Search | arXiv e-print repository

<!DOCTYPE html> <html lang="en"> <head> <meta charset="utf-8"/> <meta name="viewport" content="width=device-width, initial-scale=1"/>  <link rel="apple-touch-icon" sizes="180x180" href="https://static.arxiv.org/static/base/1.0.0a5/images/icons/apple-touch-icon.png"> <link rel="icon" type="image/png" sizes="32x32" href="https://static.arxiv.org/static/base/1.0.0a5/images/icons/favicon-32x32.png"> <link rel="icon" type="image/png" sizes="16x16" href="https://static.arxiv.org/static/base/1.0.0a5/images/icons/favicon-16x16.png"> <link rel="manifest" href="https://static.arxiv.org/static/base/1.0.0a5/images/icons/site.webmanifest"> <link rel="mask-icon" href="https://static.arxiv.org/static/base/1.0.0a5/images/icons/safari-pinned-tab.svg" color="#b31b1b"> <link rel="shortcut icon" href="https://static.arxiv.org/static/base/1.0.0a5/images/icons/favicon.ico"> <meta name="msapplication-TileColor" content="#b31b1b"> <meta name="msapplication-config" content="images/icons/browserconfig.xml"> <meta name="theme-color" content="#b31b1b">  <title>Search | arXiv e-print repository</title> <script defer src="https://static.arxiv.org/static/base/1.0.0a5/fontawesome-free-5.11.2-web/js/all.js"></script> <link rel="stylesheet" href="https://static.arxiv.org/static/base/1.0.0a5/css/arxivstyle.css" /> <script type="text/x-mathjax-config"> MathJax.Hub.Config({ messageStyle: "none", extensions: ["tex2jax.js"], jax: ["input/TeX", "output/HTML-CSS"], tex2jax: { inlineMath: [ ['$','$'], ["\$","\$"] ], displayMath: [ ['$$','$$'], ["\\[","\\]"] ], processEscapes: true, ignoreClass: '.*', processClass: 'mathjax.*' }, TeX: { extensions: ["AMSmath.js", "AMSsymbols.js", "noErrors.js"], noErrors: { inlineDelimiters: ["$","$"], multiLine: false, style: { "font-size": "normal", "border": "" } } }, "HTML-CSS": { availableFonts: ["TeX"] } }); </script> <script src='//static.arxiv.org/MathJax-2.7.3/MathJax.js'></script> <script src="https://static.arxiv.org/static/base/1.0.0a5/js/notification.js"></script> <link rel="stylesheet" href="https://static.arxiv.org/static/search/0.5.6/css/bulma-tooltip.min.css" /> <link rel="stylesheet" href="https://static.arxiv.org/static/search/0.5.6/css/search.css" /> <script src="https://code.jquery.com/jquery-3.2.1.slim.min.js" integrity="sha256-k2WSCIexGzOj3Euiig+TlR8gA0EmPjuc79OEeY5L45g=" crossorigin="anonymous"></script> <script src="https://static.arxiv.org/static/search/0.5.6/js/fieldset.js"></script> <style> radio#cf-customfield_11400 { display: none; } </style> </head> <body> <header><a href="#main-container" class="is-sr-only">Skip to main content</a>  <div class="attribution level is-marginless" role="banner"> <div class="level-left"> <a class="level-item" href="https://cornell.edu/"><img src="https://static.arxiv.org/static/base/1.0.0a5/images/cornell-reduced-white-SMALL.svg" alt="Cornell University" width="200" aria-label="logo" /></a> </div> <div class="level-right is-marginless"><p class="sponsors level-item is-marginless"><span id="support-ack-url">We gratefully acknowledge support from<br /> the Simons Foundation, <a href="https://info.arxiv.org/about/ourmembers.html">member institutions</a>, and all contributors. <a href="https://info.arxiv.org/about/donate.html">Donate</a></span></p></div> </div>  <div class="identity level is-marginless"> <div class="level-left"> <div class="level-item"> <a class="arxiv" href="https://arxiv.org/" aria-label="arxiv-logo"> <img src="https://static.arxiv.org/static/base/1.0.0a5/images/arxiv-logo-one-color-white.svg" aria-label="logo" alt="arxiv logo" width="85" style="width:85px;"/> </a> </div> </div> <div class="search-block level-right"> <form class="level-item mini-search" method="GET" action="https://arxiv.org/search"> <div class="field has-addons"> <div class="control"> <input class="input is-small" type="text" name="query" placeholder="Search..." aria-label="Search term or terms" /> <p class="help"><a href="https://info.arxiv.org/help">Help</a> | <a href="https://arxiv.org/search/advanced">Advanced Search</a></p> </div> <div class="control"> <div class="select is-small"> <select name="searchtype" aria-label="Field to search"> <option value="all" selected="selected">All fields</option> <option value="title">Title</option> <option value="author">Author</option> <option value="abstract">Abstract</option> <option value="comments">Comments</option> <option value="journal_ref">Journal reference</option> <option value="acm_class">ACM classification</option> <option value="msc_class">MSC classification</option> <option value="report_num">Report number</option> <option value="paper_id">arXiv identifier</option> <option value="doi">DOI</option> <option value="orcid">ORCID</option> <option value="author_id">arXiv author ID</option> <option value="help">Help pages</option> <option value="full_text">Full text</option> </select> </div> </div> <input type="hidden" name="source" value="header"> <button class="button is-small is-cul-darker">Search</button> </div> </form> </div> </div>  <div class="container"> <div class="user-tools is-size-7 has-text-right has-text-weight-bold" role="navigation" aria-label="User menu"> <a href="https://arxiv.org/login">Login</a> </div> </div> </header> <main class="container" id="main-container"> <div class="level is-marginless"> <div class="level-left"> <h1 class="title is-clearfix"> Showing 1–27 of 27 results for author: <span class="mathjax">Tamkin, A</span> </h1> </div> <div class="level-right is-hidden-mobile">  <span class="help" style="display: inline-block;"><a href="https://github.com/arXiv/arxiv-search/releases">Search v0.5.6 released 2020-02-24</a>  </span> </div> </div> <div class="content"> <form method="GET" action="/search/cs" aria-role="search"> Searching in archive <strong>cs</strong>. <a href="/search/?searchtype=author&query=Tamkin%2C+A">Search in all archives.</a> <div class="field has-addons-tablet"> <div class="control is-expanded"> <label for="query" class="hidden-label">Search term or terms</label> <input class="input is-medium" id="query" name="query" placeholder="Search term..." type="text" value="Tamkin, A"> </div> <div class="select control is-medium"> <label class="is-hidden" for="searchtype">Field</label> <select class="is-medium" id="searchtype" name="searchtype"><option value="all">All fields</option><option value="title">Title</option><option selected value="author">Author(s)</option><option value="abstract">Abstract</option><option value="comments">Comments</option><option value="journal_ref">Journal reference</option><option value="acm_class">ACM classification</option><option value="msc_class">MSC classification</option><option value="report_num">Report number</option><option value="paper_id">arXiv identifier</option><option value="doi">DOI</option><option value="orcid">ORCID</option><option value="license">License (URI)</option><option value="author_id">arXiv author ID</option><option value="help">Help pages</option><option value="full_text">Full text</option></select> </div> <div class="control"> <button class="button is-link is-medium">Search</button> </div> </div> <div class="field"> <div class="control is-size-7"> <label class="radio"> <input checked id="abstracts-0" name="abstracts" type="radio" value="show"> Show abstracts </label> <label class="radio"> <input id="abstracts-1" name="abstracts" type="radio" value="hide"> Hide abstracts </label> </div> </div> <div class="is-clearfix" style="height: 2.5em"> <div class="is-pulled-right"> <a href="/search/advanced?terms-0-term=Tamkin%2C+A&terms-0-field=author&size=50&order=-announced_date_first">Advanced Search</a> </div> </div> <input type="hidden" name="order" value="-announced_date_first"> <input type="hidden" name="size" value="50"> </form> <div class="level breathe-horizontal"> <div class="level-left"> <form method="GET" action="/search/"> <div style="display: none;"> <select id="searchtype" name="searchtype"><option value="all">All fields</option><option value="title">Title</option><option selected value="author">Author(s)</option><option value="abstract">Abstract</option><option value="comments">Comments</option><option value="journal_ref">Journal reference</option><option value="acm_class">ACM classification</option><option value="msc_class">MSC classification</option><option value="report_num">Report number</option><option value="paper_id">arXiv identifier</option><option value="doi">DOI</option><option value="orcid">ORCID</option><option value="license">License (URI)</option><option value="author_id">arXiv author ID</option><option value="help">Help pages</option><option value="full_text">Full text</option></select> <input id="query" name="query" type="text" value="Tamkin, A"> <ul id="abstracts"><li><input checked id="abstracts-0" name="abstracts" type="radio" value="show"> <label for="abstracts-0">Show abstracts</label></li><li><input id="abstracts-1" name="abstracts" type="radio" value="hide"> <label for="abstracts-1">Hide abstracts</label></li></ul> </div> <div class="box field is-grouped is-grouped-multiline level-item"> <div class="control"> <span class="select is-small"> <select id="size" name="size"><option value="25">25</option><option selected value="50">50</option><option value="100">100</option><option value="200">200</option></select> </span> <label for="size">results per page</label>. </div> <div class="control"> <label for="order">Sort results by</label> <span class="select is-small"> <select id="order" name="order"><option selected value="-announced_date_first">Announcement date (newest first)</option><option value="announced_date_first">Announcement date (oldest first)</option><option value="-submitted_date">Submission date (newest first)</option><option value="submitted_date">Submission date (oldest first)</option><option value="">Relevance</option></select> </span> </div> <div class="control"> <button class="button is-small is-link">Go</button> </div> </div> </form> </div> </div> <ol class="breathe-horizontal" start="1"> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2412.13678">arXiv:2412.13678</a> <span> [<a href="https://arxiv.org/pdf/2412.13678">pdf</a>, <a href="https://arxiv.org/format/2412.13678">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computers and Society">cs.CY</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Cryptography and Security">cs.CR</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> </div> </div> <p class="title is-5 mathjax"> Clio: Privacy-Preserving Insights into Real-World AI Use </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=McCain%2C+M">Miles McCain</a>, <a href="/search/cs?searchtype=author&query=Handa%2C+K">Kunal Handa</a>, <a href="/search/cs?searchtype=author&query=Durmus%2C+E">Esin Durmus</a>, <a href="/search/cs?searchtype=author&query=Lovitt%2C+L">Liane Lovitt</a>, <a href="/search/cs?searchtype=author&query=Rathi%2C+A">Ankur Rathi</a>, <a href="/search/cs?searchtype=author&query=Huang%2C+S">Saffron Huang</a>, <a href="/search/cs?searchtype=author&query=Mountfield%2C+A">Alfred Mountfield</a>, <a href="/search/cs?searchtype=author&query=Hong%2C+J">Jerry Hong</a>, <a href="/search/cs?searchtype=author&query=Ritchie%2C+S">Stuart Ritchie</a>, <a href="/search/cs?searchtype=author&query=Stern%2C+M">Michael Stern</a>, <a href="/search/cs?searchtype=author&query=Clarke%2C+B">Brian Clarke</a>, <a href="/search/cs?searchtype=author&query=Goldberg%2C+L">Landon Goldberg</a>, <a href="/search/cs?searchtype=author&query=Sumers%2C+T+R">Theodore R. Sumers</a>, <a href="/search/cs?searchtype=author&query=Mueller%2C+J">Jared Mueller</a>, <a href="/search/cs?searchtype=author&query=McEachen%2C+W">William McEachen</a>, <a href="/search/cs?searchtype=author&query=Mitchell%2C+W">Wes Mitchell</a>, <a href="/search/cs?searchtype=author&query=Carter%2C+S">Shan Carter</a>, <a href="/search/cs?searchtype=author&query=Clark%2C+J">Jack Clark</a>, <a href="/search/cs?searchtype=author&query=Kaplan%2C+J">Jared Kaplan</a>, <a href="/search/cs?searchtype=author&query=Ganguli%2C+D">Deep Ganguli</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2412.13678v1-abstract-short" style="display: inline;"> How are AI assistants being used in the real world? While model providers in theory have a window into this impact via their users' data, both privacy concerns and practical challenges have made analyzing this data difficult. To address these issues, we present Clio (Claude insights and observations), a privacy-preserving platform that uses AI assistants themselves to analyze and surface aggregate… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2412.13678v1-abstract-full').style.display = 'inline'; document.getElementById('2412.13678v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2412.13678v1-abstract-full" style="display: none;"> How are AI assistants being used in the real world? While model providers in theory have a window into this impact via their users' data, both privacy concerns and practical challenges have made analyzing this data difficult. To address these issues, we present Clio (Claude insights and observations), a privacy-preserving platform that uses AI assistants themselves to analyze and surface aggregated usage patterns across millions of conversations, without the need for human reviewers to read raw conversations. We validate this can be done with a high degree of accuracy and privacy by conducting extensive evaluations. We demonstrate Clio's usefulness in two broad ways. First, we share insights about how models are being used in the real world from one million Claude.ai Free and Pro conversations, ranging from providing advice on hairstyles to providing guidance on Git operations and concepts. We also identify the most common high-level use cases on Claude.ai (coding, writing, and research tasks) as well as patterns that differ across languages (e.g., conversations in Japanese discuss elder care and aging populations at higher-than-typical rates). Second, we use Clio to make our systems safer by identifying coordinated attempts to abuse our systems, monitoring for unknown unknowns during critical periods like launches of new capabilities or major world events, and improving our existing monitoring systems. We also discuss the limitations of our approach, as well as risks and ethical concerns. By enabling analysis of real-world AI usage, Clio provides a scalable platform for empirically grounded AI safety and governance. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2412.13678v1-abstract-full').style.display = 'none'; document.getElementById('2412.13678v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 18 December, 2024; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> December 2024. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2406.10162">arXiv:2406.10162</a> <span> [<a href="https://arxiv.org/pdf/2406.10162">pdf</a>, <a href="https://arxiv.org/format/2406.10162">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> </div> </div> <p class="title is-5 mathjax"> Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Denison%2C+C">Carson Denison</a>, <a href="/search/cs?searchtype=author&query=MacDiarmid%2C+M">Monte MacDiarmid</a>, <a href="/search/cs?searchtype=author&query=Barez%2C+F">Fazl Barez</a>, <a href="/search/cs?searchtype=author&query=Duvenaud%2C+D">David Duvenaud</a>, <a href="/search/cs?searchtype=author&query=Kravec%2C+S">Shauna Kravec</a>, <a href="/search/cs?searchtype=author&query=Marks%2C+S">Samuel Marks</a>, <a href="/search/cs?searchtype=author&query=Schiefer%2C+N">Nicholas Schiefer</a>, <a href="/search/cs?searchtype=author&query=Soklaski%2C+R">Ryan Soklaski</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Kaplan%2C+J">Jared Kaplan</a>, <a href="/search/cs?searchtype=author&query=Shlegeris%2C+B">Buck Shlegeris</a>, <a href="/search/cs?searchtype=author&query=Bowman%2C+S+R">Samuel R. Bowman</a>, <a href="/search/cs?searchtype=author&query=Perez%2C+E">Ethan Perez</a>, <a href="/search/cs?searchtype=author&query=Hubinger%2C+E">Evan Hubinger</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2406.10162v3-abstract-short" style="display: inline;"> In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be to… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2406.10162v3-abstract-full').style.display = 'inline'; document.getElementById('2406.10162v3-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2406.10162v3-abstract-full" style="display: none;"> In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration. In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments. Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering. These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2406.10162v3-abstract-full').style.display = 'none'; document.getElementById('2406.10162v3-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 28 June, 2024; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 14 June, 2024; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> June 2024. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">Make it easier to find samples from the model, and highlight that our operational definition of reward tampering has false positives where the model attempts to complete the task honestly but edits the reward. Add paragraph to conclusion to this effect, and add sentence to figure 1 to this effect</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2406.07814">arXiv:2406.07814</a> <span> [<a href="https://arxiv.org/pdf/2406.07814">pdf</a>, <a href="https://arxiv.org/format/2406.07814">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Human-Computer Interaction">cs.HC</span> </div> <div class="is-inline-block" style="margin-left: 0.5rem"> <div class="tags has-addons"> <span class="tag is-dark is-size-7">doi</span> <span class="tag is-light is-size-7"><a class="" href="https://doi.org/10.1145/3630106.3658979">10.1145/3630106.3658979 <i class="fa fa-external-link" aria-hidden="true"></i></a></span> </div> </div> </div> <p class="title is-5 mathjax"> Collective Constitutional AI: Aligning a Language Model with Public Input </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Huang%2C+S">Saffron Huang</a>, <a href="/search/cs?searchtype=author&query=Siddarth%2C+D">Divya Siddarth</a>, <a href="/search/cs?searchtype=author&query=Lovitt%2C+L">Liane Lovitt</a>, <a href="/search/cs?searchtype=author&query=Liao%2C+T+I">Thomas I. Liao</a>, <a href="/search/cs?searchtype=author&query=Durmus%2C+E">Esin Durmus</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Ganguli%2C+D">Deep Ganguli</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2406.07814v1-abstract-short" style="display: inline;"> There is growing consensus that language model (LM) developers should not be the sole deciders of LM behavior, creating a need for methods that enable the broader public to collectively shape the behavior of LM systems that affect them. To address this need, we present Collective Constitutional AI (CCAI): a multi-stage process for sourcing and integrating public input into LMs-from identifying a t… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2406.07814v1-abstract-full').style.display = 'inline'; document.getElementById('2406.07814v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2406.07814v1-abstract-full" style="display: none;"> There is growing consensus that language model (LM) developers should not be the sole deciders of LM behavior, creating a need for methods that enable the broader public to collectively shape the behavior of LM systems that affect them. To address this need, we present Collective Constitutional AI (CCAI): a multi-stage process for sourcing and integrating public input into LMs-from identifying a target population to sourcing principles to training and evaluating a model. We demonstrate the real-world practicality of this approach by creating what is, to our knowledge, the first LM fine-tuned with collectively sourced public input and evaluating this model against a baseline model trained with established principles from a LM developer. Our quantitative evaluations demonstrate several benefits of our approach: the CCAI-trained model shows lower bias across nine social dimensions compared to the baseline model, while maintaining equivalent performance on language, math, and helpful-harmless evaluations. Qualitative comparisons of the models suggest that the models differ on the basis of their respective constitutions, e.g., when prompted with contentious topics, the CCAI-trained model tends to generate responses that reframe the matter positively instead of a refusal. These results demonstrate a promising, tractable pathway toward publicly informed development of language models. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2406.07814v1-abstract-full').style.display = 'none'; document.getElementById('2406.07814v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 11 June, 2024; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> June 2024. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">ACM Class:</span> I.2.7; K.4.2 </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Journal ref:</span> Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency. 1395-1417 </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2403.05534">arXiv:2403.05534</a> <span> [<a href="https://arxiv.org/pdf/2403.05534">pdf</a>, <a href="https://arxiv.org/format/2403.05534">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> </div> </div> <p class="title is-5 mathjax"> Bayesian Preference Elicitation with Language Models </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Handa%2C+K">Kunal Handa</a>, <a href="/search/cs?searchtype=author&query=Gal%2C+Y">Yarin Gal</a>, <a href="/search/cs?searchtype=author&query=Pavlick%2C+E">Ellie Pavlick</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a>, <a href="/search/cs?searchtype=author&query=Andreas%2C+J">Jacob Andreas</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Li%2C+B+Z">Belinda Z. Li</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2403.05534v1-abstract-short" style="display: inline;"> Aligning AI systems to users' interests requires understanding and incorporating humans' complex values and preferences. Recently, language models (LMs) have been used to gather information about the preferences of human users. This preference data can be used to fine-tune or guide other LMs and/or AI systems. However, LMs have been shown to struggle with crucial aspects of preference learning: qu… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2403.05534v1-abstract-full').style.display = 'inline'; document.getElementById('2403.05534v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2403.05534v1-abstract-full" style="display: none;"> Aligning AI systems to users' interests requires understanding and incorporating humans' complex values and preferences. Recently, language models (LMs) have been used to gather information about the preferences of human users. This preference data can be used to fine-tune or guide other LMs and/or AI systems. However, LMs have been shown to struggle with crucial aspects of preference learning: quantifying uncertainty, modeling human mental states, and asking informative questions. These challenges have been addressed in other areas of machine learning, such as Bayesian Optimal Experimental Design (BOED), which focus on designing informative queries within a well-defined feature space. But these methods, in turn, are difficult to scale and apply to real-world problems where simply identifying the relevant features can be difficult. We introduce OPEN (Optimal Preference Elicitation with Natural language) a framework that uses BOED to guide the choice of informative questions and an LM to extract features and translate abstract BOED queries into natural language questions. By combining the flexibility of LMs with the rigor of BOED, OPEN can optimize the informativity of queries while remaining adaptable to real-world domains. In user studies, we find that OPEN outperforms existing LM- and BOED-based methods for preference elicitation. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2403.05534v1-abstract-full').style.display = 'none'; document.getElementById('2403.05534v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 8 March, 2024; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> March 2024. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2312.03689">arXiv:2312.03689</a> <span> [<a href="https://arxiv.org/pdf/2312.03689">pdf</a>, <a href="https://arxiv.org/format/2312.03689">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> </div> </div> <p class="title is-5 mathjax"> Evaluating and Mitigating Discrimination in Language Model Decisions </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Askell%2C+A">Amanda Askell</a>, <a href="/search/cs?searchtype=author&query=Lovitt%2C+L">Liane Lovitt</a>, <a href="/search/cs?searchtype=author&query=Durmus%2C+E">Esin Durmus</a>, <a href="/search/cs?searchtype=author&query=Joseph%2C+N">Nicholas Joseph</a>, <a href="/search/cs?searchtype=author&query=Kravec%2C+S">Shauna Kravec</a>, <a href="/search/cs?searchtype=author&query=Nguyen%2C+K">Karina Nguyen</a>, <a href="/search/cs?searchtype=author&query=Kaplan%2C+J">Jared Kaplan</a>, <a href="/search/cs?searchtype=author&query=Ganguli%2C+D">Deep Ganguli</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2312.03689v1-abstract-short" style="display: inline;"> As language models (LMs) advance, interest is growing in applying them to high-stakes societal decisions, such as determining financing or housing eligibility. However, their potential for discrimination in such contexts raises ethical concerns, motivating the need for better methods to evaluate these risks. We present a method for proactively evaluating the potential discriminatory impact of LMs… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2312.03689v1-abstract-full').style.display = 'inline'; document.getElementById('2312.03689v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2312.03689v1-abstract-full" style="display: none;"> As language models (LMs) advance, interest is growing in applying them to high-stakes societal decisions, such as determining financing or housing eligibility. However, their potential for discrimination in such contexts raises ethical concerns, motivating the need for better methods to evaluate these risks. We present a method for proactively evaluating the potential discriminatory impact of LMs in a wide range of use cases, including hypothetical use cases where they have not yet been deployed. Specifically, we use an LM to generate a wide array of potential prompts that decision-makers may input into an LM, spanning 70 diverse decision scenarios across society, and systematically vary the demographic information in each prompt. Applying this methodology reveals patterns of both positive and negative discrimination in the Claude 2.0 model in select settings when no interventions are applied. While we do not endorse or permit the use of language models to make automated decisions for the high-risk use cases we study, we demonstrate techniques to significantly decrease both positive and negative discrimination through careful prompt engineering, providing pathways toward safer deployment in use cases where they may be appropriate. Our work enables developers and policymakers to anticipate, measure, and address discrimination as language model capabilities and applications continue to expand. We release our dataset and prompts at https://huggingface.co/datasets/Anthropic/discrim-eval <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2312.03689v1-abstract-full').style.display = 'none'; document.getElementById('2312.03689v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 6 December, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> December 2023. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2310.17769">arXiv:2310.17769</a> <span> [<a href="https://arxiv.org/pdf/2310.17769">pdf</a>, <a href="https://arxiv.org/format/2310.17769">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> </div> </div> <p class="title is-5 mathjax"> Social Contract AI: Aligning AI Assistants with Implicit Group Norms </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Fr%C3%A4nken%2C+J">Jan-Philipp Fr盲nken</a>, <a href="/search/cs?searchtype=author&query=Kwok%2C+S">Sam Kwok</a>, <a href="/search/cs?searchtype=author&query=Ye%2C+P">Peixuan Ye</a>, <a href="/search/cs?searchtype=author&query=Gandhi%2C+K">Kanishk Gandhi</a>, <a href="/search/cs?searchtype=author&query=Arumugam%2C+D">Dilip Arumugam</a>, <a href="/search/cs?searchtype=author&query=Moore%2C+J">Jared Moore</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Gerstenberg%2C+T">Tobias Gerstenberg</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N+D">Noah D. Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2310.17769v2-abstract-short" style="display: inline;"> We explore the idea of aligning an AI assistant by inverting a model of users' (unknown) preferences from observed interactions. To validate our proposal, we run proof-of-concept simulations in the economic ultimatum game, formalizing user preferences as policies that guide the actions of simulated players. We find that the AI assistant accurately aligns its behavior to match standard policies fro… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2310.17769v2-abstract-full').style.display = 'inline'; document.getElementById('2310.17769v2-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2310.17769v2-abstract-full" style="display: none;"> We explore the idea of aligning an AI assistant by inverting a model of users' (unknown) preferences from observed interactions. To validate our proposal, we run proof-of-concept simulations in the economic ultimatum game, formalizing user preferences as policies that guide the actions of simulated players. We find that the AI assistant accurately aligns its behavior to match standard policies from the economic literature (e.g., selfish, altruistic). However, the assistant's learned policies lack robustness and exhibit limited generalization in an out-of-distribution setting when confronted with a currency (e.g., grams of medicine) that was not included in the assistant's training distribution. Additionally, we find that when there is inconsistency in the relationship between language use and an unknown policy (e.g., an altruistic policy combined with rude language), the assistant's learning of the policy is slowed. Overall, our preliminary results suggest that developing simulation frameworks in which AI assistants need to infer preferences from diverse users can provide a valuable approach for studying practical alignment questions. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2310.17769v2-abstract-full').style.display = 'none'; document.getElementById('2310.17769v2-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 3 December, 2023; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 26 October, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> October 2023. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">SoLaR NeurIPS 2023 Workshop (https://solar-neurips.github.io/)</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2310.17230">arXiv:2310.17230</a> <span> [<a href="https://arxiv.org/pdf/2310.17230">pdf</a>, <a href="https://arxiv.org/format/2310.17230">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> </div> </div> <p class="title is-5 mathjax"> Codebook Features: Sparse and Discrete Interpretability for Neural Networks </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Taufeeque%2C+M">Mohammad Taufeeque</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N+D">Noah D. Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2310.17230v1-abstract-short" style="display: inline;"> Understanding neural networks is challenging in part because of the dense, continuous nature of their hidden states. We explore whether we can train neural networks to have hidden states that are sparse, discrete, and more interpretable by quantizing their continuous features into what we call codebook features. Codebook features are produced by finetuning neural networks with vector quantization… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2310.17230v1-abstract-full').style.display = 'inline'; document.getElementById('2310.17230v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2310.17230v1-abstract-full" style="display: none;"> Understanding neural networks is challenging in part because of the dense, continuous nature of their hidden states. We explore whether we can train neural networks to have hidden states that are sparse, discrete, and more interpretable by quantizing their continuous features into what we call codebook features. Codebook features are produced by finetuning neural networks with vector quantization bottlenecks at each layer, producing a network whose hidden features are the sum of a small number of discrete vector codes chosen from a larger codebook. Surprisingly, we find that neural networks can operate under this extreme bottleneck with only modest degradation in performance. This sparse, discrete bottleneck also provides an intuitive way of controlling neural network behavior: first, find codes that activate when the desired behavior is present, then activate those same codes during generation to elicit that behavior. We validate our approach by training codebook Transformers on several different datasets. First, we explore a finite state machine dataset with far more hidden states than neurons. In this setting, our approach overcomes the superposition problem by assigning states to distinct codes, and we find that we can make the neural network behave as if it is in a different state by activating the code for that state. Second, we train Transformer language models with up to 410M parameters on two natural language datasets. We identify codes in these models representing diverse, disentangled concepts (ranging from negative emotions to months of the year) and find that we can guide the model to generate different topics by activating the appropriate codes during inference. Overall, codebook features appear to be a promising unit of analysis and control for neural networks and interpretability. Our codebase and models are open-sourced at https://github.com/taufeeque9/codebook-features. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2310.17230v1-abstract-full').style.display = 'none'; document.getElementById('2310.17230v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 26 October, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> October 2023. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2310.11589">arXiv:2310.11589</a> <span> [<a href="https://arxiv.org/pdf/2310.11589">pdf</a>, <a href="https://arxiv.org/format/2310.11589">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> </div> </div> <p class="title is-5 mathjax"> Eliciting Human Preferences with Language Models </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Li%2C+B+Z">Belinda Z. Li</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a>, <a href="/search/cs?searchtype=author&query=Andreas%2C+J">Jacob Andreas</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2310.11589v1-abstract-short" style="display: inline;"> Language models (LMs) can be directed to perform target tasks by using labeled examples or natural language prompts. But selecting examples or writing prompts for can be challenging--especially in tasks that involve unusual edge cases, demand precise articulation of nebulous preferences, or require an accurate mental model of LM behavior. We propose to use *LMs themselves* to guide the task specif… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2310.11589v1-abstract-full').style.display = 'inline'; document.getElementById('2310.11589v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2310.11589v1-abstract-full" style="display: none;"> Language models (LMs) can be directed to perform target tasks by using labeled examples or natural language prompts. But selecting examples or writing prompts for can be challenging--especially in tasks that involve unusual edge cases, demand precise articulation of nebulous preferences, or require an accurate mental model of LM behavior. We propose to use *LMs themselves* to guide the task specification process. In this paper, we introduce **Generative Active Task Elicitation (GATE)**: a learning framework in which models elicit and infer intended behavior through free-form, language-based interaction with users. We study GATE in three domains: email validation, content recommendation, and moral reasoning. In preregistered experiments, we show that LMs prompted to perform GATE (e.g., by generating open-ended questions or synthesizing informative edge cases) elicit responses that are often more informative than user-written prompts or labels. Users report that interactive task elicitation requires less effort than prompting or example labeling and surfaces novel considerations not initially anticipated by users. Our findings suggest that LM-driven elicitation can be a powerful tool for aligning models to complex human preferences and values. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2310.11589v1-abstract-full').style.display = 'none'; document.getElementById('2310.11589v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 17 October, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> October 2023. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">26 pages, 15 figures</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2309.13457">arXiv:2309.13457</a> <span> [<a href="https://arxiv.org/pdf/2309.13457">pdf</a>, <a href="https://arxiv.org/format/2309.13457">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computer Vision and Pattern Recognition">cs.CV</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computational Physics">physics.comp-ph</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Fluid Dynamics">physics.flu-dyn</span> </div> </div> <p class="title is-5 mathjax"> Turbulence in Focus: Benchmarking Scaling Behavior of 3D Volumetric Super-Resolution with BLASTNet 2.0 Data </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Chung%2C+W+T">Wai Tong Chung</a>, <a href="/search/cs?searchtype=author&query=Akoush%2C+B">Bassem Akoush</a>, <a href="/search/cs?searchtype=author&query=Sharma%2C+P">Pushan Sharma</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Jung%2C+K+S">Ki Sung Jung</a>, <a href="/search/cs?searchtype=author&query=Chen%2C+J+H">Jacqueline H. Chen</a>, <a href="/search/cs?searchtype=author&query=Guo%2C+J">Jack Guo</a>, <a href="/search/cs?searchtype=author&query=Brouzet%2C+D">Davy Brouzet</a>, <a href="/search/cs?searchtype=author&query=Talei%2C+M">Mohsen Talei</a>, <a href="/search/cs?searchtype=author&query=Savard%2C+B">Bruno Savard</a>, <a href="/search/cs?searchtype=author&query=Poludnenko%2C+A+Y">Alexei Y. Poludnenko</a>, <a href="/search/cs?searchtype=author&query=Ihme%2C+M">Matthias Ihme</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2309.13457v3-abstract-short" style="display: inline;"> Analysis of compressible turbulent flows is essential for applications related to propulsion, energy generation, and the environment. Here, we present BLASTNet 2.0, a 2.2 TB network-of-datasets containing 744 full-domain samples from 34 high-fidelity direct numerical simulations, which addresses the current limited availability of 3D high-fidelity reacting and non-reacting compressible turbulent f… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2309.13457v3-abstract-full').style.display = 'inline'; document.getElementById('2309.13457v3-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2309.13457v3-abstract-full" style="display: none;"> Analysis of compressible turbulent flows is essential for applications related to propulsion, energy generation, and the environment. Here, we present BLASTNet 2.0, a 2.2 TB network-of-datasets containing 744 full-domain samples from 34 high-fidelity direct numerical simulations, which addresses the current limited availability of 3D high-fidelity reacting and non-reacting compressible turbulent flow simulation data. With this data, we benchmark a total of 49 variations of five deep learning approaches for 3D super-resolution - which can be applied for improving scientific imaging, simulations, turbulence models, as well as in computer vision applications. We perform neural scaling analysis on these models to examine the performance of different machine learning (ML) approaches, including two scientific ML techniques. We demonstrate that (i) predictive performance can scale with model size and cost, (ii) architecture matters significantly, especially for smaller models, and (iii) the benefits of physics-based losses can persist with increasing model size. The outcomes of this benchmark study are anticipated to offer insights that can aid the design of 3D super-resolution models, especially for turbulence models, while this data is expected to foster ML methods for a broad range of flow physics applications. This data is publicly available with download links and browsing tools consolidated at https://blastnet.github.io. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2309.13457v3-abstract-full').style.display = 'none'; document.getElementById('2309.13457v3-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 27 October, 2023; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 23 September, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> September 2023. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">Accepted in Adv. in Neural Information Processing Systems 36 (NeurIPS 2023). Link: https://nips.cc/virtual/2023/poster/73433 . 55 pages, 21 figures. Keywords: Super-resolution, 3D, Neural Scaling, Physics-informed Loss, Computational Fluid Dynamics, Partial Differential Equations, Turbulent Reacting Flows, Direct Numerical Simulation, Fluid Mechanics, Combustion, Computer Vision</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2308.03296">arXiv:2308.03296</a> <span> [<a href="https://arxiv.org/pdf/2308.03296">pdf</a>, <a href="https://arxiv.org/format/2308.03296">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Machine Learning">stat.ML</span> </div> </div> <p class="title is-5 mathjax"> Studying Large Language Model Generalization with Influence Functions </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Grosse%2C+R">Roger Grosse</a>, <a href="/search/cs?searchtype=author&query=Bae%2C+J">Juhan Bae</a>, <a href="/search/cs?searchtype=author&query=Anil%2C+C">Cem Anil</a>, <a href="/search/cs?searchtype=author&query=Elhage%2C+N">Nelson Elhage</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Tajdini%2C+A">Amirhossein Tajdini</a>, <a href="/search/cs?searchtype=author&query=Steiner%2C+B">Benoit Steiner</a>, <a href="/search/cs?searchtype=author&query=Li%2C+D">Dustin Li</a>, <a href="/search/cs?searchtype=author&query=Durmus%2C+E">Esin Durmus</a>, <a href="/search/cs?searchtype=author&query=Perez%2C+E">Ethan Perez</a>, <a href="/search/cs?searchtype=author&query=Hubinger%2C+E">Evan Hubinger</a>, <a href="/search/cs?searchtype=author&query=Luko%C5%A1i%C5%ABt%C4%97%2C+K">Kamil臈 Luko拧i奴t臈</a>, <a href="/search/cs?searchtype=author&query=Nguyen%2C+K">Karina Nguyen</a>, <a href="/search/cs?searchtype=author&query=Joseph%2C+N">Nicholas Joseph</a>, <a href="/search/cs?searchtype=author&query=McCandlish%2C+S">Sam McCandlish</a>, <a href="/search/cs?searchtype=author&query=Kaplan%2C+J">Jared Kaplan</a>, <a href="/search/cs?searchtype=author&query=Bowman%2C+S+R">Samuel R. Bowman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2308.03296v1-abstract-short" style="display: inline;"> When trying to gain better visibility into a machine learning model in order to understand and mitigate the associated risks, a potentially valuable source of evidence is: which training examples most contribute to a given behavior? Influence functions aim to answer a counterfactual: how would the model's parameters (and hence its outputs) change if a given sequence were added to the training set?… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2308.03296v1-abstract-full').style.display = 'inline'; document.getElementById('2308.03296v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2308.03296v1-abstract-full" style="display: none;"> When trying to gain better visibility into a machine learning model in order to understand and mitigate the associated risks, a potentially valuable source of evidence is: which training examples most contribute to a given behavior? Influence functions aim to answer a counterfactual: how would the model's parameters (and hence its outputs) change if a given sequence were added to the training set? While influence functions have produced insights for small models, they are difficult to scale to large language models (LLMs) due to the difficulty of computing an inverse-Hessian-vector product (IHVP). We use the Eigenvalue-corrected Kronecker-Factored Approximate Curvature (EK-FAC) approximation to scale influence functions up to LLMs with up to 52 billion parameters. In our experiments, EK-FAC achieves similar accuracy to traditional influence function estimators despite the IHVP computation being orders of magnitude faster. We investigate two algorithmic techniques to reduce the cost of computing gradients of candidate training sequences: TF-IDF filtering and query batching. We use influence functions to investigate the generalization patterns of LLMs, including the sparsity of the influence patterns, increasing abstraction with scale, math and programming abilities, cross-lingual generalization, and role-playing behavior. Despite many apparently sophisticated forms of generalization, we identify a surprising limitation: influences decay to near-zero when the order of key phrases is flipped. Overall, influence functions give us a powerful new tool for studying the generalization properties of LLMs. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2308.03296v1-abstract-full').style.display = 'none'; document.getElementById('2308.03296v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 7 August, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> August 2023. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">119 pages, 47 figures, 22 tables</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2306.16388">arXiv:2306.16388</a> <span> [<a href="https://arxiv.org/pdf/2306.16388">pdf</a>, <a href="https://arxiv.org/format/2306.16388">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> </div> </div> <p class="title is-5 mathjax"> Towards Measuring the Representation of Subjective Global Opinions in Language Models </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Durmus%2C+E">Esin Durmus</a>, <a href="/search/cs?searchtype=author&query=Nguyen%2C+K">Karina Nguyen</a>, <a href="/search/cs?searchtype=author&query=Liao%2C+T+I">Thomas I. Liao</a>, <a href="/search/cs?searchtype=author&query=Schiefer%2C+N">Nicholas Schiefer</a>, <a href="/search/cs?searchtype=author&query=Askell%2C+A">Amanda Askell</a>, <a href="/search/cs?searchtype=author&query=Bakhtin%2C+A">Anton Bakhtin</a>, <a href="/search/cs?searchtype=author&query=Chen%2C+C">Carol Chen</a>, <a href="/search/cs?searchtype=author&query=Hatfield-Dodds%2C+Z">Zac Hatfield-Dodds</a>, <a href="/search/cs?searchtype=author&query=Hernandez%2C+D">Danny Hernandez</a>, <a href="/search/cs?searchtype=author&query=Joseph%2C+N">Nicholas Joseph</a>, <a href="/search/cs?searchtype=author&query=Lovitt%2C+L">Liane Lovitt</a>, <a href="/search/cs?searchtype=author&query=McCandlish%2C+S">Sam McCandlish</a>, <a href="/search/cs?searchtype=author&query=Sikder%2C+O">Orowa Sikder</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Thamkul%2C+J">Janel Thamkul</a>, <a href="/search/cs?searchtype=author&query=Kaplan%2C+J">Jared Kaplan</a>, <a href="/search/cs?searchtype=author&query=Clark%2C+J">Jack Clark</a>, <a href="/search/cs?searchtype=author&query=Ganguli%2C+D">Deep Ganguli</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2306.16388v2-abstract-short" style="display: inline;"> Large language models (LLMs) may not equitably represent diverse global perspectives on societal issues. In this paper, we develop a quantitative framework to evaluate whose opinions model-generated responses are more similar to. We first build a dataset, GlobalOpinionQA, comprised of questions and answers from cross-national surveys designed to capture diverse opinions on global issues across dif… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2306.16388v2-abstract-full').style.display = 'inline'; document.getElementById('2306.16388v2-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2306.16388v2-abstract-full" style="display: none;"> Large language models (LLMs) may not equitably represent diverse global perspectives on societal issues. In this paper, we develop a quantitative framework to evaluate whose opinions model-generated responses are more similar to. We first build a dataset, GlobalOpinionQA, comprised of questions and answers from cross-national surveys designed to capture diverse opinions on global issues across different countries. Next, we define a metric that quantifies the similarity between LLM-generated survey responses and human responses, conditioned on country. With our framework, we run three experiments on an LLM trained to be helpful, honest, and harmless with Constitutional AI. By default, LLM responses tend to be more similar to the opinions of certain populations, such as those from the USA, and some European and South American countries, highlighting the potential for biases. When we prompt the model to consider a particular country's perspective, responses shift to be more similar to the opinions of the prompted populations, but can reflect harmful cultural stereotypes. When we translate GlobalOpinionQA questions to a target language, the model's responses do not necessarily become the most similar to the opinions of speakers of those languages. We release our dataset for others to use and build on. Our data is at https://huggingface.co/datasets/Anthropic/llm_global_opinions. We also provide an interactive visualization at https://llmglobalvalues.anthropic.com. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2306.16388v2-abstract-full').style.display = 'none'; document.getElementById('2306.16388v2-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 11 April, 2024; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 28 June, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> June 2023. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2306.02889">arXiv:2306.02889</a> <span> [<a href="https://arxiv.org/pdf/2306.02889">pdf</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computers and Society">cs.CY</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> </div> </div> <p class="title is-5 mathjax"> Operationalising the Definition of General Purpose AI Systems: Assessing Four Approaches </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Uuk%2C+R">Risto Uuk</a>, <a href="/search/cs?searchtype=author&query=Gutierrez%2C+C+I">Carlos Ignacio Gutierrez</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2306.02889v1-abstract-short" style="display: inline;"> The European Union's Artificial Intelligence (AI) Act is set to be a landmark legal instrument for regulating AI technology. While stakeholders have primarily focused on the governance of fixed purpose AI applications (also known as narrow AI), more attention is required to understand the nature of highly and broadly capable systems. As of the beginning of 2023, several definitions for General Pur… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2306.02889v1-abstract-full').style.display = 'inline'; document.getElementById('2306.02889v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2306.02889v1-abstract-full" style="display: none;"> The European Union's Artificial Intelligence (AI) Act is set to be a landmark legal instrument for regulating AI technology. While stakeholders have primarily focused on the governance of fixed purpose AI applications (also known as narrow AI), more attention is required to understand the nature of highly and broadly capable systems. As of the beginning of 2023, several definitions for General Purpose AI Systems (GPAIS) exist in relation to the AI Act, attempting to distinguish between systems with and without a fixed purpose. In this article, we operationalise these differences through the concept of "distinct tasks" and examine four approaches (quantity, performance, adaptability, and emergence) to determine whether an AI system should be classified as a GPAIS. We suggest that EU stakeholders use the four approaches as a starting point to discriminate between fixed-purpose and GPAIS. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2306.02889v1-abstract-full').style.display = 'none'; document.getElementById('2306.02889v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 5 June, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> June 2023. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2304.08486">arXiv:2304.08486</a> <span> [<a href="https://arxiv.org/pdf/2304.08486">pdf</a>, <a href="https://arxiv.org/format/2304.08486">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computer Vision and Pattern Recognition">cs.CV</span> </div> </div> <p class="title is-5 mathjax"> BenchMD: A Benchmark for Unified Learning on Medical Images and Sensors </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Wantlin%2C+K">Kathryn Wantlin</a>, <a href="/search/cs?searchtype=author&query=Wu%2C+C">Chenwei Wu</a>, <a href="/search/cs?searchtype=author&query=Huang%2C+S">Shih-Cheng Huang</a>, <a href="/search/cs?searchtype=author&query=Banerjee%2C+O">Oishi Banerjee</a>, <a href="/search/cs?searchtype=author&query=Dadabhoy%2C+F">Farah Dadabhoy</a>, <a href="/search/cs?searchtype=author&query=Mehta%2C+V+V">Veeral Vipin Mehta</a>, <a href="/search/cs?searchtype=author&query=Han%2C+R+W">Ryan Wonhee Han</a>, <a href="/search/cs?searchtype=author&query=Cao%2C+F">Fang Cao</a>, <a href="/search/cs?searchtype=author&query=Narayan%2C+R+R">Raja R. Narayan</a>, <a href="/search/cs?searchtype=author&query=Colak%2C+E">Errol Colak</a>, <a href="/search/cs?searchtype=author&query=Adamson%2C+A">Adewole Adamson</a>, <a href="/search/cs?searchtype=author&query=Heacock%2C+L">Laura Heacock</a>, <a href="/search/cs?searchtype=author&query=Tison%2C+G+H">Geoffrey H. Tison</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Rajpurkar%2C+P">Pranav Rajpurkar</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2304.08486v2-abstract-short" style="display: inline;"> Medical data poses a daunting challenge for AI algorithms: it exists in many different modalities, experiences frequent distribution shifts, and suffers from a scarcity of examples and labels. Recent advances, including transformers and self-supervised learning, promise a more universal approach that can be applied flexibly across these diverse conditions. To measure and drive progress in this dir… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2304.08486v2-abstract-full').style.display = 'inline'; document.getElementById('2304.08486v2-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2304.08486v2-abstract-full" style="display: none;"> Medical data poses a daunting challenge for AI algorithms: it exists in many different modalities, experiences frequent distribution shifts, and suffers from a scarcity of examples and labels. Recent advances, including transformers and self-supervised learning, promise a more universal approach that can be applied flexibly across these diverse conditions. To measure and drive progress in this direction, we present BenchMD: a benchmark that tests how well unified, modality-agnostic methods, including architectures and training techniques (e.g. self-supervised learning, ImageNet pretraining),perform on a diverse array of clinically-relevant medical tasks. BenchMD combines 19 publicly available datasets for 7 medical modalities, including 1D sensor data, 2D images, and 3D volumetric scans. Our benchmark reflects real-world data constraints by evaluating methods across a range of dataset sizes, including challenging few-shot settings that incentivize the use of pretraining. Finally, we evaluate performance on out-of-distribution data collected at different hospitals than the training data, representing naturally-occurring distribution shifts that frequently degrade the performance of medical AI models. Our baseline results demonstrate that no unified learning technique achieves strong performance across all modalities, leaving ample room for improvement on the benchmark. Code is released at https://github.com/rajpurkarlab/BenchMD. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2304.08486v2-abstract-full').style.display = 'none'; document.getElementById('2304.08486v2-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 26 June, 2023; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 17 April, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> April 2023. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2302.05757">arXiv:2302.05757</a> <span> [<a href="https://arxiv.org/pdf/2302.05757">pdf</a>, <a href="https://arxiv.org/format/2302.05757">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computer Vision and Pattern Recognition">cs.CV</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> </div> </div> <p class="title is-5 mathjax"> Multispectral Contrastive Learning with Viewmaker Networks </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Bayrooti%2C+J">Jasmine Bayrooti</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2302.05757v3-abstract-short" style="display: inline;"> Contrastive learning methods have been applied to a range of domains and modalities by training models to identify similar "views" of data points. However, specialized scientific modalities pose a challenge for this paradigm, as identifying good views for each scientific instrument is complex and time-intensive. In this paper, we focus on applying contrastive learning approaches to a variety of re… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2302.05757v3-abstract-full').style.display = 'inline'; document.getElementById('2302.05757v3-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2302.05757v3-abstract-full" style="display: none;"> Contrastive learning methods have been applied to a range of domains and modalities by training models to identify similar "views" of data points. However, specialized scientific modalities pose a challenge for this paradigm, as identifying good views for each scientific instrument is complex and time-intensive. In this paper, we focus on applying contrastive learning approaches to a variety of remote sensing datasets. We show that Viewmaker networks, a recently proposed method for generating views, are promising for producing views in this setting without requiring extensive domain knowledge and trial and error. We apply Viewmaker to four multispectral imaging problems, each with a different format, finding that Viewmaker can outperform cropping- and reflection-based methods for contrastive learning in every case when evaluated on downstream classification tasks. This provides additional evidence that domain-agnostic methods can empower contrastive learning to scale to real-world scientific domains. Open source code can be found at https://github.com/jbayrooti/divmaker. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2302.05757v3-abstract-full').style.display = 'none'; document.getElementById('2302.05757v3-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 3 June, 2023; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 11 February, 2023; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> February 2023. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">Appearing in CVPR-PBVS 2023</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2212.10711">arXiv:2212.10711</a> <span> [<a href="https://arxiv.org/pdf/2212.10711">pdf</a>, <a href="https://arxiv.org/format/2212.10711">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> </div> </div> <p class="title is-5 mathjax"> Task Ambiguity in Humans and Language Models </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Handa%2C+K">Kunal Handa</a>, <a href="/search/cs?searchtype=author&query=Shrestha%2C+A">Avash Shrestha</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2212.10711v1-abstract-short" style="display: inline;"> Language models have recently achieved strong performance across a wide range of NLP benchmarks. However, unlike benchmarks, real world tasks are often poorly specified, and agents must deduce the user's intended behavior from a combination of context, instructions, and examples. We investigate how both humans and models behave in the face of such task ambiguity by proposing AmbiBench, a new bench… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2212.10711v1-abstract-full').style.display = 'inline'; document.getElementById('2212.10711v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2212.10711v1-abstract-full" style="display: none;"> Language models have recently achieved strong performance across a wide range of NLP benchmarks. However, unlike benchmarks, real world tasks are often poorly specified, and agents must deduce the user's intended behavior from a combination of context, instructions, and examples. We investigate how both humans and models behave in the face of such task ambiguity by proposing AmbiBench, a new benchmark of six ambiguously-specified classification tasks. We evaluate humans and models on AmbiBench by seeing how well they identify the intended task using 1) instructions with varying degrees of ambiguity, and 2) different numbers of labeled examples. We find that the combination of model scaling (to 175B parameters) and training with human feedback data enables models to approach or exceed the accuracy of human participants across tasks, but that either one alone is not sufficient. In addition, we show how to dramatically improve the accuracy of language models trained without large-scale human feedback training by finetuning on a small number of ambiguous in-context examples, providing a promising direction for teaching models to generalize well in the face of ambiguity. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2212.10711v1-abstract-full').style.display = 'none'; document.getElementById('2212.10711v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 20 December, 2022; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> December 2022. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2212.08378">arXiv:2212.08378</a> <span> [<a href="https://arxiv.org/pdf/2212.08378">pdf</a>, <a href="https://arxiv.org/format/2212.08378">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computer Vision and Pattern Recognition">cs.CV</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Sound">cs.SD</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Audio and Speech Processing">eess.AS</span> </div> </div> <p class="title is-5 mathjax"> Feature Dropout: Revisiting the Role of Augmentations in Contrastive Learning </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Glasgow%2C+M">Margalit Glasgow</a>, <a href="/search/cs?searchtype=author&query=He%2C+X">Xiluo He</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2212.08378v1-abstract-short" style="display: inline;"> What role do augmentations play in contrastive learning? Recent work suggests that good augmentations are label-preserving with respect to a specific downstream task. We complicate this picture by showing that label-destroying augmentations can be useful in the foundation model setting, where the goal is to learn diverse, general-purpose representations for multiple downstream tasks. We perform co… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2212.08378v1-abstract-full').style.display = 'inline'; document.getElementById('2212.08378v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2212.08378v1-abstract-full" style="display: none;"> What role do augmentations play in contrastive learning? Recent work suggests that good augmentations are label-preserving with respect to a specific downstream task. We complicate this picture by showing that label-destroying augmentations can be useful in the foundation model setting, where the goal is to learn diverse, general-purpose representations for multiple downstream tasks. We perform contrastive learning experiments on a range of image and audio datasets with multiple downstream tasks (e.g. for digits superimposed on photographs, predicting the class of one vs. the other). We find that Viewmaker Networks, a recently proposed model for learning augmentations for contrastive learning, produce label-destroying augmentations that stochastically destroy features needed for different downstream tasks. These augmentations are interpretable (e.g. altering shapes, digits, or letters added to images) and surprisingly often result in better performance compared to expert-designed augmentations, despite not preserving label information. To support our empirical results, we theoretically analyze a simple contrastive learning setting with a linear model. In this setting, label-destroying augmentations are crucial for preventing one set of features from suppressing the learning of features useful for another downstream task. Our results highlight the need for analyzing the interaction between multiple downstream tasks when trying to explain the success of foundation models. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2212.08378v1-abstract-full').style.display = 'none'; document.getElementById('2212.08378v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 16 December, 2022; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> December 2022. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2204.08491">arXiv:2204.08491</a> <span> [<a href="https://arxiv.org/pdf/2204.08491">pdf</a>, <a href="https://arxiv.org/format/2204.08491">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computer Vision and Pattern Recognition">cs.CV</span> </div> </div> <p class="title is-5 mathjax"> Active Learning Helps Pretrained Models Learn the Intended Task </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Nguyen%2C+D">Dat Nguyen</a>, <a href="/search/cs?searchtype=author&query=Deshpande%2C+S">Salil Deshpande</a>, <a href="/search/cs?searchtype=author&query=Mu%2C+J">Jesse Mu</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2204.08491v1-abstract-short" style="display: inline;"> Models can fail in unpredictable ways during deployment due to task ambiguity, when multiple behaviors are consistent with the provided training data. An example is an object classifier trained on red squares and blue circles: when encountering blue squares, the intended behavior is undefined. We investigate whether pretrained models are better active learners, capable of disambiguating between th… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2204.08491v1-abstract-full').style.display = 'inline'; document.getElementById('2204.08491v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2204.08491v1-abstract-full" style="display: none;"> Models can fail in unpredictable ways during deployment due to task ambiguity, when multiple behaviors are consistent with the provided training data. An example is an object classifier trained on red squares and blue circles: when encountering blue squares, the intended behavior is undefined. We investigate whether pretrained models are better active learners, capable of disambiguating between the possible tasks a user may be trying to specify. Intriguingly, we find that better active learning is an emergent property of the pretraining process: pretrained models require up to 5 times fewer labels when using uncertainty-based active learning, while non-pretrained models see no or even negative benefit. We find these gains come from an ability to select examples with attributes that disambiguate the intended behavior, such as rare product categories or atypical backgrounds. These attributes are far more linearly separable in pretrained model's representation spaces vs non-pretrained models, suggesting a possible mechanism for this behavior. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2204.08491v1-abstract-full').style.display = 'none'; document.getElementById('2204.08491v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 18 April, 2022; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> April 2022. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2202.12312">arXiv:2202.12312</a> <span> [<a href="https://arxiv.org/pdf/2202.12312">pdf</a>, <a href="https://arxiv.org/format/2202.12312">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> </div> </div> <p class="title is-5 mathjax"> Oolong: Investigating What Makes Transfer Learning Hard with Controlled Studies </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Wu%2C+Z">Zhengxuan Wu</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Papadimitriou%2C+I">Isabel Papadimitriou</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2202.12312v2-abstract-short" style="display: inline;"> When we transfer a pretrained language model to a new language, there are many axes of variation that change at once. To disentangle the impact of different factors like syntactic similarity and vocabulary similarity, we propose a set of controlled transfer studies: we systematically transform the language of the GLUE benchmark, altering one axis of crosslingual variation at a time, and then measu… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2202.12312v2-abstract-full').style.display = 'inline'; document.getElementById('2202.12312v2-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2202.12312v2-abstract-full" style="display: none;"> When we transfer a pretrained language model to a new language, there are many axes of variation that change at once. To disentangle the impact of different factors like syntactic similarity and vocabulary similarity, we propose a set of controlled transfer studies: we systematically transform the language of the GLUE benchmark, altering one axis of crosslingual variation at a time, and then measure the resulting drops in a pretrained model's downstream performance. We find that models can largely recover from syntactic-style shifts, but cannot recover from vocabulary misalignment and embedding matrix re-initialization, even with continued pretraining on 15 million tokens. %On the other hand, transferring to a dataset with an unaligned vocabulary is extremely hard to recover from in the low-data regime. Moreover, good-quality tokenizers in the transfer language do not make vocabulary alignment easier. Our experiments provide insights into the factors of cross-lingual transfer that researchers should most focus on when designing language transfer scenarios. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2202.12312v2-abstract-full').style.display = 'none'; document.getElementById('2202.12312v2-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 23 January, 2024; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 24 February, 2022; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> February 2022. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">EMNLP 2023</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2112.05340">arXiv:2112.05340</a> <span> [<a href="https://arxiv.org/pdf/2112.05340">pdf</a>, <a href="https://arxiv.org/format/2112.05340">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computer Vision and Pattern Recognition">cs.CV</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> </div> </div> <p class="title is-5 mathjax"> Tradeoffs Between Contrastive and Supervised Learning: An Empirical Study </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Karthik%2C+A">Ananya Karthik</a>, <a href="/search/cs?searchtype=author&query=Wu%2C+M">Mike Wu</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2112.05340v1-abstract-short" style="display: inline;"> Contrastive learning has made considerable progress in computer vision, outperforming supervised pretraining on a range of downstream datasets. However, is contrastive learning the better choice in all situations? We demonstrate two cases where it is not. First, under sufficiently small pretraining budgets, supervised pretraining on ImageNet consistently outperforms a comparable contrastive model… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2112.05340v1-abstract-full').style.display = 'inline'; document.getElementById('2112.05340v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2112.05340v1-abstract-full" style="display: none;"> Contrastive learning has made considerable progress in computer vision, outperforming supervised pretraining on a range of downstream datasets. However, is contrastive learning the better choice in all situations? We demonstrate two cases where it is not. First, under sufficiently small pretraining budgets, supervised pretraining on ImageNet consistently outperforms a comparable contrastive model on eight diverse image classification datasets. This suggests that the common practice of comparing pretraining approaches at hundreds or thousands of epochs may not produce actionable insights for those with more limited compute budgets. Second, even with larger pretraining budgets we identify tasks where supervised learning prevails, perhaps because the object-centric bias of supervised pretraining makes the model more resilient to common corruptions and spurious foreground-background correlations. These results underscore the need to characterize tradeoffs of different pretraining objectives across a wider range of contexts and training regimes. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2112.05340v1-abstract-full').style.display = 'none'; document.getElementById('2112.05340v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 10 December, 2021; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> December 2021. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">NeurIPS 2021 Workshop: Self-Supervised Learning - Theory and Practice</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2111.12062">arXiv:2111.12062</a> <span> [<a href="https://arxiv.org/pdf/2111.12062">pdf</a>, <a href="https://arxiv.org/format/2111.12062">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computer Vision and Pattern Recognition">cs.CV</span> </div> </div> <p class="title is-5 mathjax"> DABS: A Domain-Agnostic Benchmark for Self-Supervised Learning </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Liu%2C+V">Vincent Liu</a>, <a href="/search/cs?searchtype=author&query=Lu%2C+R">Rongfei Lu</a>, <a href="/search/cs?searchtype=author&query=Fein%2C+D">Daniel Fein</a>, <a href="/search/cs?searchtype=author&query=Schultz%2C+C">Colin Schultz</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2111.12062v2-abstract-short" style="display: inline;"> Self-supervised learning algorithms, including BERT and SimCLR, have enabled significant strides in fields like natural language processing, computer vision, and speech processing. However, these algorithms are domain-specific, meaning that new self-supervised learning algorithms must be developed for each new setting, including myriad healthcare, scientific, and multimodal domains. To catalyze pr… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2111.12062v2-abstract-full').style.display = 'inline'; document.getElementById('2111.12062v2-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2111.12062v2-abstract-full" style="display: none;"> Self-supervised learning algorithms, including BERT and SimCLR, have enabled significant strides in fields like natural language processing, computer vision, and speech processing. However, these algorithms are domain-specific, meaning that new self-supervised learning algorithms must be developed for each new setting, including myriad healthcare, scientific, and multimodal domains. To catalyze progress toward domain-agnostic methods, we introduce DABS: a Domain-Agnostic Benchmark for Self-supervised learning. To perform well on DABS, an algorithm is evaluated on seven diverse domains: natural images, multichannel sensor data, English text, speech recordings, multilingual text, chest x-rays, and images with text descriptions. Each domain contains an unlabeled dataset for pretraining; the model is then is scored based on its downstream performance on a set of labeled tasks in the domain. We also present e-Mix and ShED: two baseline domain-agnostic algorithms; their relatively modest performance demonstrates that significant progress is needed before self-supervised learning is an out-of-the-box solution for arbitrary domains. Code for benchmark datasets and baseline algorithms is available at https://github.com/alextamkin/dabs. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2111.12062v2-abstract-full').style.display = 'none'; document.getElementById('2111.12062v2-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 5 January, 2023; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 23 November, 2021; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> November 2021. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">NeurIPS 2021</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2108.10307">arXiv:2108.10307</a> <span> [<a href="https://arxiv.org/pdf/2108.10307">pdf</a>, <a href="https://arxiv.org/format/2108.10307">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> </div> </div> <p class="title is-5 mathjax"> C5T5: Controllable Generation of Organic Molecules with Transformers </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Rothchild%2C+D">Daniel Rothchild</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Yu%2C+J">Julie Yu</a>, <a href="/search/cs?searchtype=author&query=Misra%2C+U">Ujval Misra</a>, <a href="/search/cs?searchtype=author&query=Gonzalez%2C+J">Joseph Gonzalez</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2108.10307v1-abstract-short" style="display: inline;"> Methods for designing organic materials with desired properties have high potential impact across fields such as medicine, renewable energy, petrochemical engineering, and agriculture. However, using generative modeling to design substances with desired properties is difficult because candidate compounds must satisfy multiple constraints, including synthetic accessibility and other metrics that ar… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2108.10307v1-abstract-full').style.display = 'inline'; document.getElementById('2108.10307v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2108.10307v1-abstract-full" style="display: none;"> Methods for designing organic materials with desired properties have high potential impact across fields such as medicine, renewable energy, petrochemical engineering, and agriculture. However, using generative modeling to design substances with desired properties is difficult because candidate compounds must satisfy multiple constraints, including synthetic accessibility and other metrics that are intuitive to domain experts but challenging to quantify. We propose C5T5, a novel self-supervised pretraining method that enables transformers to make zero-shot select-and-replace edits, altering organic substances towards desired property values. C5T5 operates on IUPAC names -- a standardized molecular representation that intuitively encodes rich structural information for organic chemists but that has been largely ignored by the ML community. Our technique requires no edited molecule pairs to train and only a rough estimate of molecular properties, and it has the potential to model long-range dependencies and symmetric molecular structures more easily than graph-based methods. C5T5 also provides a powerful interface to domain experts: it grants users fine-grained control over the generative process by selecting and replacing IUPAC name fragments, which enables experts to leverage their intuitions about structure-activity relationships. We demonstrate C5T5's effectiveness on four physical properties relevant for drug discovery, showing that it learns successful and chemically intuitive strategies for altering molecules towards desired property values. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2108.10307v1-abstract-full').style.display = 'none'; document.getElementById('2108.10307v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 23 August, 2021; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> August 2021. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2108.07258">arXiv:2108.07258</a> <span> [<a href="https://arxiv.org/pdf/2108.07258">pdf</a>, <a href="https://arxiv.org/format/2108.07258">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computers and Society">cs.CY</span> </div> </div> <p class="title is-5 mathjax"> On the Opportunities and Risks of Foundation Models </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Bommasani%2C+R">Rishi Bommasani</a>, <a href="/search/cs?searchtype=author&query=Hudson%2C+D+A">Drew A. Hudson</a>, <a href="/search/cs?searchtype=author&query=Adeli%2C+E">Ehsan Adeli</a>, <a href="/search/cs?searchtype=author&query=Altman%2C+R">Russ Altman</a>, <a href="/search/cs?searchtype=author&query=Arora%2C+S">Simran Arora</a>, <a href="/search/cs?searchtype=author&query=von+Arx%2C+S">Sydney von Arx</a>, <a href="/search/cs?searchtype=author&query=Bernstein%2C+M+S">Michael S. Bernstein</a>, <a href="/search/cs?searchtype=author&query=Bohg%2C+J">Jeannette Bohg</a>, <a href="/search/cs?searchtype=author&query=Bosselut%2C+A">Antoine Bosselut</a>, <a href="/search/cs?searchtype=author&query=Brunskill%2C+E">Emma Brunskill</a>, <a href="/search/cs?searchtype=author&query=Brynjolfsson%2C+E">Erik Brynjolfsson</a>, <a href="/search/cs?searchtype=author&query=Buch%2C+S">Shyamal Buch</a>, <a href="/search/cs?searchtype=author&query=Card%2C+D">Dallas Card</a>, <a href="/search/cs?searchtype=author&query=Castellon%2C+R">Rodrigo Castellon</a>, <a href="/search/cs?searchtype=author&query=Chatterji%2C+N">Niladri Chatterji</a>, <a href="/search/cs?searchtype=author&query=Chen%2C+A">Annie Chen</a>, <a href="/search/cs?searchtype=author&query=Creel%2C+K">Kathleen Creel</a>, <a href="/search/cs?searchtype=author&query=Davis%2C+J+Q">Jared Quincy Davis</a>, <a href="/search/cs?searchtype=author&query=Demszky%2C+D">Dora Demszky</a>, <a href="/search/cs?searchtype=author&query=Donahue%2C+C">Chris Donahue</a>, <a href="/search/cs?searchtype=author&query=Doumbouya%2C+M">Moussa Doumbouya</a>, <a href="/search/cs?searchtype=author&query=Durmus%2C+E">Esin Durmus</a>, <a href="/search/cs?searchtype=author&query=Ermon%2C+S">Stefano Ermon</a>, <a href="/search/cs?searchtype=author&query=Etchemendy%2C+J">John Etchemendy</a>, <a href="/search/cs?searchtype=author&query=Ethayarajh%2C+K">Kawin Ethayarajh</a> , et al. (89 additional authors not shown) </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2108.07258v3-abstract-short" style="display: inline;"> AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their cap… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2108.07258v3-abstract-full').style.display = 'inline'; document.getElementById('2108.07258v3-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2108.07258v3-abstract-full" style="display: none;"> AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2108.07258v3-abstract-full').style.display = 'none'; document.getElementById('2108.07258v3-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 12 July, 2022; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 16 August, 2021; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> August 2021. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">Authored by the Center for Research on Foundation Models (CRFM) at the Stanford Institute for Human-Centered Artificial Intelligence (HAI). Report page with citation guidelines: https://crfm.stanford.edu/report.html</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2102.02503">arXiv:2102.02503</a> <span> [<a href="https://arxiv.org/pdf/2102.02503">pdf</a>, <a href="https://arxiv.org/ps/2102.02503">ps</a>, <a href="https://arxiv.org/format/2102.02503">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> </div> </div> <p class="title is-5 mathjax"> Understanding the Capabilities, Limitations, and Societal Impact of Large Language Models </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Brundage%2C+M">Miles Brundage</a>, <a href="/search/cs?searchtype=author&query=Clark%2C+J">Jack Clark</a>, <a href="/search/cs?searchtype=author&query=Ganguli%2C+D">Deep Ganguli</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2102.02503v1-abstract-short" style="display: inline;"> On October 14th, 2020, researchers from OpenAI, the Stanford Institute for Human-Centered Artificial Intelligence, and other universities convened to discuss open research questions surrounding GPT-3, the largest publicly-disclosed dense language model at the time. The meeting took place under Chatham House Rules. Discussants came from a variety of research backgrounds including computer science,… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2102.02503v1-abstract-full').style.display = 'inline'; document.getElementById('2102.02503v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2102.02503v1-abstract-full" style="display: none;"> On October 14th, 2020, researchers from OpenAI, the Stanford Institute for Human-Centered Artificial Intelligence, and other universities convened to discuss open research questions surrounding GPT-3, the largest publicly-disclosed dense language model at the time. The meeting took place under Chatham House Rules. Discussants came from a variety of research backgrounds including computer science, linguistics, philosophy, political science, communications, cyber policy, and more. Broadly, the discussion centered around two main questions: 1) What are the technical capabilities and limitations of large language models? 2) What are the societal effects of widespread use of large language models? Here, we provide a detailed summary of the discussion organized by the two themes above. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2102.02503v1-abstract-full').style.display = 'none'; document.getElementById('2102.02503v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 4 February, 2021; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> February 2021. </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2011.04823">arXiv:2011.04823</a> <span> [<a href="https://arxiv.org/pdf/2011.04823">pdf</a>, <a href="https://arxiv.org/format/2011.04823">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> </div> </div> <p class="title is-5 mathjax"> Language Through a Prism: A Spectral Approach for Multiscale Language Representations </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Jurafsky%2C+D">Dan Jurafsky</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2011.04823v1-abstract-short" style="display: inline;"> Language exhibits structure at different scales, ranging from subwords to words, sentences, paragraphs, and documents. To what extent do deep models capture information at these scales, and can we force them to better capture structure across this hierarchy? We approach this question by focusing on individual neurons, analyzing the behavior of their activations at different timescales. We show tha… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2011.04823v1-abstract-full').style.display = 'inline'; document.getElementById('2011.04823v1-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2011.04823v1-abstract-full" style="display: none;"> Language exhibits structure at different scales, ranging from subwords to words, sentences, paragraphs, and documents. To what extent do deep models capture information at these scales, and can we force them to better capture structure across this hierarchy? We approach this question by focusing on individual neurons, analyzing the behavior of their activations at different timescales. We show that signal processing provides a natural framework for separating structure across scales, enabling us to 1) disentangle scale-specific information in existing embeddings and 2) train models to learn more about particular scales. Concretely, we apply spectral filters to the activations of a neuron across an input, producing filtered embeddings that perform well on part of speech tagging (word-level), dialog speech acts classification (utterance-level), or topic classification (document-level), while performing poorly on the other tasks. We also present a prism layer for training models, which uses spectral filters to constrain different neurons to model structure at different scales. Our proposed BERT + Prism model can better predict masked tokens using long-range context and produces multiscale representations that perform better at utterance- and document-level tasks. Our methods are general and readily applicable to other domains besides language, such as images, audio, and video. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2011.04823v1-abstract-full').style.display = 'none'; document.getElementById('2011.04823v1-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 9 November, 2020; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> November 2020. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">NeurIPS 2020</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2010.07432">arXiv:2010.07432</a> <span> [<a href="https://arxiv.org/pdf/2010.07432">pdf</a>, <a href="https://arxiv.org/format/2010.07432">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Computer Vision and Pattern Recognition">cs.CV</span> </div> </div> <p class="title is-5 mathjax"> Viewmaker Networks: Learning Views for Unsupervised Representation Learning </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Wu%2C+M">Mike Wu</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2010.07432v2-abstract-short" style="display: inline;"> Many recent methods for unsupervised representation learning train models to be invariant to different "views," or distorted versions of an input. However, designing these views requires considerable trial and error by human experts, hindering widespread adoption of unsupervised representation learning methods across domains and modalities. To address this, we propose viewmaker networks: generativ… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2010.07432v2-abstract-full').style.display = 'inline'; document.getElementById('2010.07432v2-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2010.07432v2-abstract-full" style="display: none;"> Many recent methods for unsupervised representation learning train models to be invariant to different "views," or distorted versions of an input. However, designing these views requires considerable trial and error by human experts, hindering widespread adoption of unsupervised representation learning methods across domains and modalities. To address this, we propose viewmaker networks: generative models that learn to produce useful views from a given input. Viewmakers are stochastic bounded adversaries: they produce views by generating and then adding an $\ell_p$-bounded perturbation to the input, and are trained adversarially with respect to the main encoder network. Remarkably, when pretraining on CIFAR-10, our learned views enable comparable transfer accuracy to the well-tuned SimCLR augmentations -- despite not including transformations like cropping or color jitter. Furthermore, our learned views significantly outperform baseline augmentations on speech recordings (+9% points, on average) and wearable sensor data (+17% points). Viewmakers can also be combined with handcrafted views: they improve robustness to common image corruptions and can increase transfer performance in cases where handcrafted views are less explored. These results suggest that viewmakers may provide a path towards more general representation learning algorithms -- reducing the domain expertise and effort needed to pretrain on a much wider set of domains. Code is available at https://github.com/alextamkin/viewmaker. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2010.07432v2-abstract-full').style.display = 'none'; document.getElementById('2010.07432v2-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 29 March, 2021; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 14 October, 2020; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> October 2020. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">ICLR 2021</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/2004.14975">arXiv:2004.14975</a> <span> [<a href="https://arxiv.org/pdf/2004.14975">pdf</a>, <a href="https://arxiv.org/format/2004.14975">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Computation and Language">cs.CL</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> </div> </div> <p class="title is-5 mathjax"> Investigating Transferability in Pretrained Language Models </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Singh%2C+T">Trisha Singh</a>, <a href="/search/cs?searchtype=author&query=Giovanardi%2C+D">Davide Giovanardi</a>, <a href="/search/cs?searchtype=author&query=Goodman%2C+N">Noah Goodman</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="2004.14975v2-abstract-short" style="display: inline;"> How does language model pretraining help transfer learning? We consider a simple ablation technique for determining the impact of each pretrained layer on transfer task performance. This method, partial reinitialization, involves replacing different layers of a pretrained model with random weights, then finetuning the entire model on the transfer task and observing the change in performance. This… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2004.14975v2-abstract-full').style.display = 'inline'; document.getElementById('2004.14975v2-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="2004.14975v2-abstract-full" style="display: none;"> How does language model pretraining help transfer learning? We consider a simple ablation technique for determining the impact of each pretrained layer on transfer task performance. This method, partial reinitialization, involves replacing different layers of a pretrained model with random weights, then finetuning the entire model on the transfer task and observing the change in performance. This technique reveals that in BERT, layers with high probing performance on downstream GLUE tasks are neither necessary nor sufficient for high accuracy on those tasks. Furthermore, the benefit of using pretrained parameters for a layer varies dramatically with finetuning dataset size: parameters that provide tremendous performance improvement when data is plentiful may provide negligible benefits in data-scarce settings. These results reveal the complexity of the transfer learning process, highlighting the limitations of methods that operate on frozen models or single data samples. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('2004.14975v2-abstract-full').style.display = 'none'; document.getElementById('2004.14975v2-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 9 November, 2020; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 30 April, 2020; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> April 2020. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Comments:</span> <span class="has-text-grey-dark mathjax">Findings of EMNLP 2020</span> </p> </li> <li class="arxiv-result"> <div class="is-marginless"> <p class="list-title is-inline-block"><a href="https://arxiv.org/abs/1911.01546">arXiv:1911.01546</a> <span> [<a href="https://arxiv.org/pdf/1911.01546">pdf</a>, <a href="https://arxiv.org/format/1911.01546">other</a>] </span> </p> <div class="tags is-inline-block"> <span class="tag is-small is-link tooltip is-tooltip-top" data-tooltip="Machine Learning">cs.LG</span> <span class="tag is-small is-grey tooltip is-tooltip-top" data-tooltip="Artificial Intelligence">cs.AI</span> </div> </div> <p class="title is-5 mathjax"> Being Optimistic to Be Conservative: Quickly Learning a CVaR Policy </p> <p class="authors"> <span class="search-hit">Authors:</span> <a href="/search/cs?searchtype=author&query=Keramati%2C+R">Ramtin Keramati</a>, <a href="/search/cs?searchtype=author&query=Dann%2C+C">Christoph Dann</a>, <a href="/search/cs?searchtype=author&query=Tamkin%2C+A">Alex Tamkin</a>, <a href="/search/cs?searchtype=author&query=Brunskill%2C+E">Emma Brunskill</a> </p> <p class="abstract mathjax"> <span class="has-text-black-bis has-text-weight-semibold">Abstract</span>: <span class="abstract-short has-text-grey-dark mathjax" id="1911.01546v2-abstract-short" style="display: inline;"> While maximizing expected return is the goal in most reinforcement learning approaches, risk-sensitive objectives such as conditional value at risk (CVaR) are more suitable for many high-stakes applications. However, relatively little is known about how to explore to quickly learn policies with good CVaR. In this paper, we present the first algorithm for sample-efficient learning of CVaR-optimal p… <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('1911.01546v2-abstract-full').style.display = 'inline'; document.getElementById('1911.01546v2-abstract-short').style.display = 'none';">▽ More</a> </span> <span class="abstract-full has-text-grey-dark mathjax" id="1911.01546v2-abstract-full" style="display: none;"> While maximizing expected return is the goal in most reinforcement learning approaches, risk-sensitive objectives such as conditional value at risk (CVaR) are more suitable for many high-stakes applications. However, relatively little is known about how to explore to quickly learn policies with good CVaR. In this paper, we present the first algorithm for sample-efficient learning of CVaR-optimal policies in Markov decision processes based on the optimism in the face of uncertainty principle. This method relies on a novel optimistic version of the distributional Bellman operator that moves probability mass from the lower to the upper tail of the return distribution. We prove asymptotic convergence and optimism of this operator for the tabular policy evaluation case. We further demonstrate that our algorithm finds CVaR-optimal policies substantially faster than existing baselines in several simulated environments with discrete and continuous state spaces. <a class="is-size-7" style="white-space: nowrap;" onclick="document.getElementById('1911.01546v2-abstract-full').style.display = 'none'; document.getElementById('1911.01546v2-abstract-short').style.display = 'inline';">△ Less</a> </span> </p> <p class="is-size-7"><span class="has-text-black-bis has-text-weight-semibold">Submitted</span> 2 April, 2020; <span class="has-text-black-bis has-text-weight-semibold">v1</span> submitted 4 November, 2019; <span class="has-text-black-bis has-text-weight-semibold">originally announced</span> November 2019. </p> <p class="comments is-size-7"> <span class="has-text-black-bis has-text-weight-semibold">Journal ref:</span> Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI 2020) </p> </li> </ol> <div class="is-hidden-tablet">  <span class="help" style="display: inline-block;"><a href="https://github.com/arXiv/arxiv-search/releases">Search v0.5.6 released 2020-02-24</a>  </span> </div> </div> </main> <footer> <div class="columns is-desktop" role="navigation" aria-label="Secondary">  <div class="column"> <div class="columns"> <div class="column"> <ul class="nav-spaced"> <li><a href="https://info.arxiv.org/about">About</a></li> <li><a href="https://info.arxiv.org/help">Help</a></li> </ul> </div> <div class="column"> <ul class="nav-spaced"> <li> <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 512 512" class="icon filter-black" role="presentation"><title>contact arXiv</title><desc>Click here to contact arXiv</desc><path d="M502.3 190.8c3.9-3.1 9.7-.2 9.7 4.7V400c0 26.5-21.5 48-48 48H48c-26.5 0-48-21.5-48-48V195.6c0-5 5.7-7.8 9.7-4.7 22.4 17.4 52.1 39.5 154.1 113.6 21.1 15.4 56.7 47.8 92.2 47.6 35.7.3 72-32.8 92.3-47.6 102-74.1 131.6-96.3 154-113.7zM256 320c23.2.4 56.6-29.2 73.4-41.4 132.7-96.3 142.8-104.7 173.4-128.7 5.8-4.5 9.2-11.5 9.2-18.9v-19c0-26.5-21.5-48-48-48H48C21.5 64 0 85.5 0 112v19c0 7.4 3.4 14.3 9.2 18.9 30.6 23.9 40.7 32.4 173.4 128.7 16.8 12.2 50.2 41.8 73.4 41.4z"/></svg> <a href="https://info.arxiv.org/help/contact.html"> Contact</a> </li> <li> <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 512 512" class="icon filter-black" role="presentation"><title>subscribe to arXiv mailings</title><desc>Click here to subscribe</desc><path d="M476 3.2L12.5 270.6c-18.1 10.4-15.8 35.6 2.2 43.2L121 358.4l287.3-253.2c5.5-4.9 13.3 2.6 8.6 8.3L176 407v80.5c0 23.6 28.5 32.9 42.5 15.8L282 426l124.6 52.2c14.2 6 30.4-2.9 33-18.2l72-432C515 7.8 493.3-6.8 476 3.2z"/></svg> <a href="https://info.arxiv.org/help/subscribe"> Subscribe</a> </li> </ul> </div> </div> </div>   <div class="column"> <div class="columns"> <div class="column"> <ul class="nav-spaced"> <li><a href="https://info.arxiv.org/help/license/index.html">Copyright</a></li> <li><a href="https://info.arxiv.org/help/policies/privacy_policy.html">Privacy Policy</a></li> </ul> </div> <div class="column sorry-app-links"> <ul class="nav-spaced"> <li><a href="https://info.arxiv.org/help/web_accessibility.html">Web Accessibility Assistance</a></li> <li> <p class="help"> <a class="a11y-main-link" href="https://status.arxiv.org" target="_blank">arXiv Operational Status <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 256 512" class="icon filter-dark_grey" role="presentation"><path d="M224.3 273l-136 136c-9.4 9.4-24.6 9.4-33.9 0l-22.6-22.6c-9.4-9.4-9.4-24.6 0-33.9l96.4-96.4-96.4-96.4c-9.4-9.4-9.4-24.6 0-33.9L54.3 103c9.4-9.4 24.6-9.4 33.9 0l136 136c9.5 9.4 9.5 24.6.1 34z"/></svg></a><br> Get status notifications via <a class="is-link" href="https://subscribe.sorryapp.com/24846f03/email/new" target="_blank"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 512 512" class="icon filter-black" role="presentation"><path d="M502.3 190.8c3.9-3.1 9.7-.2 9.7 4.7V400c0 26.5-21.5 48-48 48H48c-26.5 0-48-21.5-48-48V195.6c0-5 5.7-7.8 9.7-4.7 22.4 17.4 52.1 39.5 154.1 113.6 21.1 15.4 56.7 47.8 92.2 47.6 35.7.3 72-32.8 92.3-47.6 102-74.1 131.6-96.3 154-113.7zM256 320c23.2.4 56.6-29.2 73.4-41.4 132.7-96.3 142.8-104.7 173.4-128.7 5.8-4.5 9.2-11.5 9.2-18.9v-19c0-26.5-21.5-48-48-48H48C21.5 64 0 85.5 0 112v19c0 7.4 3.4 14.3 9.2 18.9 30.6 23.9 40.7 32.4 173.4 128.7 16.8 12.2 50.2 41.8 73.4 41.4z"/></svg>email</a> or <a class="is-link" href="https://subscribe.sorryapp.com/24846f03/slack/new" target="_blank"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 448 512" class="icon filter-black" role="presentation"><path d="M94.12 315.1c0 25.9-21.16 47.06-47.06 47.06S0 341 0 315.1c0-25.9 21.16-47.06 47.06-47.06h47.06v47.06zm23.72 0c0-25.9 21.16-47.06 47.06-47.06s47.06 21.16 47.06 47.06v117.84c0 25.9-21.16 47.06-47.06 47.06s-47.06-21.16-47.06-47.06V315.1zm47.06-188.98c-25.9 0-47.06-21.16-47.06-47.06S139 32 164.9 32s47.06 21.16 47.06 47.06v47.06H164.9zm0 23.72c25.9 0 47.06 21.16 47.06 47.06s-21.16 47.06-47.06 47.06H47.06C21.16 243.96 0 222.8 0 196.9s21.16-47.06 47.06-47.06H164.9zm188.98 47.06c0-25.9 21.16-47.06 47.06-47.06 25.9 0 47.06 21.16 47.06 47.06s-21.16 47.06-47.06 47.06h-47.06V196.9zm-23.72 0c0 25.9-21.16 47.06-47.06 47.06-25.9 0-47.06-21.16-47.06-47.06V79.06c0-25.9 21.16-47.06 47.06-47.06 25.9 0 47.06 21.16 47.06 47.06V196.9zM283.1 385.88c25.9 0 47.06 21.16 47.06 47.06 0 25.9-21.16 47.06-47.06 47.06-25.9 0-47.06-21.16-47.06-47.06v-47.06h47.06zm0-23.72c-25.9 0-47.06-21.16-47.06-47.06 0-25.9 21.16-47.06 47.06-47.06h117.84c25.9 0 47.06 21.16 47.06 47.06 0 25.9-21.16 47.06-47.06 47.06H283.1z"/></svg>slack</a> </p> </li> </ul> </div> </div> </div>  </div> </footer> <script src="https://static.arxiv.org/static/base/1.0.0a5/js/member_acknowledgement.js"></script> </body> </html>

CINXE.COM

Search | arXiv e-print repository