,\n\nGibson, Carl, n\n\nGill v. Whitford, \u2013n\n\nGive Us The Ballot (Berman), \u2013\n\nGonzales, Alberto,\n\n\"The GOP is Dying Off. Literally\" (McGraw), n\n\n\"The GOP Targets State Legislatures: He who controls redistricting can control Congress\" (Rove), n\n\n\"The GOP's cyber election hit squad\" (Rosenfeld), n\n\n\"The GOP's Stealth War Against Voters: Will an anti-voter-fraud program designed by one of Trump's advisers deny tens of thousands their right to vote in November?\" (Palast), \u2013n\n\nGorsuch, Neil, \u2013\n\nGotVoterId.com, n\n\nGovernment Accountability Board,\n\n\"Governor McAuliffe Provides Update on Restoration of Rights Numbers\" (WWIR), n\n\nThe Great Suppression: Voting Rights, Corporate Cash, and the Conservative Assault on Democracy (Roth), n\n\nGreen, Joshua, , n\n\nGreen Party, xiv\u2013xvi, xxii, , \u2013, \u2013, \u2013, n\n\nGreer, Jim, \u2013\n\nGrijalva, Raoul,\n\nGrimm, Ray, n\n\nGutierrez, Isela, n\n\nH\n\nhacking. See Russian hacking; vote hacking\n\nhacking, extent of, \u2013\n\nhacking technology use,\n\n\"Hacking The Diebold Machine\" (Felten), n\n\nHajnal, Zoltan, , n\n\nHalderman, Alex, , \u2013\n\nHall, Bob, n\n\nHamilton, Alexander,\n\nHamilton Electors, \u2013\n\nHarris, Katherine, , n\n\n\"Harry Reid delivers for Hillary Clinton: Nevada's 'neutral' power players may have saved a campaign and changed history.\" (Ralston), n\n\n\"A Harsh Dose of Electoral Reality: Democrats Have Uphill Battle in '18 and Need to Elect Governors to Fight Gerrymandered GOP Monopoly\" (Rosenfeld), n\n\nHasen, Rick, \u2013, n\n\nHAVV, \u2013n\n\nHayden, Michael, ,\n\nHazen, Don, n\n\n\"Hearing Before The Committee on House Administration, Columbus Ohio, March 21, 2005.,\" \u2013n\n\nHearne, Mark \"Thor,\" \u2013, , \u2013n\n\nHelp America Vote Act (HAVA),\n\nHenrig, Jess, n\n\nHeritage Foundation, , ,\n\nA Sampling of Election Fraud Cases From Across The Country (Heritage Foundation), n\n\nHerren, T. Christian, Jr., n\n\nThe Hill, \"Ex-NSA head suggests US also hacks political parties\" (Uchill) Oct. 18, 2016, n\n\nHomeland Security, database,\n\nHorwitz, Sari, n\n\nHouse, retaking by Democrats in 2018, \u2013\n\nHouse Judiciary Committee, \u2013\n\nHouse of Representatives, n, n\n\nHouston Chronicle,\n\n\"How the GOP Is Re-segregating the South: Republicans are using the redistricting process to undermine minority voting power and ensure their party's dominance\" (Berman), \u2013n\n\nThe Huffington Post, \"Obama's DOJ Fought Texas Voter ID Law. Trump's New Civil Rights Chief Offered Tips On Writing It: That's not a great sign for voting rights\" (Reilly) Jan. 25, 2017, n\n\nHuffington Post,\n\n\"Over 220,000 Ballots Didn't Count In The Presidential Primaries; Try to cast a regular ballot, rather than a provisional one, when voting this fall\" (Lachman) June 16, 2016, \u2013n\n\n\"Snagged Votes In Los Angeles\" (Frankel) July 5, 2016, \u2013n\n\n\"Trump's Former Aide Concedes There Was No Voter Fraud In New Hampshire: Corey Lewandowski is not reading from the same script\" (Stein) Feb. 20, 2017, n\n\n\"Hunt Commission,\" n\n\nHussein, Fatima, n\n\nHusted, Jon, \u2013\n\nHusted v. A. Philip Randolph Institute, n\n\nHutchinson, Asa,\n\nI\n\nID requirements,\n\nillegal Democratic voters. See voter fraud\n\nillegal voting allegations, xix\n\n\"Inaccurate, Costly, and Inefficient: Evidence That America's Voter Registration System Needs an Upgrade\" (Pew Center), n\n\nIndependent Voter, YouTube.com, n\n\nIndependent votes, \u2013\n\nIndiana voting barriers, \u2013\n\nIndyStar.com, \"Republicans limit early voting in Marion County, letting it bloom in suburbs\" (Hussein) Aug. 10, 2017, n\n\nIndyWeek.com, \"Voter Integrity Project boot camp teaches how to challenge voters\" (Porter and Ball) Sept. 18, 2013, n\n\n\"Information Operations and Facebook\" (Weedon, Nuland and Stamos)), n\n\n\"Inside The Recount: Jill Stein and a ragtag team of computer experts decided to take America's elections to court. Here's how it all went wrong\" (Friess), n\n\n\"Inside the Trump Bunker, With Days to Go: Win or lose, the Republican candidate and his inner circle have built a direct marketing operation that could power a TV network\u2014or finish off the GOP\" (Green and Issenberg), n\n\nThe Intercept,\n\nInternational Festival of Arts and Ideas, n\n\nInterstate Voter Registration Crosscheck Program, , , , , , n\n\nintimidation tactics, , \u2013, , \u2013n\n\nIowa Caucuses, 2016 election, , \u2013, , n\n\n\"The Iowa caucuses: An accident of history,\" \u2013n\n\nIowa Democratic Party, \u2013\n\nIowa's nightmare revisited: Was correct winner called? (Jacobs), n\n\nIssenberg, Sasha, , n\n\nIssues Related to Registering Voters and Administering Elections, n\n\nIssues Related to State Voter Identification Laws, n\n\n\"It's time to bust the myth: Most Trump voters were not working class,\" n\n\nJ\n\nJackson, Simon,\n\nJacobs, Jennifer, n\n\nJacobson, Jeff, \u2013\n\nJankowski, Chris, , \u2013\n\nJayapal, Pramila,\n\nJeff Sessions, wikipedia.org, n, n\n\nJefferson, David,\n\nJim Crow South, xxiv, , , , \u2013n, n\n\n\"Judge: North Carolina counties must restore voters removed in 'insane' process,\" n\n\nJudges Find Wisconsin Redistricting Unfairly Favored Republicans (Wines), n\n\nJustice Department Suing NYC Board Of Elections Over Last Year's Pre-Primary Mass Voter Purge (Tempey), n\n\n\"Justice Dept. sides with Ohio's purge of inactive voters in case headed to Supreme Court\" (Horwitz), n\n\nJustSecurity.org, \"Connecting The Dots: Political Microtargeting and the Russia Investigation\" (Brannen) May 18, 2017, n\n\nK\n\nKansasCity.com, \"Secretary of State Kris Kobach announces first conviction of noncitizen voting in Kansas\" (Woodall) April 12, 2017, n\n\nKaplan, Eric, n\n\nKasdan, Diana, n\n\nKeith, Douglas, n\n\nKemp, Brian, \u2013, \u2013\n\nKerry, John, xiii, ,\n\nKessler, Glenn, n\n\n\"Key Question for Supreme Court: Will It Let Gerrymanders Stand?\" (Wines), n\n\nKing, Martin Luther, Jr., ,\n\nKnight, John, n\n\nKobach, Kris, \u2013, , \u2013, , \u2013, , n, n\n\nKobach et al. v. The United States Election Assistance Commission, \u2013n\n\nKozikowski, Debra, \u2013\n\n\"Kris Kobach supports Donald Trump's assertion millions voted illegally\" (Carpenter), n\n\n\"Kris Kobach\" wikipedia.org, n\n\nL\n\nLaCapria, Kim, n\n\nLachman, Samantha, \u2013n\n\nLajevardi, Nazita, , n\n\n\"The Latest Voter Fraud Lie\" (The Editorial Board), n\n\nLawrence Livermore National Library,\n\nLawyers Committee for Civil Rights Under Law and others, \u2013n\n\nLeague of United Latin American Citizens (LULAC) of Wisconsin, et, v. Judge David Deininger et al, Case no. 12-CV-00185, \u2013n\n\nLeague of United Latin American Citizens v. Perry 548 US 399 (2006), n\n\nLeague of Women Voters, ,\n\nLeague of Women Voters v. Newby, n\n\n\"Leaked Email Proves DNC Rigged Primary Debates,\" n\n\n\"At Least 50 Trump Electors Were Illegitimately Seated as Electoral College Members: More evidence surfaces as calls mount to challenge congressional ratification of Electoral College vote\" (Rosenfeld), \u2013n\n\nLee, Barbara,\n\nLee, Shiela Jackson, \u2013\n\nLerer, Lisa, n\n\nLevitt, Justin, n\n\nLewandowski, Corey, n\n\nLGBT rights and gerrymandering, \u2013\n\nLimbaugh, Rush,\n\nLindell, Chuck, \u2013n\n\nLiptak, Adam, n, n\n\n\"logic and accuracy\" tests,\n\nLopez, Tomas, n\n\nLos Angeles Times,\n\n\"Would-be independents joining the American Independent Party could blame California's voter registration card\" (Myers) April 19, 2016, n\n\nLos Angeles Times, February 23, 2016, n\n\nLucey, Catherine, n\n\nLupu, Noam, n\n\nLynch, Loretta, n\n\nM\n\nMalewitz, Jim, n, n\n\n\"The Man Behind Trump's Voter Fraud Obsession\" (Berman), \u2013n\n\nMaricopa county, AZ, \u2013\n\nMarion Three, n\n\nMark Martin, Secretary of State, et al v. Freedom Kohls, Toylanda Smith, Joe Flakes and Barry Haas [CV-14-462], Oct. 15,2014, Arkansas, \u2013n\n\nMassachusetts Party (Democrats), \u2013\n\nMassaro, Ann, n\n\nMasset, Royal,\n\nMatt, Paul,\n\nMayer, Kenneth R.,\n\nMcAuliffe, Terry,\n\nMcCain, John, ,\n\nMcConnell, Mitch,\n\nMcCrory, Pat, , \u2013, n\n\nMcCue, George, n\n\nMcDonald, Michael, , \u2013n\n\nMcElwee, Sean, n\n\nMcGovern, George,\n\nMcGovern, James, \u2013\n\nMcGraw, Daniel J., n\n\nMcGuire, Andy, \u2013\n\nMebane, Walter, Jr., \u2013\n\nmedia response to Trump win, \u2013\n\nMedium.com, \"Who Hacked the Election? Ad Tech Did. Through 'Fake News,' Identity Resolution and Hyper-Personalization\" (Albright) July 30, 2017, n\n\n\"Meet the Brains Behind the Effort to Get the Supreme Court to Rethink Civil Rights: Edward Blum has orchestrated a string of big cases targeting voting rights and affirmative action\" (Mencimer), n\n\nMerica, Dan, n\n\nMichael McDonald, July 1, 2017, n\n\nMichigan Election Reform Alliance, , n\n\nMichigan recounts, n\n\nMihalik, Lily, n\n\nMillender-McDonald, Juanita, \u2013\n\nMiller, Stephen, \u2013\n\nminority voters, \u2013, \u2013, , \u2013, , \u2013, n, n, \u2013n, n. see also Jim Crow South; voter ID laws; voter suppression\n\nMintPress News, \"Donald Trump Warned Of A 'Rigged' Election, Was He Right?\" (Simon) Jan. 24, 2017, n\n\n\"Monkey Cage,\"\n\nMook, Robby,\n\nMoomaw, Graham, \u2013n\n\nMother Jones, \"Meet the Brains Behind the Effort to Get the Supreme Court to Rethink Civil Rights: Edward Blum has orchestrated a string of big cases targeting voting rights and affirmative action\" (Mencimer) March\/April 2016, n\n\nMotz, Diana Gribbon, \u2013\n\nMoveOn.org, ,\n\n\"Debbie Wasserman Schultz: Give the voter file back to Bernie Sanders' campaign\" (Gibson), n\n\nMueller, Robert,\n\nMulvihill, Geoff, \u2013n\n\nMyers, Jon, n\n\nmyth of voter fraud. See voter fraud\n\nN\n\nNAACP, , n\n\nNAACP Legal Defense Fund, \u2013, , n\n\nNAACP v. Florida, n\n\nThe Nation, \u2013, , ,\n\n\"How the GOP Is Re-segregating the South: Republicans are using the redistricting process to undermine minority voting power and ensure their party's dominance\" (Berman) Feb. 20, 2012, \u2013n\n\n\"The Man Behind Trump's Voter Fraud Obsession\" (Berman) Oct. 15, 2013, \u2013n\n\n\"Separate and Unequal Voting in Arizona and Kansas: Two states want to create a two-tiered voting system reminiscent of the Jim Crow South\" (Berman) Oct. 15, 2013, \u2013n\n\n\"This Election Is Being Rigged\u2014but Not by Democrats: The GOP's voter-suppression efforts are the real voter fraud\" (Berman) Oct. 17, 2016, n\n\n\"Will North Carolina Lead the Way to a New South? Progressives like Reverend Barber and candidates like Deborah Ross are hoping to assemble a multiracial coalition to defeat a conservative backlash\" (Walsh) Nov. 7, 2016, n\n\nNational Conference of State Legislatures, n, n\n\n\"Voter Identification Requirements | Voter ID Laws\" (Underhill) June 5, 2017, n\n\nNational Director of Intelligence,\n\nNational Public Radio (NPR), xiii\n\n\"Citizen Scientists Comb Images To Find An 'Overexcited Planet\" (Palca) June 1, 2017, n\n\nNational Voter Registration Act, 1993 (NVRA), , \u2013, , \u2013, , \u2013n, n, n, n, n, n\n\nNational Voter Registration Act (NVRA),\n\n\"The National Voter Registration Act Of 1993 (NVRA): Questions and Answers,\" n\n\n\"NCGOP Sees Encouraging Early Voting, Obama\/Clinton Coalition Tired, Fail to Resonate in North Carolina,\" n\n\nNegron, Joe,\n\nNeilson, Lindsay,\n\nNevada Caucuses, 2016, \u2013, n, n\n\nNevada Democratic Party, \u2013, n\n\nNew Georgia Project, \u2013\n\n\"New Georgia Project releases financial records, renews voter push\" (Torres), n\n\n\"New Report: Extreme Partisan Maps Account for 16\u201317 Republican Seats in Congress,\" n\n\nNew Republic, \u2013\n\n\"New Video: Watch Wisconsin Election Officials Reject Hand Counts After Electronic Scanners Make Big Mistake,\" n\n\n\"New Voting Restrictions in America,\" n\n\n\"New Voting Restrictions in Place for the 2016 Election\" (Brennan Center for Justice), \u2013n\n\nNew York, April 2016 primary, \u2013\n\nThe New York magazine, \"Cyberwar for Sale: After a maker of surveillance software was hacked, its leaked documents shed light on a shadowy global industry that has turned email theft into a terrifying\u2014and lucrative\u2014political weapon.\" (Schwartz) Jan. 8, 2017, n\n\nNew York magazine,\n\n\"Don Yelton, GOP Precinct Chair, Delivers Most Baldly Racist Daily Show Interview of All Time\" (Coscarelli), \u2013n\n\n\"Experts Urge Clinton Campaign to Challenge Election Results in 3 Swing States\" (Sherman) Nov. 22, 2016, n\n\nThe New York Post, \"Top Democrat backpedals after saying primaries were 'rigged'\" (Steinbuch) Feb. 10, 2017, n\n\nThe New York Times, , , \u2013,\n\n\"Advocate for Tough Voting Rules to Steer Trump's Elections Commission\" (Wines and Bosman) May 14, 2017, n\n\n\"Child of the Enlightenment\" (Wills) July 2, 2017, n\n\n\"Critics See Efforts by Counties and Towns to Purge Minority Voters From Rolls\" (Wines) july 31, 2016, n\n\n\"Key Question for Supreme Court: Will It Let Gerrymanders Stand?\" (Wines) April 21, 2017, n\n\n\"The Latest Voter Fraud Lie\" (The Editorial Board) Feb. 13, 2017, n\n\n\"Plaintiffs Claim Bias During Closing Argument Against Texas Voter ID Law\" (Fernandez) Sept. 22, 2014, n\n\n\"Strict North Carolina Voter ID Law Thwarted After Supreme Court Rejects Case\" (Liptak and Wines) May 15, 2017, n\n\n\"The Success of the Voter Fraud Myth\" (Editorial Board) Sept. 20, 2016, n\n\n\"Texas Agrees to Soften Voter ID Law After Court Order\" (Wines) Aug. 4, 2016, n\n\n\"For Voting Rights Advocates, Court Decision Is 'Temporary Victory,'\" (Wines) May 16,2017, \u2013n\n\n\"When Does Political Gerrymandering Cross a Constitutional Line?\" (Liptak) May 15, 2017, n\n\nNew York Times,\n\n\"Why I Will Not Cast My Electoral Vote for Donald Trump\" (Suprun) Dec. 6, 2016, n\n\n\"Wisconsin Strict ID Law Discouraged Voters, Study Finds\" (Wines) Sept. 25, 2017, n\n\nThe New Yorker, \"Will Keith Ellison Move the Democrats Left? By running for the D.N.C. chair, the Minnesota congressman hopes to lead a populist opposition against Trump\" (Cunningham) Feb. 27 2017, n\n\nNewby, Brian, \u2013, n\n\nNewRepublic.com, \"Inside The Recount: Jill Stein and a ragtag team of computer experts decided to take America's elections to court. Here's how it all went wrong\" (Friess) Feb. 15, 2017, n\n\nThe News and Oberserver, \"Will third time be a charm for those who want NC Supreme Court to invalidate election maps?\" (Blythe) May 30, 2017, \u2013n\n\nNey, Bob, \u2013\n\nNichols, John, n\n\nNielson, Lindsay, n\n\n\"No, 1.8 million dead people aren't going to vote in November\" (Bump), n\n\nNo Party Preference (NPP), \u2013\n\nnoncitizen voting, , n, n, \u2013n, n, n\n\n\"Noncitizen Voting: The Missing Millions\" (Famighetti, Keith, Perez), n\n\nNorth Carolina Republican Party, Nov. 6, 2016, n\n\nNorth Carolina, State Board of Elections, April 2017,\n\nNorth Carolina State Conference of the NAACP et al. v. North Carolina State Board of Elections et al, n\n\nNorth Carolina State Conference of the NAACP et al. v. Pat McCrory et al., n\n\nNorth Carolina voter discrepencies,\n\nNorth Carolina voter hijacks, \u2013\n\nNorth Carolina voter purges, n, n, n, n\n\nNorth Carolina voting vigilantes, \u2013\n\nO\n\nObama, Barack, xvii, , \u2013, , , , n\n\nObamacare and gerrymandering, \u2013\n\n\"Obama's DOJ Fought Texas Voter ID Law. Trump's New Civil Rights Chief Offered Tips On Writing It\" Huffington Post,\n\n\"Obama's DOJ Fought Texas Voter ID Law. Trump's New Civil Rights Chief Offered Tips On Writing It: That's not a great sign for voting rights\" (Reilly), n\n\n\"October House Overview: GOP Risk Factors\" (Wasserman), \u2013n\n\n\"Ohio Voter Fraud Billboards Target Minorities\" (Feeney), \u2013n\n\nOhio voter purge, \u2013, \u2013, n, \u2013n, n, \u2013n, n, n\n\nOhlemacher, Stephen, n\n\nO'Malley, Martin,\n\n\"1.7 million people in 33 states and D.C. cast a ballot without voting in the presidential race\" (Bump), \u2013n\n\n\"The Other Right-Wing Tidal Wave Sweeping America: Federal and State Preemption of Local Progressive Laws: Preemption allows corporations to boost their profits by suppressing local government power, community groups and citizens\" (Hazen and Rosenfeld), n\n\nOur Revolution: A Future to Believe In (Sanders), , , \u2013\n\n\"Over 220,000 Ballots Didn't Count In The Presidential Primaries; Try to cast a regular ballot, rather than a provisional one, when voting this fall\" (Lachman), \u2013n\n\nP\n\nPadilla, Alex, n\n\nPalast, Greg, \u2013\n\nPalca, Joe, n\n\nPalm Beach Post, \u2013\n\npaper ballot-based systems,\n\nParnes, Amie, , , n\n\nparticipation barriers, \u2013, \u2013\n\npartisan ghettos, \u2013\n\nPaul, Matt, n\n\nPaxton, Ken, \u2013n\n\nPence, Mike, ,\n\npendulum swings, political, , \u2013, ,\n\nPennsylvania recounts, \u2013\n\nPerez, Evan, n\n\nPerez, Myrna, n\n\nPerez, Tom,\n\nPERMISSION TO PLACE IN THE RECORD A STATEMENT REGARDING THE JOINT SESSION OF ELECTION, Jan. 6, 2017 House of Representatives, n\n\nPesce, Anthony, n\n\nPeterson, James,\n\nPew Center, ,\n\nThe Pew Center on The States\/Election Initiatives 2012, n\n\nThe Pew Center on The States\/Election Initiatives Nov. 2010, xviii, xix, n, n\n\nA Philip Randolph Institute, Northeast Ohio Coalition for the Homeless, Larry Harmon v. Jon Husted Sept. 23, 2016, n\n\nA. Phillip Randolph Institute v. Husted, \u2013\n\nPhillips, Amber, n\n\nPhillips, Gregg,\n\n\"Plaintiffs Claim Bias During Closing Argument Against Texas Voter ID Law\" (Fernandez), n\n\nPodesta, John, , \u2013\n\npolitical action committees (PACS),\n\npolitical maps, xviii\n\npolitical microtargeting, social media, \u2013\n\npolitical power grabs, \u2013\n\nPolitico Magazine,\n\n\"The GOP is Dying Off. Literally\" (McGraw), May 17, 2015, n\n\n\"Why Republicans Can't Find the Big Voter Fraud Conspiracy: If the last federal investigation is any guide, the answer is simple: It probably doesn't exist\" (Rab) April 2, 2017, n\n\npolitifact.com, \"Sean Spicer wrongly uses Pew study to bolster claim that noncitizens vote in large numbers\" (Carroll) Jan. 25, 2017, n\n\n\"Poll: Nearly half of Americans say voter fraud occurs often\" (Guskin and Clement), n\n\nPorter, Jane, n\n\nPraxton, Ken, n\n\n\"President Trump Appoints Country's Worst Vote Suppressor to His 'Election Integrity' Commission\" (Hasen), n\n\n\"Presidential Advisory Commission on Election Integrity, Mission, Procedures and Topics for Consideration\" July 19, 2017, n\n\n\"Presidential Executive Order on the Establishment of Presidential Advisory Commission on Election Integrity,\" The White House, May 11, 2017, n\n\nPrimary Day (CA),\n\n\"Primary Elections in California\" (Padilla), n\n\n\"Prince Harry,\"\n\nProCon.org, \"Are Felons More Likely to Vote for Democrats over Republicans?\" (ProCon.org) March 7, 2014, n\n\nProgressive Democrats,\n\nProgressive voters,\n\nProjupecz, Sharon, n\n\nproof of citizenship requirements, , , \u2013, , , , \u2013, n, n, n, \u2013n, n, n\n\nproof of electronic theft of election, lack of,\n\n\"Proof of US Citizenship for Voter Registration: NOTE: This requirement is effective January 1, 2013.,\" n\n\nprovisional ballots,\n\nPurcell, Helen, \u2013\n\nPutin, Vladimir, \u2013\n\nQ\n\n\"Quick Response Evaluation: Accuracy of the Help America Vote Verification Program Responses\" (O'Carroll) June 22, 2009, \u2013n\n\nQuinnipiac University poll, xxiii, n\n\nR\n\nRab, Lisa, n\n\nrace, use when creating districts, \u2013, \u2013\n\nRalston, Jon, \u2013, n\n\nRao, Amy,\n\nRapoport, Miles, , , n, n\n\nRaskin, Jamie,\n\n\"Ratf**ked: The True Story behind The Secret Plan to Steal America's Democracy\" (Daley), n\n\nRatf**ked: The True Story Behind the Secret Plan to Steal America's Democracy (Daley), \u2013, , , n\n\nReagan, Ronald,\n\n\"The Real Story of the Recount\" (Fitrakis), n\n\n\"Recidivism Watch: Trump's claim that millions of people voted illegally\" (Kessler), n\n\nrecount process, n, n\n\nrecount requests, \u2013\n\nrecounts, aborted,\n\n\"red shift\" in exit poll,\n\n\"red shift\" in votes,\n\n\"Redistricting\" (TexasTribune.org), \u2013n\n\nredistricting, extreme, , \u2013, \u2013, \u2013n, n, \u2013n\n\nREDMAP (REDistricting MAjority Project), \u2013, , , , , , \u2013, n. see also gerrymandering; Rove, Karl\n\n\"Register to Vote and Confirm or Change Registration,\" n\n\n\"Register To Vote In Your State by Using This Postcard Form and Guide,\" n\n\nregistration, same day. See Election Day registration\n\nregistration drive errors,\n\nReid, Harry,\n\nReilly, Ryan, n\n\nRennert, Joan, xv\u2013xvi\n\nReno Gazette Journal,\n\n\"Update: Suspicious votes, long times at Dem caucus\" (Robison) Feb. 20, 2016, n\n\n\"repeaters,\"\n\n\"reprehensible acts,\" \u2013\n\nRepublican Party\/GOP, xvii, xix, xxii. see also anti-voter barriers; early voting days; e-poll books; gerrymandering; REDMAP (REDistricting MAjority Project); voter ID laws\n\n\"ALEC Is Talking About Changing the Way Senators Are Elected and Taking Away Your Vote: A proposed resolution advocates for overturning the 17th Amendment so Republican-controlled state legislatures could pick senators\" (Nichols), n\n\nand anti-voter laws, \u2013\n\nattacks on voting,\n\nand caging, \u2013\n\ncondemned by Sanders,\n\nand extreme voter redistricting, \u2013\n\nand false voter fraud allegations, \u2013\n\nGeorgia voter laws, \u2013\n\n\"The GOP is Dying Off. Literally\" (McGraw), May 17, 2015, n\n\n\"The GOP's cyber election hit squad\" (Rosenfeld), n\n\nJudges Find Wisconsin Redistricting Unfairly Favored Republicans (Wines), n\n\n\"NCGOP Sees Encouraging Early Voting, Obama\/Clinton Coalition Tired, Fail to Resonate in North Carolina,\" n\n\n\"New Report: Extreme Partisan Maps Account for 16\u201317 Republican Seats in Congress,\" n\n\npower grabs, \u2013\n\n\"Republican redistricting is taking a beating in the courts (again)\" (Phillips), n\n\nresponse to Trump win,\n\n\"Revised Redistricting Plans Face Strong Public Opposition: GOP lawmakers on a key Senate committee approved a revision Friday of controversial redistricting plans, but Democrats and others accused Republicans of lacking transparency\" (Kaplan), n\n\nsuspicion of hacking of voting machines,\n\n\"A Very Early Look At The Battle For The House In 2018: Donald Trump is unpopular enough that Republicans could lose the House, but there's a lot of uncertainty.\" (Enten), \u2013n\n\nvote rigging, xxiii\u2013xxiv\n\nand voter fraud hijacks, \u2013\n\nand voter hijacks playbook, \u2013, \u2013\n\nand voter laws,\n\nvoter laws, \u2013\n\nvoter suppression, \u2013\n\nand voter technicalities, \u2013\n\nwar on voting,\n\n\"We're Supposed to Be a Democracy, But Half the GOP Is OK with Postponing 2020's Election: A party full of delusions exposes the dark underbelly of America\" (Rosenfeld), n\n\n\"Why Republicans Can't Find the Big Voter Fraud Conspiracy: If the last federal investigation is any guide, the answer is simple: It probably doesn't exist\" (Rab) April 2, 2017, n\n\n\"Republican redistricting is taking a beating in the courts (again)\" (Phillips), n\n\n\"Republicans limit early voting in Marion County, letting it bloom in suburbs\" (Hussein), n\n\n\"Restore lawful enforcement policies in the Civil Rights Division,\" \u2013n\n\n\"Results from the 2016 Nevada GOP caucuses\" (Mihalik, Pesce and Walsh), n\n\nReuters, ,\n\n\"Special Report: Behind US race cases, a little-known recruiter\" (Biskupic), n\n\n\"Revised Debate Memo,\" April 27, 2015, wikileaks.org, n\n\n\"Revised Redistricting Plans Face Strong Public Opposition: GOP lawmakers on a key Senate committee approved a revision Friday of controversial redistricting plans, but Democrats and others accused Republicans of lacking transparency\" (Kaplan), n, \u2013n\n\nRichmond Times-Dispatch, \"Va. Supreme Court strikes down McAuliffe's order on felon voting rights\" (Moomaw) July 23, 2016, \u2013n\n\nRiley, Michael, n\n\nRoberts, John, 31,\n\nRobertson, Jordan, n\n\nRobison, Mark, n\n\nRolling Stone, \"The GOP's Stealth War Against Voters: Will an anti-voter-fraud program designed by one of Trump's advisers deny tens of thousands their right to vote in November?\" (Palast) Aug. 24, 2016, \u2013n\n\nRolling Stone, Aug. 2016, \u2013\n\nRomney, Mitt,\n\nRoosevelt, Franklin Delano,\n\nRosenfeld, Steven, n, n, n, n, n, \u2013n, n, n\n\nRoss, Deborah, \u2013, n\n\nRoth, Zachary, n\n\nRove, Karl, xxi, \u2013, , , , n\n\nRussia and US voters, xxi, xxiii\n\n\"Russian Cyber Hacks on US Electoral System Far Wider Than Previously Known\" (Riley and Robertson), n\n\nRussian hacking, , , \u2013, , \u2013, , \u2013n, n, n, n, n\n\nand GOP voter hijacks playbook,\n\nand Trump victory,\n\nRussian intelligence hackers,\n\nRuthelle Frank, et al., v. Scott Walker Case no. 11-CV-01128, \u2013n\n\nRyan, Micael,\n\nS\n\nSablosky, Bret,\n\nSales, Ruby, n\n\n\"Same-day voter registration,\" n\n\nA Sampling of Election Fraud Cases From Across The Country (Heritage Foundation), n\n\nSanders, Bernie, xiv\u2013xv, xiv, xxii\u2013xxiii,\n\n2016 election, \u2013\n\nCalifornia 2016 primaries, \u2013\n\nand Iowa Caucuses, 2016, \u2013\n\nMarch-April primaries, \u2013\n\nNevada Caucuses, 2016, \u2013, n\n\nNew Hampshire primary, 2016 election,\n\nvoter suppression,\n\n\"Sanders campaign sues DNC after database breach,\" n\n\nSanders-Clinton split, xxv\n\nSantorum, Rick,\n\nScalia, Antonin,\n\nSchneiderman, Eric,\n\nSchuette, Bill, xix\u2013xx, \u2013\n\nSchultz, Debbie Wasserman, xxii, , \u2013, n\n\nSchwartz, Mattathias, n\n\nSchwarzenegger, Arnold,\n\nScott, Rick,\n\n\"Sean Spicer wrongly uses Pew study to bolster claim that noncitizens vote in large numbers\" (Carroll), n\n\n\"Secretary of State Kris Kobach announces first conviction of noncitizen voting in Kansas\" (Woodall), n\n\nSection 5 of Voting Rights Act, , \u2013n, n\n\nsegregation. See Jim Crow South; redistricting, extreme, , n\n\n\"Senate Approves Redistricting Plan; Black Democrats Object: The Alabama Senate approved new legislative districts over the objections of black Democrats\" (Chandler), \u2013n\n\nSenate Intelligence Community, n\n\nSenate seats, \u2013\n\n\"Separate and Unequal Voting in Arizona and Kansas: Two states want to create a two-tiered voting system reminiscent of the Jim Crow South\" (Berman), \u2013n\n\nSequoia AVC Edge,\n\nSessions, Jeff, , \u2013, n, n, n\n\nSeventeenth Amendment, ,\n\nshadowproof.com, n\n\nShattered: Inside Hillary Clinton's Doomed Campaign (Allen and Parnes), , , n\n\nShelby County v. Holder, \u2013, , \u2013n, n, n\n\nSherman, Amy, n\n\nSherman, Gabriel, n\n\nSimas, David,\n\nSimon, Jonathan, \u2013\n\nSimons, Barbara, \u2013, n\n\nSmith, Grant, \u2013n\n\n\"Snagged Votes In Los Angeles\" (Frankel), \u2013n\n\nsnopes.com, n\n\nSnyder, Timothy, \u2013, n\n\nsocial media operation, Russian hackers, \u2013\n\nSocial Security Administration (SSA), \u2013\n\nSocial Security numbers,\n\n\"Some States Review Election Systems for Signs of Intrusion\" (Dalesio and Mulvihill), \u2013n\n\n\"Special Report: Behind US race cases, a little-known recruiter\" (Biskupic), n\n\nSpicer, Sean,\n\nSSNVS, \u2013n\n\nSSOLV, \u2013n\n\nStanford Law Review, June 2016, \"Three Tests for Practical Evaluation of Partisan Gerrymandering\" (Wang), n\n\n\"The State of Redistricting Litigation (March 2017 edition)\" (Brennan Center for Justice at New York University, n\n\nstate-issued ID, xvii\n\n\"Statement From A. G. Schneiderman On Voting Issues During New York's Primary Election,\" n\n\n\"Statement on Hillary Clinton's Victory in the Iowa Caucus\" (Paul), n\n\nStein, Jill, xiv\u2013xv, \u2013, \u2013, , n, n, n\n\nStein, Sam, n\n\nSteinbuch, Yaron, n\n\nStephanopoulos, Nicholas O.,\n\nStrach, Kim Westbrook, \u2013, \u2013n\n\n\"Strict North Carolina Voter ID Law Thwarted After Supreme Court Rejects Case\" (Liptak and Wines), n\n\nSublette, Constance, n\n\nSublette, Ned, n\n\n\"The Success of the Voter Fraud Myth\" (Editorial Board), n\n\nSullivan, Andy, \u2013n\n\nSullivan, Jake,\n\nSuper PACS,\n\nSuper Tuesday, ,\n\nsuperdelegate system, \u2013\n\nsuperdelegates, DNC, xxii\u2013xxiii, \u2013, , , \u2013,\n\nSupreme Court\n\nArizona Prop 200 overturned,\n\nBaker v. Carr,\n\nBlum lawsuits, n\n\nBRIEF OF FORMER ATTORNEYS OF THE CIVIL RIGHTS DIVISION OF THE UNITED STATES DEPARTMENT OF JUSTICE AS AMICI CURIAE IN SUPPORT OF PETITIONER March, 9, 2017, n\n\nBRIEF OF GEORGIA AND 14 OTHER STATES AS AMICI CURIAE IN SUPPORT OF PETITIONER March 10, 2016, \u2013n\n\ncampaign finance rulings,\n\nemergency recount petitions,\n\nand extreme voter redistricting,\n\nand filibuster rule,\n\nFlorida recount, 2000, xx\n\nand gerrymandering, \u2013, \u2013, , n, n, n, \u2013n\n\nGill v. Whitford, \u2013n\n\ngutting of Voting Rights Act, n\n\nHusted v. A. Philip Randolph Institute, n\n\nOhio voter purge, \u2013,\n\nand Ohio voter purges, \u2013\n\noverturns Arkansas voter ID law,\n\nand race redistricting, \u2013\n\nrejects voting rights case, n\n\nShelby County v. Holder, \u2013n\n\nSupreme Court,\n\nand the Voting Rights Act, , , , \u2013\n\n\"With Supreme Court appeal, Texas wants to keep congressional map intact: Attorney General Ken Paxton revealed that Texas has no plans to ask lawmakers to redraw the state's Congressional map in a fresh round of legislative overtime. Instead, Paxton is appealing the ruling to the US Supreme Court\" (Malewitz), n\n\n\"The Supreme Court Could Make It Easier for States to Purge Voters; The Court has decided to hear a new case from Ohio, where Democrats were twice as likely as Republicans to be purged from the rolls in the state's largest counties.\" (Berman), \u2013n\n\n\"Supreme Court forced to confront the 'unsavory' politics of district lines\" (Wolf), n\n\n\"Supreme Court: Virginia redistricting must be reexamined for racial bias\" (Barnes), \u2013n\n\nSuprun, Christopher, \u2013, n\n\nSwedlund, Mark,\n\nT\n\nTampa Bay Times, \"Florida halts purge of noncitizens from voter rolls\" (Bousquet and Sherman) March 27, 2014, n\n\nTea Party,\n\ntechnicality hijacks, \u2013\n\nTempey, Nathan, n\n\n\"Texas, Wisconsin Photo ID Wins Could Help Nearly 1 Million Registered Voters\" (Brennan Center for Justice), \u2013n\n\n\"Texas Agrees to Soften Voter ID Law After Court Order\" (Wines), n\n\n\"Texas Democratic primary and caucuses, 2008,\" \u2013n\n\nTexas NAACP v. Steen (consolidate with Veasey v. Abbot), \u2013n\n\nTexas Republican Party,\n\nThe Texas Tribune\n\n\"With Supreme Court appeal, Texas wants to keep congressional map intact: Attorney General Ken Paxton revealed that Texas has no plans to ask lawmakers to redraw the state's Congressional map in a fresh round of legislative overtime. Instead, Paxton is appealing the ruling to the US Supreme Court\" (Malewitz) Aug. 18, 2017, n\n\n\"Trump administration confirms plans to drop key claim against Texas Voter ID: The US Department of Justice confirmed Monday it plans to ditch its longstanding position that Texas lawmakers purposefully discriminated against minority voters by passing the nation's strictest voter identification law in 2011\" (Malewitz) Feb. 27, 2017, n\n\nTexas voter hijacks, \u2013\n\nTexas voter ID laws, \u2013,\n\nTexasTribune.org, \u2013n\n\nTheIntercept.com,\n\n\"Top-Secret NSA Report Details Russian Hacking Effort Days Before 2016 Election\" (Cole, Esposito, Biddle and Grimm) June 5, 2017, n\n\n\"There Were 5-Hour Lines to Vote in Arizona Because the Supreme Court Gutted the Voting Rights Act\" (Berman), n\n\n\"This Election Is Being Rigged\u2014but Not by Democrats: The GOP's voter-suppression efforts are the real voter fraud\" (Berman), n\n\nThomas, Chris, xviii\u2013xx, \u2013,\n\n\"Three Tests for Practical Evaluation of Partisan Gerrymandering\" (Wang), n\n\nThree-Fifths Clause, Article 1,\n\n\"Top Democrat backpedals after saying primaries were 'rigged'\" (Steinbuch), n\n\nThe Topeka Capital-Journal, \"Kris Kobach supports Donald Trump's assertion millions voted illegally\" (Carpenter) Nov. 20, 2016, n\n\n\"Top-Secret NSA Report Details Russian Hacking Effort Days Before 2016 Election\" (Cole, Esposito, Biddle and Grimm), n\n\nTorres, Kristina, n, n, n\n\nToutant-Beauregard, Pierre Gustave,\n\nTreyz, Catherine, n\n\nTribune News Services\/Associated Press, \"Judge: North Carolina counties must restore voters removed in 'insane' process\" Nov. 4, 2016, n\n\nTrue The Vote, n\n\nTrueTheVote,\n\nTrump, Donald, xiii\u2013xiv, xvi, xx, xxiii, , n, n, n\n\nand empowerment of voter suppression,\n\n\"Lock her up,\"\n\n\"Obama's DOJ Fought Texas Voter ID Law. Trump's New Civil Rights Chief Offered Tips On Writing It: That's not a great sign for voting rights\" (Reilly) Jan. 25, 2017, n\n\nPennsylvania election,\n\nPutin and Russian hacking, \u2013\n\n\"Recidivism Watch: Trump's claim that millions of people voted illegally\" (Kessler), 197n263\u2013198n263\n\nand REDMAP targeting, \u2013\n\nand Russian hacking, n\n\n\"Trump administration confirms plans to drop key claim against Texas Voter ID: The US Department of Justice confirmed Monday it plans to ditch its longstanding position that Texas lawmakers purposefully discriminated against minority voters by passing the nation's strictest voter identification law in 2011\" (Malewitz), n\n\n\"Trump's Former Aide Concedes There Was No Voter Fraud In New Hampshire: Corey Lewandowski is not reading from the same script\" (Stein), n\n\n\"Trump's voter fraud experts registered in 3 states\" (Burke) Jan. 31, 2017, n\n\nTweet July 1, 2017, n\n\nTweet \"Stop Crooked Hillary From Rigging This Election!,\"\n\nTweet: Voter Fraud Panel,\n\nvictory in 2016 election, \u2013\n\nvoter fraud,\n\nand voter hijacks,\n\nvoter ID laws, \u2013\n\nand voter machine hacking,\n\nand war on voting,\n\n\"Why I Will Not Cast My Electoral Vote for Donald Trump\" (Suprun), n\n\nwins Wisconsin,\n\n\"Trump administration confirms plans to drop key claim against Texas Voter ID: The US Department of Justice confirmed Monday it plans to ditch its longstanding position that Texas lawmakers purposefully discriminated against minority voters by passing the nation's strictest voter identification law in 2011\" (Malewitz), n\n\n\"Trump Election Observers,\"\n\nTrump-Russian probes, . see also Mueller, Robert; Russian hacking\n\n\"Trump's Former Aide Concedes There Was No Voter Fraud In New Hampshire: Corey Lewandowski is not reading from the same script\" (Stein), n\n\n\"Trump's voter fraud claims undermine the democratic process and his presidency\" (Balz), n\n\n\"Trump's voter fraud experts registered in 3 states\" (Burke), n\n\nTubbs-Jones, Stephanie, \u2013,\n\nTurner, Albert, n\n\nTwentieth Amendment,\n\n\"2011 State and Legislative Partisan Composition\" (National Conference of State Legislatures), n\n\n2004 Election and the Implementation of The Help America Vote Act, \u2013\n\n2016 Pre- and Post-Election State Legislative Control (National Conference of State Legislatures) November 9, 2016, n\n\n\"On Tyranny: Lessons From The Twentieth Century\" (Synder), n\n\nU\n\nUchill, Joe, n\n\n\"Uncovering Kris Kobach's Anti-Voting History\" (Lopez and Clark), n\n\nUnderhill, Wendy, n\n\nunions and gerrymandering, \u2013\n\n\"United States House of Representatives elections, 2012,\" \u2013n\n\n\"United States House of Representatives elections, 2016,\" n\n\n\"Update: Suspicious votes, long times at Dem caucus\" (Robinson), n\n\n\"Upgrading Democracy: Improving America's Elections by Modernizing States' Voter Registration Systems\" (Pew Center), n\n\nUS Court of Appeals for the Fourth Circuit, North Carolina State Conference of the NAACP et al. v. Pat McCrory et al. July 29, 2016, n\n\nUS Court of Appeals for the Sixth Circuit, File Name 16a0241p.of, A Philip Randolph Institute, Northeast Ohio Coalition for the Homeless, Larry Harmon v. Jon Husted Sept. 23, 2016, n\n\nUS Department of Justice, n, n\n\n\"About Section 5 of the Voting Rights Act,\" n\n\nUS District Court Eastern District of Wisconsin, April 29, 2014, \u2013n\n\nUS District Court, Nov. 4, 2016, North Carolina State Conference of the NAACP et al. v. North Carolina State Board of Elections et al, n\n\nUS Election Assistance Commission (EAC), ,\n\nUS Government Accountability Office,\n\n\"Additional Data Could Help State and Local Elections Officials Maintain Accurate Voter Registration Lists\" June 2005, n\n\nIssues Related to Registering Voters and Administering Elections June 2016, n\n\nIssues Related to State Voter Identification Laws Sept. 2014, n\n\nUS House, 2012 election,\n\nUS Supreme Court\n\nBRIEF OF FORMER ATTORNEYS OF THE CIVIL RIGHTS DIVISION OF THE UNITED STATES DEPARTMENT OF JUSTICE AS AMICI CURIAE IN SUPPORT OF PETITIONER March, 9, 2017, n\n\nBRIEF OF GEORGIA AND 14 OTHER STATES AS AMICI CURIAE IN SUPPORT OF PETITIONER March 10, 2016, \u2013n\n\nUSA Today, , ,\n\n\"Supreme Court forced to confront the 'unsavory' politics of district lines\" (Wolf) May 25, 2017, n\n\nUSA Today, February, \"Harry Reid delivers for Hillary Clinton: Nevada's 'neutral' power players may have saved a campaign and changed history.\" (Ralston) Feb. 20, 2016, n\n\nUSA.gov, n\n\n\"Use it or lose it: Occasional Ohio voters may be shut out in November\" (Sullivan and Smith), \u2013n\n\nV\n\n\"Va. Supreme Court strikes down McAuliffe's order on felon voting rights\" (Moomaw), \u2013n\n\nverification of voting machines, \u2013\n\nVerifiedVoting.org, \"The Verifier - Polling Place Equipment - Current,\" n\n\n\"The Verifier - Polling Place Equipment - Current,\" n\n\n\"A Very Early Look At The Battle For The House In 2018: Donald Trump is unpopular enough that Republicans could lose the House, but there's a lot of uncertainty.\" (Enten), \u2013n\n\nVieth v. Jubelirer [02-1580], , n, n, \u2013n\n\nVirginia voter ID laws, \u2013\n\nVirginia voting rights,\n\n\"Virginia's breakthrough on rectifying an enormous injustice\" (Editorial Board), n\n\nVolkswagen, xxi\n\nvon Spakovsky, Hans, \u2013\n\nvote certification,\n\nvote discrepencies, post 2016 election, \u2013\n\nvote hacking, n\n\nvote recounts, n, n, n, n, n\n\nMichigan, xviii\u2013xxi\n\nWisconsin, xiv\u2013xvii\n\nvote.gov, n\n\nvoter bias a.k.a. tilting, rigging, gaming, \u2013\n\nvoter caging operations, n\n\nvoter fraud, \u2013, \u2013, n, n, n, n, n, n, n, \u2013n, n, \u2013n, n, n, \u2013n, n, n, n, n, n, n, n, n\n\nvoter fraud hijacks, \u2013\n\nvoter ID laws, , , , \u2013, \u2013, \u2013, \u2013, , , , n, \u2013n, n, \u2013n, n, n, n, n, \u2013n, \u2013n, \u2013n, n, n, n\n\n\"Voter Identification Requirements | Voter ID Laws\" (Underhill), n\n\nvoter impersonation fraud,\n\nVoter Integrity Project, n\n\n\"Voter Integrity Project boot camp teaches how to challenge voters\" (Porter and Ball), n\n\nVoter Integrity Project of North Carolina, \u2013\n\nvoter laws, \u2013, n\n\nvoter machine hacking, , n, n, n, n\n\nvoter purges, , \u2013, , \u2013, , \u2013, \u2013n, \u2013n, n, \u2013n, n, n, \u2013n, n\n\nvoter registration database problems, \u2013, n, n, n, n, \u2013n\n\nvoter registration laws, \u2013\n\nVoter Registration Modernization Design Working Group,\n\nvoter registration verification, \u2013n, n, n, \u2013n, n\n\nvoter suppression, , \u2013, , n, n\n\n\"Voter Suppression in the Mirror and Looking Forward: How much damage occurred in 2016, and what's in store for 2018 and beyond?\" (Rapoport), n\n\n\"Voter Suppression Works\" (McElwee), n\n\n\"Voter Turnout in Presidential Elections: 1828\u20132012\" (American Presidency Project), n\n\nvotes, intentions of nation's founders,\n\n\"Voting Advocates Announce a Settlement of 'Exact Match' Lawsuit in Georgia: Minor Typos and Data Entry Errors will No Longer Deny Eligible Georgians the Right to Register and Vote,\" \u2013n\n\n\"Voting Law Changes: Election Update\" (Weiser and Kasden), n\n\nvoting laws, n\n\nvoting laws, restrictive, n\n\n\"Voting Laws Roundup 2013\" (Brennen Center for Justice), n\n\n\"Voting Laws Roundup 2017\" (Brennan Law Review), n\n\n\"Voting Machine Digital Ballot Images Could Let Public Recount Elections, But Many Locales Aren't Saving or Sharing This Data: Publicly verified elections are key to de-corrupting our democracy\" (Brakey), n\n\nvoting machine tampering, \u2013, \u2013, , n\n\nvoting machines, xxi\u2013xxii\n\nvoting machines, assumptions of integrity,\n\nvoting maching verification, n\n\nvoting operating rules, gaming, \u2013\n\nvoting rights, \u2013, \u2013, \u2013n\n\nVoting Rights Act, 1965 (VRA), , , \u2013, , \u2013, , , , \u2013, , n\n\n\"The Voting Rights Act of 1965\" (US Department of Justice), n\n\n\"For Voting Rights Advocates, Court Decision Is 'Temporary Victory,'\" (Wines), \u2013n\n\nVoting Section, US Department of Justice, letter, June 28, 2017, n\n\nvoting vigilantes, \u2013\n\nVR Systems hack,\n\nW\n\nWaldman, Michael, , , \u2013, n\n\nWalker, Scott,\n\nThe Wall Street Journal, \"The GOP Targets State Legislatures: He who controls redistricting can control Congress\" (Rove) March 4, 2010, n\n\nThe Wall Street Journal, March 4, 2010, \u2013, ,\n\nWalsh, Ben, n\n\nWalsh, Joan, , , n\n\nWalsh, Shushannah, n\n\nWang, Samuel S.-H, n\n\nwar on voting, \u2013. see also Republican Party\/GOP\n\nWarner, Mark, n\n\nThe Washington Post, , , , , , ,\n\n\"1.7 million people in 33 states and D.C. cast a ballot without voting in the presidential race\" (Bump) Dec. 14, 2016, \u2013n\n\n\"Do voter identification laws suppress minority voting? Yes. We did the research.\" (Hajnal, Lajevardi, Nielson) Feb. 15, 2017, n\n\n\"Donald Trump is wrong. Rigging an election is almost impossible: This is how hard it would be to pull off widespread voter fraud\" (Berman) Aug. 5, 2016, n\n\n\"It's time to bust the myth: Most Trump voters were not working class\" June 5, 2017, n\n\n\"Justice Dept. sides with Ohio's purge of inactive voters in case headed to Supreme Court\" (Horwitz) Aug. 8, 2017, n\n\n\"No, 1.8 million dead people aren't going to vote in November\" (Bump) Oct. 18, 2016, n\n\n\"Poll: Nearly half of Americans say voter fraud occurs often\" (Guskin and Clement) Sept. 15, 2016, n\n\n\"Recidivism Watch: Trump's claim that millions of people voted illegally\" (Kessler) Jan. 24, 2017, n\n\n\"Republican redistricting is taking a beating in the courts (again)\" (Phillips) May 22, 2017, n\n\n\"Trump's voter fraud claims undermine the democratic process and his presidency\" (Balz) Jan. 25, 2017, n\n\n\"Virginia's breakthrough on rectifying an enormous injustice\" (Editorial Board) June 11, 2017, n\n\nThe Washington Post, Feb. 2017,\n\nThe Washington Post\/ABC News poll, , n\n\nWasserman, David, \u2013n\n\nWasserman, Harvey, n\n\n\"Watch: Hackers Demonstrate How to Crack Into Electronic Voting Machine in Minutes: Disturbing footage from the DEF CON 25 hacker convention\" (Friesdat), n\n\nWaters, Maxine,\n\nWeaver, Jeff, \u2013,\n\nWeiser, Wendy, n\n\n\"Were 2016 vote counts in Michigan and Wisconsin hacked? We double-checked\" (Mebane and Bernhard) June 6, 2017, n\n\n\"We're Supposed to Be a Democracy, But Half the GOP Is OK with Postponing 2020's Election: A party full of delusions exposes the dark underbelly of America\" (Rosenfeld), n\n\n\"What 6 Top Election Experts Are Saying About the Next Big Step in the 2016 Recount\" (Rosenfeld), n\n\nWhat Happened in Ohio? A Documentary Record of Theft and Fraud in the 2004 Election (Fitrakis, Rosenfeld and Wasserman), n\n\n\"What We Know and Don't Know about Election 'Hacks' in 2016\" (Becker), n\n\n\"When Does Political Gerrymandering Cross a Constitutional Line?\" (Liptak), n\n\nWhitlock, Liz, xv\u2013xvi\n\n\"Who Hacked the Election? Ad Tech Did. Through 'Fake News,' Identity Resolution and Hyper-Personalization\" (Albright), n\n\n\"Who We Are\" (EricStates.org), n\n\n\"Why Did the Founding Fathers Choose a Republic?\" (DeMichele), n\n\n\"Why I Will Not Cast My Electoral Vote for Donald Trump\" (Suprun), n\n\n\"Why Republicans Can't Find the Big Voter Fraud Conspiracy: If the last federal investigation is any guide, the answer is simple: It probably doesn't exist\" (Rab), n\n\nWikiLeaks, \u2013, \u2013, ,\n\nWikipedia.org, n, n, \u2013n, n, n\n\n\"Will Keith Ellison Move the Democrats Left? By running for the D.N.C. chair, the Minnesota congressman hopes to lead a populist opposition against Trump\" (Cunningham), n\n\n\"Will North Carolina Lead the Way to a New South? Progressives like Reverend Barber and candidates like Deborah Ross are hoping to assemble a multiracial coalition to defeat a conservative backlash\" (Walsh), n\n\n\"Will third time be a charm for those who want NC Supreme Court to invalidate election maps?\" (Blythe), \u2013n\n\nWills, Garry, n\n\nWilson, Woodrow,\n\nWines, Michael, \u2013n, n, n, n, n, n, n, n\n\nWinger, Richard,\n\nWisconsin Elections Commission,\n\nWisconsin recounts, \u2013\n\nWisconsin redistricting, \u2013n\n\n\"Wisconsin Strict ID Law Discouraged Voters, Study Finds\" (Wines), n\n\nWisconsin voter hijacks,\n\nWisconsin voter ID laws,\n\nWolf, Richard, , n\n\nwomen's rights and gerrymandering, \u2013\n\nWood, Erkia, \u2013n\n\n\"Would-be independents joining the American Independent Party could blame California's voter registration card\" (Myers), n\n\nWright, Dan, n\n\nWWIR, NBC29.com, \"Governor McAuliffe Provides Update on Restoration of Rights Numbers\" Jan. 18, 2017, n\n\nY\n\nYelton, Janet, \u2013\n\nYen, Hope, n\n\nYolo, Ted,\n\nyounger voting block, \u2013\n\nYouTube.com, \"Hacking The Diebold Machine\" (Felten), n\n\nZ\n\nZeleny, Jeff, n\n\nZuckerman-Parker, Michelle, \n","meta":{"redpajama_set_name":"RedPajamaBook"}}
+{"text":"\n\nThis navy SEAL is on a top-secret mission of seduction...\n\nDavid Holloway is used to hard assignments. Now he faces his most difficult yet: cozying up to the soft curves of Swan Jamison. He wants to know the island beauty...in every way. But romancing Swan as part of his SEAL team's mission tests his honor, even as his feelings become increasingly real. Soon he must make a choice: duty...or desire?\n\nA Westmoreland Legacy Novel\nPraise for New York Times and USA TODAY bestselling author Brenda Jackson\n\n\"Brenda Jackson writes romance that sizzles and characters you fall in love with.\"\n\n\u2014New York Times and USA TODAY bestselling author Lori Foster\n\n\"Jackson's trademark ability to weave multiple characters and side stories together makes shocking truths all the more exciting.\"\n\n\u2014Publishers Weekly\n\n\"There is no getting away from the sex appeal and charm of Jackson's Westmoreland family.\"\n\n\u2014RT Book Reviews on Feeling the Heat\n\n\"What is it with these Westmoreland men? Each is sexier and more charming than the one before.... Hot, sexy, smart and romantic, this story has it all.\"\n\n\u2014RT Book Reviews on The Proposal\n\n\"Jackson has a talent for creating the sexiest men and pairing them off against feisty females. This story has everything a hot romance should have.\"\n\n\u2014RT Book Reviews on Hot Westmoreland Nights\n\n\"Is there anything more irresistible than a man so in love with a woman that he's willing to give her what she believes is her heart's desire? The Westmoreland clan will claim even more fans with this entry.\"\n\n\u2014RT Book Reviews on What a Westmoreland Wants\nDear Reader,\n\nI am proud to be back with another Westmoreland Legacy novel. The Westmorelands exemplify what a strong family is all about when it comes to sharing love and giving support. Not only to each other\u2014that love and support extends to their friends, as well.\n\nI enjoy introducing you to Bane Westmoreland's friends who are his navy SEAL teammates. You've met Gavin \"Viper\" Blake in The Rancher Returns and Laramie \"Coop\" Cooper in His Secret Son. Now it's time for David \"Flipper\" Holloway to take center stage and flex his muscles.\n\nFlipper's story takes you to beautiful Key West, Florida, where he meets a woman who will not only catch his eye, but also capture his heart. However, the last thing Swan Jamison wants is to give her heart to another man who could break it again. The heat is definitely on with Flipper and Swan.\n\nGet ready for a romantic adventure where true love prevails in the end.\n\nHappy reading!\n\nBrenda Jackson\n\nAn Honorable Seduction\n\nBrenda Jackson is a New York Times bestselling author of more than one hundred romance titles. Brenda lives in Jacksonville, Florida, and divides her time between family, writing and traveling.\n\nEmail Brenda at authorbrendajackson@gmail.com or visit her on her website at brendajackson.net.\n\nBooks by Brenda Jackson\n\nHarlequin Desire\n\nThe Westmorelands\n\nZane  \nCanyon  \nStern  \nThe Real Thing  \nThe Secret Affair  \nBreaking Bailey's Rules  \nBane\n\nThe Westmoreland Legacy\n\nThe Rancher Returns  \nHis Secret Son  \nAn Honorable Seduction\n\nVisit her Author Profile page at www.Harlequin.com, or brendajackson.net, for more titles!\n\nJoin Harlequin My Rewards today and earn a FREE ebook!  \nClick here to Join Harlequin My Rewards  \nhttp:\/\/www.harlequin.com\/myrewards.html?mt=loyalty&cmpid=EBOOBPBPA201602010002\nTo the man who will forever be the love of my life, Gerald Jackson, Sr.\n\nTo all of my readers who asked for Flipper's story. This one is for you!\n\nTo the Brenda Jackson Book Club\/Facebook   \nfans. Over 4,000 strong and after fourteen   \nyears, you guys still rock!\n\nMany waters cannot quench love;   \nrivers cannot sweep it away.\n\n\u2014Song of Solomon 8:7\nContents\n\nPrologue\n\nChapter One\n\nChapter Two\n\nChapter Three\n\nChapter Four\n\nChapter Five\n\nChapter Six\n\nChapter Seven\n\nChapter Eight\n\nChapter Nine\n\nChapter Ten\n\nChapter Eleven\n\nChapter Twelve\n\nChapter Thirteen\n\nChapter Fourteen\n\nChapter Fifteen\n\nChapter Sixteen\n\nEpilogue\n\nExcerpt from _Reunited...with Baby_ by Sara Orwig\nPrologue\n\nThe Naval Amphibious _Base  \n_ Coronado, San Diego, California\n\n\"What kind of trouble have you gotten into?\"\n\nDavid Holloway, known to his Navy SEAL teammates as Flipper, glanced at the four men surrounding him. They were like brothers to him. More than once they'd risked their lives for each other and they would continue to have each other's backs, on duty or off. That bond was what accounted for the concerned looks on their faces. He wondered how they'd known he'd been summoned to the admiral's office.\n\n\"Let's hope I'm not in any trouble, Mac,\" Flipper said, rubbing a hand down his face.\n\nHe had to admit he was wondering what was going on, just like they were. Usually, you were only summoned to a meeting with the admiral when you were getting reprimanded for some reason, and he never got into trouble. At least he _rarely_ did. As the son of a retired SEALs commanding officer and the youngest of five brothers\u2014all Navy SEALs\u2014he knew better.\n\n\"Maybe there's an event on the base and he wants you to escort his daughter now that you're the single one among us,\" Coop said, grinning.\n\nFlipper didn't grin back. They'd seen Georgianna Martin, the admiral's twenty-three-year-old daughter. She was beautiful, but they'd heard the horror stories from other teammates who'd been ordered to take her out on dates. According to them, those evenings had been the dates from hell. The young woman was spoiled rotten, selfish as sin and had an attitude that sucked. That's why Flipper didn't find Coop's comment at all amusing. He hoped that wasn't why the admiral wanted to see him.\n\nIt didn't surprise Flipper that it was Mac who'd asked if Flipper had gotten into trouble. Thurston McRoy\u2014code name Mac\u2014was older than the other four men on the team, who had all started their careers as SEALs around the same time. Mac had been a SEAL five years before the rest of them. Mac seemed to like to think he was the big brother looking out for them, almost like he figured they couldn't take care of themselves. He was forever giving them advice\u2014even when they didn't ask for it.\n\nIn addition to Mac and Flipper, their SEAL team included Brisbane Westmoreland, code name Bane; Gavin Blake, whose code name was Viper; and Laramie Cooper, whose code name was Coop.\n\nFlipper checked his watch. \"Since I have a couple of hours to spare before meeting with the admiral, let's grab something to eat,\" he suggested.\n\n\"Sounds good to me,\" Bane said.\n\nLess than an hour later, Flipper and his four teammates shared burgers, fries and milkshakes at one of the most popular eating places on base. They decided to sit outside at one of the caf\u00e9 tables in the front instead of inside where it was crowded since it was such a beautiful May day.\n\nNo one brought up his meeting with the admiral again or the notion of him taking the admiral's daughter on a date. He was glad. Instead, the guys had more important things to talk about, namely their families.\n\nBane's wife, Crystal, had given birth to triplets last year and he had new photos to share, so they passed Bane's cell phone around.\n\nViper's wife, Layla, was expecting with only a few months to go before Gavin Blake IV would be born. Viper was excited about becoming a father, of course.\n\nLike Bane, Mac had plenty of photos to share; he was married and the father of four.\n\nAnd Coop had a two-year-old son he hadn't known about until he'd run into his old girlfriend about six months ago. They'd reconnected, gotten married and were now a happy family.\n\nEarlier in the week, the teammates had gotten word from their commanding officer that next week was the start of a four-month leave. For Flipper, that meant heading home to Dallas and he couldn't wait. His mother had a birthday coming up and he was glad he would be home to celebrate.\n\n\"I don't care what plans you all are making for your leave, just as long as you remember my mom's birthday celebration. I understand you not showing up, Viper, with a baby on the way. The rest of you guys, no excuses.\"\n\n\"We hear you,\" Bane said, grinning. \"And we will be there.\"\n\nWhen Viper ordered another hamburger, everyone teased him about being the one to eat for two instead of his wife. And then everyone talked about what they planned to do with their four months off.\n\nIt was two hours later when Flipper walked into the admiral's office. He was surprised to find Commanding Officer Shields there as well. Flipper saluted both men.\n\n\"At ease. Please have a seat, Lieutenant Holloway.\"\n\n\"Thank you, sir,\" he said, sitting down. He was used to being under his commanding officer's intense scrutiny, but there was something in the sharp green eyes of Admiral Norris Martin that was making him feel uncomfortable.\n\n\"You come highly recommended by your commanding officer here, Lieutenant Holloway. And the reason I asked to meet with you is that we need you. Your country needs you.\"\n\nFlipper was happy to step up. He was a Navy SEAL, and the reason he'd enlisted, like his father and brothers, was to protect his country. \"And what am I needed to do, sir?\" he asked.\n\n\"Our investigators have provided intelligence and a preliminary report that says acts of espionage are happening in Key West. Someone is trading valuable government secrets to China.\"\n\nFlipper didn't respond immediately.\n\nThe one thing he hated was a traitor, but he'd discovered that for the right price, a number of American citizens would perform acts of treason. He understood that. However, what he didn't understand was why he'd been singled out for this meeting. He was part of a SEAL team. He didn't work in naval intelligence.\n\nConfusion must have shown on his face because Admiral Martin continued, \"The report was given to me, but I don't believe it.\"\n\nFlipper raised a brow. \"You don't believe a report that classified documents are being traded in Key West, sir?\"\n\n\"Oh, I believe all that, but what I refuse to believe is that this suspect is guilty of anything.\"\n\n\"Is there a reason why, sir?\"\n\n\"Here is the information,\" said Commanding Officer Shields, speaking for the first time as he handed Flipper a folder.\n\nFlipper opened it to find a picture of a very beautiful woman. She looked to be around twenty-four, with dark, sultry eyes and full, shapely lips. Then there was her mass of copper-brown spiral curls that flowed to her shoulders, crowning a cocoa-colored face. A pair of dangling gold earrings hung from her ears and a golden pendant necklace hung around her neck.\n\nHe knew he was spending too much time studying her features, but it couldn't be helped. The woman was strikingly beautiful.\n\nReluctantly he moved his gaze away from her face to check out the background of the photo. From the tropical vegetation captured by the photographer, she seemed to be on an island somewhere. She stood near a body of water that showed in the corner of the eight-by-ten photo. Scribbled across the bottom were the words:\n\n_Miss you, Godpop 1_\n\n_Love, Swan_\n\nSwan? It was an unusual name, but it fit.\n\nHe moved to the next document in the file. Attached to it was a small family photo that showed a tall Caucasian man with sandy-brown hair and brown eyes standing beside a beautiful woman who closely resembled Swan. Her mother. In front of the couple was a beautiful little girl who looked to be around eight.\n\nFlipper studied the child's face and knew that child had grown up to be the gorgeous woman in the first photo. The shape of her face was the same, as were her eyes. Even as a child, she'd had long curly hair.\n\nThe family photo was clipped to a profile of the young woman. As he'd guessed, she was twenty-four. Her name was Swan Jamison. She was an American, born in Key West. Presently, she owned a jewelry store on the island. That was all the information the document provided.\n\nFlipper lifted his gaze to find his commanding officer and the admiral staring at him. \"I assume this is the person naval intelligence believes is the traitor.\"\n\n\"Yes,\" Admiral Martin said. \"She's my goddaughter. I am Godpop 1.\"\n\n\"She's my goddaughter as well,\" added Commanding Officer Shields. \"I am Godpop 2.\"\n\nFlipper's gaze moved from one man to the other. \"I see, sirs.\"\n\nAdmiral Martin nodded. \"Her father was part of our SEAL team and our best friend. His name was Andrew Jamison.\"\n\nFlipper had heard that Commanding Officer Shields and Admiral Martin were part of the same SEAL team a number of years ago.\n\n\"Andrew was the best. He lost his life saving ours,\" said Commanding Officer Shields. \"He didn't die immediately, and before he died, he made us promise to look after his wife, Leigh, and his daughter, Swan.\" The man paused and then said, \"Over twenty-eight years ago, when we were taking some R & R in Jamaica, Andrew met Leigh, who was a Jamaican model. They married a year later, and he moved her to Key West, where our team was stationed. After Andrew was killed, Leigh returned to Jamaica. When Swan graduated from high school, she returned to the Keys and moved into her parents' home.\"\n\n\"How old was she when her father was killed?\" Flipper asked.\n\n\"She was fifteen,\" Admiral Martin said. \"Swan was close to her dad. Leigh was so broken up over Andrew's death that she didn't want to live in the States without him, which was why she returned to Jamaica. She passed away two years ago.\"\n\nFlipper's commanding officer then took up the tale. \"Leigh sent for us before she died of stomach cancer, asking us to look out for Swan after she was gone. We would have done that anyway, since we always kept in touch with both Leigh and Swan. In fact, Swan rotated summers with us and our families even after Leigh returned to Jamaica. We took our roles as godfathers seriously. We were even there when Swan graduated from high school and college.\"\n\n\"Did Swan have any American grandparents?\" Flipper asked.\n\nHe saw both men's lips tighten into frowns. \"Yes. However, her paternal grandparents didn't approve of their son's marriage to Leigh,\" said Commanding Officer Shields.\n\n\"So they never accepted their granddaughter.\" It was more of a statement than a question.\n\n\"No, they never did,\" Admiral Martin confirmed. As if it was a topic he'd rather change, the man added, \"We've been given some time to find out the truth, but not much. Luckily, Swan's Godpop 3 has a high-level position at naval intelligence. Otherwise, we wouldn't know about the investigation. We have thirty days to prove Swan is not a traitor and identify the person who is. That's where we need your help. Instead of releasing you to go home as we're doing for the other members of your team, we are assigning you to a special mission, Lieutenant Holloway. You are being sent to Key West.\"\nOne\n\nKey West, Florida\n\nSwan Jamison was beside herself with excitement as she opened the huge box on her desk. Although it contained only her jewelry-making supplies, the package served as affirmation that while rebuilding was still taking place in certain areas, the majority of the island had recovered from the hurricane that had hit eight months ago.\n\n\"Anything for me?\" Rafe asked, sticking his head through the office door.\n\nHer shop was in a very trendy area so she could capitalize on the tourists visiting the island. To help with high operating costs, she leased out one of the large rooms in the back. Rafe was her tenant, who'd converted the back room into a tattoo shop. On some days, he got more customers than she did.\n\n\"Nothing for you, Rafe, just supplies for me.\" She checked her watch. \"You're early today.\" Usually he didn't open up until noon.\n\n\"I have a special appointment at ten thirty and I need to ready my ink.\" And then he was gone. Rafe didn't say a whole lot except to his customers.\n\nThe door chime alerted her that _she_ had a customer. Jamila, who worked part-time and usually only in the mornings, had taken time off for a day of beauty\u2014hair, nails, pedicure, bikini wax, the works. Her boyfriend worked on a cruise ship that was due in port tomorrow. Swan was happy for Jamila and happy for herself as well. The cruise ships always brought in tourists who wanted to purchase authentic handmade jewelry.\n\nShe walked out of her office as a man perused her jewelry display case near the door. That was good. While he checked out her jewelry, she would check him out.\n\nHe had a nice profile. Tall, broad shoulders that looked good in a T-shirt and a pair of muscular thighs that fit well in his jeans. He had diamond-blond hair that was neatly trimmed and his hands were the right size for his body.\n\nThere was something about the way he stood, straight and tall, that all but spelled out _military man_. And the way his legs were braced apart, as if he had to maintain his balance even on land, spelled out _navy_.\n\nToo bad. She didn't do military men. In all honesty, lately she hadn't done men at all. Too busy.\n\nAnd then there was the issue of Candy's divorce. Swan knew she shouldn't let what had happened to her best friend darken her own view, but Swan was known to claim whatever excuse suited her and that one did at the moment.\n\nAnd speaking of the moment, she had looked her fill. She needed to make her first sale of the day. \"May I help you?\"\n\nHe turned and looked at her, and every cell in her body jolted to attention.\n\nWow! She'd seen blue eyes before, but his were a shade she'd never seen. They were laser blue; the intense sharpness of the pupils captured her within their scope. And his features... Lordy! The man had such gorgeous bone structure! There was no way a woman had ever passed by him and not taken a second look. Even a third, while wiping away drool.\n\n\"Yes, you can help me.\"\n\nAnd why did he have to sound so good, too? The sound of his voice\u2014a deep, husky tone\u2014made her throat go dry.\n\n\"All right,\" she said, walking over to him. She knew she had to get a grip. Her store had been closed for two months due to the hurricane, and now that the tourists were returning, she needed to catch up on sales.\n\n\"And how can I help you?\" She didn't miss the way he was looking at her. She saw interest in his eyes. There was nothing wrong with that. She took pride in her appearance because she had been raised to do so. Leigh Rutledge Jamison, who'd been a Jamaican model, had taught her daughter that your appearance was everything.\n\nPain settled in Swan's heart. She missed her mom so much.\n\n\"I'm looking for a gift for someone.\"\n\nSwan nodded as she came to stand beside him. Not only did he look good and sound good, but he smelled good as well. She glanced down at his hand and didn't see a wedding ring. He was probably buying a gift for his girlfriend or soon-to-be fianc\u00e9e.\n\n\"What do you have in mind?\"\n\n\"What do you suggest?\" he asked her.\n\n\"Well, it depends,\" she said, looking into those gorgeous eyes.\n\n\"On what?\"\n\n\"What the person likes. I make jewelry from stones, but as you can see, there are a number of them, in various shades, colors and styles.\"\n\nHe smiled and Swan felt a tingling in the pit of her stomach when a dimple appeared in one of his cheeks. \"I honestly don't know what she likes. Her tastes change from year to year. It's hard to keep up.\"\n\nSwan nodded. \"Oh. Sounds like the two of you have known each other for a while.\"\n\nHis smile widened even more. \"We have. I would have to say I've known Mom all my life.\"\n\n\"Your mom?\"\n\n\"Yes. Her birthday is next month. I was passing by your shop and thought I would drop in to see what you had.\"\n\n_A racing heart for starters_ , Swan thought. So the woman he was thinking about buying jewelry for was his mother. \"Well, I'm glad you came in. Let me show you what I have.\"\n\n\"All right. There looks to be a lot of nice pieces.\"\n\nShe appreciated the compliment. \"Thanks. I made most of them myself.\"\n\n\"Really? Which ones?\"\n\nShe led him to the area set aside for Swan Exclusives. \"These. Most of the stones come from India, Argentina and Africa.\"\n\nHe leaned in to look. \"You did an excellent job.\"\n\nWhoever said flattery, especially coming from a good-looking man, would get you anywhere knew just what they were talking about. \"Thank you.\"\n\n\"I'm David, by the way. David Holloway.\" He offered her his hand.\n\nShe took it and tried to ignore the sensations that suddenly flowed through her from the contact. \"Nice to meet you, David.\" She quickly released his hand. \"And I'm Swan.\"\n\n\"The name of the shop.\"\n\n\"Yes.\"\n\n\"It's a unique name.\"\n\n\"Yes, my parents thought so. On their first date, my father flew Mom from Jamaica to New York to see _Swan Lake_.\"\n\n\"Some date.\"\n\n\"Yes, he was trying to impress her.\"\n\n\"I take it he did.\"\n\nSwan chuckled. \"Yes, because he actually flew them there. He had his pilot's license.\"\n\n\"Now I'm impressed.\"\n\nShe didn't like bragging about her father but there were times when she just couldn't help it. \"He served in the air force\u2014that's where he learned to fly. And then he went into the navy after deciding he wanted to be a SEAL. That's when he met Mom, while he was a SEAL. She hadn't known about his stint in the air force until the night he rented a plane to fly them to New York.\"\n\nWhy was she telling him all this? Usually she wasn't chatty. \"What about this one?\" she asked as they moved to another glass case. \"I call this piece _Enchantment_.\"\n\n\"Why?\"\n\n\"Look at it,\" she suggested, leaning closer to the glass. He followed suit. \"This is one of my favorite pieces because the teardrop gemstone necklace is pretty similar to my very first piece.\" No need to tell him that she'd made that one for her own mother.\n\n\"It is beautiful.\"\n\nSomething in his tone made her glance over at him, and she found him staring at her and not at the jewelry in the case. His eyes held her captive and their gazes met for a minute too long before she broke eye contact with him.\n\nShe swallowed. \"So are you interested...in this piece?\" She wanted to ignore the way her stomach seemed to be filled with all kinds of sensations, but she could not.\n\n\"I'm interested in a lot of pieces, Swan, but I'll start with this one.\"\n\n* * *\n\nSwan Jamison was even more beautiful than the photograph he'd seen last week.\n\nThe photographer hadn't fully captured the rich creaminess of her skin. And the shade of red lipstick she wore today seemed to make her lips plumper, more well-defined. Luscious.\n\nHe had read the dossier on her. He knew his commanding officer and Admiral Martin were operating based on a personal connection with her. He was not. If Miss Jamison was guilty of any wrongdoing, he would find out. And if she wasn't the one handing out classified data to China, then he would discover who was.\n\n\"So you want to buy this particular piece?\"\n\nHer question brought his thoughts back to the present. \"Yes.\"\n\n\"Wonderful. I think your mother will like it.\"\n\n\"I'm sure she will. What about earrings?\"\n\nShe lifted a brow. \"Earrings?\"\n\n\"Yes. Do earrings come with the necklace?\"\n\n\"No, but I can make you some.\"\n\nHe'd been hoping she'd say that. \"When?\"\n\n\"It will take me a couple of days. The cruise ship docks tomorrow, so the shop will be busy. Two days from now will work for me, unless you need them sooner.\"\n\n\"No, I can wait. My mother's birthday is next month.\"\n\nHe would have an excuse to return to her shop.\n\nFlipper watched her open the case and pull out the necklace. He knew his mother was going to love it.\n\n\"If you don't mind, please complete this ticket,\" she said. \"And I will need full payment for the earrings before I make them.\"\n\n\"That's no problem,\" he said, taking the document from her.\n\nAfter he completed the form, he handed it back to her. She glanced at it. \"So you're from Texas?\"\n\n\"Yes. Dallas. Ever been there?\"\n\n\"Yes, once. I thought it was a nice city.\"\n\n\"It is. I was born and raised there.\"\n\n\"And what brought you to Key West?\" she asked him.\n\n\"Work, at least for the next thirty days.\" That wasn't a total lie.\n\n\"Hurricane relief?\"\n\n\"Something like that.\"\n\n\"You're military?\"\n\n\"At one point but not now.\" He would let her think he was no longer military.\n\n\"I knew immediately.\"\n\nHe lifted a brow. \"How?\"\n\nShe shrugged. \"Military men are easily recognized, at least by me.\"\n\n\"Because your dad is military?\"\n\n\"He _was_ military. Dad died years ago in the line of duty.\"\n\n\"I'm sorry.\" Flipper was always sorry whenever a fellow soldier lost their life.\n\n\"Thank you. Your package will be ready in two days, David. Your mobile number is on the form you completed. If I get to it sooner, I will call you.\"\n\n\"Two days is fine. I'll be back.\"\n\n\"'Bye, David.\"\n\n\"'Bye, Swan.\" He then turned and walked out of the shop.\n\nAs much as he wanted to invite her out to lunch today, he knew he couldn't rush things. He needed to earn her trust, even though he had less than thirty days to prove her innocence and determine who had no qualms about making her look guilty.\n\n* * *\n\nSwan was cheerful that night as she let herself into her home. Sales today had been better than normal. A tour group from New York had converged on the island and they'd come to spend money. She'd been happy to oblige.\n\nOpening a jewelry shop had been a risky business move, but one that had paid off. She'd earned a degree in business management from the University of Miami and returned to the island after college to work as a manager at one of the elite hotels on the island. She'd enjoyed her job but had felt something was missing in her life. She hadn't been using her jewelry-making talent.\n\nShe'd promised her mother on her deathbed that she would find a way to use that talent.\n\nEven after taking care of all her mother's funeral expenses, there had been more than enough money left to buy a little storefront. It had been a good investment because of its location. Some days were busier than others. This had been one of those busy days.\n\nNow she was ready to wind down for the evening. She pulled her hair back in a ponytail and eased her feet into her favorite flats before heading to the kitchen for a glass of wine. As she did so, she couldn't help but think about her first customer of the day.\n\nDavid Holloway.\n\nHe was a cutie, she had to give him that. And the memory of those eyes had stayed with her most of the day.\n\nDavid Holloway had come into her shop to buy a birthday gift for his mother. How sweet. His mother was lucky. A lot of men didn't even remember their mothers' birthdays. She'd dated quite a few of those men and never developed lasting relationships with any of them. She figured if a man didn't treat his mother right, then there was no hope for a girlfriend.\n\nAs she opened the French doors to step out on the patio, she again remembered those blue eyes and how she'd felt whenever she'd looked into them. No man's eyes had ever made her feel that way before.\n\nThe effect was unsettling.\n\nOkay, so what was wrong with her? Cutie or no cutie, she normally didn't get caught up over a man. She dated when it suited her, but she would admit that no one had suited her lately. At least not since her best friend, Candy, had left Key West to go live in Boston. Candy had refused to live on the island with her ex and his new wife\u2014the one he'd married before the ink had even dried on the divorce papers.\n\nRefusing to dwell on how shabbily Donald Knoll had treated Candy, Swan looked out at the water. It was calm tonight. When she had evacuated due to the hurricane, she hadn't known what to expect when she returned. Between her home and her shop, there had been some damage, but not as much as she'd feared.\n\nThe thought of losing her home had been devastating. This was where her father had brought her mom after they'd married. This home held so many childhood memories\u2014of her father leaving on his missions as a Navy SEAL, of how happy she and her mother would be whenever he returned.\n\nBut then he hadn't returned.\n\nSwan felt a knot in her throat as she recalled that day. She'd never seen that sparkle in her mother's eyes again. Swan recalled her mother telling her once that when you met a man who could put that sparkle in your eyes, then you knew he was a keeper.\n\nSwan often wondered if she would ever find her keeper.\n\nShe had plenty of time. Besides, she needed to rethink her opinion about men first. If what Don had done to Candy wasn't enough to keep her single, all Swan had to do was remember William Connors, the businessman she had met while working at the hotel.\n\nAt the time, he had convinced her he was a bachelor without a care in the world but claimed that he wanted to make her Mrs. William Connors one day.\n\nFor some reason, Candy hadn't trusted him. She had a friend who worked for a private investigator check him out. Swan had been devastated when the investigation revealed there was already a Mrs. William Connors, along with three Connors children.\n\nWilliam had been playing her. He had been a lesson well learned. Her only regret was that she'd shared her body with him. She'd been young, naive and impressionable. He had been her first and he should not have been.\n\nShe was not naive now and she went into relationships with caution and even a little mistrust. Her mother once told her that being mistrustful wasn't a good thing. Swan knew she would have to learn how to trust again.\n\nShe took another sip of wine. Unfortunately, she hadn't gotten there yet.\n\n* * *\n\n\"So how did things go, Flipper?\"\n\n\"Have you met her yet?\"\n\n\"Does she have a traitorous face or just a pretty one?\"\n\n\"Do you think you'll be able to prove she's innocent?\"\n\nFlipper heard the questions coming at him nearly all at once. While unpacking, he had placed his mobile call on speaker to engage in a five-way conversation with his SEAL teammates.\n\n\"I think things went rather well, Mac. And yes, I met Swan Jamison today, Viper. I went into her jewelry store to purchase Mom a birthday gift.\"\n\nFlipper eased open the dresser drawers to place his T-shirts inside. \"She doesn't have a traitorous face or just a pretty one, Coop. The woman is simply gorgeous. Beautiful beyond belief. And yes, I hope to prove she's innocent, Bane, because Commanding Officer Shields and Admiral Martin truly believe she is.\"\n\n\"What do you believe?\" Viper asked.\n\nFlipper leaned against the dresser for a minute and thought about Viper's question. \"Too early to tell.\"\n\n\"Did you ask her out on a date?\" Coop wanted to know. They could hear Coop's two-year-old son, Laramie, chattering in the background.\n\n\"No, not yet.\" Flipper's attraction to her had been instant. He'd felt it the moment he looked into her face. Discussing her now wasn't helping matters. All it did was force him to recall what a beautiful woman she was\u2014a woman he would have to spend time with in order to discover the truth.\n\n\"Then how do you plan to see her again if you don't ask her out?\" Mac wanted to know, interrupting Flipper's thoughts.\n\n\"I ordered a pair of earrings to go with the necklace I bought for Mom. She has to make the earrings and I'll make my move when I pick up my purchases in two days.\"\n\n\"And if she turns you down?\" Viper asked.\n\n\"Not an option. I now have less than thirty days to get this all straightened out.\"\n\n\"We should be there with you, watching your back,\" Bane said.\n\n\"No, you guys are just where you need to be, which is home with your families. I've got this.\"\n\n\"Well, some of our families don't appreciate us being home,\" Mac grumbled.\n\nFlipper rolled his eyes. They'd all heard the complaints from Mac before. After every extended mission, their teammate went home to an adjustment period, where he would have to get to know his wife all over again and reclaim his position as head of the house. Sometimes the adjustment didn't go over well. Mac had a strong personality and so did Mac's wife, Teri. \"Do we have to send both you and Teri into the time-out corners?\"\n\n\"Hell, I didn't do anything,\" Mac exclaimed.\n\nFlipper chuckled. \"Yeah, right. You better get your act together, Mac. No other woman is going to put up with your BS.\"\n\n\"Whatever. So what did you notice about the place today?\"\n\nMac was changing the subject and Flipper decided to let him. \"Everything matched the architectural report I was given. Even with the repairs due to the hurricane, there were no major changes. Front door. Back door. High windows. Glass storefront. No video cameras outside. There are several rooms in back. One is being used as a tattoo parlor. I didn't see the person who runs it. I think I'll go out tonight and do a little more investigating,\" he said, sliding into a black T-shirt.\n\n\"Be careful, Flipper,\" Viper said. \"Although you might not have seen any video cameras, that doesn't mean there aren't any.\"\n\n\"I know. That's why I'm wearing my Pilf gear.\"\n\nEverybody knew how much Flipper liked digital technology. In addition to all the futuristic developments the military used, Flipper had created a few of his own high-tech gadgets behind the scenes. Some had been so impressive the federal government had patented them as Pilf gear to be used by the military. Pilf was the name Flip spelled backward. On more than one occasion, Flipper had been offered a position with the Department of Defense's Research and Development Department and had turned down each offer, saying he loved being a Navy SEAL more.\n\n\"We don't give a damn if you plan to parade around naked tonight, Flipper. Be careful.\"\n\nHe knew Mac was in his big-brother mode. \"Okay, Mac. I hear you and I will be careful.\"\n\n\"Call to check in when you get back to the hotel tonight,\" Bane said.\n\n\"It will be late and I wouldn't want to wake up any babies, kids or a pregnant woman. I'll text everyone.\"\n\nA short while later, wearing undetectable military gear under his clothing, Flipper left his hotel using the stairs.\nTwo\n\nTwo days later, Swan didn't leave the shop for lunch. Instead she accepted Jamila's offer to bring her something back from the sandwich shop on the corner. Although she'd tried convincing herself her decision to hang around had nothing to do with the fact that David Holloway would be returning today to pick up his items, she knew it did.\n\nAnd her anticipation was so bad that every time the door chimed, her heartbeat would kick up a notch, only to slow back down when someone other than him walked in. She checked her watch. The shop would be closing in an hour. What if he didn't make it before closing time? What if...?\n\nThe door chimed, and her heart nearly stopped when David Holloway walked in.\n\nShe'd told herself the man hadn't _really_ looked as good as she remembered from that first day, but now she saw that he did. In fact, today he looked even better than she remembered. Maybe it had something to do with the unshaven look. Men with a day-old beard had sex appeal. But it could also be his tan, which indicated he'd probably spent the last couple of days lying in the sun.\n\nIf he'd been at the beach, there was a good chance he hadn't been there alone. But didn't he say he was in the Keys working?\n\nWhy did she care?\n\nShe quickly dismissed all those questions from her mind as she continued to watch him walk toward her in a strut that had blood rushing through her veins. His blond hair and blue eyes seemed brighter against his tanned skin. He was deliciousness with a capital _D_.\n\nBut then that capital _D_ could also stand for _dangerous_ if she wasn't careful. Or it could stand for _delusional_ if she didn't get control of her senses. Right now, she would play it safe and claim the capital _D_ stood for _David_. She couldn't allow herself to think any other way for now, no matter how tempting.\n\nShe smiled. \"Hello, David.\"\n\n\"Hi, Swan.\"\n\n\"Your tan looks nice.\"\n\nHe chuckled. \"So does yours.\"\n\nShe grinned. \"Yes, but mine's permanent.\"\n\n\"I know and I like it.\"\n\nShe didn't say anything to that because she understood what he was implying. He was letting her know he had no problem with interracial dating. She didn't have a problem with it either. Neither had her father, although his family had had conniptions about his marriage to Swan's mother. She pushed that thought to the back of her mind, refusing to dwell on an extended family that had never accepted her or her mother.\n\nShe reached behind the counter and retrieved a box. \"I hope you like the way the earrings came out.\" She opened it to show him the final earrings.\n\n\"Wow!\" He ran his finger over the stone that came closest to matching the color of his eyes. \"You're very gifted.\"\n\n\"Thank you, and I believe your mother will love them.\"\n\n\"I'm sure she will. I think I've outdone my brothers this time.\"\n\nShe closed the box and placed it, along with the one containing the necklace, into a shopping bag. \"You have brothers?\"\n\n\"Yes, four of them. I'm the youngest.\"\n\n\"My goodness. Any sisters?\"\n\n\"Not a one. Three of my four brothers are married, so I have sisters-in-law. They are the best.\"\n\n\"And the fourth brother is still single?\"\n\n\"He's divorced but has a beautiful little girl. And she's my parents' only granddaughter. They have six grandsons.\"\n\n\"Sounds like a nice family. Is your father still alive?\"\n\n\"Yes, Dad is still alive. He and Mom own a medical supply store.\"\n\nShe nodded as she offered him the bag. \"Here you are, David. Thanks again for your business.\"\n\nHe accepted the bag. \"Thanks. Now that this is taken care of, there's something I want to ask you, Swan.\"\n\nShe lifted a brow. \"What?\"\n\n\"Would you go out to dinner with me tonight?\"\n\n* * *\n\nNormally Flipper was good at reading people, but he was having a hard time reading Swan. He definitely needed to remedy that. Although both Commanding Officer Shields and Admiral Martin were convinced of her innocence, the jury was still out for him. He had to remain impartial and deal with the facts, not speculations.\n\nFor two nights, he'd searched the area around her shop. Getting inside without triggering her alarm hadn't been easy, but he'd done it. Once he'd picked up the location of the interior security cameras, it was a small matter to make sure he stayed out of their range and within a certain perimeter until he could deactivate them and do what he needed to.\n\n\"Go to dinner with you?\"\n\n\"Yes.\"\n\nShe was apparently mulling over his invitation in her mind and he would give her time to do that. He had no problems studying her while he waited for her answer. Today she looked even prettier than the other day. He figured it had to be the lighting in this place.\n\n\"Yes, David. I'll go to dinner with you. You name the restaurant and I'll meet you there.\"\n\nShe wasn't going to give him her address and he had no problem with her being cautious. Little did she know he already knew where she lived and had visited yesterday while she'd been here at her shop. She had a beautiful home on the ocean. Inside it was as neat as a pin with no clutter. She'd even made up her bed before leaving.\n\n\"I noticed a restaurant off the pier. Summer Moon. I've heard only good things about it since I've been here.\" And he knew the place was within walking distance from her home.\n\n\"Everything you've heard is true. Summer Moon is fabulous and one of my favorite eating places. I'd love to join you there. What time?\"\n\n\"What about seven? Will that be a good time for you?\" He figured since it didn't get dark until close to nine, he wouldn't have to worry about her walking to the restaurant in the dark. After dinner, he would walk her home or put her in a cab regardless of the fact that she lived only a few blocks away.\n\n\"Seven is perfect.\"\n\n\"Good. I'll see you then.\"\n\n* * *\n\nSwan watched him walk out of the shop.\n\nDavid had the kind of tush that made a woman want to squeeze it...after doing all kinds of other things with it.\n\nShe jumped when fingers snapped in her face. Frowning, she looked at Jamila. \"What did you do that for?\"\n\n\"To keep you from having an orgasm in the middle of your shop.\"\n\nSwan rolled her eyes. Jamila, the attractive twenty-two-year-old green-eyed blonde, evidently thought reaching a climactic state was that easy. \"It would take more than ogling a man for that to happen, Jamila.\"\n\n\"I don't know. Your eyes were about to pop out of their sockets and your breathing sounded funny.\"\n\n\"You're imagining things.\"\n\n\"Denial can be good for the soul, I guess. So who is he?\"\n\nSwan and Jamila had more than an employer-and-employee relationship. Their friendship had started when Jamila first moved to the island a couple of years ago and patronized Swan's. It didn't take long to discover that Jamila liked nice things and decided Swan's was one of her favorite places to shop. Last year, Jamila had been looking for work after she lost her job as a day cruise ship captain.\n\nAs far as Swan was concerned, it hadn't been Jamila's fault when an intoxicated customer had tried coming on to her and she'd kicked him in the balls. Surgery had to be performed and the man had sued the ship company. They'd settled out of court but not before firing Jamila for all the trouble she'd caused.\n\nJamila had gotten an attorney herself so she could not only sue her former employer for an unfair firing but also sue the intoxicated customer. To avoid negative publicity, her former employer wanted to settle out of court with her as well. The intoxicated customer was also trying to settle since the woman he'd been with on the ship hadn't been his wife. If things worked out in Jamila's favor, she wouldn't need a job at Swan's much longer.\n\n\"He is a customer who came into the shop a couple of days ago to buy a gift for his mother.\"\n\n\"His mother and not his wife?\"\n\n\"He says his mother.\"\n\nJamila snorted. \"Men lie all the time.\"\n\nHow well she knew, Swan thought. Then she wondered why Jamila was men-bashing today. This wasn't the first comment of that type she'd made since arriving to work. Her boyfriend had come to town a couple of days ago with the cruise ship, right? So what was going on?\n\nSwan decided not to ask. She didn't want to hear another sad story about a man that would ruin her date tonight with David. It was a date she was definitely looking forward to. She figured going out to dinner with him wouldn't be risky as long as she kept things in perspective.\n\nShe knew what could happen if she let her guard down when it came to a man.\n\n* * *\n\nFlipper deliberately arrived at Summer Moon early so he could see when Swan arrived. His stomach felt floaty the moment she turned the corner from the street where she lived.\n\n_Be still, my...everything_.\n\nShe was wearing a printed sundress and a pair of high-heeled sandals, but what caught his attention\u2014and was still holding it tight\u2014were her long shapely legs that seemed to go on forever. He would love to see where they stopped under that dress. He forced that thought to the back of his mind.\n\nBut the closer she got, the more that thought wiggled back to the forefront. He shouldn't let it. He was on assignment and she was the subject of an investigation. He shouldn't see her as temptation. Letting his guard down around her could be a dangerous and costly mistake. He had to keep his head screwed on straight, no matter how innocent she seemed and how beautiful she was, and she was definitely one gorgeous woman.\n\nMen, even some with female companions, were giving Swan second looks, and Flipper tried to downplay his anger. He had no right to be upset about other men checking her out when he was checking her out himself. The best thing to do to control his crazy reaction was to stop looking at her, so he glanced down at his bottle of beer and thought about the reports he'd finished reading a short while ago on her employee and her tenant.\n\nJamila Fairchild had worked for Swan for a year. He knew all about her former job as a captain of a day cruise ship, why she'd gotten fired and her litigation against not only her former employer but also the man who'd caused the ruckus in the first place. Naval intelligence hadn't left any stone unturned in Ms. Fairchild's report and she'd come up clean. Flipper would verify that she was.\n\nThen there was Rafe Duggers, the tattoo artist. Although his parlor was located inside Swan's shop, there was a back door for his customers to use without entering through the jewelry shop. Flipper hadn't gotten a chance to look around the tattoo parlor and he intended to do another visit in a few days. Rafe was too squeaky-clean to be true.\n\nNo wonder naval intelligence was trying to point the finger at Swan. After all, it was her shop and they had somehow traced activity as originating there. But how? When? He hadn't found anything.\n\nHe had searched Swan's office, the small kitchen in the back, the bathrooms and another room that she used as a workshop where she made her jewelry. He'd come up with nothing, even after checking out her computer. So what were the grounds for accusing her?\n\nFlipper's mind flicked back to Swan and he stood when the waiter escorted her to his table. \"Hello, Swan. You look nice.\"\n\n\"Thanks and so do you. I was trying to be early and you still beat me here,\" she said, sitting down across from him.\n\n\"I was thirsty,\" he said, sitting back down and indicating the beer. Now that she was here and sitting directly across from him, he was more than thirsty. If he wasn't careful, he could have a full-fledged attack of desire. She had a pair of beautiful shoulders and her skin appeared soft and smooth to the touch.\n\nThen his mind drifted to wanting her and he quickly snatched it back. \"You walked here. Does that mean you live close by?\" he asked, deciding it was best to keep the conversation moving.\n\n\"Yes, not too far,\" she said. He knew she was deliberately being evasive.\n\nThe waiter handed him another beer and gave them both menus. \"What would like to drink, miss?\" the waiter asked her.\n\n\"A glass of Moscato please.\"\n\nWhen the waiter left, she glanced over at Flipper before picking up her menu. \"You're not working so hard that you're not enjoying the Keys, are you?\"\n\n\"I'm doing a bit of both. I admit the ocean is beautiful tonight.\"\n\nShe smiled. \"I think it's beautiful every night.\"\n\nHe nodded as he took another sip of his beer, straight from the bottle. \"So are you a native or a transplant?\"\n\n\"A native. I was born and raised right here on the island in the same house I live in now. My mother never made it to the hospital before I was born.\"\n\nHe raised a brow. \"She didn't?\"\n\n\"No. Mom came from a part of Jamaica where the belief was that when it comes to delivering a baby, a midwife is better than a medical doctor. My father promised to find her a midwife here. Otherwise she would have insisted that I be born in Jamaica and he didn't want that. He wanted me born in America.\"\n\n\"So he was able to find a midwife?\"\n\n\"Yes, but I was born a few weeks early and the midwife wasn't here.\"\n\n\"So who delivered you?\"\n\n\"My dad, with the help of three of his closest military friends. They were stationed at the base here and were visiting, watching a football game at the time. Needless to say, over the years I've gotten four different versions of what happened that night. My mother didn't remember a thing other than it took four men to deliver me. Although Godpop 1 claims my father passed out trying to cut the umbilical cord.\"\n\nFlipper laughed. He then asked, \"Godpop 1?\"\n\n\"Yes, my father's three closest friends, the ones who assisted that night, became my godfathers. That's how I distinguish them. Godpop 1, Godpop 2 and Godpop 3.\"\n\nFlipper nodded. No wonder the three men felt such strong ties to her. \"You're lucky to have three godfathers. I don't have a one.\"\n\n\"Yes, I'm lucky,\" she said, after the waiter set the glass of wine in front of her. \"They were my and Mom's rocks after we lost Dad, especially when my grandparents showed up at the funeral trying to cause problems.\"\n\nThen, as if she realized she might have shared too much, she asked, \"So what do you plan to order?\"\n\n* * *\n\nSwan thought David had picked the right place for them to have dinner. When he asked for recommendations on what to order, she suggested Summer Moon's crab cakes and, as usual, they were delicious. The mango salad was superb, and after dinner they enjoyed listening to the live band.\n\nWhen the band played their last song, she glanced over at David to discover him staring at her. The intensity in his gaze nearly scorched her and she took a sip of her wine. \"Thanks for dinner, David.\"\n\n\"Thank you for accepting my invitation. The place is about to close. Are you ready to go?\" he asked her.\n\n\"Yes.\" Because she knew he would suggest that he walk her home, she added, \"If you still have a little bit of energy, I'd like to treat you to something.\"\n\nHe lifted a brow. \"What?\"\n\n\"A laser show that officially kicks off the summer season. It's a short walk from here.\" Since it was in the opposite direction from where she lived, she would have no problem catching a cab back later\u2014alone.\n\nHe smiled as he beckoned for the waiter to bring their check. \"Then by all means, let's go.\"\n\nOnce the show began, it didn't take Swan long to decide that David was wonderful company. She could tell he was enjoying the laser lights as much as she was.\n\nShe attended the event every year and it seemed the displays only got better and better. Each year, they honored a different state and tonight that state was New York. The New Yorkers in the crowd showed their happiness with whistles and shouting. And when a huge display of the Statue of Liberty flashed across the sky in a brilliant variety of colors, Swan caught her breath.\n\nAfter that, the showrunners took the time to honor the servicemen in attendance with a flag salute. She couldn't hold back her tears as she remembered how much her father had loved his country and how, in the end, he'd given his life for it and for her.\n\nDavid must have detected her weepy state. He pulled her closer to his side.\n\n\"Sorry,\" she said. \"I get all emotional about our servicemen and servicewomen, especially those who sacrifice their lives.\"\n\n\"You sound very patriotic.\"\n\nShe pulled back and looked up at him. \"Of course I'm patriotic. Aren't you? You did say you used to be in the military, right?\"\n\n\"Yes, I'm very patriotic,\" he said, wrapping his arms around her. She wished she didn't think the arms around her felt so strong and powerful.\n\n\"I thought you would be, but you said I sounded patriotic as if you thought that perhaps I wasn't.\"\n\n\"I apologize. I didn't mean it that way. I'm glad you're so patriotic.\"\n\nShe nodded, accepting his apology. Scanning the area around them, she said, \"They are serving complimentary wine coolers over there. Let's go grab a couple.\"\n\n\"Sure thing.\" He placed his hand on the small of her back.\n\nThe contact sent a rush of desire through her that was so strong she had to force herself to breathe. Swan quickly glanced up at him and noticed he'd been affected by the feeling as well. However, he hadn't removed his hand.\n\nInstead, he pressed his hand more firmly into her back and she felt him urging her away from the crowd and toward a cluster of low-hanging palm trees. Once they stood in the shadows, he turned her in his arms, stared down at her for a long moment and then lowered his mouth to hers.\n\nThe moment their lips touched, he slid his tongue inside her mouth, and she recalled her thoughts from earlier that day. He was delicious\u2014and dangerous\u2014with a capital _D_. And it wasn't just because he tasted of the beer he'd consumed at Summer Moon, but because he tasted like a hot-blooded man. All the sexiness she'd seen in him was reflected in his kiss.\n\nWhen she began kissing him back, he wrapped his arms around her and deepened the exchange by crushing his mouth to hers.\n\nShe didn't mind his eagerness. In fact, she welcomed the pleasure of his hunger, his taste, which was getting more provocative with every stroke of his tongue. It had been a while since she'd been kissed, and certain parts of her were reminding her of just how long it had been. Not only that, those certain parts were goading her to keep up with the forceful demands of his mouth. She hadn't been kissed so thoroughly or possessively before in her life. Or so passionately.\n\nSwan wasn't sure how long they stood there kissing. It wasn't until they heard the sound of fireworks that they disengaged their mouths. She glanced up as more fireworks exploded in the sky. Instead of looking up, David trailed his tongue along her neck and collarbone with wet licks.\n\n\"Say you'll go out with me again, Swan.\"\n\nThere was no way she wouldn't say it. She looked at him and saw deep desire in the eyes looking back at her. \"Yes, I'll go out with you again.\"\n\n\"Good.\"\n\nAnd then he lowered his head and kissed her again.\n\n* * *\n\nFlipper had tried everything possible to get to sleep. He'd counted sheep, counted backward, rolled his eyes for a full thirty minutes and had even tried hypnotizing himself. None of those things helped.\n\nHe couldn't remember ever feeling this tight with need. So here he was, close to four in the morning, and still wide awake. Nothing he did could erase the taste of Swan from his mouth and the act of kissing her from his mind.\n\nThe kiss would complicate his mission, but it hadn't been an act. It had been the most real thing he'd done in a long time. He had wanted that kiss. Needed it. It had been inevitable.\n\nSitting across from her at dinner and watching the movement of her mouth had caused a throbbing need to erupt in his gut, making him rock hard. There had been no way to ignore the delicious heat of carnal attraction spiking between them.\n\nAnd the patriotism he'd seen in her eyes when she'd gotten teary-eyed in support of servicemen, and then when she'd told him about her work with the city to find lodging for homeless vets, hadn't helped. Neither had the fact that she'd looked stunning and had smelled irresistibly hot tonight.\n\nKissing her had made his entire body feel alive. Had revved up his passion to a degree that his libido had him tied in knots and had his pulse tripping. He could feel himself riding the fine edge of intense desire heightened by more sexual energy than he'd felt in a long time.\n\nWhile kissing her, he hadn't cared that they could have been seen in spite of the low-hanging trees. He'd been beyond the point of caring. He'd been tempted to drag her to the ground right there.\n\nDamn. How was he going to clear her of anything when the only thing he'd wanted to clear her of was her clothes?\n\nHe had access to women whenever he needed them. There were always women who went bonkers for men in uniform and he had no problem engaging in one-night stands. Those types of relationships had always been the way to go for him. He liked being single, coming and going as he pleased, with no one to worry about but himself.\n\nIt had been a long time since any woman had kept him up at night and that wasn't cool.\n\nGrabbing his phone he texted the message: If anyone is awake. Call me.\n\nWithin seconds, his phone rang. It was Bane. \"What's going on, Flipper?\"\n\n\"Why are you up?\" Flipper asked his friend.\n\n\"Feeding time. Crystal and I rotate.\"\n\n\"Oh? You're breastfeeding now?\"\n\n\"No, smart-ass. The trio are on bottles now. What are you doing up?\"\n\nFlipper stretched out across the bed. \"I couldn't sleep. I tried everything. I even tried to hypnotize myself.\"\n\nBane chuckled. \"I guess it didn't work.\"\n\n\"No, it didn't work.\"\n\n\"So why can't you sleep, Flip?\"\n\nHe wasn't one to kiss and tell, no matter who the listener was, so he said, \"I still haven't figured out anything about the situation down here and the CO and the admiral are depending on me.\"\n\n\"Maybe they're going to have to accept naval intelligence's report that she's guilty.\"\n\n\"I don't think so.\" Flipper paused. \"She cried tonight.\"\n\n\"What do you mean, she cried?\"\n\n\"Today was the first day of summer and there's an annual laser show to commemorate the change in season. One of the laser displays was a salute to New York, where they did an awesome light replica of the Statue of Liberty and American soldiers. She got emotional and cried. Dammit, Bane, a person who is betraying their country doesn't cry for those in the service. Call me a sucker for tears but I don't believe she has a traitorous bone in her body.\"\n\n\"Then it's up to you to prove it. What about those two people who hang around her shop?\"\n\n\"The woman who works for her and the tattoo guy? Both seem clean. But I will dig further. I have to.\"\n\n\"Okay, but make sure while you're digging for answers that you're not burrowing yourself into something you can't handle.\"\n\n\"What do you mean?\"\n\n\"I think you know what I mean, Flip. You were sent there to prove her innocence\u2014not to prove she has a passionate side. Remember that. Good night.\"\n\nFlipper clicked off the phone and rubbed a hand down his face. Little did Bane know that after the kiss with Swan tonight, Flipper was driven to do more than prove her innocence, or her passion.\n\nHe wanted to possess Swan completely.\n\nAnd he had a feeling the desire wasn't one-sided. He'd seen the look in her eyes during dinner. He'd felt how her body had responded to his touch. He was certain the same sensations that rushed through him had affected her, too. Kissing her had been inevitable, something they both wanted and needed.\n\nThe genie called desire was out of the bottle and Flipper honestly didn't know how to get it back inside.\nThree\n\nSwan pushed away from her desk and took another huge gulp of ice-cold lemonade. It had been that way for her all day. Instead of concentrating on the online orders she needed to fill and ship out, her mind was wrapped around that kiss from last night.\n\nAll she had to do was close her eyes to remember every single detail, specifically every sensuous lick of his tongue inside her mouth. Even now, the memory sent multiple sensations coursing through her body, causing pleasure the likes of which she'd never encountered before.\n\nShe looked up at the sound of a knock on her door. \"Yes?\"\n\nJamila stuck her head in. \"Mr. Make-you-have-an-instant-orgasm is back.\"\n\nSwan didn't need to ask Jamila what she meant or who she was talking about. \"Any reason you can't wait on him?\"\n\nJamila smiled naughtily. \"I could use the pleasure but he specifically asked for you.\"\n\nSwan nodded. \"I'll be out in a minute.\"\n\n\"Okay, I will let him know.\"\n\nSwan reached over and took another gulp of her lemonade. She didn't want to admit it, but after that kiss last night, David could become an addiction. Besides putting down a gallon of lemonade, she'd been twitching in her seat most of the day, thinking that if his tongue could do that to her mouth, then Lordy...she could only imagine what else he would be able to do...\n\nShe quickly stood, refusing to go there even as a naughty part of her mind wished that he would. Leaving her office, she rounded the corner and stopped.\n\nDavid stood in the middle of her shop wearing a pair of khaki shorts and a muscle shirt. The sight of his muscled abs and strong legs made Swan bite back a groan. Just when she thought he couldn't get any sexier, he'd proved her wrong.\n\nHe must have heard the sound of her footsteps because he turned and smiled.\n\nAs if on cue, she smiled back. \"Hello, David, you came to make more purchases?\" Hopefully he would take the hint that she didn't expect him to just drop by without a reason.\n\n\"Yes. I'm buying jewelry for my three sisters-in-law and would love for you to offer suggestions.\"\n\nSwan couldn't help but smile since she liked making sales. What store owner wouldn't? \"I'd love to help you pick out pieces of jewelry for them.\"\n\nAn hour later, Swan stood at the cash register to ring up all of David's purchases. With her assistance, he'd selected some really nice pieces, with a number of the stones chosen specifically because that's what he'd said they would like. Then he wanted earrings to complement the necklaces, which he paid for in advance. They decided to select stones for the earrings tomorrow since they'd spent a lot of time on the necklaces today and her shop would be closing in less than an hour.\n\nFrom their conversation, she knew the Holloways were a close-knit family. He'd even pulled out his phone to show her pictures of his young niece and nephews.\n\n\"No pressure for you to marry?\" she asked when he tucked his phone back into the pocket of his shorts.\n\n\"None. My parents have been married for more than forty years and are still very much in love. They make sure their kids and grandkids know that. They believe we will know when it's time for us to marry without any pressure from them. We'll be the ones to have to live with the people we choose. They just want all their children to be happy.\"\n\nShe nodded. \"I like the way your parents think. I want to believe that, had my parents lived, they would have a similar philosophy. Dad used to tell me all the time that he wanted me to grow up and be whatever I wanted to be and do whatever I wanted to do, and that he and Mom would always have my back.\"\n\nShe suddenly felt a deep sense of loss. \"Appreciate your parents, David. You never know how truly great they are until they're gone. But in all honesty, I think I've always known I had great parents.\"\n\nAt that moment, he did something she wouldn't have expected from him\u2014he reached out and took her hand. \"They sound great and I know they're proud of your accomplishments.\"\n\n\"Thanks.\" That was a nice thing for him to say. To avoid thinking about just how nice he was, she slid the bag with his purchases toward him and gave him the credit card slip. He signed it and gave it back to her.\n\n\"How would you like to go to happy hour at Danica's with me?\"\n\nAfter talking about her parents and missing them like crazy, she could use more than just an hour of happiness. She would love to be able to have a lifetime of that feeling.\n\nIt wasn't that she was _unhappy_ , because she wasn't, but there were times when she wondered if maybe there was more out there for her than what was currently in her life. Perhaps she was shortchanging herself on some things. What those things were, she had no idea.\n\n\"I would love to go but good luck getting a table at Danica's. They have the best hot wings and are always crowded, _especially_ for happy hour. I think the entire island heads over there at five.\"\n\n\"Since I know you don't close your shop until five, how about if we meet over there at five-thirty? I guarantee we'll have a place to sit.\"\n\n\"Um, sounds like you might have connections, David Holloway.\"\n\n\"We'll see.\" He took the bag and turned to leave, and just like before, she watched his movements until he was no longer in sight.\n\n* * *\n\n\"Wow. You do have connections, don't you?\" Swan said, sliding into a stool at the bar. \"I've been here a number of times and the best seat I've ever gotten is at one of those tables outside.\"\n\nFlipper smiled. Like at Summer Moon, he'd arrived early and was waiting for her. He liked seeing her stroll down the sidewalk looking as beautiful as ever.\n\nToday she was wearing a pair of shorts and a pretty top. Her legs were long and shapely and he could imagine them wrapped around him while...\n\nWhoa, he didn't need to go there. Ever since that kiss, he'd been trying _not_ to go there\u2014no matter how tempted he was to do so. Quickly, he changed the direction of his thoughts.\n\n\"I know Danica personally,\" he said, trying hard to keep his naughty thoughts in check.\n\nShe lifted a brow. \"Really? How?\"\n\nThere was no way he would tell her the whole story. Danica was the godmother of former SEAL team member Nick Stover. Nick had given up being a SEAL a few years ago to take a job with Homeland Security after his wife had triplets. Instead of the whole history, Flipper gave her a modified version. \"Her godson and I used to work together.\"\n\n\"Oh.\" The bartender chose that moment to take their drink order.\n\n\"I know you used to be in the military at one point but what do you do now?\" she asked once the bartender had walked away.\n\nFlipper had expected that question sooner or later and had a prepared answer. \"I travel a lot and my job deals with ocean marine work. I guess you can say I'm a specialist in that area.\"\n\n\"Sounds interesting.\"\n\nHe chuckled. \"Trust me, it is.\"\n\nThe bartender set their beers in front of them along with a huge plate of hot wings. They dug in.\n\n\"Your assistant at the store seems nice,\" Flipper commented. \"I hope she didn't get offended when I asked specifically for you.\"\n\n\"No, very little offends Jamila, trust me.\"\n\n\"You've known her a long time?\"\n\nIf his question seemed odd, she didn't mention it. \"We met a couple of years ago when she moved to the island. The first time she came into my shop she nearly bought out the place. Like you, she has a huge family living up north and wanted to buy holiday gifts for everyone. Thanks to her, I made my month's quota in that one day. She earned a friend for life.\"\n\nFlipper took a long swig of his beer. What Swan had just told him was interesting. Based on the naval intelligence report he'd read, Jamila didn't have any family. No parents, siblings, aunts, uncles or cousins. She'd been adopted and her adopted parents had been killed in a car accident in her last year of high school. And they hadn't lived in the north but out west in California.\n\nWhy had Jamila lied?\n\n\"So you hired her that day?\" he asked, grinning, trying to make a joke of what she'd told him.\n\n\"No, she had a job as a ship captain at one of the day cruise companies in town. When things didn't work out for her there, I hired her on part-time.\"\n\nHe'd read the report and knew why Jamila had been let go and knew about her pending lawsuits. There was a big chance both cases would be settled out of court in her favor. \"Is the reason she's part-time because she's a student?\"\n\n\"Sort of. She saw how much money Rafe makes and\u2014\"\n\n\"Rafe?\" He knew who Rafe was, but Swan didn't know that.\n\n\"Yes, Rafe. He rents space in my shop where he operates a tattoo parlor. He's good and always has a steady stream of customers. Some are so pleased with his work that they recommend him to others. I've known people to fly in just to use his services.\"\n\nShe took a sip of her beer, grinned and added, \"Jamila decided to give him some real competition by becoming a tattoo artist as well. I have to admit she's pretty good. But Rafe doesn't seem worried. He even allows her to assist him sometimes. I guess you can say he's taken her under his wing. I think that's nice of him.\"\n\nFlipper took another swig of his beer. \"Yes, that is nice of him. Real nice.\"\n\n* * *\n\nLater that night, as they waited for a car at the taxi stand, Swan turned to face David. \"I had a wonderful time this evening.\"\n\nOnce again, she had enjoyed his company and hated that their time together was about to end. It didn't come as a surprise to her that the sexual chemistry between them was more explosive than ever. The kiss they'd shared the night before had ignited something within her. From the way she'd noticed him looking at her, she believed something had ignited within him as well.\n\nMore than once, her smooth bare legs had brushed against his hairy ones. The sensual contact had sent a gush of desire through her.\n\nThe first few times it happened, she'd pulled away. But finally, she'd decided not to pull her legs back and he'd given her one of those _I know you did that on purpose_ looks and she had smiled innocently and sipped her beer.\n\nHe had initiated the next physical contact and she could envision his mind at work trying to decide how to push her sensual buttons. She doubted he could push them more than he was already.\n\n\"I'm glad I got to meet Ms. Danica. After all the years I've been living here, this was my first time meeting her. She's nice.\"\n\n\"Yes, she is.\"\n\n\"And I definitely appreciate this,\" she said, holding up the bag of hot wings the older woman had given Swan to take home.\n\n\"I think she appreciated how much you enjoyed them.\"\n\nShe chuckled. \"You're probably right.\"\n\n\"What do you have planned for later?\" he asked in a deep, husky tone that seemed to have dropped a purposeful octave.\n\nHe had taken her hand when they left Danica's to walk to the taxi stand. The feel of his fingers entwined with hers had stirred something within her, something that grew with every step they took. She was aware of every detail about him as they walked together. Because of his long legs, more than once he had to slow his pace so she could keep up with him.\n\nSwan could have walked home but figured he would suggest walking there with her. She was still cautious about letting him know where she lived. When she left Jamaica to begin living on her own, her mother had drilled into her the danger of letting a man know where you lived too soon. In her heart, Swan felt David was safe, but still...\n\n\"It's near the end of the month and I need to work on the books for my accountant.\" No need to mention she had tried doing that very thing today at work and hadn't been able to concentrate for remembering their kiss from last night.\n\n\"How about dinner tomorrow night?\" he asked her.\n\nShe didn't answer right away. Instead, she broke eye contact with him and glanced down at the sidewalk. Hadn't they seen each other enough these last few days? Where was this leading? Wasn't he leaving the Keys in less than a month?\n\nShe glanced back at him. \"Why? We've gone out twice already. I wouldn't want to dominate your time.\"\n\n\"You're not. And the reason I want to take you out again is because I enjoy your company.\"\n\nShe certainly enjoyed his. \"Can I ask you something, David?\"\n\nHe nodded. \"Yes?\" Considering her history with William, it was something she probably should have asked David before going out on their first date. She'd discovered the hard way that a man not wearing a wedding ring didn't mean anything these days.\n\n\"What do you want to ask me, Swan?\"\n\nShe met his gaze and hoped she would be able to see the truth in his eyes. \"Do you have a wife or a significant other?\"\n\nInstead of guilt flashing in his eyes, she saw surprise. \"No. I'm not married and I've never been married. I dated a woman for years but because of my frequent travels, she decided to end things. That was over six years ago.\" He then leaned against a light post and asked, \"What about you, Swan? Have you ever been married or is there a significant other?\"\n\n\"Of course not.\"\n\nHe nodded slowly. \"Then I assume there is a reason you thought that maybe I was in a relationship?\"\n\n\"I needed to be sure.\"\n\nHe didn't say anything. Instead, he looked at her as if tumbling her answer around in his head. \"But like I said, I assume there is a reason you needed to know.\"\n\n\"Yes.\" However, she didn't intend to go into any details.\n\n\"Well, rest assured there is not a Mrs. David Holloway out there anywhere. Nor is there any woman wearing my ring. Satisfied?\"\n\n\"Yes.\"\n\nAt that moment, a taxi pulled up. \"Thanks for dinner again.\" She was about to move toward the taxi when he reached out, took hold of her hand and tugged her to him. He lowered his mouth to hers and kissed her quickly but soundly on the lips.\n\n\"I'll see you tomorrow,\" he said, his words a soft whisper against her wet lips.\n\n\"Tomorrow?\" she asked in a daze from his kiss.\n\n\"Yes, we're supposed to go over designs for the earrings, remember?\"\n\nIt was hard to remember anything after a kiss like that. \"Yes, I remember,\" she said.\n\n\"Then I'll see you tomorrow.\"\n\nShe nodded, and when he opened the door for her, she quickly got into the taxi and headed home alone.\n\n* * *\n\nThe moment Flipper entered his hotel room he went to the small refrigerator beneath the wet bar and pulled out a beer. Just then it didn't matter that he'd already drank a couple at Danica's. He needed another. There was just something about Swan that was getting to him, touching him on a level he wasn't used to when it came to women. He had truly enjoyed her company tonight.\n\nHe and his SEAL teammates had just returned from a two-month mission in South Africa and more than anything he had needed to unwind. He would be home in Texas doing just that had he not been summoned to the admiral's office.\n\nSo here he was, and although he was in Key West on official military business and he was supposed to be investigating Swan, he loved spending time with her.\n\nTonight, when she'd met Danica, it had been priceless. You would have thought Swan had met a Hollywood celebrity. He had sat there while the two women conversed, immediately as comfortable as old friends.\n\nThe sound of Swan's voice had been maddeningly sexy with a tinge of sweetness that had stroked his senses. For the first time since returning to the States, he had allowed himself to uncoil, to loosen up and relax while appreciating the richness of her personality. Her persona was uniquely hers and the sensuality of her very being called to him in a primitive way.\n\nAnd that wasn't good.\n\nTaking a huge swig of his beer, he switched his thoughts to what he should be focused on\u2014what she'd told him about Jamila and Rafe. Remembering what she'd said, he pulled his phone out of the pocket of his shorts and with one click he connected to his friend Nick Stover.\n\n\"This better be good, Flipper. Natalie is taking an art class at the university tonight and I have babysitting duties.\"\n\nFlipper couldn't help but smile. Like Bane, Nick had triplets and from the sound of the noise in the background, the triplets had him. \"Stop whining. Taking care of a trio of three-year-olds can't be too bad.\"\n\n\"Then you come do it.\"\n\n\"Sorry, I'm on assignment.\"\n\n\"So I hear. In the Keys, right?\"\n\nHe figured for Nick to know that much meant he'd either talked to Bane, Viper, Mac or Coop. \"Yes, I'm in Key West.\"\n\n\"While you're there, be sure to stop by Danica's. Give her a hug for me.\"\n\n\"I did that already. Tonight, in fact.\"\n\n\"Good.\"\n\n\"I think she has more photos of the triplets than you do.\"\n\n\"I wouldn't doubt it. So if you can't be a backup babysitter, why are you calling?\"\n\n\"When you arrive at your cushy job at Homeland Security tomorrow, there are two people I need you to check out for me. I've read naval intelligence reports on them, but something isn't adding up. Call me a suspicious bastard, but after that situation with Bane, when those traitors within the agencies were exposed, I'm not taking any chances.\"\n\nHe then told Nick about the discrepancies between what the reports said and what Swan had told him. \"Somebody is lying. Either Jamila lied to Swan or someone falsified the report, and I want to know which it is.\"\nFour\n\n\"He's _baaack_ ,\" Jamila said.\n\nSwan pushed away from her desk. She didn't have to ask who Jamila was talking about. \"I was expecting him,\" she said in what she hoped was a professional tone. \"He needs to look at designs for earrings.\"\n\n\"If you say so. I'll send him in here.\"\n\nSwan was about to tell Jamila they could use the computer out front, but Jamila was gone after closing the door behind her.\n\nStanding, Swan inhaled deeply. How she had finished the books last night, she wasn't sure. Thoughts of David had been stronger than ever after their night out. When she'd gone to bed, she had dreamed about him. Okay, she'd dreamed about him before, but the dreams last night had been so hot it was a wonder they hadn't burned her sheets. She had been tempted to do something she hadn't done in a long time, reactivate her vibrator.\n\nShe drew in a deep breath when she heard the knock on her door. \"Come in.\"\n\nAnd in walked David, looking sexier than he had the other times she'd seen him. Last night, to stay focused, she had come up with every reason she could think of for why she shouldn't be attracted to him and why a relationship with him wouldn't work.\n\nShe'd even thrown in the race card. But of course that was thrown out when she remembered her parents and how happy they had been together. Yet she also couldn't forget how her father's family had ostracized him for his choice in love. Would David's family be the same way? There was no reason to think they wouldn't. And wasn't she getting ahead of herself for even throwing love in the mix?\n\n\"Hello, David.\"\n\n\"Swan.\" He glanced at her desk, taking in all the folders spread across it. \"You're busy.\"\n\n\"That's fine. Besides, I need to get those earrings ready for you.\"\n\nNow that he'd seen her desk, it would make perfect sense for her to suggest they use the computer out front to design the earrings. But now that she had him behind closed doors, she liked it.\n\nNot that she planned on doing anything about having him here.\n\n\"Please have a seat while I clear off my desk.\" Today he was wearing jeans and she couldn't help but watch how fluidly his body eased into the chair. How the denim stretched across a pair of muscular thighs. She quickly switched her gaze before he caught her looking.\n\n\"Nice office.\"\n\n\"Thanks.\" She closed the folders and placed them in her inbox tray. She then glanced over at him and caught him looking at her. She followed his gaze and soon figured out why he was staring.\n\nShe was wearing a skirt with a V-neck blouse, and when she'd leaned over to place the folders in the tray, her shirt had shown a portion of her cleavage. Instead of calling him out for trying to cop a view of her breasts, the fact that he was interested sparked a distinct warmth between her legs.\n\nShe quickly sat down. \"Now if you would roll that chair over here, I am ready.\" Too late, she realized how that sounded and quickly added, \"To look at designs.\"\n\nHe smiled. \"I know what you meant.\"\n\nHe rolled his chair behind her desk to place it right next to hers. When he sat down, their legs touched. Moving away would be pretty obvious so she let the denim of his jeans rub against her bare legs.\n\n\"Now, then,\" she said, trying not to notice how good he smelled. \"What do you think of these?\" she asked, bringing up a few designs on the computer screen.\n\nWhen he didn't answer, she glanced over at him and found him staring at her. Sitting so close to him, she could look directly into his laser-blue eyes. It was as if his gaze was deliberately doing something to her, causing a surge in her breath and arousal to coil in her core. She saw the dark heat in his eyes and desire clawed at her even more.\n\n\"May I make a suggestion?\" he asked in a voice that seemed to wobble in a sexual way.\n\n\"It depends on what that suggestion is,\" she heard herself say.\n\nHe leaned in a little closer and whispered, \"I want to kiss you again. Only problem is that I don't want to stop until I get enough. And I'm not sure I would.\"\n\nShe had been staring at his lips, watching how they moved while he talked. She slowly dragged her gaze back up to his eyes. She saw need flare in his gaze at the same time that anticipation for his kiss thickened the air flowing in and out of her lungs.\n\n\"I don't know what to say.\"\n\n\"Don't say anything, Swan. Just bring your mouth closer to mine.\"\n\nShe knew she shouldn't, but she found herself doing it anyway.\n\n* * *\n\nFlipper drew in a deep breath when Swan's lips were almost touching his. He flicked out his tongue and she gave a sharp intake of breath when he began licking her lips from corner to corner with the tip of his tongue.\n\n\"What are you doing?\" she asked on a wobbly breath.\n\n\"Since you asked...\" He captured her mouth and when she closed her eyes on a moan, he reached up and cradled her face in his hands while he kissed her with a greed he didn't know was in him.\n\nWhat was there about her that made him accept the primitive part of himself that wouldn't be satisfied until he made love to her? Was it because she crept into his dreams at night and into his every waking thought? Or was it because an arrow of liquid heat shot straight to his groin whenever he saw her? Or could he blame it on the fact that whenever she touched him, he burned? She made him edgy and aroused him as no other woman could.\n\nIt was all of those things and more.\n\nRight now, he didn't know how to back away. So he didn't. Instead he accepted the stream of heat in his gut and the crackle of energy passing between them.\n\nTheir lips were copulating in a way that sent blood coursing through his veins like a raging river. It was raw, hot and explosive, causing a hot ache to erupt in his gut. It wouldn't take much to lose control and take her here on her desk. At that moment, his entire body was tight with need, totally entranced by everything about her.\n\nThe phone rang and they quickly broke off the kiss, drawing in deep breaths of air. He watched as she reached across her desk to press the speaker button. \"Thank you for calling Swan's.\"\n\nAt first, no one said anything and then a deep male voice said, \"Swan? Are you okay? You sound out of breath.\"\n\nHe watched as she pulled in another deep breath before a smile touched her lips. \"I'm fine, Godpop 1. How are you?\"\n\nKnowing who she was talking to on the phone was like a pail of cold water drenching Flipper. He was quickly reminded why he'd been sent to Key West. His admiral would have him court-martialed if he knew what Flipper had just done with his goddaughter. If the man had any idea how many times Flipper had kissed her already and how each time he'd wished they had gone even further...\n\nShe turned off the speaker so he heard only one side of the conversation, and from the sound of her voice, he knew she was happy about receiving the call.\n\nFeeling a tightness in his crotch from his still-aroused body, he got up from the chair and walked to the window. If she could have this sort of effect on him just from a kiss, he didn't want to think about what would happen if he were to make love to her. Just the thought of easing his body into hers had his stomach churning and caused an ache low in his gut.\n\nKnowing he needed to think of something else, he glanced up into the sky. It was a beautiful day. Monday was Memorial Day and he wondered if Swan had made any plans to celebrate. He'd heard there would be a parade and unlike some places in the States, where stores remained open on Memorial Day, the laid-back businesses in the Keys closed up for one big party.\n\nHe liked the Keys. When he retired from being a SEAL, he could see himself moving here to live out the rest of his days. The island was surrounded by the ocean and they didn't call him Flipper for nothing. He loved water. Being in it and being a part of it. Living this close to the sea would certainly be a plus for him. But then there was the question of how he would deal with Swan if he chose to retire here. Even if he could prove she was not guilty of espionage, there was always that possibility she would hate his guts regardless of the outcome, because he had not been truthful with her.\n\n\"Sorry about that, David.\"\n\nHe turned, not caring that she could see his still-hard erection. It was something he couldn't hide even if he had tried. Was she sorry they'd been interrupted or was she regretting that they'd kissed in the first place? He hoped it was the former because he doubted he could ever regret kissing her. \"I take it that was one of your godfathers?\" he asked, knowing it had been.\n\nShe was staring at him below the waist, but after his question, her gaze slowly moved upward to his face. \"Ah, yes, that was one of my godfathers. The other two will be calling sometime today as well. It always works out that they all call within twenty-four hours of each other.\"\n\nHe nodded and slowly walked back over to his chair to sit down. \"I know you're busy so let's look at the designs.\"\n\nHad he just seen a flash of disappointment in her eyes? Did she want them to continue what they'd been doing before she'd gotten that call? Didn't she know how close they'd both been to going off the edge and falling into waters too deep to swim out of? Even for him, a SEAL master swimmer.\n\nSomehow they got through the next half hour looking at earring designs. Just as each one of the necklaces were different, he wanted the earrings to be different as well and reflect each one of his sisters-in-law's personalities.\n\nWhen he was satisfied with his choices, he stood, convinced he needed to rush back to the hotel and take a cold shower. Sitting beside Swan and inhaling her scent without touching her was one of the most difficult things he'd had to do in a long time.\n\nShe was so female that the maleness in him couldn't help responding to everything about her. A part of him felt drugged by her scent and the intense physical awareness of her. Even now, desire was racing through his bloodstream.\n\n\"I owe you additional monies, right?\" he asked. A couple of the designs he'd selected cost more than what she'd originally estimated.\n\n\"Yes. I'll let you know the difference after I finish designing them, when you pick up everything.\"\n\nHe hadn't missed the fact that when he stood her gaze had immediately latched on to his crotch once again. Was she still hoping to see him with a hard-on? If that was true, then she wasn't disappointed. He could get aroused just from looking at her.\n\nAnd why did she choose that moment to lick her lips? She had no idea that seeing her do such a thing sent the pulse beating in his throat and desire hammering against his ribs.\n\nOn unstable legs and with an erection the size of which should be outlawed, he moved around her desk and looked at her. \"Yesterday I asked you to go to dinner with me again, but you never gave me an answer.\"\n\nHe figured that seeing how aroused he was, she probably wouldn't give him an answer now either. She surprised him when she said, \"Yes, we can dine together this evening.\"\n\nHe nodded. \"Okay, you get to pick the place.\"\n\nShe took a slip of paper off her desk, wrote something on it and handed it to him. He looked at it and he must have stared at it too long, because she said, \"It's my address, David. I'm inviting you to dine with me this evening at my home.\"\n\nHe broke eye contact with her to glance back down at the paper she'd given him. He looked back at her while trying to downplay the heat rumbling around in his gut.\n\n\"Do you need me to bring anything?\" he asked her.\n\n\"No, just yourself.\"\n\n* * *\n\nSwan glanced around her home and felt the knots beginning to twist in her stomach. She hoped she hadn't made a mistake inviting David here.\n\nToday marked a week since they'd met and if she was going to continue to see him while he was on the island, she couldn't take advantage of his thoughtfulness and expect him to invite her out without ever returning the kindness. However, more than anything else, she needed to keep things in perspective. She needed to remember he was someone she could have a good time with and that's it.\n\nShe didn't want anything more than that.\n\nOne day, she would be ready to explore her options and consider a future with a man, but that time wasn't now. She liked being single and responsible only for herself.\n\nShe knew from Candy that a serious relationship was hard work. And on top of all that hard work, you could assume you had the right person in your life only to discover you didn't. By then, you would have opened yourself up to hurt and pain in the worst possible way.\n\nThe thought that a man had caused her best friend that kind of agony bothered Swan whenever she thought about it. Candy loved Key West as much as Swan did, and for a man to be the reason she had moved away was disheartening.\n\nSwan tried telling herself that not all men were like Candy's ex, Don, or like William. On days when Swan wanted to think all men were dogs, all she had to do was remember her dad.\n\nAndrew Jamison was the yardstick she used to measure a good man. She'd watched how he had treated her mother, had seen the vibrant and sincere love between them. She had not only seen it, but she'd felt it as well. Both her parents had been demonstrative individuals and Swan had often interrupted them sharing a passionate kiss or embrace.\n\nShe still felt it here, within the walls of her house and in the very floor she walked on. All the love that had surrounded her while growing up was in this house she now called home.\n\nShe was glad her mother hadn't sold it after her father died, when Leigh had made the decision to move back home to Jamaica. Instead, she had kept the house, knowing one day Swan would want to return. It was almost too spacious for one person but Swan knew she would never sell it or move away. This house had everything she needed.\n\nShe could see the water from any room, and at night, whenever she slept with the window open, the scent of the ocean would calm her.\n\nHer favorite room in the house was her parents' old bedroom, even though she had not moved into it. It had floor-to-ceiling windows and a balcony she liked sitting on while enjoying her coffee each morning. A couple of years ago, she'd had the balcony screened in to keep the birds from flying into her house, although she loved waking up to the sound of them chirping every morning.\n\nAlthough neither one of her parents would tell her the full story, Swan knew her father had come from a wealthy family. And she knew he had been disowned by them when he had fallen in love with her mother and refused to give her up. Before dying, Leigh had given Swan a beautiful leather-bound diary to read after her death. That's what had helped keep Swan sane, reading the daily account of her mother's life and love for her father and believing they were now back together.\n\nFor weeks following her mother's death, Swan had wanted to be alone to wallow in her pity and read about what she thought was the most beautiful love story that could exist between two people. Her mother had always been expressive with the written word and Swan enjoyed reading what she'd written.\n\nIt had made Swan long for such a man, such a love. Maybe that's why she had been so quick to believe in William and why, once she'd found out about his duplicity, she'd been so reluctant to get serious with a man since.\n\nFrom her mother's diary, Swan discovered her mother's appreciation for her husband's agreement to make Key West their home. The people on the island embraced diversity and tolerated different lifestyles.\n\nSwan had read the account of when her father had been stationed at a naval base in Virginia and had sent for her mother to join him there. In the diary, her mother had written about the hateful stares they would receive whenever they went out together. The unaccepting and disapproving looks. The cruel words some people had wanted them to hear.\n\nHer father hadn't tolerated any of it and hadn't minded confronting anyone who didn't accept his wife. But to avoid trouble, Leigh had preferred to live in Key West, where people's issues with an interracial marriage were practically nonexistent.\n\nHowever, people's attitudes never kept Leigh from leaving the island to join Andrew whenever he would send for her. Oftentimes, Leigh would take Swan along and they would both join Andrew in different places for weeks at a time.\n\nWhen she heard the sound of the doorbell, Swan drew in a deep breath. The time for memories was over. The only plans she had for _this_ evening were for her and David to enjoy the meal she'd prepared and later enjoy each other's company.\n\nShe had no problem with them deciding what the latter entailed when that time came.\n\n* * *\n\n\"Hello, David. Welcome to my home.\"\n\nFlipper pushed from his mind the thought of how Swan would feel if she knew this wasn't his first time here. How she would react if she knew he had invaded her space without her knowledge. If she ever found out the truth, would she understand it had been done with the best of intentions? Namely, to keep her from wasting away in a federal prison after being falsely accused of a crime?\n\nHe forced those thoughts to the back of his mind as he smiled down at her. She looked absolutely stunning in a wraparound skirt and yellow blouse. \"Hi. I know you said I didn't have to bring anything, but I wanted to give you these,\" he said, handing her both a bottle of wine and a bouquet of flowers.\n\nHe had decided on the wine early on, but the flowers had been a spur of the moment thing when he'd seen them at one of those sidewalk florist shops. Their beauty and freshness had immediately reminded him of Swan.\n\n\"Thank you. The flowers are beautiful and this is my favorite wine,\" she said, stepping aside to let him in.\n\nHe chuckled. \"I know. I remember from the other night.\" There was no way he would also mention having seen several bottles of Moscato in the wine rack the time he had checked out her house.\n\nHe glanced around, pretending to see her home for the first time. \"Nice place.\"\n\n\"Thanks. I thought we would enjoy a glass of wine and some of my mouthwatering crab balls out on the patio before dinner.\"\n\n\"Mouthwatering crab balls?\"\n\n\"Yes, from my mom's secret recipe. You won't be disappointed,\" she said, leading him through a set of French doors. The first thought that came to his mind when he stepped out on her patio, which overlooked the Atlantic Ocean, was that it was a beautiful and breathtaking view. This had to be the best spot on the island to view the ocean in all its splendor.\n\nHe recalled how, as a boy, he would visit his cousins in California and dream of one day living near the beach. Over the years, being stationed in San Diego had been the next best thing. He owned an apartment close to base that was within walking distance of the beach.\n\nHowever, his view was nothing like this. All she had to do was walk out her back door and step onto the sand. It was right there at her door. If he lived here, he would go swimming every day.\n\nHe glanced over at her. \"The view from here is beautiful.\"\n\n\"I love this house and appreciate my mother for not selling it when she decided to move back to Jamaica after Dad died. She got a lot of offers for it, believe me. So have I. Mom said being here without Dad was too painful, but she knew I'd feel differently. For me, it was just the opposite. Being here and recalling all the memories of when the three of us shared this place makes me happy.\"\n\nHearing how the loss of her parents affected her made Flipper appreciate his own parents even more. Colin and Lenora Holloway had always been their sons' staunch supporters. Their close and loving relationships had been the reason none of their sons had had any qualms about settling down and marrying. All the marriages had worked out, seemingly made in heaven, except for his brother Liam's.\n\nWhen Bonnie had gotten pregnant, Liam had done the honorable thing by marrying her. Bonnie had always been a party girl and didn't intend to let marriage or being a mommy slow her down. While Liam was somewhere protecting his country as a Navy SEAL, Bonnie was conveniently forgetting she had a husband.\n\nNo one, not even Liam, had been surprised when he returned from an assignment one year and she asked for a divorce. Liam had given it to her without blinking an eye. Since then, Bonnie had remarried, which had introduced another set of issues for Liam. He was constantly taking Bonnie to court to enforce visitation rights to see his daughter because the man Bonnie married didn't like Liam coming around.\n\nFlipper had no qualms about marriage himself, but he had too much going on right now. Namely, resisting the temptation of Swan while he continued his investigation. That was his biggest challenge. The more he was around Swan the more he liked her and the more he wanted to prove her innocence. It was hard staying objective.\n\n\"Here you are,\" she said, handing him a cold bottle of beer. \"I figured you would like this instead of the wine.\"\n\nHe smiled. Like he had picked up on her drinking preferences, she had done the same with him. \"Thanks. I've never been a wine man.\"\n\nShe chuckled. \"Neither was my dad. That's how I knew when it was time for him to come home because Mom would have his favorite beer in the fridge.\"\n\nHe opened the bottle, took a sip and noticed her watching him. He licked his lips, liking the taste of the beer, which was the brand he'd chosen the other night at Summer Moon. When he took another sip and she continued to watch him, he lifted a brow. \"Is anything wrong?\"\n\nShe smiled. \"No, nothing is wrong. I just love to watch how you drink your beer.\"\n\nHe chuckled. That was a first. No woman had ever told him that before. \"And how do I drink it?\"\n\n\"First there's the way your mouth fits on the beer bottle. I find it very sensuous.\"\n\nHe tried ignoring the quiver that surged through his veins at the tone in her voice. \"Do you?\"\n\n\"Yes. And then there's the way you drink it like you're enjoying every drop.\"\n\n\"I am.\"\n\n\"I can tell.\" Then, as if she thought perhaps she'd said too much, she took a step back. \"I'll go get those crab balls for you to try.\"\n\nWhen she turned to leave, he reached out and touched her arm. He couldn't help it. The air all but crackled with the sexual energy between them. \"Come here a minute before you go,\" he said, setting his beer bottle aside. \"Although I do enjoy drinking beer, I've discovered I enjoy feasting on your mouth even more.\"\n\nAnd then he lowered his mouth to hers.\n\n* * *\n\nPerfect timing, Swan thought, because she needed this. She'd wanted it the moment he tilted his beer bottle to his mouth and she'd watched him do so. And now he was doing her. Showing her that he was enjoying her mouth more than he'd enjoyed the beer. Just like he'd said.\n\nThere was a certain precision and meticulousness in how he mastered the art of kissing. First, as soon as his tongue would enter her mouth, he would unerringly find her tongue, capture it with his own and begin gently sucking in a way that made the muscles between her legs tighten. Then he would do other things she didn't have a name for. Things that made desire flow through her like sweet wine, kindling heated pleasure and burning passion within her.\n\nHe rocked his thighs against her and she felt him pressed against her. His arousal was massive. Instinctively, she moved her hips closer, wanting to feel him right there, at the juncture of her thighs.\n\nWhen he finally pulled his mouth away, she released a deep, satisfied breath. Her mouth was still throbbing and there was an intense ache in her limbs. Right now, their heavy breathing was the only sound audible, and the laser-blue eyes staring down at her sent a tremor to her core.\n\nShe licked her lips when she took a step back. \"Ready for a few crab balls?\"\n\n\"Yes,\" he said, after licking his own lips. \"For now.\"\nFive\n\nH _e wanted her_.\n\nFlipper knew he shouldn't, but he did. All through the delicious dinner Swan had prepared and while engaging in great conversation with her, the thought of just how much he wanted her simmered to the back of his mind. Now with dinner coming to an end, desire was inching back to the forefront. Images of her naked tried to dominate his mind, the thoughts made him shift in his chair to relieve the ache at his crotch.\n\n\"Ready for dessert, David? I made key lime pie.\"\n\nRight now, another kind of dessert was still teasing his taste buds. \"Yes, I would love a slice, and dinner was amazing by the way. You're a good cook. My mother would absolutely love you.\"\n\nToo late, he wondered why he'd said such a thing. From the look on her face, she was wondering the same thing. So he decided to clean up his mess by adding, \"She admires other women who can cook.\"\n\nSwan smiled. \"You don't have to do that, David.\"\n\n\"Do what?\"\n\n\"Try to retract the implications of what you said so I won't get any ideas.\"\n\nHe _had_ done that, but not for the reason she thought. He'd done so because it wasn't right for either of them to think something was seriously developing between them. More than likely, she would hate his guts when she learned why he was really in Key West, when she discovered she was his assignment and nothing more. He couldn't tell her the truth, but he could certainly set her straight on what the future held for them.\n\n\"And what ideas do you think I wanted to retract?\"\n\n\"The ones where I would think we were starting something here, the ones that meant I would be someone you'd take home to meet your mother.\"\n\nHe sat down his glass of ice tea, which she had served with dinner. \"Any reason why I wouldn't want to take you home to meet my mother _if_ we shared that kind of a relationship, Swan?\" Although he didn't think he needed to let her know\u2014again\u2014that they didn't share that kind of relationship, he did so anyway.\n\n\"Honestly, David, do I really have to answer that?\"\n\n\"Yes, I think you do.\"\n\nShe stared at him for a minute. \"I'm well aware when it comes to interracial relationships that not all families are accepting.\"\n\nHe chuckled. \"My family isn't one of them, trust me. Interracial or international, we couldn't care less. My brother Brad met his wife, Sela, while working in Seoul, South Korea, and my brother Michael met Gardenia in Spain. Like I told you, my parents would accept anyone who makes us happy, regardless of race, creed, religion, nationality or color.\"\n\nShe didn't say anything to that. Then she broke eye contact with him to glance down into her glass of tea. Moments later, she raised her gaze back to him.\n\n\"My father's parents didn't. They threatened him with what they would do if he married Mom and they kept their word. They disowned him. Still, my mother reached out to them when Dad died to let them know he'd passed. They came to his funeral but had no qualms about letting Mom know they still would not accept her. They would only tolerate me since I was biracial. They even tried forcing Mom to let me go back with them. That's when my godfathers stepped in.\"\n\nFlipper shook his head, feeling the pain she refused to acknowledge, the pain she'd obviously felt because of her grandparents' actions. But he'd heard it in her voice nonetheless.\n\n\"It's sad that some people can be such bigots. At the risk of this sounding like a clich\u00e9, some of my closest friends are black,\" he added, immediately thinking of Bane, Viper and Coop. Like her, Mac was of mixed heritage and had a white mother and black father.\n\n\"I'm sure some of your closest friends are, David.\"\n\nHe wondered if she believed him. One day, she would see the truth in his words. Then it suddenly occurred to him\u2014no, she would not. There would be no reason for her to ever meet the four guys who were just as close to him as his biological brothers.\n\n\"I'll be back in a minute with the pie,\" she said. Then she stood and left the room.\n\nFlipper watched her leave, feeling that he hadn't fully eradicated her doubts the way he'd wanted to do. That bothered him. He didn't want her to think he was one of those prejudiced asses who believed one race of people was better than another. What her grandparents had done to her father and mother, as well as to her, was unforgivable. Regardless of how she'd tried to come across as if their actions hadn't hurt her, as if they still didn't hurt her, he knew better.\n\nShe needed a hug right now.\n\nHe pushed back his chair and left the dining room to enter her kitchen. Instead of getting the pie like she'd said she would do, she was standing with her back to him, looking out the kitchen window at the ocean. And he could tell from the movement of her shoulders that she was crying.\n\n\"Swan?\"\n\nShe quickly turned, swiping at her tears. \"I'm sorry to take so long, I just had one of those miss-my-daddy-and-mommy moments.\"\n\nHe crossed the room to her, knowing that her tears were about more than that. He knew it and he intended for her to know he knew it. \"Not wanting to get to know you\u2014that was your grandparents' loss, Swan.\"\n\nShe gazed into his eyes and nodded. \"I know, David, but their actions hurt Dad, although he never said it did. I knew. Mom knew, too. I think that's one of the reasons she loved him so much, because of all the sacrifices he'd made for her. That's why she did anything she could to make him happy so he would never regret choosing her. But it wasn't fair. He was a good man. Mom was a good woman. They deserved each other and should have been allowed to love freely and without restrictions, reservations or censure. It just wasn't fair, David.\"\n\nAnd then she buried her face in his chest and cried in earnest. Wrapping his arms around her, he held her, leaning down to whisper in her ear that things would be all right. That her parents had had a special love, one she should be proud of, one the naysayers had envied.\n\nEmotions Flipper hadn't counted on flowed through him as he continued to stroke her hair and whisper soothing words next to her ear. Inwardly, he screamed at the injustice of trying to keep someone from loving the person they truly wanted to love. It was something he'd never understood and figured he never would. And never would he accept such a way of thinking from anyone.\n\n* * *\n\nSwan knew she should pull out of David's arms, but found she couldn't do it. Being held by him felt good. His fingers, the ones that were stroking through the strands of her hair, seemed to electrify her scalp. They sent comforting sensations all through her\u2014and something else as well. A need that he was stroking to fruition. As a result, instead of pulling out of his arms, she closed her eyes and enjoyed being held by him while inhaling his masculine scent.\n\nShe wasn't sure how long they stood there, but it didn't take long for her to notice his breathing had changed. But then so had hers. His touch had shifted from comforting to passionate. He was using the same strokes, but now the feelings within her were beginning to build to an insurmountable degree of desire.\n\nOpening her eyes, she lifted her head to stare up at him. The minute she did, she caught her breath at the intense yearning she saw in his gaze. That yearning reached out to her, jolted her with a level of throbbing need she hadn't known existed. She'd heard of raw, make-you-lose-your-senses passion, but she had never experienced it for herself.\n\nUntil now.\n\n\"David...\" She said his name as something burst to life in the pit of her stomach. It made a quivering sensation rise at the back of her neck. He implored her with his eyes to follow this passion, as if letting her know he understood what she was experiencing even if she didn't.\n\n\"Tell me what you want, Swan,\" he said in a deep voice while gently caressing the side of her face. \"Tell me.\"\n\nThe intensity in his eyes was burning her, scorching her with the sexual hunger that was coming to life inside her. She wanted more than his erection pressing hard against the apex of her thighs. She wanted him on top of her. She wanted him to slide into her body and begin thrusting in and out. She needed to lose herself in more than just his arms.\n\nSuddenly, she felt emboldened to tell him just what she wanted. \"I want you, David. In my bed.\"\n\n* * *\n\nFlipper wanted to be in her bed as well. Lord knows he shouldn't want it, but he did. He would have to deal with the consequences later. He felt too tight and hot to try to fight the demands his body was making. Sweeping her into his arms, he quickly walked out of the kitchen and headed toward her bedroom.\n\n\"You think you know where you're going, David?\"\n\nHe slowed his pace, remembering that she had no idea that he knew the layout of her home. Not only did he know where her bedroom was located, he knew the blueprint of the plumbing underneath her floor. He looked down and met her gaze, grateful she wasn't suspicious. \"I figured you would stop me if I went in the wrong direction.\"\n\n\"Yes, I would have stopped you, but you're going the right way.\"\n\n\"Good.\" When he resumed his swift pace, it didn't take him long to reach her bedroom.\n\nSwan had gotten next to him in a way he hadn't counted on happening. Seducing her had not been part of the plan and he should not have allowed things to get this far. He didn't want to think of the major complications involved, and not just because she was the goddaughter of three top naval officers.\n\nBut something was happening that he hadn't counted on. His mind and body were in sync and a rare sexual aura was overtaking him. He could no more stop making love to her than he could stop being a SEAL. For him to even make such a comparison was pretty damn serious.\n\nInstead of placing Swan on the bed, he eased her to her feet, loving the feel of her soft body sliding down his hard one. \"If you're having second thoughts about this, Swan, now's the time to say so.\"\n\nShe shook her head and then in a wobbly voice, she said, \"No second thoughts, David.\"\n\nHearing her affirmation spoken with such certainty, Flipper released a low, throaty groan as he lowered his mouth to kiss her again, needing the connection of her lips to his as much as he needed to breathe. Wrapping his arms around her waist, he pulled her body closer to him as he deepened the kiss, wanting her to feel his erection, the hard evidence of his need for her.\n\nHe had never wanted a woman with this much intensity in his life, and he had no idea why Swan was having this kind of an effect on him.\n\nWhy she, and no other woman before her, had tempted him to cross a line during a mission. His mind didn't function that way. He had yet to prove her innocence, so technically, she was still naval intelligence's prime suspect, but at the moment that didn't matter. For all he knew, he could be about to sleep with the enemy.\n\nBut right now, that didn't matter either because deep down, a part of him believed she was innocent.\n\nWhat was happening between them was definitely out of the realm of normal for him. He'd known he would have to get close to her, but he hadn't counted on this\u2014his intense desire to do inappropriate, erotic and mind-blowing things to Swan Jamison.\n\nBut he wanted her and there would be no regrets. At least not for him, and based on what she'd just said, there would be no regrets for her either.\n\nThe moment he ended the kiss, his hands were busy removing her skirt, followed by her blouse, and when she stood in front of him in her lacy panties and bra, he couldn't help but growl his satisfaction. She looked sexy as hell and the rose-colored ensemble against the darkness of her skin was stunningly beautiful. _She_ was beautiful.\n\nHe reached up and traced a finger along the material of her boxer-cut panties. This style on a woman had never done anything for him. Until now.\n\n\"You should have been a model,\" he said in a deep, throaty voice, filled with profound need and deep appreciation. She had such a gorgeously shaped body.\n\n\"My mother used to be a model. I was satisfied with being a model's daughter.\"\n\n\"And a strikingly beautiful one at that,\" he said, lowering to his knees to rid her of her panties. He couldn't wait to touch her, taste her and do all those erotic things to her he had dreamed of doing over the past few nights. He breathed in deeply, getting more aroused by the second while easing her panties down a pair of long, beautiful legs.\n\nAfter tossing her panties aside, he leaned back on his haunches and gazed at her, seeing her naked from the waist down. Her small waist, her stomach, the shape of her thighs and longs legs were perfect. She was perfect.\n\nAfter looking his fill, he leaned forward and rested his forehead against her stomach, inhaling her luscious scent. He loved the way it flowed through his nostrils, opening his pores and causing his body to become even more erect.\n\nAnd then he did something he'd wanted to do since their first kiss. He used the tip of his tongue to kiss her stomach, loving the indention around her naval and tracing a path around the area. Then he shifted his mouth lower, licking his way down and enjoying the sound of her moans.\n\nWhen he came to the very essence of her, he licked around her womanly folds before leaning in to plant a heated kiss right there. It was as if sampling her special taste was as essential to him as breathing. His hands held firm to her thighs when he slid his tongue inside of her, loving the sound of his name from her lips.\n\nThen he went deeper, using his tongue to taste her, claim her and brand her. The latter gave him pause but not enough to stop what he was doing. He'd never claimed a woman as his own and had never thought about doing so. But with Swan, it seemed such a thing wasn't just desired but was required.\n\nAnd he didn't want it any other way. She was the first woman he wanted to claim. Forcefully, he pushed to the back of his mind what it could mean to make any woman his and decided he would dwell on that aspect of things at a later time. For now, he wanted to focus on the delicious, succulent, enjoyable taste that was Swan.\n\nHe took his time, wanting her to know just how much he loved doing this to her. He wanted her to feel the connection his tongue was making with her flesh. However, he wanted her to do more than feel it, he wanted this connection absorbed into her senses, into her mind, into every part of her body.\n\nMoments later, Flipper knew he'd achieved his goal when he felt her fingers dig into his shoulder blades, followed by the quivering of her thighs. Tightening his hold on her hips, he knew what would be next and he was ready.\n\nShe screamed his name when she was thrown into an orgasmic state. Her fingernails dug deeper into his skin, but he didn't feel the pain because knowing he was giving her pleasure made him immune to it. What he felt was a desire to take things to the next level, to slide into her body and go so deep it would be impossible to detect where his body ended and hers began.\n\nHe finally pulled his mouth away and looked up at her, saw the glazed look in her eyes. Without saying a word, he traced his fingers around the womanly mound he'd just kissed before inserting his finger inside of her. She was ultra-wet and mega-hot and he had every intention of capitalizing on both. The orgasm she'd just experienced would be small in comparison to the one he intended to give her.\n\nPulling his finger from her, he licked it clean, knowing she was watching his every move. \"Sweet,\" he said softly, holding her gaze.\n\nHe slowly eased to his feet and reached behind her to remove her bra. When she stood totally naked in front of him, he feasted his gaze on her. \"And I'm about to show you just how sweet I think you are, Swan.\"\nSix\n\nSwan was having difficulty breathing and the blue eyes staring at her made getting air to flow through her lungs even more difficult. Never had she felt this energized from a sexual act. And when David got to his feet and leaned in to kiss her, letting her taste herself on his lips, she felt weak in the knees. But he held her around the waist, holding her up as he kissed her more deeply, making her wish the kiss could last forever.\n\nShe released a low disappointed groan in her throat when he pulled his mouth away.\n\n\"Don't worry, there's more coming.\"\n\nHe swept her off her feet and carried her over to the bed, placed her on it and joined her there.\n\n\"You still have clothes on,\" she said, reaching out to touch his shirt.\n\n\"I know and they will be coming off. Right now, I just want to lie here with you and hold you in my arms.\"\n\nShe smiled at him. \"You're not going to fall asleep on me, are you?\"\n\nChuckling, he said, \"Asleep? With you lying beside me without a stitch of clothes on? Sleep is the last thing I'd be able to do, trust me.\"\n\nHe'd already pleasured her with his mouth, so she couldn't help wondering what was next. She soon discovered his intent when he reached over and cupped her breasts.\n\n\"You are perfect,\" he said in a deep husky voice.\n\nThe words triggered a memory of overhearing her father whisper the same compliment to her mother, after she surprised him with a special dinner after he returned home from one of his missions.\n\nSwan knew she was far from perfect. Those were just words David was speaking. But still, hearing them filled her with joy. Maybe she shouldn't let them, but they did.\n\nThen any further thoughts dissolved from her mind when David eased a nipple between his lips. She moaned at the pleasure she felt all the way to her toes. Just when she thought she couldn't stand anymore, he began torturing her other nipple.\n\nWhen he finally eased away, she opened her eyes to watch him undress. When he removed his shirt, she saw the tattoos covering his tanned skin on both of his upper arms\u2014huge dolphins emerging from beautiful blue ocean waters. Another tattoo of even more dolphins was painted across his back in beautiful vivid colors. She'd never been into tattoos but she thought his were stunning.\n\n\"I like your tattoos,\" she said.\n\nHe glanced over at her and smiled. \"Thanks.\"\n\nWhen he lowered his shorts, her gaze moved to the area between a pair of masculine thighs. His shaft was massive and marvelously formed. Just the thought of him easing that part of himself inside of her sent her pulse skyrocketing.\n\n\"You okay?\"\n\nShe lifted her gaze to his. She wasn't sure if she was okay. A thickness had settled in her throat when she saw how he was looking at her. Not only did he intend to join his body with hers, she had a feeling he planned to keep them connected for a while.\n\n\"Yes, I'm okay.\"\n\nSwan continued to check him out, thinking he had a mighty fine physique. His body was all muscle and it was obvious that he worked out regularly. A man didn't get those kinds of abs if he didn't.\n\nShe watched as he pulled a condom from his wallet and sheathed himself in a way that was so erotic, she felt herself getting wetter between the legs just watching him. Then he was strolling back toward the bed. To her.\n\n\"I'm about to make sure you feel more than okay,\" he said, reaching down and easing her up to rest her chest against his. Her breasts were still sensitive from his mouth and rubbing them against his chest caused a multitude of arousing sensations to swamp her.\n\n\"What are you doing to me?\" she asked in a ragged breath, barely able to get the words out.\n\n\"Anything you can imagine,\" he whispered, lowering her back on the mattress and then straddling her. He stared down at her as he gently moved her legs apart. She felt him, that part of him, lightly touch her feminine folds and then he was rubbing back and forth across them, sending even more sensations racing through her bloodstream.\n\n\"Trying to torture me, David?\"\n\n\"No, trying to pleasure you. Ready for me?\"\n\nThe movement of his manhood against her was making it impossible for her to concentrate. \"What do you think?\"\n\n\"You're wet and hot, so I think you're ready.\" And then he entered her in one deep thrust.\n\nShe gasped at the fullness and was glad he'd gone still for a minute. This gave her the chance to feel him fully embedded deep within her. It had been a long time for her and her inner muscles were greedily clamping on to him, tightening their hold.\n\n\"You're big,\" she whispered.\n\n\"You're tight,\" was his response. \"But we're making this work.\"\n\nAnd he did. First he began moving again, gently sliding in and out of her. That only lasted a few seconds before he picked up the pace and began thrusting harder.\n\nShe responded by wrapping her legs around his waist. Then he lifted her hips to receive more of him. When he established a slow and deep rhythm, touching areas in her body that hadn't been touched in a long time, or ever, she fought back a scream. She grabbed hold of his hair and pulled it, but he didn't seem to mind.\n\n\"Rule number one, Swan. Don't hold back.\"\n\nWas he kidding? It wasn't a matter of holding back. It was more like she was trying to keep her sanity. David was so powerfully male that he was pushing her over the edge with every deep stroke. Every cell within her vibrated in response to his precise thrusts.\n\n\"Hold on, baby. Things are about to get wild.\"\n\n* * *\n\nFlipper had given Swan fair warning. When he began pounding harder, making strokes he'd never attempted with another woman\u2014going deep, pulling out and then going deep again\u2014he felt a quivering sensation start at the base of his testicles and move toward her womb with each and every thrust. He had to hold on tight to her to keep them on the bed. He was determined to show her wild.\n\nSimultaneously, he leaned down to have his way with her mouth, licking it corner to corner and then inserting his tongue inside with the same rhythm he was using below.\n\nWhat he was feeling right now was more than off the charts, it was out of the atmosphere. When she finally let go and screamed his name, the sound vibrated in every part of his body, especially in her inner muscles. They clamped down on him, trying to pull everything out of him while her hands tightened even more in his hair.\n\n\"David!\"\n\nShe screamed his name again. The sound drove him. He wanted more of her. Wanted to go deeper. Throwing his head back, he felt the veins in his neck strain. There was pain but not enough to dim the pleasure.\n\nAnd he knew at that moment Swan had gotten under his skin in a way no other woman had.\n\nHe began rocking hard into her with an intensity that made him go deeper with every thrust. Then he was the one hollering out in pleasure, saying her name as an explosion ripped through him. Then like a crazed sexual maniac, he leaned in to feast on her mouth and breasts. It was like his desire for her could not end.\n\n\"David!\"\n\nHe knew she was coming again and, dammit to hell, so was he. Marveling at such a thing, he tightened his hold on her. His control had not only gotten shot to hell and back but had died an explosive death as the result of the most powerful orgasms he'd ever endured.\n\nThis was what real lovemaking was about. No holds barred. No restrictions. Every part of him felt alive, drained, renewed. The room had the scent of sex and more sex. But that wasn't all. Emotions he'd never felt before touched him and swelled his heart.\n\nHe quickly forced those emotions back, refusing to go there. Knowing he couldn't go there.\n\n* * *\n\nThe husky sound of deep, even breathing made Swan open her eyes. She was still in bed with David. Their limbs were entangled and his head was resting on her chest as he slept.\n\nThis man had been the most giving of lovers. He didn't come until he made sure she came first. He had kissed every part of her body, some parts more than others, and he had stoked passion within her in a way that had made her reach the boiling point. No man had ever made love to her with such intensity.\n\nHe had warned her about them getting wild. As far as she was concerned, they had gotten more than wild, they had gotten uninhibited, untamed. She hadn't known she had so much passion within her. He had brought it out and made her do more than own it. He had made her so aware of it that she doubted she could undo what he'd done.\n\nDavid Holloway had done more than push a few of her buttons. He had turned on all the lights.\n\nThat thought made her smile and pull him closer. Feeling exhausted, she closed her eyes and drifted into sleep.\n\n* * *\n\nFlipper slowly opened his eyes, taking in the sight and scent of the woman lying beside him, snuggled close to his body. He was so sexually contented, he could groan out loud. He didn't. Instead he tightened his arms around her.\n\nThings had gotten wild. They had finally fallen asleep after four rounds of the most satisfying lovemaking possible.\n\nWhile making love with Swan, he had discovered there was a vast difference in making love to her versus making love to other women. He'd known it before but she had made that point crystal clear tonight.\n\nWith other women, he'd usually had one goal in mind\u2014seeking sexual pleasure and making sure she got hers. With Swan it had been about that, too, but it had also been about finding closeness. No other woman had made him want to stay inside her. It had only been the need to replace condoms that had forced him from Swan's side. And then he had been back inside her in a flash...like that was where he belonged. Hell, he was still thinking that way and the twitch in his aroused manhood was letting him know just what he desired.\n\nFlipper was known to have a robust sexual appetite. When you lived your life on the edge, engaging in covert operations as his team did, then you needed a way to release.\n\nUsually, unerringly, he found his release in some woman's bed. He made sure she knew it was one and done. Due to the nature of his occupation, he didn't have time for attachments or anything long-term. Some SEALs did; he didn't. He'd tried it once and it hadn't worked out. Now he preferred being a loner. It didn't bother him that he was the lone single guy among his close friends. To each his own.\n\nSo how could one night in Swan Jamison's bed have him thinking things he shouldn't be thinking, especially considering why he was in Key West in the first place?\n\nIt had everything to do with the woman he'd had mind-blowing sex with for the past four hours or so. Now he saw her as more than an assignment. Now she was also a woman who had the ability to match his sexual needs one-on-one, something he found invigorating and energizing on all levels. He was a totally physical male and Swan Jamison was wholly, utterly female. Almost to the point that she'd blown his ever-loving mind.\n\nNow she was sleeping peacefully while he was lying here thinking, knowing his honor was being tested. As a military man, he always did what was honorable. On top of that, his mother had drilled into all five of her sons that honor was not just for their country but extended to humans just as much, especially women. Why had that thought settled deep into his mind now?\n\nOne reason might be that he'd read the report on her. He knew about those elderly people residing at the senior living complex that she visited on her weekends off and how she'd championed so hard for the homeless. She was working with the mayor to help find funding to build a housing complex for them.\n\nShe was a caring person. He'd witnessed her love for her country, for her father, that night at the fireworks and the more he was around her, the more he believed in her innocence.\n\nShe made a sound now and he glanced down and met beautiful brown eyes staring at him. Immediately his senses connected with those eyes. She trusted him. He could see it in the gaze staring back at him. Otherwise he would not be here in her bed.\n\nWhat would she think when she learned the truth? Would she still trust him? He pushed the thought to the back of his mind.\n\nShe gave him a beautiful, sleepy smile that melted his insides. Made him wish he had come here to the Keys for a real vacation, a much-needed one. He wished he had entered her shop with no ulterior motive but to do as he claimed, which was to buy his mother a birthday gift. He would still have tried his hand at seducing her, but things would have been different. Specifically, he wouldn't feel as if his honor was being compromised.\n\n\"You didn't try my key lime pie,\" she whispered.\n\n\"We can get up and eat some now if you want,\" he said.\n\n\"No, I like being just where I am. We can always eat some later...or even for breakfast.\"\n\nHe leaned down and brushed a kiss across her lips. \"Um, breakfast sounds nice. Is that an invitation to stay the night?\"\n\n\"Only if you want to.\"\n\nHe wanted to. And when he brushed another kiss across her lips, he slid his tongue inside her mouth to kiss her deeply and let her know how much he wanted to stay.\n\nHe knew at that moment that his commanding officer and the admiral weren't the only ones with a personal interest in Swan Jamison. He now had a personal interest in her as well.\nSeven\n\nThe next morning, Swan woke up to bright sunlight flowing in through her window and a powerfully male body sleeping beside her.\n\nLast night was rated right up there with _Ripley's Believe It or Not_. It had been just that spectacular. They'd made love a couple more times before getting up after midnight to eat the pie she'd prepared for dessert. After clearing off the table and loading the dishes into the dishwasher, he'd suggested they walk on the beach.\n\nSo at one in the morning, they had strolled hand in hand along the water's edge. It had been a beautiful night with a full moon in the sky. The breeze off the ocean had provided the perfect reason for him to pull her close while walking barefoot in the sand. He told her more about his family; namely about his parents' medical supply company.\n\nThen at some point, they began talking about her company and she found herself telling him just about everything about jewelry making. He was curious about her stones and complimented how beautiful they were and inquired how she was able to create so many pieces.\n\nNo man had ever taken an interest in her work before and she was excited that he thought what she did for a living was important. She had found herself explaining the day-to-day operations of Swan's. He couldn't believe how she found the time to handcraft a number of the items sold in her shop.\n\nDavid also thought it was great that Rafe, through his connections with a huge distributor in California, was able to get some of Swan's more expensive stones at a lower cost and had even helped her save on shipping by including them in the packaging with his ink.\n\nShe glanced over at him now as he shifted in bed. He kept his arms wrapped around her while he slept. She studied his features and saw how relaxed he looked.\n\nShe drew in a deep breath, still amazed at the depth of what they had shared last night. It had been the most profound thing she'd ever experienced with a man. Making love with David had touched her in ways she hadn't thought possible. He had made her feel things she hadn't ever felt before and those things weren't just sexual in nature. While in his arms, she had felt safe and secure. Protected.\n\nAs far as she was concerned, what they'd shared last night was more compelling and meaningful than any other time she'd shared with a man, even more meaningful than the time she'd spent with William. She'd never really allowed herself to fully let go with William. Now she could admit to herself that she'd known in the back of her mind that something didn't add up with him.\n\nYet she'd been so desperate for companionship after losing her mother that she had wanted to believe William was honest, even though he'd seemed too good to be true. She was glad Candy had become suspicious when he'd never wanted them to be photographed together or when he'd insisted that they spend the night at Swan's place instead of the hotel.\n\nAt the time, his requests hadn't bothered her because she hadn't wanted her employer or her coworkers to get in her business. But Candy had seen through all that and knew something in the milk wasn't clean, as she would often say. It had been Candy who'd unveiled her own husband's secret affair with a flight attendant. And once confronted, Don hadn't denied a thing. He'd said he was glad she'd found out because he wanted a divorce.\n\nPushing thoughts of Don's and William's betrayals to the back of her mind, Swan continued to study David. She couldn't help but recall the number of times he'd made her climax. Now that was simply amazing all by itself.\n\nShe was enjoying her time with him, even knowing it wouldn't last. Later this month, he would leave the island and she would probably not hear from him again. She knew that, accepted it. She had long ago learned to live for the now and not sweat the small stuff. Especially those things she couldn't change.\n\n\"You're awake.\"\n\nShe couldn't help but smile at the slumberous blue eyes staring at her. The dark shadow on his chin made him look even sexier. \"Yes, I'm awake. I guess I should be a good host and prepare breakfast before you leave.\"\n\n\"Um, I've overstayed my welcome?\"\n\n\"No, but today is Sunday and I have a lot to do.\"\n\n\"Maybe I can help you.\"\n\n\"You don't know what I'll be doing.\"\n\nHe reached out and pushed her hair back off her shoulders so he could completely see her face. \"Then tell me.\"\n\nShe gazed into his eyes. \"The shop is closed on Sundays and I use my day off to visit Golden Manor Senior Place. My mom used to do volunteer work there when we lived on the island years ago. I would go with her on Sundays to visit everyone. I guess you could say it's become a family tradition that I decided to continue.\"\n\n\"I think that's a wonderful thing you're doing. I'm sure the residents there appreciate it.\"\n\n\"Yes, they do, although those who knew my mom are no longer there. They've passed on. I'm establishing new relationships and friendships.\"\n\n\"Good for you. I'd love to join you.\"\n\n\"You would?\" she asked, surprised.\n\n\"Yes, and don't worry about preparing breakfast. I'll go home and refresh and be back here within an hour. We can grab breakfast somewhere before heading over. Afterward, we can spend more time on the beach. I enjoyed the walk last night with you.\"\n\nAnd she had enjoyed it, too. The thought that he wanted to spend more time with her made her feel really good inside. \"Okay, that sounds wonderful.\"\n\n\"I'm glad you think so, and before I leave...\"\n\n\"Yes?\"\n\nHe leaned over to kiss her and she knew where things would lead from there. She looked forward to getting wild again with him.\n\n* * *\n\nFlipper clicked on his phone the minute he walked into his hotel room. He noted several missed calls since he'd deliberately cut off his phone last night. One was from the admiral, who was probably calling for an update. But first Flipper would return the call to Nick.\n\n\"Flipper, should I ask why I couldn't reach you last night?\"\n\n\"No, you shouldn't,\" Flipper said, flopping down in the nearest chair.\n\n\"You have heard the saying that you shouldn't mix business with pleasure, right?\"\n\n_Too late for that_ , Flipper thought, running a hand down his face. Instead of responding to Nick's comment, he said, \"I hope you have something for me. There's another angle I want you to check out.\"\n\n\"Okay, and yes, I have something for you. I found out the initial investigation was handed off to a group of civilian investigators, which means naval intelligence didn't rank it at the top at first.\"\n\nFlipper was very much aware of the part government bureaucracy played in certain investigations. If someone thought a case should be under naval intelligence's radar, then they made sure it got there. \"Why?\"\n\n\"Not sure yet, but first, let's talk about Jamila Fairchild.\"\n\nFlipper leaned forward in his chair. \"Okay, let's talk about her. What do you have?\"\n\n\"Not what you obviously think. What she told Swan was the truth. She does have a huge family who lives in the north.\"\n\nFlipper raised a brow. \"Then who made the error in the report from naval intelligence?\"\n\n\"Don't know, but it's worth checking out, although I don't think it's anything suspicious on Ms. Fairchild's end. Especially when I tell you who her family is.\"\n\n\"And who is her family?\"\n\n\"Her mother's brother is Swan's grandfather.\"\n\nA frown covered Flipper's face. \"The grandfather who disowned Swan's father?\"\n\n\"Yes, from what I've gathered. But I can find no record of her grandfather ever reaching out to her.\"\n\n\"Interesting.\"\n\n\"Yes, it is. I take it Swan Jamison doesn't know about the family connection.\"\n\n\"No, she doesn't.\" Flipper decided not to try to wrap his head around this bit of news just yet. Instead he asked, \"What about Rafe Duggers?\"\n\n\"Personally, I think something is going on with him.\"\n\nFlipper lifted a brow. \"What?\"\n\n\"First of all, certain aspects of his info are sealed.\"\n\n\"Sealed?\"\n\n\"Yes. I would think if naval intelligence was checking into something related to Swan and her story, they would see that sealed record for her tenant as a red flag. For them not to have flagged it raises my own suspicions about a few things.\"\n\nThat raised Flipper's suspicions as well. Was Rafe a double agent? Someone working undercover? Was someone in naval intelligence deliberately setting Swan up as the traitor? If so, why?\n\n\"You weren't able to find out anything about him?\"\n\nThere was a husky chuckle. \"I didn't say that. There are ways to find out anything you want when you know how to do it.\"\n\nAnd Nick knew how to do it. He'd been an amazing SEAL, but as far as Flipper was concerned, Nick's natural investigative talents were better served at Homeland Security. \"When will you let me know something?\"\n\n\"Give me a couple of days. In the meantime, don't say anything to anyone about my suspicions about Duggers.\"\n\n\"Not even the CO and admiral?\"\n\n\"Not even them for now. You mentioned Swan had a third godfather who was someone high up at naval intelligence. Was his identity revealed to you?\"\n\n\"No, it wasn't but then I didn't ask,\" Flipper said.\n\n\"It wasn't hard to find out,\" Nick replied. \"All you have to do is find out who Andrew Jamison's SEAL teammates were at the time he died and do a little research to determine where they are now.\"\n\n\"I take it you've done that.\"\n\n\"Yes, and would you believe Swan Jamison's third godfather is Director of Naval Intelligence Samuel Levart?\"\n\nFlipper would not have considered Director Levart in a million years, but it all made sense now. In order for someone to have delayed making formal charges against Swan, that person would have to be someone in power. The admiral had alluded to as much. \"Swan doesn't know how favored she is to have three powerful men in her corner.\"\n\n\"Yes, but we both know it wouldn't matter if one of her godfathers was the President. If naval intelligence believes they have enough evidence to prosecute her, they will,\" Nick said.\n\nFlipper knew that to be true. Now more than ever he had to find the person intent on framing Swan. To him, it was beginning to look like an inside job.\n\n\"So what else do you have for me to check out?\" Nick asked, reclaiming Flipper's attention.\n\n\"It's about something Swan told me.\" He then shared with Nick the information about Rafe Duggers's association with some huge distributor in California. \"I need you to check that out.\"\n\n\"I will. I know time is of the essence so I'll get back to you soon, Flipper.\"\n\n\"Thanks, I appreciate it.\"\n\n* * *\n\n\"If you're so concerned about me, then why not return to the Keys and keep an eye on me, Candy?\" Swan asked. She moved around her bedroom getting dressed while talking to her friend on speakerphone.\n\n\"You know why I won't return, just yet. But I did hear something that's interesting.\"\n\n\"What?\" Swan asked as she shimmied into her skirt.\n\n\"I talked to Francola the other day and she said Marshall mentioned to her that Don is thinking about moving away from the island.\"\n\nSwan paused. Francola and her husband, Marshall, had been close friends of Don and Candy's while they were married. The two couples often did things together. Personally, Swan didn't care much for Francola because the woman had been aware Don was cheating on Candy but hadn't told her friend. \"I would take anything Francola says with a grain of salt these days,\" Swan said as she continued dressing.\n\n\"I know you still fault her for not telling me about Don and I admit I was angry with her, too, but now I understand her not doing so.\"\n\n\"Do you?\"\n\n\"Yes. Her relationship with me is not like our relationship, Swan. You and I have been best friends since grade school and we have no secrets. You would have told me about Don had you suspected anything.\"\n\n\"Darn right.\"\n\n\"Well, Francola and I didn't have that kind of relationship. We only met through our husbands, who worked together. Besides, I'm not sure I would have believed her even had she told me. I would have been in denial.\" Candy paused. \"Now, enough about me. Tell me more about this David Holloway.\"\n\nSwan smiled while putting hoop earrings into her ears. \"He's a real nice guy. Thoughtful. Considerate. Handsome as sin.\" She glanced over at her made-up bed. Although there were no signs of anyone sleeping in it last night, it didn't take much for her to remember all the wild action she and David had shared under the sheets. \"And he's great in bed. More than great. He's fantastic.\"\n\n\"Just be careful, Swan. Protect your heart.\"\n\nSwan slipped her feet into her sandals. \"My heart? It's not like I'm falling for the guy, Candy.\"\n\n\"Aren't you? I can hear it in your voice. You like him a lot.\"\n\nYes, she did like him a lot. \"It won't go beyond me liking him,\" she said, trying to convince herself of that more so than Candy.\n\n\"Can you honestly say that?\"\n\n\"Yes, because I can't let it. His work brought him to the island and he'll be leaving soon. In less than thirty days.\"\n\n\"Doesn't matter.\"\n\nSwan knew for her it _did_ matter. She only wanted short-term. The last thing she wanted was to do long-term with any man.\nEight\n\n\"I enjoyed my time with you today, Swan,\" Flipper said, looking down at her.\n\nThey'd had brunch at Summer Moon before heading to the senior living complex where they spent the next four hours. She assisted the staff by reading to groups of people and even taking a few of the seniors for walks around the complex. Some, she'd explained, had family who rarely visited so she had become like their surrogate granddaughter.\n\nOn the flip side, considering what she'd missed out having in her life, he couldn't help but wonder if they had become her surrogate grandparents.\n\nFrom the moment she walked into the facility, everyone brightened up when they saw her. It was amazing to him. She knew just what to say to elicit a smile or to get them to engage in more conversation. The majority of the seniors knew her by name and he couldn't help noticing a number of the women wearing what looked like necklaces she'd made.\n\nWhen he inquired about the necklaces Swan confirmed they were her designs but she had taught the women to make them from stones she'd given them. It had taken longer than normal since a lot of the older women's hands weren't as nimble as they used to be.\n\nAfter leaving the nursing home, they'd grabbed lunch at a sidewalk caf\u00e9 before returning to her house where they'd spent the rest of the day on the beach. Later, after ordering takeout for dinner from Arness, they were back at her place.\n\nNo matter how tempting Swan was making it for him to stay longer at her place, he would leave when it got dark. The information about Rafe Duggers's sealed records bothered him and he'd decided to poke around in the tattoo parlor later that night to see what he could find.\n\nHe glanced over at Swan as she sat across the table from him eating dinner. Earlier today, when he had returned to take her to breakfast, she had opened the door looking fresh and perky and dressed simply in a pair of shorts and a tank top. Seeing her dressed that way reminded him of just what a gorgeous pair of legs she had, as well as how those same legs had wrapped around him while they'd made love that morning and the night before.\n\nWhen they had gone swimming, she'd worn one of the sexiest two-piece bathing suits he'd ever seen. He had totally and completely enjoyed his day with her. They would be attending the Memorial Day festivities together tomorrow in town, which included a parade.\n\nBecause he needed some investigative time, he'd come up with an excuse for why he couldn't see her a couple of days this week. Time was moving quickly and he had yet to find anything to clear her of wrongdoing.\n\nBecause of Nick's warning, Flipper hadn't told Admiral Martin everything when he'd called him back yesterday. Namely, he'd left out the discrepancies between what Nick had found out and the actual reports from naval intelligence. Until Flipper discovered what was going on, he would follow Nick's advice and keep that information to himself for now.\n\n\"Although I won't be seeing you for a few days because of work, will you still be on the island?\" Swan asked.\n\nIt was hard not to be totally truthful about why she wouldn't be seeing him. She was the last person he wanted to be dishonest with but he had no choice. His goal had always been to prove her innocence and now that was doubly true. He would check out the tattoo shop tonight and look around in both Rafe's and Jamila's homes this week while they were here at work. Although it had been established that Jamila was Swan's relative, as far as Flipper was concerned, she was still a suspect.\n\n\"Yes, I'll still be on the island but I have to concentrate on this project I was sent here to do.\" No need to tell her that the project involved her.\n\n\"I understand how things are when work calls.\"\n\nHe reached up and caressed the side of her face. \"We still have a date for the parade tomorrow, right?\"\n\n\"Yes.\"\n\n\"What about dinner on Friday evening?\" he asked her.\n\nHer smile touched something deep within him. \"I'd love that, David.\"\n\n\"Good. I'll swing by your shop at closing time Friday and we can go to dinner directly from there. You pick the place.\"\n\n\"All right.\"\n\n\"What time do you want me to come get you for the parade tomorrow?\"\n\n\"It starts at ten in the morning and we need to get there early to get a good spot. How about if I prepare pancakes for us in the morning around eight?\"\n\n\"You sure? I wouldn't want you to go to any trouble.\"\n\nShe waved off his words. \"No problem. I told you I enjoy cooking.\"\n\nAfter they finished dinner, he told her he needed to leave to read some reports for work, which wasn't a lie. She walked him to the door. He leaned down to kiss her, intending for it to be a light touch of their lips.\n\nBut the moment his mouth touched hers and she released a breathless sigh, it seemed the most natural thing to slide his tongue inside her mouth and deepen the kiss. Wrapping his arms around her waist, he pulled her tight against him and knew the exact moment the kiss had changed to something more.\n\nIt was no longer a _goodbye and I'll see you later_ kiss. Instead it was one of those _I need to have you before I go_ kind. And Swan seemed to be reciprocating those feelings as she returned the kiss with equal fervor.\n\nThe next thing Flipper knew, he was sweeping her off her feet and moving quickly toward her bedroom. When he placed her on the bed, they began stripping off their clothes.\n\nFor him, she'd become an itch he couldn't scratch and a craving that wouldn't go away. There was something about making love to her that made every part of his body ache with need. She had imprinted herself on his soul and in every bone in his body and there was nothing he could do about it but savor what they had for as long as he could.\n\nWhen she was completely naked, his pulse kicked up a notch and his breathing was forced from his lungs when he looked at her. She was beautiful and perfectly made.\n\nHe pulled a condom from his wallet in the shorts she'd helped him remove and toss aside. Knowing she was watching his every move, he rolled it over his aroused manhood.\n\n\"I want to do that for you the next time.\"\n\nHe looked at Swan. \"All right.\" So she was letting him know she intended there to be a next time for them. He was glad because he wanted a next time, too.\n\nThere was a big chance when she found out the truth about why he was here on the island that she wouldn't want to have anything to do with him again. But he forced the thought from his mind.\n\n\"You don't have all evening, you know,\" she teased.\n\nShe was right, he didn't and it was a good thing she didn't know why. He moved toward her. \"Impatient?\"\n\nShe smiled up at him. \"Yes, you could say that.\"\n\n\"In that case, I can help you with that problem.\" He leaned in. \"I've got to taste you again,\" was all he said just seconds before his mouth came down on hers.\n\nSwan automatically lifted her arms around his neck the moment his lips touched hers.\n\n* * *\n\nCapturing her tongue, David drew it into his mouth. Blood rushed fast and furious to Swan's head, making her feel both light-headed and dazed as his tongue began mating with hers. His technique was rousing her passion to a level that electrified every part of her. Insistent need rushed up her spine, spinning her senses and mesmerizing her with his delectable taste.\n\nHe suddenly broke off the kiss and they both panted furiously, drawing deep gulps of air into their lungs.\n\nShe rested her head against his chest and inhaled his scent as she continued to catch her breath. She knew she was losing herself to passion again when she felt the hardness of his erection brushing against her thigh, energizing the area between her legs.\n\nThen she heard him whispering erotic details of what he wanted to do to her. His words spread fire through her body and when he gently cupped the area between her legs, she moaned.\n\n\"You're torturing me, David,\" she said, before twisting to push him down on his back so she could straddle him. Before he could react, she lowered her head between his masculine thighs and eased his erection into her mouth.\n\n\"Ah, Swan,\" he growled huskily, gripping her hair. She was fully aware of him expanding and felt a sense of triumph in her ability to get him even more aroused than he already was. The feel of his hands locked in her hair sparked even more passion within her and motivated her to use her mouth in ways she'd never done before.\n\n\"Swan!\"\n\nShe felt his thighs flex beneath her hands before he bucked forward. She wasn't prepared when he quickly switched their positions so that she was the one on her back. The blue eyes staring down at her flared with a passion that sent tremors through her.\n\nBefore she could whisper his name, he slid inside her. He kept going deeper and deeper, stretching her in ways she didn't know she could be spread, inch by inch.\n\n\"Wrap your legs around me, baby,\" he whispered in a throaty voice.\n\nWhen she did as he asked, he began thrusting hard. It was as if his total concentration was on her, intent on giving her pleasure. She felt every inch of him as he rode her hard, not letting up.\n\n\"David!\"\n\nShe screamed his name as he continued to make love to her, throwing her into a euphoric state that seemed endless. He was using her legs to keep their bodies locked while relentlessly pounding into her. Her world was spinning and she couldn't control the need to moan, moan and moan some more.\n\nShe was unable to hold anything back when her body erupted into an orgasm so powerful it propelled her toward utter completeness. She screamed his name once again as a deep feeling of ecstasy ripped through her entire body.\n\n* * *\n\nFlipper eased off Swan to lie beside her. Pulling her into his arms, his nostrils flared as he inhaled the scent of sex. The scent of woman, this woman. A woman he still desired even now.\n\nHe was not new to lust. Been there, done that and he figured he would be doing it some more. A lot more. With Swan lying in in his arms, snuggled close to him, close to his heart, he knew something had changed between them.\n\nBottom line, Swan Jamison was not only intoxicating, she was addictive.\n\n\"I don't think I'll be able to move again, David.\"\n\nA smile touched the corners of his lips. He definitely knew how she felt, but he knew he had to move. He had somewhere to be tonight and as sexually drained as he was, he intended to be there.\n\n\"Then don't move. Just lie there. I'll let myself out,\" he said, reluctant to go, although he knew he must.\n\n\"You sure?\" she asked in a lethargic voice.\n\n\"Positive. I'll be back in the morning for the parade and then we have a date on Friday.\"\n\n\"Yes. I'm going to need it. I'll be working late Wednesday doing inventory. I probably won't leave work until around ten.\"\n\n\"With your worker's help?\"\n\n\"No. The cruise ship comes in Wednesday.\"\n\nHe released her to ease out of bed and put on his clothes. \"What does a cruise ship have to do with anything?\" He could feel her gaze on his body. He couldn't disguise the impact of knowing she was watching him. He was getting aroused all over again.\n\n\"Jamila dates a guy who works on the cruise ship and they only see each other whenever the ship comes to port. She always requests the day off to spend with him. I guess they made up.\"\n\nHe had planned to check out Jamila's house when he'd assumed she would be at work. Good thing he now knew otherwise.\n\n\"Made up?\" he asked, pulling his shirt over his head.\n\n\"Yes. I got the impression they weren't on good terms last week. Not sure what happened but it's all good now since they've apparently kissed and made up.\"\n\nHe nodded. \"How long have they been together?\"\n\n\"About six months now.\"\n\nFlipper didn't say anything as he continued dressing. There hadn't been any mention of a boyfriend for Jamila Fairchild in the report he'd read. Another discrepancy. There were too many inconsistencies for his liking and he was determined to find out why. One thing was certain, he didn't like the idea of Swan being at her shop alone late at night.\n\nHe moved back to the bed, leaned down and brushed a kiss across her lips. \"I'll see you in the morning.\"\n\n\"Looking forward to it.\"\n\nHe smiled down at her and then turned and left.\n\n* * *\n\nLater that night, Flipper, dressed all in black, moved in the shadows, careful to avoid streetlights and security cameras. He had scoped out the area and was familiar with where the cameras were located. More than once, he'd had to dart behind a shrub when people were out for a late-night stroll.\n\nHe reached the area where Swan's shop was located and when he heard voices, he darted behind a building to hide in the shadows.\n\nTwo men stood not far away. One of them was Rafe. Neither of the men saw Flipper. The other guy was a little taller and appeared to be a foreigner. Their conversation sounded like an argument and was in a language Flipper wasn't familiar with and he spoke four. Most SEALs spoke at least that many, except for Coop, who had mastered seven.\n\nWhen the men lapsed into English, they lowered their voices and could barely be heard. Flipper did make out the words _ink_ and _roses_. Was someone getting a tattoo of roses painted on their body? If so, did it mean anything?\n\nFlipper was glad when the men finally moved on. More than ever, he was determined to check out the tattoo parlor. He waited a half hour to make sure the men didn't return. When he was certain they had gone, he went to work bypassing the security alarms and cameras.\n\nUsing a sort of skeleton key, he opened the back door and walked inside the tattoo parlor. Using night goggles, he glanced around.\n\nThe place looked like a typical tattoo parlor. He should know since he and his brothers had frequented a number of them. He was proud of the images on his body. Luckily, Swan hadn't asked him about them. He was glad because the last thing he wanted to do was lie about why he was into dolphins.\n\nPulling off the camera attached to his utility belt, he replaced the night goggles with a high-tech camera, which was his own creation. This particular piece of equipment detected objects underground and under water. Looking through the lens, he scanned the room. It wasn't long before the camera light began blinking.\n\nHe moved toward the area and aimed the camera lower, toward the floor, and the blinking increased. Evidently something was buried beneath the wooden floor, a portion covered by a rug. The architectural report he'd been given of Swan's shop had not exposed any secret rooms or closets.\n\nPutting the camera aside, he moved the rug and felt around to find a latch. He opened the trapdoor to find a small compartment beneath the floor. He saw more containers of ink. Why? There was a supply case full of ink on the opposite side of the room. Why was this ink hidden?\n\nThe first thing he noticed was the difference in the labeling. Was there something different about this particular ink? There was only one way to find out, he thought, taking one out of the cubby. He would overnight one of the containers to Nick instead of naval intelligence.\n\nAt this stage of the game, he wasn't taking any chances about who could be trusted.\nNine\n\nSwan had just finished the last of her inventory when she heard the knock on her shop's door. Crossing the room, she peeped through the blinds to see who it was. A smile touched her lips as she unlocked the door. \"David, I didn't think I'd see you until Friday.\"\n\nHe glanced around her empty shop before looking back at her. \"I finished work early and remembered you saying you were working late tonight doing inventory. I wanted to make sure you got home okay.\"\n\nThat was really nice of him. \"You didn't have to do that.\" But she was glad he had. They had spent Monday together celebrating Memorial Day. He had arrived at her place for breakfast and then they'd walked to where the start of the parade would take place.\n\nAfter the parade, they'd gone to the island festival marketplace where various vendors had lined the streets with booths and a huge Ferris wheel. They had taken one of the boat rides around the islands and had ended up eating lunch on Key Largo.\n\nShe had thought about him a lot since Monday, remembering in explicit detail how he'd made love to her before leaving.\n\n\"I know you said Jamila would be off today,\" David said. \"What about your tattoo guy? Is the parlor closed on Wednesdays as well?\"\n\n\"Yes, but Rafe dropped by earlier. He was expecting a shipment of more ink to come in today but it didn't. He wasn't happy about that.\"\n\n\"He wasn't?\"\n\n\"No. He said there was a particular shade of blue he was expecting.\"\n\nFlipper nodded and checked his watch. \"Ready to go?\"\n\n\"Yes, I just need to grab my purse from my office.\" She was about to turn to get it when there was another knock at the door.\n\n\"Expecting anyone?\" David asked her.\n\n\"No. I'll see who it is.\"\n\nShe walked to the door and David went with her. After glancing out of the blinds, she turned back to David and smiled. \"It's Jamila and Horacio.\"\n\nShe unlocked the door. \"Jamila, hi.\"\n\n\"Hey, Swan. Horacio and I were in the neighborhood and I remembered you would be here late. I thought we'd drop by to say hello.\"\n\nSwan smiled at the man with Jamila. \"Horacio, it's good seeing you again.\"\n\n\"Same here, Swan,\" he said in a heavy accent that Swan always loved hearing.\n\n\"And this is my friend David Holloway. David, you already know Jamila. This is her friend Horacio Jacinto,\" Swan said, making introductions.\n\nThe two men shook hands. Swan wondered if she'd imagined it but she thought David had tensed up when he'd seen Jamila and Horacio. \"Nice meeting you, Horacio,\" David said. \"I can't place your accent. Where are you from?\"\n\n\"Portugal.\"\n\n\"Nice country,\" David said.\n\n\"Thanks.\"\n\n\"I hope you'll leave before it gets too late, Swan,\" Jamila was saying.\n\n\"I will. David came to make sure I got home okay.\" Usually whenever she worked late, either doing inventory or making her jewelry, she would catch a cab home even though she lived only a few blocks away. But since David was here, she would suggest they walk. It was a nice night and she would love to spend more time with him.\n\n\"We'll see you guys later,\" Jamila said. \"We had dinner at Marty's Diner and now we're going to Summer Moon for drinks and live music.\"\n\n\"Okay. Enjoy. And I hope to see you again the next time the ship ports, Horacio,\" Swan said.\n\nHoracio smiled. \"I hope to see you as well.\"\n\nAfter they left, Swan went to her office to get her purse. She returned and noticed David was standing in the same spot where she'd left him, staring at the door. \"Are you all right?\"\n\nHe turned to her. \"Yes, it's just that Horacio looks familiar and I was trying to remember when I might've seen him. Maybe I've run into him before, here on the island.\"\n\nShe nodded. \"That's possible. He's a chef on the Century Cruise Line that docks here once a week. Whenever it does, he comes ashore and meets up with Jamila. I think I mentioned that to you.\"\n\n\"You did, but I could have sworn I saw him a few nights ago. Sunday. After leaving your place.\"\n\nSwan shook her head. \"It wasn't him. The ship didn't arrive in our port until today. But you know what they say about everybody having a twin.\"\n\nHe chuckled. \"You're probably right, but I'm sure you don't have one. I'm convinced there's not another woman anywhere who is as beautiful as you.\"\n\nSwan knew better than to let such compliments go to her head, but she couldn't help the smile that spread across her lips. \"You, David Holloway, can make a girl's head swell if she's inclined to believe whatever you say.\"\n\n\"I hope you do believe it because I spoke the truth.\" He took her hand in his as they headed for the door.\n\n* * *\n\nFlipper pulled out his phone the minute he walked into his hotel room later that night. He'd felt it vibrate in his pocket when he was walking Swan home but figured it would be a call he needed to take in private.\n\nSwan had invited him inside but he'd declined, telling her he had a ton of paperwork waiting on him back at his hotel. That wasn't a lie. He'd begun rereading all those naval intelligence reports to see if he could determine why those investigators had failed to do their job and instead intentionally went after Swan as a scapegoat.\n\nHe checked his phone and saw Nick had called and Flipper quickly returned the call. \"What do you have for me?\"\n\n\"More than you counted on. All I can say is whoever handled that investigation did a botched-up job.\"\n\nOr they did the job they'd been expected to do, Flipper thought. \"I guess there's a reason you feel that way.\"\n\n\"Yes. That ink you sent to be analyzed isn't what it's supposed to be.\"\n\n\"It's not ink?\"\n\n\"Yes, it's ink, but coded ink. When applied to the skin as a tattoo, it can be decoded by a special light. It's my guess that's how the classified information is leaving Swan Jamison's shop\u2014with people's tattoos and not with any of her jewelry. Guess where the ink is being shipped from.\"\n\n\"Swan mentioned from some place in California.\"\n\n\"Yes, that's right and the distribution company is a few miles from the naval base in San Diego. That means someone on the base must be passing classified information that's being shipped in the ink.\"\n\nFlipper frowned. \"And because Rafe Duggers is conveniently including Swan's stones with each shipment, it makes sense for her to be suspect.\"\n\n\"Right,\" Nick agreed. \"Someone is setting her up real good, Flipper. They are definitely making her the fall guy.\"\n\nFlipper wondered who in naval intelligence had targeted Swan and why. \"I have another piece of the puzzle I need you to check out.\"\n\n\"What?\"\n\n\"The guy who was with Rafe Duggers two nights ago. The one I told you he was arguing with. I saw him today.\"\n\n\"You did?\"\n\n\"Yes. He came into the shop when Swan was closing up. His name is Horacio Jacinto and he's Jamila Fairchild's boyfriend.\"\n\n\"That's interesting. I'll find out what I can about him,\" Nick said. \"I wonder if Ms. Fairchild knows what's going on or if she's being used as a pawn.\"\n\n\"I don't know, but I'm going to make sure I keep an eye on all of them.\"\n\n\"Be careful, Flipper.\"\n\n\"I will.\"\n\nA few hours later, after taking a shower, Flipper was sitting at the desk in his hotel room suite when his cell phone went off. Recognizing the ringtone, he clicked on and said, \"What's going on, Coop?\"\n\n\"You tell us.\"\n\n_Us_ meant Bane, Viper and Mac were also on the phone. \"I guess Nick called you guys.\"\n\n\"Yes, he called us earlier today,\" Bane said. \"What's going on with Swan Jamison sounds pretty damn serious. Don't you think it's time to call the CO?\"\n\nFlipper ran a hand down his face. He glanced at the clock on the wall. It was close to three in the morning. \"If Nick told you everything, then you know it's an inside job at the base. There's a traitor somewhere and until I know who I can trust, then\u2014\"\n\n\"You know as well as we do that you can trust our CO, Flipper,\" Viper said. \"Once you tell Shields what you've found out, if he suspects Martin or Levart of any wrongdoing, he will know what to do.\"\n\n\"Yes, however, the three of them share a close friendship. What if the CO is blinded due to loyalty?\"\n\n\"We're talking about our commanding officer, Flipper. Shields would turn his own mother in if he thought she was betraying our country. You know that.\"\n\nYes, he knew it. But still... \"I don't know if Martin or Levart is really involved. Like Shields, they are Swan's godfathers and I would hate to think they are shady. I just know it's an inside job and right now I'm suspicious of just about everybody.\"\n\n\"We figured you would be, so open the damn door,\" Mac said.\n\nFlipper frowned. \"What?\"\n\n\"We said open the door,\" Coop said, knocking.\n\nFlipper heard the knock, clicked off his phone, quickly went to the door and snatched it open. There stood his four best friends.\n\n\"What are you guys doing here?\"\n\n\"What does it look like?\" Mac asked as the four moved passed Flipper to enter the hotel room.\n\n\"We figured ten pairs of eyes were better than two,\" Bane said, glancing around. \"Besides, we need to keep you objective.\"\n\n\"But what about your families? Viper, your wife is having a baby!\"\n\nViper chuckled. \"And I plan to be there when she does. According to Layla's doctor, we still have a couple of months, so I'm good.\"\n\n\"And our families are good, too,\" Coop said. \"They know we look out for each other and they agreed we should be here for you.\"\n\n\"Teri is glad I'm gone,\" Mac said, grumbling. \"Maybe when I go back, she'll have a new attitude.\"\n\n\"Or maybe you'll have one,\" Bane said, frowning at Mac.\n\n\"Whatever,\" Mac said, picking up the hotel's restaurant menu book. \"Is it too late for room service?\"\n\nFlipper closed the door and drew in a deep breath as he watched the men gather around the table, already rolling up their sleeves, ready to help him figure things out. They worked together as a team and he would admit that whenever they did so, good things happened.\n\n\"There's something all of you should know,\" he said, getting their attention.\n\nThey glanced over at him. \"What? No room service at this hour?\" Mac asked in a serious tone.\n\n\"That, too.\"\n\n\"What's the other thing we should know, Flipper?\" Viper asked, sitting back in the chair he'd claimed as soon as he came in.\n\nFlipper leaned against the closed door. \"Investigating Swan Jamison is no longer just an assignment for me. It's become personal.\"\n\nThe men nodded. \"And you think we don't know that, Flipper?\" Coop asked in a steely tone. \"That's why we're here. Someone is trying to frame your woman and we're going to help you find out who and why. But first things first. You know what you have to do, right?\"\n\nFlipper stared at the four men. Yes, he knew. Instead of answering Coop, he picked up his cell phone from the table and placed a call to his CO.\nTen\n\nAs far as Swan was concerned, Friday hadn't arrived fast enough. With every passing hour, she would glance at her shop's door expecting to see David walk in. One would think his surprise visit Wednesday would have sufficed. Unfortunately, it hadn't.\n\nShe'd had two days to think about how irrational her thoughts about David were becoming. He didn't come across as a forever sort of guy and she wasn't looking for a forever kind of relationship, so what was up with this urgency to see him?\n\nThe only reason she could give herself was that she'd been alone and without a man's attention for so long that now that she had it, she was in greedy mode, lapping it up like a desperate woman. And she had never done the desperate thing before.\n\nThe door chimed and she looked up to see that it was Rafe who walked in. Lately she'd noticed him using the front door a lot more, instead of the back door to his parlor. They had decided at the beginning of his lease that the entrance to her shop was off-limits so his customers wouldn't trounce back and forth through her shop on the days Rafe worked late.\n\n\"Did your box of ink finally arrive?\"\n\nHe stopped and looked over at her. \"Why would you be asking about my ink?\"\n\nNow that, she thought, was a silly question. Did the man have a short memory? \"Because you came by Wednesday looking for the shipment and left in a tiff when it hadn't arrived.\"\n\n\"I wasn't in a tiff and yes, I did get my box of ink.\"\n\nYes, he had been in a tiff, but if he wouldn't acknowledge it, then she would leave it alone. \"Good. I'm glad you got it.\"\n\nShe watched him walk off toward his parlor. He hadn't been in a good mood lately. But then, maybe she'd been in such an extremely good mood that she had a distorted view. In fact, come to think of it, it was pretty normal for him to be moody.\n\nMoments later, while she worked with a customer, Swan watched as Rafe walked back through her shop and toward the front door. She decided if he did that again she would remind him of their agreement about which door he should use whenever he went in or out of his tattoo parlor.\n\nAfter her customer left, she glanced at her watch. Her shop would be closing in a couple of minutes. David usually arrived early. It would be understandable if he'd gotten detained, but she hoped he hadn't been. She was so anxious to see him.\n\nThe thought of how much she was looking forward to being with him should bother her, but for some reason it didn't. Like she'd told Candy, Swan wasn't expecting anything from her relationship with David. There had been no promises made, so none would be broken. The only thing she was expecting was exactly what she was getting\u2014a good time. He was excellent company and great in bed.\n\nIt had been almost three years since William, and during that time, although she'd dated, she hadn't allowed herself to get serious over a man. Instead she had concentrated on opening her shop and making it a success.\n\nShe had put her mind, heart and soul into Swan's. Especially her heart, deciding that if she put it into her business, she wouldn't run the risk of placing it elsewhere. Now it seemed there might be a risk after all and that risk had a name. David Holloway.\n\nA part of her wanted to protect herself from another possible heartbreak by calling David and canceling any plans for tonight and then to stop sharing any time with him after that. He had given her his number so she could reach him. She could certainly come up with a plausible excuse. But did she really want to do that?\n\nNo, she didn't.\n\nDavid would be her test. If she could handle a casual affair with him, then she would ace the test with flying colors.\n\nThe door chimed and she glanced up and there he was. She watched him lock her door and put the Closed sign in place before pulling down the blinds. Then he slowly sauntered toward her wearing a pair of khaki pants and an open white shirt and holding her within the scope of those laser-blue eyes. There was his too-sexy walk and a smile that made her heart beat rapidly.\n\nSuddenly seeing him, when she'd been thinking of him all day, took complete control of her senses. Without much effort, the man had turned the sensuality up more than a notch. He had his own barometer of hotness.\n\nFinally moving her feet, she strolled across the floor to meet him halfway and walked straight into his arms. The moment he pressed his body to hers, she reached up and looped her arms around his neck. He responded by wrapping his arms around her waist, drawing her even closer so she fit against him.\n\n\"I missed you, Swan.\"\n\nShe shouldn't let his words affect her, but they did\u2014to the point where she was having difficulty replying.\n\n\"I missed you, too.\"\n\nAnd she had, although they'd seen each other Wednesday. Even when she'd tried to convince herself that missing him to such a degree meant nothing. Now, as she stood wrapped in his arms, with her body pressed tight against his, hip to hip and thigh to thigh, she knew it meant everything.\n\n\"That's good to know, sweetheart,\" he said in a throaty voice.\n\n_Sweetheart?_ The endearment left her defenseless. She was trying to summon all her senses to regroup. And it wasn't helping matters that his arousal was cradled in the apex of her thighs. Good Lord, he felt so good there.\n\n\"Ready?\" she found the voice to ask him.\n\nHis gaze studied her face as if he was seeing her for the first time. As if he was trying to record her features to memory. And then a mischievous smile touched his lips. \"I'm ready for whatever you have in mind, Swan.\"\n\nShivers of desire skittered down her spine and Swan wished his words hadn't given her ideas, but they had. Ideas that were so bold, brazen and shameless she felt her cheeks staining just thinking about them. But at that moment, she didn't care. She could and would admit to wanting him.\n\nShe should wait until later to act on her desires. That would be the safe thing to do. But she knew she would be tortured during dinner whenever she looked at him. The way his mouth moved when he ate, or the way his hands\u2014those hands that could turn her on just by looking at them\u2014gripped his beer bottle. There were so many things about David Holloway that would do her in if she were to wait until later.\n\n\"You sure about that, David?\"\n\n\"Positive. Do you want me to prove it?\"\n\nDid she? Yes, she did. \"Where?\"\n\n\"I will prove it anywhere you want. Right here in the middle of the floor if you like,\" he said. \"But I suggest your office.\"\n\n* * *\n\nFlipper could tell by the way she was looking at him that she was giving his offer serious thought. He had no problems tilting the scale in his favor and he decided to do so. Lowering his head, he kissed her, trying to be gentle and finding gentleness hard to achieve. Especially when her taste made him greedy for more.\n\nHe knew she'd ceased thinking when she responded to his kiss by sinking her body farther into his embrace and tightening her arms around his neck.\n\nSome things, he decided then and there, were just too mind-blowingly good, and kissing Swan was one of them. What they'd shared these last few days was a dimension of pleasure he hadn't felt in a long time\u2014or maybe ever\u2014while devouring a woman's mouth. And when his hands shifted from around her waist to cup her backside, he groaned at the feel of her body pressed tightly against his erection.\n\nWhen he finally broke off the kiss, he buried his face in the curve of her neck and drew in a deep breath. This woman was almost too much. She looked good, tasted good and as he drew in another deep breath, he concluded that she smelled good, too.\n\n\"You want to come with me, Mr. Holloway?\" she asked, stepping out of his arms.\n\n\"Yes.\" The answer was quick off his lips.\n\nShe took his hand. \"Then follow me.\"\n\nHe had no problem following her and the minute he crossed the threshold into her office, he recalled the last time he'd been in here. Namely, when they'd shared a kiss that had nearly brought him to his knees.\n\n\"It appears dinner will have to wait.\"\n\nHe glanced over at her. She had stepped out of her sandals. After locking the office door, he leaned against it and watched her undress. She was wearing a burnt-orange sundress with spaghetti straps. It looked good on her and the color of the dress seemed to highlight her hair and skin tone.\n\nHe had gotten little sleep since his friends had arrived in the Keys. But then they hadn't come here to rest. They had left their families to come here and help him solve a sinister plan of espionage against the country they loved.\n\nAnd to protect the woman _he_ loved.\n\nHe suddenly swallowed deep when that last thought passed through his mind. As he watched Swan remove her panties, he knew without a doubt that he had fallen in love with her. He wouldn't try to figure out how it happened but just accept that it had. Now more than ever he was determined to make sure whoever was trying to screw her over didn't succeed.\n\n\"Are you going to just stand there?\" she asked, standing before him completely naked.\n\n\"No, that's not my intention at all,\" he said, moving away from the door to stand a few feet from her in what he considered his safe zone. If he got any closer, he would be tempted to take her with his clothes on. He removed his shirt and eased both his khakis and briefs down his legs at the same time. Quick and easy.\n\n\"I love your dolphins,\" she said. \"I meant to ask you about them a number of other times, but always got sidetracked. So I'm asking you now. Any reason you chose dolphins?\"\n\nHe decided to be as truthful with her as he could. One day he would have to explain to her why he'd lied about so many things. \"Like the dolphins, I love being in the water. But this isn't just any dolphin.\"\n\n\"It's not?\"\n\n\"No. This dolphin's name is Flipper. Surely you've heard of him.\"\n\n\"Not as much as I know Willy from _Free Willy_.\"\n\nHe chuckled as he moved toward her. \"Willy was a whale. Flipper was a dolphin. That's what my friends call me. Flipper.\"\n\n\"Flipper?\"\n\n\"Yes. Like I said, I love being in the water.\"\n\n\"You don't look like a Flipper.\"\n\nHe came to a stop in front of her. \"Don't tell that to my family and friends. They wouldn't agree with you.\"\n\nShe reached out and touched the tattoo of the dolphin on his arm. Her fingers felt like fire as she traced along the design with her fingertips. \"Beautiful. Not just your tattoos but all of you, David.\"\n\n\"Thanks.\" And in one smooth sweep, he picked her up and sat her on the desk, spreading her legs in the process.\n\n\"Did I tell you how much I missed you?\" he asked, running his hands over her arms.\n\n\"Yes. Just a few moments ago when you arrived here and I told you I missed you, too. You also told me that you missed me when you walked me home Wednesday night and I invited you to stay.\"\n\nFlipper heard the disappointment in her voice. If only she knew how much he'd wanted to stay. But once he'd found out Jamila would be out for a while with Horacio, he needed that time to check out her place. \"I couldn't, but I intend to make it up to you when we have more time.\"\n\nHe was letting her know this little quickie didn't count. He had something planned for her when all this was over and he could sit her down and tell her everything.\n\n\"Not here and not now? What do you call this?\" she asked when he reached up and cupped her breasts in his hands, marveling at just how beautiful they were.\n\n\"This is an I-can't-wait-until-later quickie.\"\n\n\"Interesting.\"\n\nShifting his gaze from her breasts to her eyes, he said, \"Let me show you, Swan Jamison, just how interesting it can be.\" He leaned forward and kissed her while placing the head of his erection against her wet opening. The contact sent heat spiraling through him.\n\nWhile his tongue mated greedily with hers, he entered her in one hard stroke. Pulling his mouth from hers, he let out a guttural moan when her muscles clamped down on his throbbing erection. That made him push harder and sink deeper.\n\nAnd when she moaned his name, he knew she could feel the fire of passion spreading between them as much as he could.\n\n* * *\n\nSwan wrapped her legs completely around Flipper, loving the feel of him moving inside her. He was giving her body one heck of a workout on her desk. She could feel the heat in his eyes as he stared at her.\n\nHe used his hands to lift her hips off the desk's surface for a deeper penetration. When his erection hit a certain part of her, she gasped and arched her back.\n\n\"David...\"\n\nShe whispered his name when she felt him going deeper and deeper. The intensity of their joining sent emotions skyrocketing through her.\n\nShe needed this. She wanted this. Like him, she needed it now, not later. This was more than interesting. This was a hot, frenzied, torrid mating. More than a quickie. David was thorough, meticulously so, and not to be rushed. It was as if he intended to savor every stroke.\n\nSuddenly, she felt herself falling. Not off the desk but out of reality when an orgasm rammed through her at the same time as he shuddered with the force of his own release.\n\nThey stared at each other, realizing something at the same time. Wanting to make sure he didn't stop, she whispered, \"Pill.\"\n\nIt seemed that single word triggered another orgasm and she felt him flooding her insides again while his deep, guttural groan filled the room. His release sparked another within her. His name was torn from her lips when her body shattered in earth-shaking and mind-blowing ecstasy.\n\nAs the daze from Swan's orgasmic state receded, she felt David slowly withdraw from inside her. That's when she forced her eyes open to stare at him and accepted the hand he extended to help her off the desk. Once on her feet, she wrapped her arms around his waist, feeling weak in the knees.\n\n\"It's okay, baby, I got you. I won't let you fall,\" he whispered close to her ear as he leaned down.\n\nToo late, she thought. She'd already fallen. Head over heels in love with him. The very thought suddenly sent her mind spinning.\n\nHadn't she just given herself a good talking to moments before he'd arrived? Told herself he was someone she could enjoy, both in and out of bed and nothing more? That he was someone she knew better than to give her heart to because she hadn't wanted to take the risk?\n\nWhat on earth had happened?\n\nShe knew the answer as she moved closer into the comfort of his warm naked body. David Holloway had happened. As much as she hadn't meant to fall in love with him, she had.\n\nIt didn't matter that she had known him less than three weeks. Somehow he had come into her world and turned it upside down, whether that had been his intent or not. When his work on the island was finished, he would move on and not look back. But still, knowing that he would leave hadn't stopped him from winning her heart.\n\n\"Ready?\"\n\nShe lifted her head and look up at him. \"You know, David, that lone question will get us in trouble.\"\n\nHe held her gaze for a long moment and then caressed the side of her face. \"Or take us to places we really want to go and inspire us to do things that we really want to do.\"\n\nThen he lowered his mouth to hers and kissed her.\nEleven\n\n\"Great work finding out about that ink, Lieutenant Holloway. I knew there was no way Swan would have betrayed her country.\"\n\n\"Yes, sir. Those are my thoughts as well,\" Flipper said. He had placed his CO on speakerphone so his SEAL teammates could listen to the call. \"There's no doubt in my mind the persons naval intelligence should be concentrating on are Rafe Duggers and Horacio Jacinto.\"\n\n\"I agree. I met with Admiral Martin and Director Levart this morning and they concur there's a mole within the organization.\"\n\n\"By meeting with them, sir, does that mean you feel certain they can be totally trusted as well?\" Flipper felt he had to ask.\n\n\"Yes, Lieutenant Holloway. I do. I know that because of what you discovered and what went down with Lieutenant Westmoreland a few years ago involving those moles at Homeland Security, you're not sure who you can trust. I understand that. However, I assure you that you can trust the three of us to protect Swan with our lives if we have to. We knew she was innocent, which was why we sent you there to prove we were right. You have. Now it's up to us to find out who's behind this and bring them to justice.\"\n\n\"And in the meantime?\"\n\n\"In the meantime, Lieutenant, you are free to consider this assignment completed. Go home to Texas and enjoy the remainder of your leave.\"\n\nThere was no way he could consider this assignment completed, although under normal circumstances it would be once the CO said so. \"I think I'll hang around Key West for a while.\"\n\n\"Why?\" Commanding Officer Shields asked. \"Do you think Swan's life might be in immediate danger?\"\n\n\"As long as Duggers and Jacinta don't know they're suspected of anything, then no. However...\"\n\n\"However what, Lieutenant Holloway?\"\n\nFlipper had no problem being truthful to his CO. \"However, Swan has come to mean a lot to me, sir.\"\n\n\"Oh, I see.\"\n\nFlipper figured since his CO knew him so well, he did see. \"In that case, Lieutenant Holloway, how you choose to spend the rest of your leave is your decision. But keep in mind, since this is an ongoing investigation, you cannot tell Swan anything, including your reason for being in the Keys in the first place. That in itself will place you in what might be perceived by her as a dishonorable situation.\"\n\n\"I'm aware of that, sir, but I refuse to leave her until I have to. How long do you think it will take to wrap up the investigation?\"\n\n\"Not sure. We will not only be investigating the original investigators but we'll have to restart the entire case, making Duggers and Jacinta the primary suspects. If you remain in the Keys and notice anything I need to know, don't hesitate to bring it to my attention.\"\n\nIn other words, Commanding Officer Shields was pretty much giving Flipper the green light to do his own thing, unofficially. \"Yes, sir.\"\n\nWhen Flipper clicked off the phone, he glanced up at his friends. \"So what do you guys think?\"\n\n\"Personally, I think you're doing the right thing not leaving here until you're certain Miss Jamison's life is not in any danger,\" Bane said.\n\n\"And since we don't plan to leave until you do, it's time we figure out just who is behind this,\" Coop added.\n\n\"I agree with all the above,\" Viper tacked on.\n\nThey all looked at Mac, who rubbed his chin as if contemplating something. Then he said, \"Someone needs to play devil's advocate, so I guess it has to be me.\"\n\n\"No surprise there,\" Bane said.\n\nMac shot Bane a glare and then glanced back at Flipper. \"Think about what the CO said. You can't tell Miss Jamison anything. Once she finds out the truth, that she was nothing more than an assignment to you, she's not going to like it, no matter how noble or honorable your intentions might have been.\"\n\nFlipper drew in a deep breath. He knew Mac's words to be true. Although Swan had yet to tell him anything about her affair with William Connors, it had been in the report. The man had betrayed her and there was a chance she would probably see Flipper as doing the same. \"So, Mr. Know-It-All, what do you suggest I do?\" he asked.\n\n\"Start drawing a line in your relationship and don't cross it. In other words, stop seducing her,\" Mac said.\n\n_Too late for that_ , Flipper thought. All he had to do was remember what they'd done yesterday in Swan's office and again when he'd taken her home after dinner. Especially when he'd sat in one of her kitchen chairs and she'd straddled his body. The memories of what had started out in that chair and ended up in her bedroom made him feel hot. He hadn't left her place until dawn this morning. There was no way he could put a freeze on his relationship with Swan like Mac was suggesting.\n\n\"That's not an option, Mac. I'm going to do what I have to do now and worry about the consequences later.\"\n\n* * *\n\n\"What's this about you having a boyfriend? I can't leave you alone for one minute.\"\n\nSwan smiled when she glanced up at Rosie McCall, one of her frequent customers. Rosie, an older woman in her midforties who'd been away for the past three months visiting her family in Nevada, had returned to the Keys just yesterday. \"I see Jamila has been talking again.\"\n\n\"Doesn't matter. So tell me, who is he?\"\n\nSwan closed the jewelry case. \"First of all, he's not my boyfriend. He's just someone I'm seeing while he's here on the island working, which won't be much longer.\"\n\n\"Um, short meaningless flings are the best kind. What's his name?\"\n\n\"David. David Holloway.\"\n\n\"Where he is from?\"\n\n\"Texas.\"\n\n\"You said he's here working. What does he do for a living?\"\n\n\"Whoa, time-out,\" Swan said, using her hand for the signal. \"You don't need to know all that. David's a nice guy and that's all you really need to know.\"\n\nShe knew how Rosie liked to play matchmaker. She'd been the one who'd introduced Jamila to Horacio. Rosie had met him at one of the nightclubs and thought he was cute, too young for her but just the right age for Swan or Jamila. Swan hadn't been interested in a blind date but Jamila had. Horacio and Jamila met, hit it off and had been an item ever since.\n\n\"You can't blame me for being curious, Swan. You seldom date.\"\n\n\"My choice, remember? Besides, you do it enough for the both of us.\" And that was the truth. After her second divorce, Rosie had made it known she would never marry again but intended to date any man who asked her out as long as they were the right age. Not too old and not too young.\n\nRosie smiled. \"Yes, I do, don't I? But that doesn't mean you shouldn't go out and have fun every once in a while. There's more to life than this shop, Swan. I hope you're finally finding that out.\"\n\n\"Whatever.\" Swan had heard it before and all from Rosie. She liked the older woman and thought she was a fun person who had a zeal for life. There was never a dull moment around her.\n\nAt that moment, the shop's door chimed and Swan knew without looking in that direction that David had walked in. She also knew when he saw her with a customer that he would wait until she finished before approaching her.\n\nRosie leaned in. \"Looks like you have a customer. Let's hope he buys something since he came in a minute before closing.\"\n\nSwan inwardly smiled. \"We can only hope, right?\"\n\n\"But then he's such a cutie. Look at him.\"\n\nSwan didn't have to look at David to know what a cutie he was, but she did so anyway. He was browsing around the store wearing a pair of shorts and a sleeveless T-shirt with flip-flops on his feet. He looked laid-back and sexy as sin. \"You're right, he is a cutie.\"\n\n\"I love those tattoos on his upper arms. Nice.\"\n\n\"Yes, they are.\" She knew Rosie was into tattoos and was one of Rafe's frequent customers. The woman had them everywhere, visible and non-visible.\n\n\"You need to go wait on him. See what he wants. If he's not sure, offer him a few things.\"\n\nSwan smiled. Little did Rosie know, but she intended to offer David a lot. \"I will. Come on, I'll walk you to the door. I'm officially closed now,\" Swan said, coming from around the counter.\n\n\"You honestly want me to leave you here with him?\" Rosie whispered. \"For all you know he's not safe.\"\n\nSwan chuckled and decided it was time for her to come clean. \"He's safe, Rosie. That's David and he's here to walk me home. I'll introduce you on your way out the door.\"\n\n\"You mean that gorgeous hunk is your guy?\"\n\nSwan glanced over at David again. He was definitely a gorgeous hunk but she couldn't claim him as her guy. \"Yes, he's the guy I've been seeing a lot of lately.\"\n\n\"Smart girl.\"\n\nDavid glanced up when they approached and gave her a huge smile. \"Hi,\" he greeted.\n\n\"Hi, David. I'd like you to meet Rosie McCall. A friend who has been away for the past few months and just returned back to the island. Rosie, this is David Holloway.\"\n\nDavid extended his hand. \"Nice meeting you, Rosie.\"\n\n\"Same here, David. I like your tattoos.\"\n\n\"Thanks and I like yours,\" he responded.\n\n\"Thanks. Well, I'll be going. I hope you guys enjoy yourselves.\"\n\n\"We will,\" Swan said, smiling up at David. \"I'll be back after seeing Rosie out,\" she told him.\n\nHe nodded. \"Nice meeting you, Rosie.\"\n\n\"Same here.\"\n\nSwan returned to David a few moments later, after putting up the Closed sign, locking the door and pulling down the shades. She turned and studied him as he stood across the room, looking so amazingly sexy. She felt a lump in her throat. She loved everything about him, especially the muscles beneath his shirt, the masculine thighs and his tanned skin.\n\n\"Got more sun today, I see.\"\n\n\"Yes, I had to go out on the boat today.\"\n\n\"One day you're going to have to explain to me in detail just what your ocean duties entail.\"\n\n\"I will. But for now, come here. I missed you today.\"\n\nShe crossed the room to walk into his arms. \"I missed you, too.\"\n\n\"That's good to know. Rosie seems like a nice person.\"\n\n\"She is.\"\n\n\"She has a lot of tattoos.\"\n\nSwan chuckled. \"Yes, she does. She's one of Rafe's best customers.\"\n\n\"Is that right?\" David asked, still smiling. \"He did an awesome job.\"\n\nShe checked her watch. \"We can leave as soon as I grab my purse.\" They would be having dinner at Nathan Waterway and afterward would attend an art show. \"I'll be back in a second.\"\n\n* * *\n\nFlipper watched Swan walk off toward her office while thinking of what she'd told him about Rosie McCall. He recalled what he'd overheard Rafe and Horacio arguing about that night behind this building. Ink and roses. Or had they said Rosie? Was she a part of the group? If she was, that meant she had an ulterior motive for befriending Swan.\n\nPulling his phone from his pocket he texted Nick. Check out Rosie McCall.\n\nHe received an immediate reply. Will do.\n\nHe then texted Bane. Excursion tonight.\n\nThe reply was quick. On it.\n\nMost of today he and Viper had pretended to go fishing after Mac, who'd been tailing Rafe for the past two days, reported that Rafe had rented a boat and headed in the direction of another island close by. Today Flipper and Viper had also rented a boat, making sure they stayed a good distance behind Duggers.\n\nThe man had docked in Fleming Key. Bane and Cooper, who'd arrived ahead of them, picked up the tail on Rafe. It seemed the man had gone into a sports shop where he'd stayed for three hours.\n\nPretending to be two guys enjoying their time out on their boat, Flipper and Viper had waited at the pier and knew when Rafe had left the island to return to Key West. Mac had been there to pick up the tail and reported that the man had been carrying a package when he went inside his tattoo shop. A package Rafe had gotten from the sports shop, according to Bane and Coop.\n\n\"I'm ready.\"\n\nHe looked up and when Swan met his gaze, she quickly clarified, \"I'm ready _for dinner_.\"\n\nHe placed his phone back into his pocket and smiled. \"That's all?\"\n\nThe smile she returned made his insides quiver in anticipation. \"For now, Mr. Holloway.\"\nTwelve\n\nIt was getting harder and harder to leave Swan's bed, Flipper thought as he and his teammates docked at Fleming Key close to two in the morning. But at least he'd left her sleeping with the most peaceful smile on her face.\n\nWithout waking her, he had brushed a kiss across her lips and whispered that he loved her, knowing she would remember neither. But he decided to tell her how he felt when he saw her later today. He couldn't hold it inside any longer. She deserved to know. He wanted her to know. And when all this was over and she knew the truth, he would do whatever he needed to do to win her forgiveness and her love.\n\nHe jumped when fingers snapped in his face. He glared at Mac, who glared back. \"Stay focused. You can daydream later.\"\n\n\"I wasn't daydreaming,\" Flipper countered. He then realized he was the only one still in the boat. The others had already gotten out.\n\n\"Then night-dreaming. Call it what you want\" was Mac's reply. \"Just get out of the damn boat.\"\n\nFlipper didn't have to wonder why Mac was in a rotten mood. Teri had texted him earlier in the day to say the new washer and dryer had been delivered. They were new appliances Mac hadn't known they were buying.\n\nMoments later, dressed in all black military combat gear, the five of them circled around to the back of the sports shop Rafe had frequented lately. Being ever ready and not taking any chances, Glocks were strapped to their hips and high powered tasers to their thighs. Due to Viper's hypersensitive ears\u2014known to pick up sound over long distances away\u2014he would stay outside as the lookout. Flipper, Bane, Cooper and Mac bypassed security cameras to enter the building.\n\nOnce inside, they used Flipper's cameras and it didn't take long to find a hidden room. Making swift use of their time, they took pictures of everything. It was obvious this was the group's operation headquarters. More tattoo ink was stored here along with several specific tattoo designs. One design Flipper quickly recalled seeing on the side of Rosie's neck.\n\nFlipper scanned the room with his camera and then opened several drawers in the huge desk and took photos of the contents. When he came across a photo in one of the drawers, he suddenly froze. \"Damn.\"\n\n\"What is it, Flipper?\" Bane asked.\n\nInstead of saying anything, he motioned his head to the photograph he'd found. Mac, Bane and Coop came around him to see it as well. They looked back at him and Mac said, \"We've been royally screwed.\"\n\nAn uneasy feeling settled in the pit of Flipper's stomach. \"I need to get back to Swan as soon as possible.\"\n\n* * *\n\nSwan was awakened by the knocking on her door. She glanced at the clock on her nightstand and wondered who on earth would be at her house at four in the morning. Was it David returning? She didn't recall when he'd left but knew it was the norm for him to leave her place around midnight to return to his hotel because of his work. Usually she would be awake when he left but tonight sexual exhaustion had gotten the best of her.\n\nPulling on her robe, she tied it around her waist as she headed for the door. Looking out the peephole, she saw it was Jamila and Horacio. What were they doing out so late and why were they at her place? She found it odd that Horacio was on the island when the cruise ship wasn't due back in port again until next week.\n\nFrom the look on Jamila's face, it appeared she wasn't happy about something. In fact, from her reddened eyes, it appeared that she'd been crying. Swan wondered what on earth was wrong. Had something happened?\n\nSuddenly filled with concern, she quickly opened the door. The minute she did so, Jamila was shoved inside, nearly knocking Swan down.\n\n\"Hey, wait a minute. What's going on?\" Swan asked, fighting to regain her balance.\n\n\"Shut the hell up and don't ask questions,\" Horacio said, quickly coming inside and closing the door behind him.\n\nSwan frowned. \"Horacio? What do you mean, I can't ask any questions?\"\n\n\"Just do what you're told,\" he barked.\n\nSwan glanced over at Jamila and saw the bruise on the side of her face. \"Did he do this to you?\" Swan demanded, getting enraged. At Jamila's nod, Swan then turned to Horacio. How could he have done this when he adored Jamila? \"I want you to leave now.\"\n\n\"If I don't, what are you going to do? Call the police? Or call that SEAL you're sleeping with?\"\n\nSwan frown deepened. \"I don't know what you're talking about. Now leave or I _will_ call the police.\"\n\n\"You won't be doing anything other than what I tell you to do. When I get the word, the two of you will be coming with me.\"\n\nSwan placed her hands on her hips. \"We're not going anywhere with you.\"\n\nA cynical smile touched Horacio's lips as he pulled out a gun from his back pocket. \"This says you will.\"\n\nSwan stared at the gun, not believing Horacio had it pointed at both her and Jamila. She was about to say something when Horacio added, \"I'm giving you five minutes to go into your bedroom and put on clothes. Bring me your phone first. I don't want you to get any crazy ideas.\"\n\nSwan had no idea what was going on, but from the pleading look in Jamila's eyes, she knew it was best to do as she was told. She went and got her cell phone and handed it to him, but not before she noticed several missed calls from David. Why had he been trying to call her? Her mind was filled with so many questions.\n\n\"You got five minutes to get dressed. If you're not back in five minutes or try some kind of funny business, your cousin here will pay for it.\"\n\n_Cousin?_ Why did he refer to Jamila as her cousin? At what was obviously a confused look on her face, he said, \"That's right. Secrets. There are plenty more where those came from, Swan, and you'll be finding out about them later. Now go.\"\n\nSwan got dressed in less than five minutes. If she hadn't thought Horacio was serious about hurting Jamila, she would have escaped through her bedroom window. That bruise along the side of her friend's face indicated the man was serious.\n\nSwan was walking out of her bedroom fully dressed when Horacio's phone rang. Instead of answering it, he said, \"That's my signal that things are ready. We'll go out your back door to the beach. The boat is waiting.\"\n\n\"What boat?\"\n\n\"Please don't ask him anything, Swan,\" Jamila pleaded, reaching out and grabbing her arm. \"All of them are crazy.\"\n\nSwan wondered just who were _all of them_. But she decided not to ask.\n\n\"Move!\"\n\nFollowing Horacio's orders, she and Jamila walked toward Swan's kitchen to go out the back door.\n\n* * *\n\nAs soon as their boat docked, Flipper raced through the streets of Key West toward Swan's home with his teammates fast on his heels. He had tried reaching her on the phone but didn't get an answer. He immediately knew something was wrong because she kept her phone on the nightstand next to her bed and the ringing would have woken her up. He had tried several more times with no luck, which was why his heart was beating out of control and fear was gripping his insides, especially now that he knew who was involved.\n\nThey had contacted their CO and told him what they'd discovered. He was as shocked as they'd been and they knew Shields would be taking the necessary actions on his end. Flipper hadn't had to tell the man there would be hell to pay if anyone hurt one single hair on Swan's head.\n\nWhen they reached her house, they found the door unlocked. Her cell phone had been tossed on a living room table and a quick search of her bedroom indicated she'd change clothes.\n\n\"Take a look at this, Flipper,\" Mac called out.\n\nWhen he reached them in the kitchen, Mac pointed out the window. Flipper saw lights from a boat that was sitting idle in the ocean as if waiting to rendezvous with another vessel.\n\n\"I traced footprints in the sand that led to the water. A small watercraft probably took them out to that boat,\" Viper was saying. \"There were three sets of shoe prints belonging to two women and a man. And they left around thirty minutes ago.\"\n\nFlipper raced out Swan's back door and after putting on his night-vision eyewear, he stared out at the ocean.\n\n\"Intercept with our boat,\" he shouted over his shoulder to the others. Quickly dropping to the sand, he began removing his shoes, T-shirt and pants, leaving his body clad in a pair of swimming trunks.\n\n\"Don't try it, Flipper. The boat's too far out,\" Mac said. \"It's too dangerous for anyone, even you.\"\n\nFlipper glanced up at them while putting the waterproof military belt that contained combat gear around his waist. He then put a pair of specially designed water goggles over his eyes. \"The woman I love is on that boat and I have no idea what they plan to do, so I have to try. Even if I die trying.\"\n\nWithout saying anything else, he raced toward the water and dived in.\n\n* * *\n\nHoracio had tied their hands before forcing them into a small boat, which carried them out into the ocean to a much bigger boat. Now they were sitting idle in the waves.\n\nSwan wondered why. She glanced around and noticed that, other than the lights on the boat, there was only darkness. They were so far from land she couldn't see the lights from the homes where she lived anymore.\n\nAs if Horacio realized she was trying to figure out what was happening, he said, \"I'm waiting for the rest of the gang, then we'll decide what we will do with the two of you.\"\n\nWhat he said didn't make much sense. \"Will someone please tell me what's going on?\" Swan asked, getting angrier by the minute. None of this made any sense.\n\n\"I'll let your cousin go first since Jamila has a lot of explaining to do,\" Horacio said, grinning.\n\nSwan turned to Jamila, who was sitting on a bench beside her. \"What is he talking about? Why does he keep referring to you as my cousin?\"\n\nAt first Jamila didn't say anything. In fact, it seemed she was refusing to meet Swan's gaze, but then she finally met Swan's eyes and said, \"Because we are cousins, Swan. My mother is your grandfather's youngest sister.\"\n\n\"My grandfather?\"\n\nJamila nodded. \"Yes, Lawrence Jamison is my uncle. I knew for years that Uncle Lawrence disowned your father but I didn't know why until I was much older. Then I thought the reason was downright stupid and told the family what I thought. Everyone else in the family thought the same thing but were too afraid to stand up to Uncle Lawrence.\"\n\nSwan didn't say anything. She was still trying to dissect the fact that she and Jamila were related. She'd known from her father that Lawrence had a sister and another brother. That was all she'd known.\n\n\"When I turned twenty-one and finished college, I decided to come find you. Uncle Lawrence didn't like it but I told him I didn't care. I'm one of the few who stands up to him. He said the family would disown me if I came here.\"\n\n\"Yet you came anyway,\" Swan said.\n\n\"Yes, I came anyway.\"\n\nSwan glanced over at Horacio. He wasn't saying anything and didn't appear to be listening to what they were saying. Instead he stood at the bow of the ship looking through binoculars as if he was searching for someone. He'd said they were waiting for another boat with the gang and Swan couldn't help but wonder who the gang was.\n\nShe wanted to ask Jamila how much she knew and why they were being held hostage but figured that although Horacio was pretending not to listen to their conversation, he probably was.\n\nSwan glanced over at Jamila. \"Why didn't you tell me who you were when you first came into my shop that day? Why did you keep it a secret all this time?\"\n\n\"Because I knew how my family had treated you and your mother. I figured the last thing you'd want was to meet a relative from that side of the family. I decided to let you accept me as a friend and then later I would tell you the truth that we were cousins.\"\n\n\"Now isn't that a touching story?\" Horacio said, strolling back over to where they sat.\n\n\"Yes, it is touching,\" Swan said, defiantly lifting her chin. \"Why are we here?\"\n\nHe smiled. \"You'll find out soon enough. And I hope you're not holding out any hope that your SEAL boyfriend will be coming to rescue you because he won't.\"\n\n\"Why do you keep saying David is a SEAL when he's not? He was in the military once but he was never a SEAL.\"\n\n\"Sounds like you've been conned by him just like your cousin here was conned by me,\" he said as if it was something to brag about. \"Your lover boy _is_ a SEAL and he was sent here to get the goods on you. Whether you know it or not, you've been his assignment.\"\n\nSwan shook her head. \"No, that's not true. I don't believe you.\"\n\n\"I don't care if you believe me or not but it's true. I only found out today what he's been doing and why he was sent here by naval intelligence.\"\n\n_Naval intelligence?_ Swan glanced over at Jamila, who said, \"I don't know whether what he's saying is true or not, Swan, but he told me the same thing tonight.\"\n\n\"Why would naval intelligence suspect me of anything? It doesn't make sense.\" And more than that, she refused to believe David wasn't who and what he said he was.\n\nAt that moment, they heard the sound of a boat approaching. Horacio drew his gun and pointed the flashlight toward the oncoming boat. He put his gun back in place. \"Hold on to that question, sweetheart. The person who will explain everything just arrived.\"\n\nSwan kept her gaze trained on the boat that pulled up beside theirs and saw two people onboard. Both of them she knew. What in the world...?\n\nShe watched in shock as Rafe and Rosie came aboard. She was so focused on staring at them that she almost missed the third person who also came on board.\n\nShe gasped in shock when the person said, \"Swan, you look well.\"\n\nSuddenly losing her voice, Swan couldn't do anything but sit there and stare. There had to be some mistake. A very big mistake. There was no way the person standing before her was a part of this craziness.\n\nNo way.\n\nShe finally found her voice. \"Georgianna? What are you doing here? What is this about?\"\nThirteen\n\nFlipper reached the boat and attached himself to the ladder on the side. Lucky for him, no one had thought to pull it up. Taking slow, deep breaths, he pulled air into his lungs while ignoring the pain in his arms and legs. He didn't want to think about just how far he'd swum, but like he'd told his friends, he'd had to try.\n\nHe quickly eased back into the water when he heard the sound of an approaching boat and was grateful the vessel pulled up on the other side from where he was hiding. He glanced at his watch. It was synchronized with the ones worn by his teammates, and he knew they would do their best to get here soon. In the meantime, there was something he had to do.\n\nPulling a micro audio recorder off his belt, he moved back up the ladder to peek over the railing and into the boat.\n\nGood. Everyone's attention was on the approaching vessel and no one saw him when he attached the audio recorder that was no bigger than a dime to the interior wall of the boat. He saw Swan and Jamila seated on a bench with their hands tied behind their backs and Horacio was standing not far away. Other than a man in the cockpit, there was no one else onboard. Flipper knew that was about to change when he heard voices.\n\nSatisfied that the conversations would be recorded, he eased back down the ladder. When his watch began vibrating, he glanced down at the text message from Bane. On our way. Had 2 take care of a little problem 1st.\n\nFlipper wondered what kind of problem his friends had to take care of. No matter. They were on their way and that's what counted. He listened to the conversation going on in the boat as he began pulling items from his belt. He intended to be ready to crash this little party when the time came.\n\nHe shook his head, knowing Admiral Martin would be heartbroken to discover his own daughter had sold out their country.\n\n* * *\n\n\"I hate you,\" Georgianna said, glaring at Swan.\n\nSwan was taken aback by the woman's words. \"Why? What have I ever done to you? To any of you?\" she asked, glancing around at the people she'd assumed were friends\u2014Rafe, Horacio and Rosie. She hurt more at seeing Rosie than the others because she'd believed the woman had been a good friend.\n\n\"They work for me and did what they were told,\" Georgianna said.\n\n\"Work for you?\" Swan was even more confused.\n\n\"Yes. I'm in charge of the entire operation. But I'll tell you all about that later. First, let me tell you why I despise you so much. I've waited a long time to get this out in the open. When your father died and your mother would send you to us for the summer, my parents thought you were golden. They put you on a pedestal, especially my father. Did you know he called you his little island princess?\"\n\nYes, Swan knew but she also knew her godfather hadn't meant anything by it. It was just a nickname he'd given her when she was born. All three of her godfathers called her that sometimes. \"It was just a nickname, Georgianna.\"\n\n\"For you, it might only have been a nickname, but for me, it was Dad shifting his attention from me to you.\"\n\n\"Godpop 1 loves you. He wasn't shifting his attention to me, he was just being nice.\"\n\n\"Too nice, and I despised you for it. He had a daughter, yet any time your mother would call, he and Mom would drop everything and take off. Just because your father saved Dad's life\u2014that meant nothing. They were all SEALs and your dad was doing his job when he died. But it was as if Dad blamed himself and he needed to make it up by being nice to you, like you were somebody special. So, with the help of some friends, I decided to change everyone's opinion of you.\"\n\nIt was hard for Swan to believe what she was hearing. She'd never known that Georgianna harbored such feelings. Granted she hadn't always been overnice and had a tendency to be moody, but Swan hadn't detected animosity like this.\n\n\"What did you do?\" Swan asked her.\n\nGeorgianna smiled like she was proud of what she was about to say. \"I set up an espionage operation out of your shop with the help of Rafe and Horacio. Then, with Rosie's assistance, I made it appear that the secret information being sent to China was being done through your jewelry.\"\n\n\"What!\" Swan couldn't believe it. Her head was spinning from all the shocks she'd received tonight.\n\n\"I have to admit I put together a perfect plan. This guy I was sleeping with at the time assisted me by tipping off naval intelligence with what you were supposedly doing. They did their own investigation and my team and I made sure everything pointed at you. It should have been an open and shut case and you were to be arrested and charged with espionage.\"\n\nAs if she was tired of standing, Georgianna moved to sit on one of the benches. She frowned over at Swan. \"Everything was going according to plan until the final thread of the investigation reached Director Levart's desk.\"\n\n\"Godpop 3?\"\n\n\"See what I mean? You have three godfathers and I don't have a one,\" Georgianna said in a loud voice, pointing a finger at her. \"You don't deserve such love and loyalty, and I intended to tarnish their image of you.\"\n\n* * *\n\nFlipper's watch vibrated and he glanced down at the text message. Here. N place. Coop got layout of boat.\n\nHe texted them back. 1 N cockpit. 4 others. 2 hostages.\n\nHe quickly received a reply from Viper. Eliminating cockpit.\n\nGot 4 in scope. That particular text came from Bane, a master sniper.\n\nFlipper knew that although everyone was in place, timing was everything. Georgianna had no idea her words were being recorded so he wanted to let her talk before making his move. Then there would be no way she could deny anything.\n\nFrom listening to what the woman was saying, it was obvious she had mental issues. That could be the only reason to have such a deep hatred of Swan that Georgianna would go to such extremes. Georgianna assumed she'd had the perfect plan for Swan's downfall and Flipper was glad things hadn't turned out the way Georgianna intended.\n\nBefore inching up the ladder to listen to what else she was confessing, he texted the message: Will give signal.\n\n* * *\n\nSwan shook her head. It was obvious Georgianna's jealousy had blinded her senses and fueled her hatred. Didn't the others see it? Why were they following her blindly? Swan glanced over at Jamila and could tell by the look in her cousin's eyes that she was wondering the same thing.\n\n\"When Director Levart saw the report, he refused to believe you could be guilty of anything, especially betraying your country.\"\n\n_Thank God for that_ , Swan thought.\n\n\"He requested a thirty-day delay before agreeing to take any actions against you. Even after we made sure the investigation clearly spelled out your role in everything. There was no reason for you not to be charged,\" Georgianna said.\n\nShe paused a moment before continuing. \"Unknown to me and the others, Director Levart went to your other two godfathers and they put their heads together to see what they could do to prove your innocence. They decided to send one of their top SEALs to find out what he could and to prove your innocence.\"\n\nSwan drew in a deep breath. _Oh no, please don't let what Horacio said tonight be true. Please don't let David turn out to be someone other than what he said he was._\n\nGeorgianna's next words ripped into Swan's heart.\n\n\"The SEAL they sent was Lieutenant David Holloway. I guess you didn't know that all the time he spent with you was nothing more than an assignment. You meant nothing to him, Swan.\" Georgianna laughed as if she found the entire thing amusing while Swan's heart broke.\n\n\"Imagine how amused I was to find out just how taken you were with him, while not knowing the true purpose as to why he showed up here on the island. You were played, Swan,\" Georgianna was saying in between laughter. \"But don't worry. I sent some other members of my group to take care of him for you. I think he's dead by now.\"\n\nSuddenly a deep voice at the back of the boat said, \"As you can see, I'm very much alive.\"\n\n* * *\n\nSwan gasped just as the others did. Standing with legs braced apart and wearing only a pair of swim trunks with a utility combat belt around his waist, David looked like a mad badass. It was obvious everyone was shocked to see him, especially Georgianna, who had assumed he was dead.\n\n\"Drop your gun,\" he ordered Horacio, who was still holding his weapon on Swan and Jamila.\n\n\"How the hell did you get here?\" Horacio asked, enraged.\n\n\"I swam from Swan's home.\"\n\n\"That's impossible!\"\n\n\"Not if you're a SEAL master swimmer,\" David said. \"Now, do like I said and put your gun down.\"\n\n\"And if I don't? It will be a shame if I kill Swan or Jamila before you can get to anything on that belt you're wearing,\" Horacio sneered.\n\n\"Don't try it, Horacio. One of my team members who's a master sniper has all four of you within his scope. Before you could get off the first shot, you'd be dead.\"\n\n\"I don't believe you. There's no one else out here,\" Horacio said. When he lifted his gun to take aim at Swan, a shot rang out, hitting the man in the chest. The impact toppled him to the floor.\n\nJamila screamed and Swan understood. Jamila had fallen in love with a man who'd betrayed her and then gotten shot right before her eyes.\n\nSuddenly, Rafe dived for the gun that had dropped from Horacio's hand. Before he could reach it, another shot rang out that hit him in the side. He fell to the floor as well.\n\n\"Either of you ladies want to join them?\" Flipper asked Georgianna and Rosie.\n\nRosie looked like she was in shock and ready to pass out.\n\nHowever, Georgianna looked furious. \"You won't get away with this. No matter what you tell my father, he will never believe you over me,\" she said with absolute certainty. \"I'll tell him that you decided to team up with Swan and she turned you against your country.\"\n\n\"I figured you would lie. That's why I've recorded your little confession to Swan detailing everything. I can't wait for your father to listen to it.\"\n\nSuddenly the boat was surrounded by several naval vessels and sharp beams of light shined on them. A voice through a foghorn said, \"Lieutenant Holloway, we are coming aboard.\"\n\nA dozen men wearing SEAL gear rushed on board with their guns drawn, immediately taking Georgianna and Rosie into custody. Bane, Viper, Coop and Mac boarded the boat as well. Mac rushed over to check Horacio and Rafe. There really was no need since they were both dead.\n\nFlipper rushed over to Swan and Jamila to untie their hands. More than anything, he wanted to pull Swan into his arms and tell her he loved her. He wanted her to put out of her mind what Georgianna had said about him until she'd heard his side of things. However, he knew when she pulled away from him to give her cousin a hug that she didn't want to give him a chance to explain.\n\nHe didn't intend to let her walk away.\n\n\"We need to talk, Swan,\" he said, looking down at her.\n\nShe glared up at him. \"We have nothing to say to each other. Your assignment is over, Lieutenant Holloway. Now leave me alone.\"\nFourteen\n\nTwo weeks later\n\n\"How long are you planning to be mad at the world, Swan?\"\n\nSwan glanced over at Candy. Her best friend had returned to the Keys after hearing about what happened and she'd decided to stay. Swan was glad Candy had returned home but she was saddened by what had brought her back.\n\n\"I am not mad at the world,\" Swan said, taking a sip of her orange juice.\n\n\"But you are still mad at one particular man,\" Candy said, coming to sit beside Swan on the sofa.\n\nSwan couldn't deny that was true so she didn't. \"And what if I am?\"\n\n\"He had a job to do, Swan. He was given orders. Surely you understand that.\"\n\nSwan glared at Candy. \"I'm sure none of my godfathers' orders included sleeping with me.\"\n\n\"I'm sure they didn't but David didn't force himself on you.\"\n\n\"No, but he deceived me.\"\n\n\"So did the others.\"\n\nDid Candy have to remind her? \"And I'm not talking to them either.\"\n\nThat wasn't totally true since Swan had reached out to Georgianna where she was being held at a federal prison in Orlando. The woman had refused to see her. Swan knew Georgianna was undergoing psychiatric evaluations to see if she was fit to stand trial.\n\nSwan's godparents were heartbroken and she understood how they felt. Like her, they'd had no idea Georgianna harbored such hatred toward Swan, enough to do what she'd done. With both Rafe and Horacio dead, it was Rosie who was singing like a bird, telling everything she knew for a lessened sentence.\n\nAccording to Rosie, Georgianna had manipulated a number of the men at naval intelligence into doing whatever she wanted them to do. When you were the admiral's daughter, you could wield that kind of power. She had even threatened a few with blackmail. She'd deliberately recorded several of the men having sex with her and then threatened to give the tape to her father and accuse them of rape.\n\nSome of the men were not only married but a number were high-ranking military officers. Fearful of court-martial, the men had done whatever Georgianna asked, including falsifying records. So far, more than twelve men had been named in the scandal.\n\n\"I take it David hasn't called.\"\n\nSwan drew in a deep breath. She had seen him last week when they'd had to show up at the naval station to give statements. \"Yes, he's called. Several times. But I refuse to answer. Like I told him, we have nothing to say to each other. His assignment is over.\"\n\n\"And do you honestly think that's all you were to him, Swan?\"\n\n\"Yes, but it doesn't matter.\"\n\n\"I think it does,\" Candy countered.\n\n\"And you think too much,\" Swan said, easing off the sofa.\n\nThe first week after the incident on the boat, she had closed her shop while naval investigators did a thorough search of Rafe's tattoo parlor. She had used that time to take care of Jamila, who was still broken up over Horacio. Jamila had loved him and in a single night had seen him become an abusive monster, a man she hadn't known. Then in the end, Jamila had watched him die before her eyes.\n\nSwan knew Jamila was going through something that only time could heal. That's why when Swan had reopened the shop this week and Jamila had asked for extra work hours, Swan had given them to her.\n\n\"So what are you going to do?\" Candy asked her.\n\nSwan glanced over at her. \"About life? Work?\"\n\n\"No, about David.\"\n\nSwan just couldn't understand why Candy couldn't accept that David was no longer in the equation. \"I'm a survivor, Candy. Although it was hard, I made do after my parents' deaths and I will make do now.\" She glanced at her watch. \"I'm getting dressed to go into the shop today. The cruise ship comes into port tomorrow, so business will pick up. I want to make sure most of my new pieces are on display.\"\n\nAnother thing they had found out was that Horacio had been fired from the cruise ship months ago but hadn't told Jamila. He had moved into Rosie's place while the woman had been gone. The duplicity of the people she'd thought she knew simply amazed Swan.\n\n\"And I need to be on my way,\" Candy said. \"I promised my folks we would go out to dinner tonight. You can join us if you like.\"\n\n\"Thanks for the invite, but I'll pass. I just want to have a relaxing evening here tonight. I might go swimming on the beach later.\"\n\nSwan had called Jamila and told her she would bring lunch from their favorite sandwich caf\u00e9. However, there were no clients in her shop when Swan got there, so she decided to do something she usually didn't do, which was close for lunch.\n\nNormally, the shop remained open and she and Jamila would alternate lunch duties. But today she wanted to check on Jamila, talk to her to see how she was faring. Although Swan had been there for Jamila last week, they hadn't had a real honest-to-goodness talk since Jamila had admitted to being her cousin.\n\n\"What are you doing?\" Jamila asked when Swan put up the Closed sign and pulled down the blinds.\n\nSwan smiled over at her. \"New store policy. From here on out, we will close at noon for lunch.\"\n\n\"What about the sales you'll lose?\"\n\nSwan shrugged. \"Sales aren't everything. Besides, it's just for an hour. Come join me in my office.\"\n\n\"All right, let me grab some sodas out of the refrigerator.\"\n\nA few minutes later, she and Jamila were enjoying their lunch when Swan gave Jamila a long look. \"How are you doing?\"\n\nJamila shrugged. \"Okay, I guess. Trying to move on. I loved Horacio so much only to find out he wasn't the man I thought he was.\"\n\n\"I know the feeling.\"\n\n\"No, you don't.\"\n\nSwan snatched her head up, frowning. \"Excuse me?\"\n\n\"I said you don't know the feeling, Swan. David Holloway was nothing like Horacio. David intended to save you and Horacio would have killed me if that woman had ordered him to do so. Big difference.\"\n\n\"But like you, I was betrayed.\"\n\n\"How?\" Jamila countered. \"Your godfathers sent David Holloway here to prove your innocence and he did.\"\n\nJamila put her soda can down and then added, \"And another thing. What man takes a chance and swims across the ocean to save a woman? Do you know how far from land we were? Think about that.\"\n\nSwan had news for her\u2014she _had_ thought about it. She could never forget how David had appeared seemingly out of nowhere on that boat, looking tough and ready to kick asses while wearing nothing more than an outlandishly tight pair of swim trunks with a military belt around his waist. Even when she'd been in what seemed like a dire situation, that hadn't stopped the woman in her from noticing how dangerously sexy he'd looked at that particular moment.\n\n\"When I mentioned what an astounding feat he'd accomplished to his friends,\" Jamila said, reclaiming Swan's attention, \"they said that's why they call him Flipper. Did you know that's his code name as a SEAL?\"\n\nSwan wiped her mouth with a napkin. \"Yes, I knew he was called Flipper. But no, I didn't know it had anything to do with him being a SEAL because I didn't know he was one. I assumed Flipper was his nickname.\"\n\nSwan forced from her mind the day she'd asked him about those dolphin tattoos. He'd told her then they represented Flipper. That had been the day they'd made love in this office. Right here on this desk.\n\nShe wished she wasn't thinking so hard about that now.\n\nShe looked over at Jamila. \"Why are we talking about me instead of you?\"\n\n\"Because I think you should and because I think I should,\" Jamila continued. \"Talking about your situation actually helps me believe that not all men are jerks and that there are some who still possess real honor, Swan. Whether you want to admit it or not, David Holloway is an honorable man. He couldn't help being attracted to you any more than you could help being attracted to him.\"\n\nSwan stuffed the wrappings from her sandwich into the empty bag. \"Now you sound like Candy.\"\n\n\"Maybe there's a reason why I do,\" Jamila said, stuffing her own wrappings into a bag. \"It might be because Candy and I can see things that you refuse to see. I often think about what could have happened to us had David and his friends not shown up when they did. Do you ever think of that?\"\n\nSwan drew in a deep breath. \"I try not to.\"\n\n\"I think you should,\" Jamila said, standing. \"Thanks for bringing lunch. It will be my treat the next time.\" She then walked out of the office.\n\nSwan stayed in her office after Jamila left, trying to put their conversation out of her mind. She was working on her computer, verifying inventory, when her office phone rang. \"Thank you for calling Swan's. How may I help you?\"\n\n\"Hello, island princess.\"\n\nShe smiled upon hearing her godfather's voice. \"Godpop 2. How are you?\"\n\n\"I'm fine. I just wanted to check on you. So much has happened and I wanted to make sure you're okay.\"\n\nShe had spoken to each of her godfathers and had thanked them for believing in her. They had taken a risk with their individual careers to do that. \"I'm fine. How is Godpop 3?\"\n\n\"He's fine but as the director of naval intelligence, he has his hands full with the investigation. It seems that more names are popping up in this scandal each day.\"\n\n\"And how are Godpop 1 and Barbara?\"\n\n\"They are as well as can be expected under the circumstances. Learning about Georgianna was a shocker for all of us. We had no idea. When we decided to send Lt. Holloway to prove your innocence, the three of us weren't sure just what he would uncover. The only thing we knew for certain was that you weren't guilty of anything.\"\n\n\"Thanks for believing in me.\"\n\n\"You have Andrew's blood in your veins. You could no more be a traitor to your country than he could. Considering all that happened, I'm glad Holloway remained in Key West when he could have left.\"\n\nSwan sat up straight. \"Wasn't David on assignment?\"\n\n\"Not the entire time. His assignment officially ended when he sent that ink in to be analyzed and we discovered it was tainted. I told him that he no longer had to stay in the Keys since by then we knew you weren't involved and we would take over the investigation from there.\"\n\n\"Then why did he stay?\"\n\n\"To protect you.\"\n\n\"He told you that?\" she asked.\n\n\"Yes. I remember the conversation like it was yesterday. I told him he could consider his job assignment complete and go home to Texas and enjoy the remainder of his leave. But he said he wanted to hang around Key West for a while.\"\n\nHer godfather paused. \"I asked him if the reason he wanted to stay was because he thought your life might be in danger. He said he felt that as long as Duggers and Jacinto didn't know they were suspects, then no, your life wasn't in any immediate danger. He informed me that the reason he wanted to stay was because you had come to mean a lot to him. I told him in that case how he spent the rest of his leave was his decision. And, Swan?\"\n\nShe drew in a deep breath. \"Yes?\"\n\n\"As his commanding officer, I felt the need to remind him that although he was no longer on assignment, since the issue that had started with you was an ongoing investigation, he could not tell you anything.\"\n\nWhen Swan didn't reply, her godfather asked, \"You're still there, Swan?\"\n\n\"Yes, Godpop 2, I'm still here.\"\n\n\"Did you not know how Holloway felt about you?\"\n\n\"No. I thought I was just an assignment.\"\n\n\"You were at first and I'm glad you were. Otherwise you would be in jail wrongly accused of a crime you hadn't committed. But on the flip side, I'm also glad that when you stopped being an assignment, Holloway had the insight to stay and look out for you because he cared for you.\"\n\nLong after her telephone conversation with her godfather ended, Swan remained seated at her desk, leaning back in her chair and sitting in silence while thinking about what Candy, Jamila and her godfather had said.\n\nSome people never got betrayed, but she had been, a lot. William, Rafe, Horacio, Rosie, Georgianna and even Jamila. No one had been who she'd thought.\n\nShe remembered David and replayed in her mind all the time she'd spent with him since that day he'd first walked into her shop.\n\nWas anything he'd told her true? Did he really come from a huge family? Was his mother even celebrating a birthday? Did he honestly have three sisters-in-law?\n\nOne thing was for certain, both Candy and Jamila were right. David hadn't pushed her into sleeping with him. In fact, Swan was the one who'd invited him to dinner at her place with the full intent of having sex with him.\n\nShe got up from her desk and walked over to the window. She knew from Jamila that David had left the island with his friends after that first week, after he'd completed all the questioning by naval intelligence. Was he back home in Texas? Did his parents really own a medical supply company? What parts of what he'd told her were true and what parts were fabricated for his assignment?\n\nAnd why did she still love him so much it hurt...even when she didn't want to love him? Even when she didn't know how he felt about her? He might have told her godfather he cared for her but David hadn't told her anything. Shouldn't he have? But then, had she given him a chance to do so?\n\nThe answer to that flashed in her mind quickly. No, she hadn't.\n\nHe had saved her life that night, swam across the ocean to do so, and then she'd told him she didn't want to talk to him. And he had honored her wishes...for that one night. Then he had called her almost every single day since, and yet she had refused to take his calls.\n\nHe hadn't called today.\n\nDid that mean he'd given up and wouldn't try contacting her again? Was she ready to put her heart on the line and contact him?\n\nShe wasn't sure. But what she _was_ certain of was that they needed to do what they hadn't done before. They needed to get to know each other. She needed to know which parts of what he'd told her about himself were true and which were false.\n\nShe wanted to get to know the real David Holloway.\n\n_Then what?_\n\nHadn't she convinced herself she wanted no part of a man in the military? And what about her decision to never to get seriously involved in an interracial relationship like her parents had? Why did all of that no longer matter to her when she thought about her and David deciding to have a future together?\n\nMaybe that's how love worked. It made you see the possible instead of the impossible. It made you want things you told yourself were not good for you because you were afraid to reach beyond your comfort zone.\n\nTaking a deep a cleansing breath, she decided to call David tonight before going to bed. She had no idea what she would say to him but the words would come.\n\nShe doubted he would want to come back to the Keys anytime soon, so she would let him know she would come to him if he still wanted to talk. She would see what he said before asking Jamila if she could take care of the shop while Swan was gone. David might very well tell her that it was too late, that they had nothing to talk about. But there was a chance he would embrace her words. Embrace her.\n\nHer mood suddenly lightened, knowing that was a possibility.\n\n* * *\n\nFlipper entered the hotel room and tossed his luggage on the bed. Different hotel but same city. He had given Swan two weeks and now he was back. They needed to talk and clear up some things. She hadn't accepted his calls, but now he was here and he wouldn't be ignored.\n\nHe shook his head when his cell phone rang. \"Yes, Coop?\"\n\n\"Have you seen her yet?\"\n\n\"No, I just got here. In fact, I walked into my hotel room less than five minutes ago.\"\n\n\"Okay. And there's another reason I called. Bristol is pregnant.\"\n\n\"Wow, man. Congratulations. I didn't know you guys were trying.\"\n\nCoop laughed. \"We're always trying. But seriously, we figured it was time Laramie had a playmate.\"\n\n\"Sounds good to me.\"\n\n\"I hope things work out with you and Swan, Flipper.\"\n\n\"I hope so, too.\"\n\n\"And do me a favor.\"\n\n\"What?\"\n\n\"For once, open up. Tell her how you feel. Don't beat around the bush. You have a tendency to do that. Women love a man to get straight to the point and share their feelings. I hate to say it, bro, but you're not good at doing that.\"\n\nCoop was right, he wasn't. \"I never had to do that before. I've never truly loved a woman before Swan.\"\n\n\"I understand. But you do love her, so make sure she knows it. A woman has to believe she's loved.\"\n\nFlipper chuckled.\n\n\"What's so funny?\" Coop asked him.\n\n\"You're giving relationship advice. Do you know how much like Mac you sound?\"\n\nCoop chuckled as well. \"You would have to point that out. I guess it comes with loving a woman.\"\n\n\"I guess so.\"\n\n\"No guess in it, remember it. Know it. Feel it. Take care and good luck.\"\n\nAfter ending the call with Coop, it wasn't long before Flipper got calls from Bane, Viper and Mac as well, all letting him know they hoped things worked out for Flipper and Swan. All giving him advice. They were married men who had the women they loved and they wanted him to have the woman he loved as well.\n\nHe appreciated good friends who not only watched his back but who also cared about the condition of his heart. They knew about the pain he had lodged there and it got worse every day he and Swan were apart.\n\nFlipper glanced at his watch. Swan's store would be closing in less than an hour. He would give her time to get home and relax before paying her a visit. He refused to let her put things off any longer. They needed to talk.\n\nHe loved her and it was damn time she knew it.\nFifteen\n\nSwan had just poured a glass of wine to enjoy while sitting on the patio when she heard the knock at her door.\n\nShe knew Candy had gone out to dinner with her family and Jamila had mentioned she would just stay in tonight and chill. Swan had invited Jamila to join her so maybe her cousin had changed her mind.\n\n_Her cousin_.\n\nThat was taking a lot of getting used to but Swan knew her parents would want a family connection for them. Jamila was the only family Swan had and she appreciated their friendship more than ever.\n\nShe reached the door and glanced through the peephole. Her heart nearly stopped.\n\nWas it really David? She blinked and looked again and saw it was really him. Back in Key West. And he was standing in front of her door looking like he always did, sexy as hell.\n\nDrawing in a deep breath, she removed the security lock and opened the door. \"David? I thought you'd left the island.\"\n\n\"I had but I returned today. May I come in?\"\n\nShe nodded and stepped aside. The moment he passed her, she caught a whiff of his masculine scent, the same one she was convinced still lingered in her bedroom.\n\nSwan closed the door and stood to face him. He was standing in the middle of her living room wearing a pair of shorts and a sleeveless shirt with a huge picture of a dolphin. _Flipper_. Her gaze moved beyond the shirt to his face to find his laser-blue eyes staring at her.\n\nShe cleared her throat. \"I was about to sit out back and drink a glass of wine while enjoying the view. Would you like to have a beer and join me?\"\n\nShe could tell he was surprised by her invitation. She hadn't bothered to ask why he was there.\n\n\"Yes, I'd like that.\"\n\nMoments later, they were sitting side by side on a bench that overlooked the beach. They had been sipping their drinks for a few moments when he said, \"I told you that night two weeks ago that we needed to talk, Swan. I think we still do.\"\n\nYes, they did. She would let him go first. \"Okay, I'm listening.\"\n\n\"I want you to do more than just listen, Swan. I want you to engage by asking questions, giving me feedback, and I would like to be able to do the same with you.\"\n\n\"Okay, that seems fair because I do have some questions for you.\"\n\n\"Ask away.\"\n\nShe took another sip of her drink. \"You didn't tell me you were a SEAL and I'd\u2014\"\n\n\"I couldn't tell you I was a SEAL, Swan,\" he interrupted to say. \"That's why I lied and said I was no longer in the military.\"\n\n\"Yes, I know that now. I want to tell you, because of what happened to my father, I had made up my mind never to get serious about a military man...especially not a SEAL.\"\n\n\"Oh, I see.\"\n\nShe wouldn't tell him yet that she'd changed her mind about that. \"Is your mother's birthday really next month and do you have four brothers?\"\n\n\"Yes. Everything I told you about my family is true. I never lied to you about anything pertaining to them. I just omitted some details and couldn't elaborate on certain things.\"\n\nShe then put her wineglass down and turned toward him. \"Why did you sleep with me, David, when I was just an assignment to you?\"\n\n* * *\n\nFlipper knew this was the time of reckoning and what he told her would have an impact on their relationship for the rest of their days. He needed her to understand.\n\n\"You were supposed to be an assignment, Swan. But honestly, I don't think you ever were. From the moment I walked into your shop and saw you, a part of me knew I had to fight hard to be objective and do the job I'd been sent here to do.\"\n\nHe paused. \"I tried to keep my attraction to you out of the picture but found it harder and harder to do. Each time I saw you while getting to know you, I fell deeper under your spell. It was hard pretending with you.\"\n\nHe decided to be totally honest with her. \"Just so you know, that day you invited me to your place for dinner wasn't my first time there. I'd been to your home without you knowing anything about it. But at the time, it was just a house I was checking out as part of an investigation. The day you invited me to dinner, I saw it through another pair of eyes. Yours. And for me, it then became your home.\"\n\nShe drew in a deep breath. \"You invaded my privacy by letting yourself into my home, but that's not why I'm upset. I accept that you had a job to do and I was your assignment but...\"\n\n\"But what?\"\n\n\"You still haven't fully answered my question, David. Why did you make love to me?\"\n\nDavid frowned, realizing that he _hadn't_ answered her question. His teammates often teased him about beating around the bush, sometimes providing too much context instead of just sticking with the facts.\n\n\"The reason I made love to you, Swan, was because I desired you. Everything about you turned me on. Your looks, your scent, your walk...and then after our first kiss, it was your taste. Fighting my desire for you was no longer an option, although I tried being honorable enough not to seduce you.\"\n\n\"But then I seduced you,\" she said quietly.\n\nHe smiled. \"No, I think that night we seduced each other. Everything we did was mutual.\"\n\n\"Yes, it was.\" She took another sip of her wine. \"I spoke with Godpop 2 today and he told me your assignment ended but you decided to stay. Why?\"\n\nOkay, no beating around the bush this time, Flipper decided. \"The reason is that by then I had fallen in love with you. In all honesty, in my heart you stopped being an assignment the first time I made love to you. I crossed the line of what was honorable, and I knew why. Because I felt you here, Swan,\" he said, pointing to his heart. \"I felt you here in a way I've never felt before. No woman has ever been here, Swan. But during the one time you shouldn't have, you got there anyway.\"\n\n\"And now? How do you feel now?\"\n\nHe placed his beer bottle aside and turned toward her. \"Now you are still in my heart. Even more so. I love you so much I ache on the inside when I'm not with you. I love you so much I think of you even at times I shouldn't.\"\n\nHe reached out and took her hand in his. \"Now I need to know, Swan, just how you feel about me.\"\n\n* * *\n\nSwan felt the gentle tug on her hand and, surprising even herself, she moved to sit in his lap.\n\nWhen he wrapped his arms around her, she felt comfort flow through her. She turned in his lap to look down at him. He'd given her answers to all her questions, now she intended to give him answers to his.\n\n\"I love you, David. I fought it at first. I didn't know about you being present-day military, but I also had a problem...not with interracial dating...but with allowing anything to come of it. I saw how others saw my parents at times. Not as a beautiful couple in love but as an interracial couple in love. There should not have been a difference. I never wanted to deal with what they had to deal with in the name of love.\"\n\nShe paused. \"But then I moved beyond thinking that way after I fell in love with you. Then I realized how my parents must have felt, believing nothing mattered but their love. Even if the world was against them, as long as they had each other, that's what truly counted.\"\n\n\"So you do love me?\" he asked her as if for clarity.\n\nSwan didn't have a problem clarifying anything for him or anyone else. \"Yes, I love you, David Flipper Holloway.\" And then she lowered her mouth to his.\n\nShivers of profound pleasure shot through every part of Swan's body when David slid his tongue into her mouth. Sensations bombarded her as she concentrated on his taste, his scent and the way he pulled her tongue into his mouth to mate with his. And when she felt his hands inch upward and slide beneath her top, his touch made her purr.\n\nBoth his taste and his touch were awakening parts of her, making her feel alive in a way she hadn't felt since the last time they'd been together. Here at her house. In her bed.\n\nWhen his fingers touched her bare breasts, using his fingertips to draw circles around her nipples, she oozed deeper into the kiss, almost feeling like melted butter in his arms.\n\nHe slowly pulled his mouth from hers and looked at her. His blue eyes were sharp and filled with the same desire she felt. \"Any reason why we can't take this inside?\"\n\nShe wrapped her arms around his waist. \"No, there's no reason.\"\n\n\"Good.\"\n\nAnd then standing with her in his arms, he carried her into the house.\n\n* * *\n\n\"Just so you know, David, I didn't ask you all my questions,\" Swan said when he placed her on the bed.\n\nDavid glanced down at her. \"You didn't?\"\n\n\"No, but I can wait. None are more earth-shattering than this is going to be. And I need this.\"\n\nHe caressed the side of her face with his finger. \"I need this, too. I know why I need it, tell me why you do.\"\n\nShe met his gaze and held it while she said, \"I love the feel of you inside of me. I've never felt anything so right before. So pleasurable.\" She smiled. \"Do you know I retired my sex toy?\"\n\nHe chuckled. \"That's good to know.\"\n\n\"Um, not too much information?\"\n\n\"No. Nice to know what used to be my competition,\" he said as he began removing her clothes. Lucky for him, she wasn't wearing much. Just a top, shorts and a thong. Flipper had discovered outside earlier that she wasn't wearing a bra. He'd noticed more than once that she liked her breasts being free and so did he.\n\nShe reached out and tugged at his T-shirt and he assisted by removing his own shorts. Then he rejoined her on the bed. Reaching out, he lifted her by the waist.\n\n\"Wrap your legs around me, Swan. I'm about to join our bodies, to make us one.\"\n\nAs soon as her legs were settled around his waist, his shaft touched her core. She was wet and ready. Tilting her hips, he whispered the words, \"I love you,\" before thrusting hard into her.\n\n\"David!\"\n\nArching her back off the bed, she provided the prefect position for his penetration to go deeper. They were a perfect fit. They always would be. Not just in lovemaking but in everything they did from here on out. They had become a team.\n\nHe began moving, slowly at first and then harder and deeper, over and over again. The only constant sounds in the room were their breathing and flesh slapping against flesh. The air surrounding them was filled with the aroma of sex.\n\nHe felt on fire, like his entire body was burning and the flames fueled his need, his desire and his love. She was looking up at him, holding his gaze, and he hoped Swan saw the depth of love in his eyes.\n\nHe clenched his jaw when he felt it, the stirring of pleasure in his groin. The feeling was slowly spreading through his body and when Swan gripped his shoulders and dug her fingers into his skin, he continued to thrust inside of her like his life depended on it.\n\nAnd when she screamed out his name, he knew the same sensations that were taking him were taking her.\n\nHe drew in a sharp breath only moments before calling out her name. Multiple sensations tore into him, causing an explosion inside of him that had him bucking his body in an all-consuming orgasm. The sensations kept coming until he let go and his release shot deep inside of her.\n\nHe knew right then that he wanted her to have his baby. If not this time, another time. One day, he intended to make it happen.\n\nMoments later, he slumped down beside her and wrapped his arms around her as he tried to get his breathing under control. After recovering from his explosive orgasm, when he was able to talk, he said, \"I feel like I've been burned to a crisp.\"\n\n\"Hmm, speaking of burning, do you know what I thought was hot?\" she asked, drawing in deep breaths of air into her lungs.\n\n\"No, what?\"\n\n\"You on that boat wearing nothing but swim trunks and that military belt around your waist. Now, that was hot.\"\n\nHe grinned. \"You liked that, huh?\"\n\nShe smirked up at him when he straddled her body again. \"I liked it.\" Then her features became serious. \"I still can't believe you swam all that way to save me.\"\n\nHe leaned in and brushed a kiss across her lips. \"Believe it.\"\n\nHe then pulled back and looked down at her. His expression was serious. \"I'm a damn good swimmer. I'm known to be able to hold my breath underwater for long periods of time. Longer than what most would consider normal. But I wasn't sure I was going to make it to the boat, Swan. I told my friends I had to try even if I died trying because the woman I love was on that boat. That's what kept me going. That's what fueled every stroke I made into the ocean waters. And when my body felt tired, like I couldn't possibly swim another lap, I would think of a life without you and for me that was unacceptable.\"\n\nHe drew in a sharp breath. For a quick minute, he relived the feel of the cold water as he swam nonstop to the boat to save her, not knowing if he would make it in time. \"I had to save you.\"\n\n\"And then I rebuffed you. I refused to have that talk you wanted.\"\n\n\"I understood. I had been listening to what Georgianna Martin was saying, the picture she painted. I told myself that once I talked to you and told you the truth that you would believe me. I was just giving you time to think about everything. I figured you would realize that I did care for you.\"\n\nShe reached up and caressed the side of his face. \"You never told me you cared.\"\n\n\"I did. Our last night together, when you were asleep, I told you before I left that I loved you. I had planned to tell you the next day when we were together but that's when you were taken.\"\n\n\"And you came back,\" she said.\n\n\"That was always my plan, Swan. I never intended to let you go. I love you that much. And just so you know, my entire family is rooting for me. I told them about you and they can't wait to meet you. My brothers and I are giving Mom a party for her birthday next week. Will you go to Texas with me?\"\n\nWhen she hesitated, he added, \"What I told you about them is true. My parents accept people for who they are and not how they look. Will you trust me about that?\"\n\nShe met his gaze and nodded. \"Yes, I will trust you and yes, I will go.\"\n\nA huge smile spread across his face. \"I can't wait to introduce you to everyone. And I've got the perfect thing for you to wear.\" He quickly eased off the bed.\n\nShe pulled herself up. \"What's going on? You plan on dressing me that night?\"\n\nHe glanced over his shoulder, chuckling as he pulled a small white box out of his shorts. \"Something like that.\"\n\nHe returned to the bed and pulled her up to stand her on her feet beside the bed. Then he got down on one knee and looked up at her. \"I love you, Swan. I know we have a lot of things we still need to overcome. But I believe we will do so together. Forever. Will you marry me?\"\n\nHe saw tears form in her eyes when she nodded. \"Yes, I will marry you.\"\n\nHe slid the ring on her finger and at that moment Flipper knew he was halfway to having his world complete.\n\nHe would get the other half the day she became his wife.\nSixteen\n\nSwan glanced down at the ring David had put on her finger last week. Seeing it gave her strength and she definitely needed strength now, she thought as she entered a huge ballroom on his arm. It was his mother's sixtieth birthday party.\n\nThey had flown into Dallas last night so this would be the first time she met his family. Nervous jitters had tried taking over her stomach but a smile from David was keeping most of them at bay. He was convinced his family would love her and he had told her over and over that she was worrying for nothing. She was the woman he wanted and his family would love his choice.\n\n\"There's Mom and Dad,\" he said, with his arms around her shoulders as she carried his mother's gift. The same gift he'd purchased that first day he'd come into her shop.\n\nA man she knew had to be one of David's brothers whispered something to the older couple and they turned with huge smiles on their faces.\n\nAt that moment, Swan knew David had inherited his father's eyes and that the smiles on the couple's faces were genuine. She could actually feel their warmth. David's mom was beautiful and did not look like she was sixty or that she had five grown sons.\n\nWhen they reached his parents, David made the introduction. \"Mom. Dad. I want you to meet the woman who has agreed to be my wife, Swan Jamison.\"\n\n\"It's an honor to meet you,\" Swan said, extending her hand to his mother.\n\nInstead of taking it, the older woman engulfed Swan in a huge hug. \"It's wonderful meeting you as well, Swan, and welcome to the family.\"\n\n\"Thank you. Here's your gift. Happy birthday.\"\n\n\"Thank you.\"\n\nShe received a hug from David's father as well. Then suddenly she was surrounded and a laughing David made introductions. All his brothers had those same blue eyes and like David, they were very handsome men. She could see why when she looked at the older Holloways; they were a beautiful couple. And Swan could tell from the way Mr. Holloway looked at his wife and the way Mrs. Holloway would look back at him that the couple was still very much in love.\n\nA few nights ago, David had shared the fact that because his mother had been married to a Navy SEAL for over forty years and had five sons who were SEALs, she counseled a number of SEAL wives who had difficulties with the frequency and longevity of their spouses' missions. Swan had been glad to hear that since she would become a SEAL's wife soon.\n\nBecause David would be leaving in less than four months on anther mission, they hoped to marry within a year. Surprisingly, David wanted a big wedding. She agreed as long as the wedding took place in the Keys.\n\nThe logistics of having a big wedding were enormous, given he had four brothers who were SEALs on different teams. Not to mention his closest four friends were SEALs as well. That meant Swan and David had to make sure everyone would be on leave in the States at the same time.\n\nDavid also introduced Swan to her future sisters-in-law and they loved her engagement ring. The three were friendly and she liked them immediately. She was also introduced to other members of David's family\u2014his grandparents, his niece, nephews, cousins, aunts, uncles\u2014it was obvious the Holloway family was a huge one.\n\n\"Now I want to reintroduce you to four guys who are just as close to me as brothers. As you know, they came to the Keys to assist me in proving your innocence. And even when my assignment with you ended, they didn't leave. They stayed.\"\n\nShe had met his four friends that night after the incident on the boat, when they'd had to give statements. She had thanked them for their help but they hadn't been officially introduced.\n\n\"Did I tell you how beautiful you look tonight, sweetheart?\"\n\nShe smiled up at him as they walked across the ballroom floor to the four men and their wives. \"Yes, you told me. Thank you.\" She, Candy and Jamila had gone shopping in Miami. She'd known this was the perfect dress when she'd seen it on a store mannequin.\n\nWithin a few minutes, she had been introduced to Brisbane \"Bane\" Westmoreland and his wife, Crystal; Gavin \"Viper\" Blake and his very pregnant wife, Layla; Laramie \"Coop\" Cooper and his wife, Bristol; and Thurston \"Mac\" McRoy and his wife, Teri.\n\nAfter spending time with the couples, Swan felt that just like her future sisters-in-law, the four women were genuinely friendly and Swan looked forward to getting to know them better. They loved her engagement ring as well and told David he'd done a great job in picking it out.\n\n\"So what do you think?\" David leaned down to ask, taking her hand in his and leading her to where his parents, siblings and their spouses were getting ready to take a group picture.\n\nShe grinned up at him. \"Um, for starters, I think I need to start calling you Flipper, since everyone else does. And then, _Flipper_ , I think I am one of the luckiest women in the world right now. I love you.\"\n\nHe chuckled as he pulled her to the side of the room and wrapped his arms around her waist. \"And I, Swan Jamison, think I'm the luckiest man in the world, and I love you.\"\n\n\"A very wise woman, my mother, once told me that when you meet a man who puts that sparkle in your eyes then you'd know he was a keeper. You, Flipper, are a keeper.\"\n\nHe smiled. \"You, my beautiful Swan, are a keeper as well.\"\n\nFlipper then lowered his mouth to hers.\nEpilogue\n\nA year later in June\n\nBane Westmoreland leaned close and whispered to Flipper, \"Don't get nervous now. You wanted a big wedding and you got it.\"\n\nFlipper couldn't say anything because Bane was right. He stood flanked by his father, who was his best man, and twelve groomsmen\u2014namely his brothers, best friends and cousins.\n\nOnly his SEAL teammates knew Flipper had a tendency to tap his finger against his thigh when he was nervous. He stopped tapping but not because he noticed that Viper, Mac and Coop were grinning over at him. But then he figured both Viper and Coop had reasons to grin since they'd both become fathers this year. Viper was the proud father of a son, Gavin IV, and Coop had a beautiful daughter they'd named Paris, since that was where he'd first met his wife.\n\nIt was a beautiful day for a beach wedding and so far everything was perfect and going according to plan. Swan had hired one of the local wedding planners and the woman had done an awesome job. She had thought of everything, including the super yacht that could hold their five hundred guests that they'd be using for the wedding reception. It was anchored in the ocean near Swan's beachfront home. A fleet of passenger boats had been chartered to transport the wedding guests out to the yacht.\n\nA ten-piece orchestra sat beneath towering balustrades draped from top to bottom in thin white netting. Chairs were set up on the beach, auditorium style, facing the decorative stage where Flipper and the men in the wedding party stood waiting.\n\nSuddenly, the music began and all the ladies strolled down the beach and up the steps.\n\nSwan had chosen her wedding colors of purple and yellow and Flipper had to admit the combination was striking. It took all twelve women long enough to do their stroll. His niece was a flower girl and Coop's son and one of Flipper's nephews were the ring bearers.\n\nFlipper almost held his breath when what looked like a huge forty-foot golden swan was rolled onto the beach. When the orchestra changed their tune for the \"Wedding March,\" the swan opened and his Swan appeared in a beautiful, dazzling white gown. She looked beautiful, stunning and breathtaking all rolled into one.\n\nFlipper stared at the woman who would be his wife and felt so much love in his heart. He hadn't known until now just how much he could feel for one woman. They had spent the past year deepening their friendship and their love. He looked forward to returning from his covert operations, knowing she would be there waiting on him.\n\nHe watched as she slowly strolled toward him. All three of her godfathers participated in walking her up the aisle, passing her off to the other so many feet along the way. Then all three of them gave her away. When Swan reached his side and extended her hand to him, he accepted it while thinking she was _his_ Swan.\n\nHis beautiful Swan.\n\nThe wedding ceremony began. What Flipper would remember most when he looked back was when the minister announced them husband and wife and told him he could kiss his bride.\n\nFlipper pulled Swan into his arms and lowered his mouth to hers. She was his and he intended to keep her happy for the rest of his days. They would be flying to Dubai for a two-week honeymoon and then return to the Keys where they planned to make their permanent home.\n\nWhen David released Swan's mouth, the minister said, \"I present to everyone David and Swan Holloway.\"\n\nFlipper knew they were supposed to exit by walking down the golden steps that led to the boat that would transport them to the yacht. But at that moment, he couldn't deny himself another kiss and lowered his mouth to his wife's again.\n\n* * * * *\n\nDon't miss any of the sexy alpha heroes of Bane Westmoreland's SEAL team!\n\nBANE  \nTHE RANCHER RETURNS  \nHIS SECRET SON\n\nOnly from New York Times bestselling author Brenda Jackson and Harlequin Desire!\n\nKeep reading for an excerpt from REUNITED...WITH BABY by Sara Orwig.\n\nJoin Harlequin My Rewards today and earn a FREE ebook!  \nClick here to Join Harlequin My Rewards  \nhttp:\/\/www.harlequin.com\/myrewards.html?mt=loyalty&cmpid=EBOOBPBPA201602010003\n\nWe hope you enjoyed this Harlequin Desire story.\n\nYou want to leave behind the everyday! **Harlequin Desire** stories feature sexy, romantic heroes who have it all: wealth, status, incredible good looks...everything but the right woman. Add some secrets, maybe a scandal, and start turning pages!\n\nEnjoy six new stories from Harlequin Desire every month!\n\nConnect with us on Harlequin.com for info on our new releases, access to exclusive offers, free online reads and much more!\n\nOther ways to keep in touch:\n\nHarlequin.com\/newsletters\n\nFacebook.com\/HarlequinBooks\n\nTwitter.com\/HarlequinBooks\n\nHarlequinBlog.com\n\n**Join Harlequin My Rewards and reward the book lover in you!**\n\nEarn points for every Harlequin print and ebook you buy, wherever and whenever you shop.\n\nTurn your points into **FREE BOOKS** of your choice\n\nOR\n\n**EXCLUSIVE GIFTS** from your favorite authors or series.\n\nClick here to join for FREE\n\nOr visit us online to register at\n\nwww.HarlequinMyRewards.com\n\nHarlequin My Rewards is a free program (no fees) without any commitments or obligations.\nReunited...with Baby\n\nby Sara Orwig\n\nOne\n\nOn a Wednesday, the first of August, Luke Weston gazed out the window of his Gulfstream jet as his pilot flew east to Royal, Texas. Looking down at the wide-open, mesquite-tree-covered landscape, it struck him that he was flying directly over the Double U. The ranch that he had grown up on and was now returning home to save.\n\nAs he stared out the window, he wondered if he should let the Texas ranch go once and for all. It was no longer home to him. Ever since graduating from Stanford six years ago, he'd been living in Silicon Valley, where he'd been working tirelessly to build his company into the tech-world juggernaut that it was today. He had no desire to live anywhere else.\n\nEven so, he couldn't bear the thought of the Texas family ranch being auctioned off because his drunken father had mortgaged the Double U again and then couldn't keep up with the payments. Thank heavens, he wasn't like his dad. But when he was young, his dad hadn't been a drunk\u2014a scary fact that was impossible to forget. And his grandfather on his dad's side had been a drunk later in life, too. A drunk who was killed in a bar fight. Bad blood.\n\nLuke's thoughts shifted to his hometown friends, especially Will Sanders. Unable to attend Will's funeral because he had been in Europe, Luke had heard about Will crashing his own funeral in Royal and all hell breaking loose. Who had been impersonating Will? That had been the big question, but answers were pointing in one direction.\n\nThis morning, before Luke left California, Will had phoned to inform him that his lifelong friend Richard Lowell had been the one passing himself off as Will. But then Rich had disappeared and no one had a clue where he was now. Another friend had also gone missing\u2014Jason Phillips. At present, only the police and those close to Will knew about Rich, and they were still keeping it quiet while they tried to piece together all that had happened. Will had confided in Luke that Rich had tried to kill him on their annual fishing trip, pushing Will overboard and leaving him to drown. Will was rescued and survived to come home to his own funeral. The body had been cremated, so all they had left were ashes and bones to try to identify.\n\nAdding to the mystery, Will told Luke that funds had been pilfered from the Texas Cattleman's Club in Royal and they believed Lowell was the responsible party. Upon hearing that, Luke told Will he thought he could help him find the missing imposter via the antifraud software that he and his team at his company, West-Tech, recently created that could aid in following money trails. Will had been interested in learning more, and they agreed to discuss it in person once Luke arrived in Royal.\n\nOnce again, Luke reviewed his three purposes in going back to Texas. First, he felt compelled to save the ranch that had been in his family for more than a century. He wondered in what condition he would find the family ranch. How were the animals faring, and how many cowboys still worked on the ranch? Did his dad owe them back pay?\n\nSecond, he wanted to do everything he could to help Will find Rich Lowell. The first order of business was to contact the PI who had been hired to aid in the investigation.\n\nAnd the third reason for returning to Texas was that he felt duty-bound to go visit his dad in the assisted-living home where he had been residing ever since he had been diagnosed a year earlier with cirrhosis of the liver, a debilitating disease. As always, when he ticked off his plans, he had a deep awareness of another Royal resident, Scarlett McKittrick.\n\nScarlett was one resident he should avoid at all costs, but he suspected that wasn't going to happen. She was the best vet in Royal, Texas, so he would most likely need her professional expertise.\n\nAs he pictured her in his mind, the memories assailed him. Memories of holding a naked Scarlett in his arms, kissing her, her intense, instant response to his every caress. Luke drew a deep breath. He was _not_ the man for her, and when they parted, she had been furious with him for walking out on her instead of marrying her. Even so, what he wouldn't give for one blissful night with Scarlett before he returned to his carefree, no-strings-attached existence in Silicon Valley.\n\nBut he couldn't afford to be distracted with thoughts of his ex. He had to pour all his time and attention into trying to salvage the Double U. Luke still felt a low-burning anger at his dad because he had long ago paid off the ranch and owned it outright, but Luke had made the mistake of letting his father keep everything in his name. Luke hadn't realized what was going on and that his father was pawning things off, selling livestock, mortgaging the ranch to the hilt. Luke felt partially responsible for turning a blind eye and not coming home the moment his dad had been moved to the assisted-care facility. But he'd put on a good show, and Luke had believed him when he'd said that everything was fine.\n\nHowever, when Luke hadn't been able to contact anybody at the ranch, his internal alarm bells had gone off. Then he got a call from Nathan Battle, sheriff in Royal, who broke the news that he was going to have to put the ranch up for auction.\n\nLuke told Nathan he would be there in two days to pay off the ranch and take care of all the outstanding bills. As soon as he ended the call with the sheriff, Luke had made arrangements with his pilot to fly to Texas the next day.\n\nHe just hoped and prayed he would be able to clean up the mess his father had left in his wake.\n\nAfter landing in Royal, Luke called Cole Sullivan, the PI, as he'd told Will he would do, and made an appointment to talk to Cole tomorrow afternoon in Brinkly, Texas, a small town near Royal. He then phoned Will to tell him he had arrived, and the two men made arrangements to meet tomorrow before Luke met with the PI.\n\nAs soon as Luke left the airport in the new black pickup he had purchased by phone, he headed to his family ranch. As he reached the Double U and drove up the ranch road toward the house, he gazed out at the front pasture. It was worse than he had imagined. The first two horses he came across were so severely malnourished their ribs were showing, and they stood listlessly with their heads hanging down. Luke feared they would not live one more night.\n\nHe drove on to the house, passing fences that were down and a stock tank shot full of holes. Nearing the homestead, he saw a large part of the exterior wall had been ripped away, and it felt as if a knife had plunged into his gut.\n\nSwearing harshly, he realized he'd had no concept of the full extent of the disaster at the ranch. After taking several deep breaths to help himself calm down, he placed a call to Scarlett McKittrick's veterinary clinic and felt his frustration rise yet again when he learned she was out of the office.\n\nLuke turned his pickup around and sped toward the McKittrick place. When he crossed the cattle guard, he slowed down. As he drove up the McKittrick's ranch road, and drank in the familiar surroundings, it finally felt as if he was coming home.\n\nBombarded by memories, Luke gazed at the gravel road, but all that he could see were Scarlett McKittrick's thickly lashed hazel eyes. It had been a decade since he had last seen his high school sweetheart, and his life had changed beyond his wildest imaginings. Yet, no matter how much time had passed, there was no way to ever forget her.\n\nHe swore under his breath, every part of him aching with bittersweet longing for Scarlett. He had felt certain he had gotten over her, but if he had, why were memories rushing at him like floodwaters from an open dam?\n\nA sudden wave of nostalgia crashed through him as he thought about all that had transpired between them. Back in high school, Scarlett had said she was in love with him. And for a while during his junior year, he had allowed himself to get swept up in dating her and had returned her love fully. But then, in his senior year, reality had set in.\n\nHe had a rotten background, while Scarlett had a good, solid one. Her father died when she was young and her brother, Toby, stepped up and filled in as much as he could, while Scarlett's mother quietly took over and ran the ranch with Toby's help. Luke didn't want to mess up Scarlett's life. He was afraid of the bad blood in his family showing up in him.\n\nAs his high school graduation approached, Scarlett knew he was leaving for college, but she didn't want him to go. And even though she made it clear she wanted them to have a future together, he never expected her to tell him she wanted them to marry right after he graduated. As far as he was concerned, she was too young and inexperienced to know what she wanted for the rest of her life. She could set him on fire with a kiss, but he still saw her as a kid at sixteen. She seemed far younger than his eighteen years.\n\nHe didn't want to marry for years, if ever. His parents' marriage had been unhappy all his life. At first, they fought. Later, they drank and fought. He didn't think either one had been faithful to the other. He didn't want to pass his genes on or marry someone like Scarlett and ruin her life. Only sixteen and hopelessly in love, she didn't understand. Consequently, they didn't part on good terms\u2014something which he deeply regretted to this day.\n\nThey'd both moved on. After graduation from Stanford, Luke built his West-Tech company and he struck it rich when he invented a revolutionary\u2014and affordable for the masses\u2014smartphone that left his competitors in the dust.\n\nMeanwhile, Scarlett had pursued a career in veterinary medicine. Which didn't surprise him in the least. A real softie for animals, she always tried to help any creature that needed it, loving little kids and animals as much as he loved electronics and the challenges in the tech world. As she'd been building up her vet business, she had apparently gotten engaged but was now estranged from her fianc\u00e9, Tanner Dupree, some oil heir who'd left her stranded at the altar. In Luke's eyes, no one would ever be good enough for Scarlett\u2014definitely not himself. The oily scumbag who had deserted her didn't deserve her, either. Walking out on her on her wedding day\u2014the guy had to be selfish and rotten to the core.\n\nSighing, Luke knew he was hardly one to throw stones since he'd left her, too. He still believed that was the right thing to do because, given his tainted family history, he would never be good enough for her. But honorable intentions aside, if he was being completely honest with himself, Scarlett was still by far the sexiest woman he had ever known. Even after all this time, he could get hot just thinking about her.\n\nHe had been the first guy she'd ever been with, when they were still in their teens. He remembered holding her close, her slender body melting against his. While he was here, could he entice her back into his arms for a night down memory lane...? Groaning, he quickly squelched those illicit thoughts. When he had gone to California, he'd spent too many sleepless nights in college lying awake wanting her, fantasizing about her, fighting the urge to call her because he didn't feel worthy of her because of his family background. He didn't want to mess up Scarlett with his bad genes. He finally had put that behind him, and he didn't want to stir all those feelings up again.\n\nIt wouldn't be fair...to either of them.\n\nWith a shift of his shoulders, he forced his thoughts back to the present, determined to focus on the here and now.\n\nAs Luke approached the McKittrick house, dogs of all sizes ran toward the car. He knew they had to be strays taken in by Scarlett, and he couldn't keep from smiling while he felt a twist in his heart. He stepped out, speaking softly to the barking dogs that quieted down, the friendliest ones already wagging their tails and letting him scratch their heads.\n\nScarlett walked out onto the porch and stopped at the top step. His heart thudded. For an instant he couldn't speak or breathe and felt as if he was in a dream, except he knew she was real and only a few feet away. He had to curb the impulse to close the distance between them, sweep her into his arms and kiss her endlessly. She was absolutely breathtaking. When he looked into her wide, hazel eyes, he had the impact of a punch to his gut, and it was obvious she, too, drew a deep breath. As she inhaled, her blue blouse grew taut over her figure that had filled out into lush, gorgeous curves. He remembered a kid, a naive, fun young girl, but this was a woman who made his blood hot and fanned desire into flames.\n\nWhile his gaze locked with hers, he lost his breath again. The urge to crush her against him was overwhelming, and he knotted his fists and focused on staying where he was. His heart pounded as his gaze swept from her head to her toes. She was wearing a blue cotton short-sleeved shirt, tight jeans and boots, and her pixie hairdo complemented her high cheekbones and big hazel eyes. Eyes that were now filled with fury.\n\nHe was dazed, stunned by the reaction she stirred in him. He had thought he was over her long ago. If he was, what was going on right now to his heart, his breathing and his lower extremities? And it was obvious from her irate expression that she also was having some kind of reaction.\n\nExcept not the kind he particularly wanted.\n\n\"You get off the McKittrick property, Luke Weston,\" she snapped. \"And you can just go straight to hell.\"\n\n\"Scarlett, I need your help,\" he said, talking fast before she cut him off. \"My dad is in an assisted-living facility and he's let the ranch go. The animals are dying and need attention\u2014\"\n\nHe knew when he mentioned the dying animals he had her. The anger left her expression, replaced by worry. She never could hide her emotions, and she was a sucker for any animal in trouble. The yard filled with dogs was proof of that.\n\nShe clamped her lips together and stared at him.\n\n\"I saw a few horses, and they look so severely malnourished that they can't even hold up their heads.\"\n\nShe closed her eyes for an instant as if in pain. When she opened them, he knew he had gotten through to her. \"I took an oath to help animals. I'll get my instruments.\"\n\n\"You can ride with me, and I'll bring you back when you're through. Just save some of the horses or let's put them out of their misery. I don't know which ones to put down,\" he said, only half meaning it because he was certain that would convince her to help.\n\n\"You don't put any down. I'll take care of them.\"\n\n\"There's no feed in the vicinity. I need to get some. I'll take you with me to get supplies.\"\n\n\"This place is the same as it was when you lived here, and you know where to find feed and hay. Go load your pickup with whatever you need for your livestock tonight. While you do that, I'll get my things and then I'll join you,\" she said.\n\n\"Thanks, Scarlett. I appreciate it because I need a good vet. Those horses are in dire shape. You'll see.\"\n\nNodding, she turned away. He drove to the barn and hurried inside. One glance at the loft and memories bombarded him. The most persistent memory was of making love with Scarlett, but he had to stop torturing himself because they had no future. He wasn't the man for Scarlett. He had done well in business, but that wasn't all there was to life. Always, he came back to thinking about his parents. His dad did well enough in business for years even after the alcohol began to cloud his judgment.\n\nDamn, Scarlett looked good. As angry as she had looked when she first saw him, he didn't think he would have any choice except to remain cool and impersonal if he wanted her help. But that was easier said than done. He better do that for his own good. He went through hell leaving her before. Now they were adults and the stakes were higher. He didn't want to get involved and have to go through another goodbye and that's all he could hope for with her.\n\nWhen they reached his family ranch, was he really going to be able to keep his hands off her?\n\n* * *\n\nScarlett went inside to speak to her mother, who was bathing little Carl. For an instant worries fled as she smiled at her precious, adopted baby boy. Her heart squeezed when Carl smiled at her and held out his little arms. \"I can't take him, Mom. Luke Weston is here. He's in town and said his dad let the ranch go and the animals are sick, maybe dying now. He came to ask for help. Mom, I have to help those animals.\"\n\nHer mother frowned and shook her head. \"I know you're not going to ignore the livestock, so do what you have to and then come home. You don't need to get involved with Luke Weston again. He broke your heart, Scarlett. Don't let him come back and hurt you again.\"\n\n\"I won't. He lives in a different world now and he'll go right back to it,\" she said, thinking about the big, strapping man standing in her yard, instead of the young boy she remembered from their high school days. He was wickedly handsome, and her heart had pounded to such an extent that there was no way to ignore what she felt.\n\nAs Scarlett talked to her mother, she cut up apples from a bowl her mom kept on the counter. She bagged the apples and smiled at Carl.\n\nKissing her little boy's chubby cheek, she dodged when he grabbed for her hair. She and her mother both laughed, but she saw the worry in her mother's eyes as she left.\n\nScarlett fought the urge to tell Luke she couldn't go, but when she thought about the horses that might be hurt or hungry, she knew she had to help. She didn't want Luke putting any animal down unless it was hopelessly suffering and she couldn't save it.\n\nShe felt a tangle of emotions\u2014shock because her pulse had raced at the sight of Luke when she had convinced herself that she was completely over him. And then there was the anger. It was always churning beneath the surface when she thought of him. Of how he'd left her behind. She didn't want to react to him or remind herself that he was more handsome than ever. He was a man now, not a boy, and so incredibly hot.\n\nScarlett closed her eyes and shook her head. \"No, no, no,\" she whispered. She didn't want to find him better looking than ever, more appealing. Breathtakingly sexy. She didn't want her heartbeat to race. All of that manifested itself, leaving her gasping as if she were sixteen again, lusting like a starry-eyed schoolgirl over the most irresistible boy in Texas.\n\nLittle reminders of Luke still popped up in her life, but for all these years since he'd left Royal, she had ignored them. So she had _thought_ she was totally over him, but how wrong that was. All he had to do was step out of his pickup and stride up to the porch, and she was ready to either melt into a quivering mess, or run and throw herself into his arms.\n\nShe didn't want him to come back to town and cause that kind of reaction simply by laying eyes on him. When he looked at her, he felt something, too. She knew him well enough to know he'd had a reaction to seeing her, which just compounded her desire for him.\n\nCould she work a few hours with him on his ranch and keep a wall around her feelings? She never wanted to suffer through another heartbreak over Luke like she had when he'd left Texas all those years ago. She had cried herself to sleep every night for more than a month.\n\nScarlett hurried to the closet to grab her new jeans and a shirt she liked, and then she looked down at the clothes in her hands. Whoa...what had gotten into her? Dressing better because of Luke was just asking for trouble.\n\n\"Nope, not happening,\" she said aloud and shoved the jeans and blouse back in the closet. She couldn't resist looking at herself in the mirror, though, and running a comb through her short hair. She guessed there weren't any women he took out in California who had freckles and pixie haircuts. She sighed because it wouldn't matter how she looked, she wasn't the woman for Luke. She wasn't risking her heart a second time because he would never make her a permanent part of his life. She had little Carl to think of now and how what she did would affect him.\n\nHurrying to her office, she tried to focus on what she might need at his ranch as she grabbed her bag.\n\nBefore she left she paused, pressing her forehead against the wall. \"Don't let him break your heart again. Take care of the animals and then come home. Treat him as if you're with a stranger,\" she whispered and then shook her head as she hurried out of her room. Who was she kidding? She knew she couldn't heed her own advice, but if she could just remember he was totally off-limits and keep her guard up, she might avoid more heartache. He hadn't loved her before\u2014now he definitely never would since he was completely out of her league. She knew he was the newest addition to _Forbes_ 's billionaire list. Luke could have any woman he wanted, she was sure. In fact, she had seen his occasional picture in magazines or the news and knew he dated gorgeous celebrities and some very beautiful socialites.\n\nHolding her bag of instruments, medications and ointments, she rushed out. Luke leaned against his pickup and was looking down at his phone. She couldn't keep from stealing a glance down the length of him, admiring his broad shoulders, his narrow waist and his long legs. When he saw her, he jammed his phone into a back pocket. He straightened and his gaze drifted slowly over her as she approached him, and all her advice to herself to pay little attention to him evaporated.\n\nEverywhere his gaze drifted over her, she tingled. Part of her wanted to turn around and go right back to the house and lock the door. Part of her wanted to yell at him to get off their property and go straight to hell because he had hurt her badly. Yet another, more urgent, part of her just wanted to rush into Luke's arms, pull his head down and kiss him senselessly. She sucked in a breath, and her hand tightened on the handle of her bag while she struggled to think about something besides Luke, his hands and mouth and her pounding heart.\n\nHe opened the door to the truck for her, and when she came close, he reached to take the black bag from her. When his big, warm hand closed on hers, she thought her knees would buckle. It was the first time in years\u2014since he'd left for college\u2014that he touched her, and the slight contact sent a sizzling current racing to settle low inside her. Longing rocked her, and she had to take a deep breath and clench her fists.\n\n\"I'll put your things in back,\" he said, his voice raspy, which happened when he was aroused. She couldn't answer him and merely nodded. They knew each other so well. He knew she was having a reaction to being with him, and she knew he was having his own reaction to her. That made the moment hotter and more intense, and kept dredging up memories of their lovemaking when he had lived in Texas.\n\nAgain, he took her arm to help her into the pickup\u2014help she didn't want or need. Help that made her quiver and have to fight more memories of his hands on her. When he closed the door and walked around to the driver's side, she inhaled deeply and watched him. A breeze tousled his dark blond hair over his forehead. He looked sexy, more handsome than ever\u2014something she didn't want to acknowledge. Luke was tall, with scruffy stubble on his jaw, and he had gorgeous blue-green eyes, broad, powerful shoulders and well-shaped hands. Hands that could carry her to paradise. In short, Luke was a fantasy come to life.\n\nHe slid behind the wheel, closed the door and started the pickup, glancing at her. He sat too close, looked too enticing. \"Thanks, Scarlett,\" he said in a husky voice that still wasn't his normal speaking voice and she knew it.\n\nShe nodded. \"Let's get this over with,\" she said curtly, staring out the front window, fighting to ignore him as much as possible. What was happening to her? She was over him, over the hurt he caused when he left for California and said goodbye without a second thought. The old familiar anger and pain made her sit up straight and look out the window as he drove away from the house.\n\n\"Stop at the barn, and we can get some bales of hay,\" she said.\n\n\"I did. They're in the back of the truck. I loaded up hay and feed, and then drove back to your house to wait for you.\" He spared her a quick glance. \"I'll reimburse you for everything. I appreciate not having to go back to town to get supplies. I'm guessing there's nothing at the ranch\u2014just on the drive in, the place looked abandoned. There were signs of vandals, and the animals have been left to die,\" he said gruffly. \"I was just there a few minutes, but it's clear I have a catastrophe on my hands. I want to save what animals we can.\"\n\nScarlett knew Luke so well that she could tell he was angry with his father. When they drove past the barn on her family ranch, she stared ahead, sitting stiffly, fighting yet another wave of memories.\n\n\"You still have the big barn,\" he rasped.\n\n\"We're not going down memory lane,\" she snapped without looking at him. But she was already down it. Her fingers knotted and she fought the urge to glance again at the barn she saw every day of her life, yet it held special memories of an unforgettable night.\n\nHer whole family had been away for a barn dance. Early in the evening, Luke had coaxed her to leave with him. They had gone back to her place because everyone had gone to the party. Instead of driving to the house, Luke had stopped at the barn. The minute they stepped inside, he pulled her into his arms to kiss her. Later, he spread a blanket on the hayloft and drew her to him again to make love to her, her first time.\n\nLooking away from the barn, she tried to think of something else and forget that night so long ago, forget memories of his slow, sweet kisses that made her want him with her whole being, memories of his strong arms around her, his mouth on her, his seductive hands all over her body.\n\n\"How are your mom, and Toby and his wife?\"\n\n\"They're fine,\" she answered, glancing at him. \"Toby and Naomi have a little girl, Ava.\" While Luke watched the road, her gaze swept over him, once again taking in the short stubble that covered his jaw, his tousled, dark blond hair that she could remember running her fingers through too many times to count. His shoulders were broader now, thicker. Desire rocked her and she took a deep breath. Realizing once again where her thoughts were going, she turned swiftly to stare out the window, not really seeing the landmarks they passed, but remembering being held in his arms, her head back against his shoulder.\n\nShe gave a tiny shake and struggled to get her attention off of her ex. He would leave as suddenly as he had come, and she didn't want one tiny bit more hurt in her life because of Luke Weston, especially now that she had a son to care for.\n\nShe looked at familiar land, places she had grown up, and in seconds Luke dominated her thoughts yet again.\n\nShe had to resist his appeal. In no time he would be back in his private plane, headed to California, back to his ritzy life, back to glamorous models, celebs and rich socialites, eventually marrying one who could give him the children he'd want.\n\n\"You've done well in California. You did the right thing to move out there. It suits you as much as all this suits me,\" she said, knowing the Silicon Valley world was his world.\n\n\"I guess you're right, Scarlett. It's my real home,\" he said without looking at her. He sounded casual, but his hand was tight on the steering wheel, so obviously he felt something, too. \"Common sense says to sell the ranch and forget it. I won't live in Texas again. But...I can't sell it. I just can't let the family place go. It's been in our family since the 1800s.\" He blew out a frustrated breath. \"I paid the house off three years ago, and damned if he didn't go out and mortgage it to the hilt again. He hasn't kept up his payments\u2014no surprise there. He's let the help go. I just found that out before I came.\"\n\n\"Sorry, Luke,\" she said, again without looking at him. How polite and cool they were being with each other. \"So you're going to keep the ranch, even though you'll go back to California? You think you'll come back to the ranch someday?\" she asked, watching him and curious about his answer even though she knew she shouldn't care at all. They would never again mean anything to each other. Unfortunately, the jump in her pulse today showed she still had to work at getting him out of her system.\n\n\"No, I never will, but at this point in my life, I just don't want to let it go. I know that doesn't make sense, because California is absolutely my forever home.\"\n\n\"You don't need to be in a hurry. Your dad is still around. It may mean something to him.\"\n\n\"Booze is the only thing that means anything to him,\" Luke said, and she heard the anger and bitterness in his reply. \"He'll never be able to live alone again.\"\n\nAfter they left the McKittrick ranch, they rode quietly. Her thoughts were in turmoil because she couldn't lose that intense awareness she had of Luke. She never had been able to ignore him, and she definitely couldn't now. Why couldn't she ever see him as just another guy? She had to get over him or get hurt again. She could never be the woman for him because of her fertility problems. One man who loved her enough to ask her to marry him had already walked out on her. Luke hadn't been interested when he had never been out of Texas and was getting ready to leave the family ranch. Now, he wouldn't have any permanent interest in a small-town female vet who couldn't bear his children. If she got involved with him, he would love her and leave her and in doing that he would get to know her baby. If she let Luke in her life again, when he said goodbye, he would not only break her heart again, he would break little Carl's heart. That could be a lifetime hurt for her and her baby.\n\nScarlett tried to avoid remembering Luke's kisses, but whenever she glanced at his handsome profile or his sexy mouth, the memory was vivid, tantalizing, still painful after all this time. She looked at his big, masculine hands on the steering wheel, but shifted her attention swiftly because she could remember those hands on her body, working their magic. An undercurrent of longing taunted her.\n\nShe released a quavering breath. Why did her heart race when he had merely brushed her fingers with his? She remembered how much she'd hurt when he left when she was sixteen. She didn't want a bigger hurt now.\n\nShe couldn't understand her own reactions to him. She wasn't in love with him\u2014she barely liked him because of the bitter fight before he left for California. How could he set her pulse pounding just by reappearing? She had to get over him. She didn't want to spend years longing for a guy she knew as a boy in high school. A guy who didn't want her.\n\nThey rode in silence until he turned and headed up the road toward the house where he had spent his boyhood.\n\nThe first sign of neglect was a rusty pickup smashed against a tree. She saw bullet holes where kids had probably placed bottles on it or just shot out the windows and used the truck for a target. The wheels were gone. Weeds grew up in the road that was barely visible in spots.\n\n\"Evidently, after Dad let the hands go, he sold some of the horses to subsidize his alcohol addiction.\" Luke scowled. \"I used to send money home, but he just bought liquor with it, so I stopped. I'll get a crew out here as fast as I can, but right now I wanted you to see if we can't save some of these horses. But honestly? I don't know how the horses I saw can last through the night. No one works here. The damn ranch is deserted\u2014the animals left to starve and die.\"\n\nShe could hear the anger and pain in his voice and couldn't blame him for his reaction. She was equally shocked by the terrible conditions.\n\nIn minutes, Luke approached a pasture with half a dozen horses standing near a stock tank that needed water. The windmill had broken boards and wasn't working. She gasped. \"Oh, no,\" she whispered without knowing she had spoken when she saw the horses with ribs painfully revealed and two with their heads hanging. All the horses looked severely malnourished. The stock tank had holes in the side.\n\n\"Sorry to pull you into this because I know it'll tear you up, but I need your help here,\" Luke said.\n\n\"Oh, my heavens. Look at the horses,\" she lamented. \"It breaks my heart. You know I'll help these animals,\" she said, horrified to look at the condition of the horses. She felt sympathy for Luke, even though she didn't want to get caught up in his problems. But what he had come home to was ghastly, and he had tried to help his dad to keep the ranch in good shape.\n\nShe could certainly understand his anger and disappointment, and gave a silent prayer of thanks for her own family. They helped each other and did the best they could and always could be counted on. \"Oh, Luke, this is unbelievable. I had no idea this ranch had just been abandoned. We're neighbors. Our ranch adjoins yours, and nobody in the area has said a word about it. Why didn't someone speak up? The last hands that left here\u2014why didn't one of them contact you? How could your dad neglect everything so badly?\"\n\n\"Because he's a sick old drunk who doesn't care about anybody or anything except his next drink,\" Luke bit out, and she was sorry for saying anything because Luke was obviously suffering over finding his home in shambles.\n\n\"I'll get the pasture gate,\" Luke said, getting out to drag open a sagging, battered gate made with barbed wire. He returned to drive the pickup in and close the gate.\n\n\"Sorry, Luke,\" she said stiffly when he was behind the wheel again. She spoke without looking at him, trying to avoid thinking about what he was going through. \"We'll start. Let's get to work.\"\n\n\"I'll patch those holes enough to get water in that tank so they can drink. I sent Dad money to get fiberglass tanks and look what we've got\u2014the old corrugated metal the cows have pushed against and bent years ago. Damn, I wasn't sure what I'd find here, but I didn't expect it to be this bad. Every dime I sent home must have gone for booze.\"\n\nShe looked around and saw three horse carcasses. The live horses had moved away from them and they were decomposing, probably torn by predators and birds.\n\n\"You have dead animals.\"\n\nHe sighed. \"Damn. I can get a temporary crew out here to help.\" He parked near the horses and a few watched them while two slowly moved toward them. Luke was already on his phone, calling someone who worked for him to start trying to hire a crew of cowboys to do temp work.\n\nWhen she approached the horses, her sympathy shifted to the animals, and she could hardly blame Luke for being so upset at his father for letting this happen. When Luke was a kid, the Double U had been a fine ranch. His dad was a good rancher, and he knew what he was doing to his livestock when he neglected them. At least he had to have known when he was sober. She spoke softly and got her bag of apples, but the horses couldn't raise their heads. She knelt to open her bag and get a needle to give shots that would help more than anything else.\n\n\"I'll get these horses to the pasture by the barn. There's water there. I'll get halters on them and lead them back, and you drive the pickup. You can follow the road here to the house,\" he said. \"If there are any horses we can't move, we'll try to take care of them here.\"\n\nIt was almost an hour later when they climbed back into his pickup and drove toward the house.\n\n\"I came home every year for the first three years while I was in college, and it was never like this. Things were messy at the house, but otherwise, he kept things in relatively good shape.\" He scrubbed a hand over his jaw. \"We had some good hands and a good foreman. I never stayed more than a night or two, so he must have pulled himself together.\n\n\"Several years ago at Christmas, I sent a plane for him and brought him to California. He said everything here was fine. He couldn't wait to get back here and cut short his stay. Gradually, we've grown more apart than ever, and I haven't been home. If I did make contact with him, he always said everything was going okay.\" Luke worked his jaw back and forth. \"I should have kept up with him better and maybe I could have prevented some of this. I could have hired someone to come out here and run the ranch.\"\n\n\"You didn't know.\"\n\n\"I _should've_ known. He always could carry on a decent conversation when he was dead drunk. I should have guessed what was going on.\"\n\n\"Luke, I'm sorry. This is a disaster.\"\n\n\"We'll just have to hunt for the animals. I doubt if there are any cattle left. I'm sure they've all been stolen. The horses probably were passed over at first for cattle. By the time anyone turned attention to the horses, they may have been in such bad shape no one wanted them. I just barely glanced at the house, but I'll walk through in case there are any animals in it.\"\n\nScarlett cringed when the house came into view. One wall was shattered, as if someone had tried to drive through it. Windows were smashed. Steps to the porch had collapsed. The front door was missing. Bullet holes dotted the walls, and boards had been ripped from the porch floor. Someone had thrown black paint at the house, and a big splash of paint had spilled down a wall. A living room chair was upside down in the yard, one leg broken, another leg missing.\n\n\"Oh, my heavens, Luke...\" she commiserated softly. Certainly it gutted him to look at his childhood home so badly damaged.\n\n\"While neglect did a lot to the house and barn and outbuildings, vandals and thieves caused the rest,\" he said grimly. \"My dad, because of his damned drinking, has just let our home\u2014a damn fine ranch\u2014go to hell.\"\n\nKnowing how she would feel if it had been her home, she ached for him. \"I'm sorry, Luke. How awful for you to come home to this.\" Impulsively, she squeezed his wrist and Luke turned, his blue-green __ eyes intent on her, causing a chemical reaction. The minute she touched him, the moment changed. Sympathy vanished, replaced by sizzling desire. But she didn't want to be swept off her feet by him again. She'd been through too much heartache because of Luke to go through more.\n\nWhen his gaze locked with hers, she drew a deep breath, conscious of Luke and nothing else. Worse, she was absolutely certain he felt something, too.\n\n\"I'm sorry for you, and I'm sorry for your dad. Have you seen him yet?\" she asked, her words tumbling out too fast as she tried to get back to anything less intimate. But that slight touch of his wrist brought a truckload of memories pouring over her, and she felt her anger with Luke lose a bit of its intensity.\n\nShe felt sympathy for him. It would be devastating if she came home to find the McKittrick ranch in ruin. She tried to pay attention to what he was saying about seeing his dad.\n\n\"No, that's on my list of things to do while I'm here. What he's done\u2014or more accurately, _not_ done\u2014is going to make seeing him again even more difficult than I expected. He must be in terrible shape to let all this happen.\"\n\n\"Well, let's look for the horses or whatever livestock that's still here,\" she said, struggling to get back to business.\n\nHe nodded. \"I stopped here briefly before coming to get you. I want the house torn down. I can't stand to see it in ruin. The memories from there weren't all that great anyway,\" he said, and her heart lurched at the bitterness in his voice. She curbed the impulse to reach out and squeeze his wrist again. It was obvious he hurt badly.\n\n\"There's a half bath in the barn, so at least we have a little in the way of facilities for us. There may be running water and electricity in the bathrooms in the house. Right now, however, we better find what animals we can while it's daylight. I'll try to get them back to the pasture by the barn, where you can do what you have to do and I can feed and water them. Can you stay longer?\" he asked.\n\n\"Yes, I'll stay. I want to save as many horses as I can,\" she said.\n\n\"We'll take the pickup now. Later, I'll probably have to search on horseback because there are places on the ranch where I can't drive. I may have to go buy a horse because none of these can carry me on its back.\" He released a breath. \"But for now, I've got rope in the back of the truck, some feed and a saddle if I need it, all sorts of supplies.\"\n\n\"All right,\" he said, \"let's get started.\" He turned his truck and as he drove she looked for any livestock. They hadn't driven a half mile when she gasped. \"Luke, stop. There's an animal. It's a dog, and it's dead. I think it looks like it might be Mutt.\" With a pang, she remembered the dog that followed Luke around when he was home.\n\nThey got out of the pickup and walked closer. Luke knelt and ran his hand over the dog's head. \"Oh, dammit to hell. That's Mutt. He was old and weak, and I guess coyotes got him.\"\n\nShe knelt to look over the carcass more closely, and she hurt even more for Luke because this was the ranch dog that he claimed as his.\n\n\"I left him here when I went to California because the ranch was up and running and in good shape,\" he said, his voice raspy with regret.\n\n\"The ranch was in good shape because you were here,\" she said quietly, still looking at the dog.\n\n\"The guys liked him and he was happy here. I thought he'd be better off. He looks starved. He was old and weak, but something's really torn him up.\"\n\n\"Luke, he's been shot. Someone shot him, and they may have done it because he was old and he may have been sick. There's one shot and it's a killing shot, so this wasn't random or someone being mean. I think he was torn up by buzzards and coyotes after he was shot.\"\n\nLuke leaned closer to look as she pointed to the wound. \"I hope he didn't suffer. I loved the old mutt. He was a good dog.\" He released a shaky breath. \"I'm going to bury him. I have a shovel, and I'll wrap him in a tarp and bury him back at the house.\"\n\nShe heard the catch in Luke's voice, and a lump rose to her own throat. They both stood, and she looked up at him. Without thinking about it, she touched his wrist again. \"I'm sorry. I know you loved him.\" The minute her hand rested on his, she knew she shouldn't have touched him, even though it was obvious he was hurting badly. His wrist was twice the size of hers, warm, his wrist bone hard. Something flickered in the depths of his eyes, and he gazed at her intently.\n\n\"I haven't loved much in my life, but I loved him,\" he said roughly, his voice grating and a muscle working in his jaw. She couldn't get her breath, and she couldn't understand the intensity of his gaze or his remark that he hadn't loved much in his life. Was he just talking about a dog\u2014or was there more to his statement?\n\nShe wanted his arms around her so badly it frightened her, and she stepped away quickly, going back to the truck. \"Tell me if I can help,\" she called over her shoulder.\n\nShe was breathing hard as he walked to the back of his pickup. Pulling work gloves from his pocket, he got a tarp to wrap the dog. In minutes, he slid behind the steering wheel and drove on in silence.\n\nWhy did it feel as if Luke had been away only days instead of years? Those empty years had vanished in too many ways. The worst part was the realization that she had never really gotten over him, something she had struggled desperately to do.\n\nShe had never felt this way for Tanner, and she had been engaged to him. Was it because Luke had been the first love in her life? Or did it go deeper than that?\n\nThey rode together in silence, only a couple of feet of space separating them in the pickup, but there was still a permanent, deep chasm dividing them.\n\nHe hadn't loved her and he never would, so why couldn't she forget him? She clearly meant nothing to him\u2014that alone should stop the volatile reaction she had to him and the desire that still steadily simmered through her veins.\n\nThey bounced over the rough ground, and she looked around carefully, trying to see any sign of livestock. In another twenty minutes she spotted horses to the east. \"Luke, over there.\"\n\n\"Yeah, I see. They're in a fenced pasture, so let's keep looking and see if we can find some more and get them back with these. They may be in bad shape, but hopefully we can get them to the barn.\" He continued to drive, and she gazed around, looking for any more signs of life.\n\n\"Luke, I see horses through the trees,\" she said a few minutes later, and he swung the pickup in the direction she pointed.\n\nFor the next six hours they worked\u2014rounding up horses, finding a few cattle, getting them back to the barn\u2014trying to do it all while it was still daylight. When the horses were finally in the corral by the barn, the cattle in a pasture, Luke closed the corral gate and turned to her.\n\n\"You start checking the horses while I get feed to them. They've got water now in that tank. Shortly, it'll be dark, so I'll get lanterns out now and have them ready, and we can keep working if you can stay. If not, I'll take you home. I'd appreciate your help if you can.\"\n\n\"I can stay.\"\n\nHe looked at her and reached out to hug her. \"Thanks, Scarlett,\" he said.\n\nAs his arms wrapped around her and pulled her against his solid, hard body, her heart thudded. His strong embrace made her tremble and want to wrap her arms around him and hold him tightly against her heart. How was she going to work with him into the night without stirring all those old feelings she had for him?\n\nCopyright \u00a9 2018 by Harlequin Books S.A.\nISBN-13: 9781488091971\n\nAn Honorable Seduction\n\nCopyright \u00a9 2018 by Brenda Streater Jackson\n\nAll rights reserved. By payment of the required fees, you have been granted the nonexclusive, nontransferable right to access and read the text of this ebook on-screen. No part of this text may be reproduced, transmitted, downloaded, decompiled, reverse engineered, or stored in or introduced into any information storage and retrieval system, in any form or by any means, whether electronic or mechanical, now known or hereafter invented, without the express written permission of the publisher, Harlequin Enterprises Limited, 22 Adelaide St. West, 40th Floor, Toronto, Ontario M5H 4E3, Canada.\n\nThis is a work of fiction. Names, characters, places and incidents are either the product of the author's imagination or are used fictitiously, and any resemblance to actual persons, living or dead, business establishments, events or locales is entirely coincidental. This edition published by arrangement with Harlequin Books S.A.\n\n\u00ae and \u2122 are trademarks of the publisher. Trademarks indicated with \u00ae are registered in the United States Patent and Trademark Office, the Canadian Intellectual Property Office and in other countries.\n\nwww.Harlequin.com\n","meta":{"redpajama_set_name":"RedPajamaBook"}}