PageSourceSearch

https://okitouni.github.io/

html okitouni.github.io collected 2026-10-03 09:42:31 UTC 38,844 bytes, 77 lines download raw bytes

1<!doctype html><html lang=en-us><head><meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1"><meta http-equiv=X-UA-Compatible content="IE=edge"><meta name=generator content="Wowchemy 5.0.0-beta.1 for Hugo"><meta name=author content="Ouail Kitouni"><meta name=description content="Ph.D. Student"><link rel=alternate hreflang=en-us href=https://okitouni.github.io/><link rel=preconnect href=https://fonts.gstatic.com crossorigin><meta name=theme-color content="#2962ff"><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/academicons/1.9.0/css/academicons.min.css integrity="sha512-W4yqoT1+8NLkinBLBZko+dFB2ZbHsYLDdr50VElllRcNt2Q4/GSs6u71UHKxB7S6JEMCp5Ve4xjh3eGQl/HRvg==" crossorigin=anonymous><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.14.0/css/all.min.css integrity="sha256-FMvZuGapsJLjouA6k7Eo2lusoAX9i0ShlWFG6qt7SLc=" crossorigin=anonymous><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/fancybox/3.5.7/jquery.fancybox.min.css integrity="sha256-Vzbj7sDDS/woiFS3uNKo8eIuni59rjyNGtXfstRzStA=" crossorigin=anonymous><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/highlight.js/10.2.0/styles/github.min.css crossorigin=anonymous title=hl-light><link rel=stylesheet href=https://cdnjs.cloudflare.com/ajax/libs/highlight.js/10.2.0/styles/dracula.min.css crossorigin=anonymous title=hl-dark disabled>
1<script src=https://cdnjs.cloudflare.com/ajax/libs/lazysizes/5.2.2/lazysizes.min.js integrity="sha512-TmDwFLhg3UA4ZG0Eb4MIyT1O1Mb+Oww5kFG0uHqXsdbyZz9DcvYQhKpGgNkamAI6h2lGGZq2X8ftOJvF/XjTUg==" crossorigin=anonymous async></script>
1<link rel=stylesheet href="https://fonts.googleapis.com/css?family=Montserrat:400,700%7CRoboto:400,400italic,700%7CRoboto+Mono&display=swap"><link rel=stylesheet href=/css/wowchemy.min.f51aa7b8e7cf2863ae003a3e92431343.css>
1<script>(function(e,t,n,s,o){e[s]=e[s]||[],e[s].push({"gtm.start":(new Date).getTime(),event:"gtm.js"});var a=t.getElementsByTagName(n)[0],i=t.createElement(n),r=s!="dataLayer"?"&l="+s:"";i.async=!0,i.src="https://www.googletagmanager.com/gtm.js?id="+o+r,a.parentNode.insertBefore(i,a)})(window,document,"script","dataLayer","G-PSYH3QTWGB")</script>
1<script src=https://identity.netlify.com/v1/netlify-identity-widget.js></script>
1<link rel=alternate href=/index.xml type=application/rss+xml title="Ouail Kitouni"><link rel=manifest href=/index.webmanifest><link rel=icon type=image/png href=/images/icon_hu78093045f097574c5772539a9c4bdc89_10083_32x32_fill_lanczos_center_3.png><link rel=apple-touch-icon type=image/png href=/images/icon_hu78093045f097574c5772539a9c4bdc89_10083_192x192_fill_lanczos_center_3.png><link rel=canonical href=https://okitouni.github.io/><meta property="twitter:card" content="summary"><meta property="og:site_name" content="Ouail Kitouni"><meta property="og:url" content="https://okitouni.github.io/"><meta property="og:title" content="Ouail Kitouni"><meta property="og:description" content="Ph.D. Student"><meta property="og:image" content="https://okitouni.github.io/images/icon_hu78093045f097574c5772539a9c4bdc89_10083_512x512_fill_lanczos_center_3.png"><meta property="twitter:image" content="https://okitouni.github.io/images/icon_hu78093045f097574c5772539a9c4bdc89_10083_512x512_fill_lanczos_center_3.png"><meta property="og:locale" content="en-us"><meta property="og:updated_time" content="2024-06-01T00:00:00+00:00">
1<script type=application/ld+json>{"@context":"https://schema.org","@type":"WebSite","potentialAction":{"@type":"SearchAction","target":"https://okitouni.github.io/?q={search_term_string}","query-input":"required name=search_term_string"},"url":"https://okitouni.github.io/"}</script>
1<title>Ouail Kitouni</title></head><body id=top data-spy=scroll data-offset=70 data-target=#navbar-main class=page-wrapper>
1<script src=/js/wowchemy-init.js></script>
1<aside class=search-results id=search><div class=container><section class=search-header><div class="row no-gutters justify-content-between mb-3"><div class=col-6><h1>Search</h1></div><div class="col-6 col-search-close"><a class=js-search href=#><i class="fas fa-times-circle text-muted" aria-hidden=true></i></a></div></div><div id=search-box><input name=q id=search-query placeholder=Search... autocapitalize=off autocomplete=off autocorrect=off spellcheck=false type=search class=form-control></div></section><section class=section-search-results><div id=search-hits></div></section></div></aside><div class=page-header><nav class="navbar navbar-expand-lg navbar-light compensate-for-scrollbar" id=navbar-main><div class=container><div class="d-none d-lg-inline-flex"><a class=navbar-brand href=/>Ouail Kitouni</a></div><button type=button class=navbar-toggler data-toggle=collapse data-target=#navbar-content aria-controls=navbar aria-expanded=false aria-label="Toggle navigation">
2<span><i class="fas fa-bars"></i></span></button><div class="navbar-brand-mobile-wrapper d-inline-flex d-lg-none"><a class=navbar-brand href=/>Ouail Kitouni</a></div><div class="navbar-collapse main-menu-item collapse justify-content-start" id=navbar-content><ul class="navbar-nav d-md-inline-flex"><li class=nav-item><a class=nav-link href=/#about data-target=#about><span>Home</span></a></li><li class=nav-item><a class=nav-link href=/#posts data-target=#posts><span>Posts</span></a></li><li class=nav-item><a class=nav-link href=/#publications data-target=#publications><span>Publications</span></a></li><li class=nav-item><a class=nav-link href=/#contact data-target=#contact><span>Contact</span></a></li></ul></div><ul class="nav-icons navbar-nav flex-row ml-auto d-flex pl-md-2"><li class=nav-item><a class="nav-link js-search" href=# aria-label=Search><i class="fas fa-search" aria-hidden=true></i></a></li><li class="nav-item dropdown theme-dropdown"><a href=# class=nav-link data-toggle=dropdown aria-haspopup=true aria-label="Display preferences"><i class="fas fa-moon" aria-hidden=true></i></a><div class=dropdown-menu><a href=# class="dropdown-item js-set-theme-light"><span>Light</span>
3</a><a href=# class="dropdown-item js-set-theme-dark"><span>Dark</span>
4</a><a href=# class="dropdown-item js-set-theme-auto"><span>Automatic</span></a></div></li></ul></div></nav></div><div class=page-body><span class="js-widget-page d-none"></span><section id=about class="home-section wg-about"><div class=home-section-bg></div><div class=container><div class=row><div class="col-12 col-lg-4"><div id=profile><img class="avatar avatar-circle" src=/author/ouail-kitouni/avatar_hu2181cbef5d7399608346d2294bd81f2a_2386892_270x270_fill_q75_lanczos_center.jpg alt="Ouail Kitouni"><div class=portrait-title><h2>Ouail Kitouni</h2><h3>Ph.D. Student</h3><h3><a href=https://www.mit.edu/ target=_blank rel=noopener><span>Massachusetts Institute of Technology</span></a></h3></div><ul class=network-icon aria-hidden=true><li><a href=/#contact aria-label=envelope><i class="fas fa-envelope big-icon"></i></a></li><li><a href="https://scholar.google.com/citations?user=ngbMs6sAAAAJ&amp;hl=en&amp;oi=ao" target=_blank rel=noopener aria-label=google-scholar><i class="ai ai-google-scholar big-icon"></i></a></li><li><a href=https://github.com/okitouni target=_blank rel=noopener aria-label=github><i class="fab fa-github big-icon"></i></a></li><li><a href=https://www.linkedin.com/in/ouail-kitouni-645804187/ target=_blank rel=noopener aria-label=linkedin><i class="fab fa-linkedin big-icon"></i></a></li><li><a href=https://x.com/WKitouni target=_blank rel=noopener aria-label=twitter><i class="fab fa-twitter big-icon"></i></a></li><li><a href=/media/resume.pdf aria-label=cv><i class="ai ai-cv big-icon"></i></a></li></ul></div></div><div class="col-12 col-lg-8"><h1>About</h1><p><a href=media/pronounce_name.mp3>[wa-ill kitoonee] 🔊</a></p><p>I am interested in the science of deep learning. Recently, I&rsquo;ve been very excited about topics like reasoning, multi-modal foundation models, and safe and scalable deep learning.
5During my time at Microsoft Research, I worked on developing a knowledge base generative model towards a knowledge-augmented LLM approach to improve interpretability and limit hallucination. At FAIR, I worked on 
5new pre-training objectives to make LLMs more data-efficient (learn more with less) and improve their knowledge storage and planning capabilities.</p><div class=row><div class=col-md-5><h3>Interests</h3><ul class=ul-interests><li>Science of Deep Learning</li><li>(Mechanistic) Interpretability</li><li>Reasoning in Foundation Models</li><li>Safety and Robustness</li></ul></div><div class=col-md-7><h3>Education</h3><ul class="ul-edu fa-ul"><li><i class="fa-li fas fa-graduation-cap"></i><div class=description><p class=course>Interdisciplinary Ph.D. in Physics and Statistics, 2019 - Present</p><p class=institution>Massachusetts Institute of Technology</p></div></li><li><i class="fa-li fas fa-graduation-cap"></i><div class=description><p class=course>BSc in Physics and Mathematics, 2019</p><p class=institution>University of Rochester</p></div></li></ul></div></div></div></div></div></section><section id=experience class="home-section wg-experience"><div class=home-section-bg></div><div class=container><div class=row><div class="col-12 col-lg-4 section-heading"><h1>Experience</h1></div><div class="col-12 col-lg-8"><div class="row experience"><div class="col-auto text-center flex-column d-none d-sm-flex"><div class="row h-50"><div class=col>&nbsp;</div><div class=col>&nbsp;</div></div><div class=m-2><span class="badge badge-pill border">&nbsp;</span></div><div class="row h-50"><div class="col border-right">&nbsp;</div><div class=col>&nbsp;</div></div></div><div class="col py-2"><div class=card><div class=card-body><h4 class="card-title exp-title text-muted mt-0 mb-1">Research Scientist Intern</h4><h4 class="card-title exp-company text-muted my-0"><a href=https://ai.facebook.com target=_blank rel=noopener>Meta AI</a></h4><div class="text-muted exp-meta">Jan 2024 –
6May 2024
7<span class=middot-divider></span>
8<span>NYC, NY</span></div></div></div></div></div><div class="row experience"><div class="col-auto text-center flex-column d-none d-sm-flex"><div class="row h-50"><div class="col border-right">&nbsp;</div><div class=col>&nbsp;</div></div><div class=m-2><span class="badge badge-pill border">&nbsp;</span></div><div class="row h-50"><div class="col border-right">&nbsp;</div><div class=col>&nbsp;</div></div></div><div class="col py-2"><div class=card><div class=card-body><h4 class="card-title exp-title text-muted mt-0 mb-1">Research Intern</h4><h4 class="card-title exp-company text-muted my-0"><a href=https://www.microsoft.com/en-us/research/lab/microsoft-research-cambridge/ target=_blank rel=noopener>Microsoft Research</a></h4><div class="text-muted exp-meta">May 2023 –
9Aug 2023
10<span class=middot-divider></span>
11<span>Cambridge, UK</span></div></div></div></div></div><div class="row experience"><div class="col-auto text-center flex-column d-none d-sm-flex"><div class="row h-50"><div class="col border-right">&nbsp;</div><div class=col>&nbsp;</div></div><div class=m-2><span class="badge badge-pill border">&nbsp;</span></div><div class="row h-50"><div class=col>&nbsp;</div><div class=col>&nbsp;</div></div></div><div class="col py-2"><div class=card><div class=card-body><h4 class="card-title exp-title text-muted mt-0 mb-1">Machine Learning Researcher Intern</h4><h4 class="card-title exp-company text-muted my-0"><a href=https://frontierdevelopmentlab.org target=_blank rel=noopener>NASA/SETI Frontier Development Lab</a></h4><div class="text-muted exp-meta">May 2022 –
12Aug 2022
13<span class=middot-divider></span>
14<span>Mountain View, CA</span></div></div></div></div></div></div></div></div></section><section id=featured class="home-section wg-featured"><div class=home-section-bg></div><div class=container><div class=row><div class="col-12 col-lg-4 section-heading"><h1>Featured Publications</h1></div><div class="col-12 col-lg-8"><div class=card-simple><div class=article-metadata><div><span><a href>Ouail Kitouni</a></span>, <span><a href>Niklas Nolte</a></span>, <span><a href>Diane Bouchacourt</a></span>, <span><a href>Adina Williams</a></span>, <span><a href>Mike Rabbat</a></span>, <span><a href>Mark Ibrahim</a></span></div><span class=article-date>June 2024
15</span><span class=middot-divider></span>
16<span class=pub-publication></span><span class=middot-divider></span>
17<a href=/publication/mlmu/#disqus_thread></a></div><a href=/publication/mlmu/><div class=img-hover-zoom><img src=/publication/mlmu/featured_huf9d75f51e263cfeff1269972301e62f8_151194_4757dba6957d1faef1ea586e5ed7380a.png data-src=/publication/mlmu/featured_huf9d75f51e263cfeff1269972301e62f8_151194_808x455_fill_lanczos_center_3.png class="article-banner lazyload" alt="The Factorization Curse: Which Tokens You Predict Underlie the Reversal Curse and More"></div></a><h3 class="article-title mb-1 mt-3"><a href=/publication/mlmu/>The Factorization Curse: Which Tokens You Predict Underlie the Reversal Curse and More</a></h3><a href=/publication/mlmu/ class=summary-link><div class=article-style><p>Training transformers to predict &ldquo;any-to-any&rdquo;
17 as opposed to just next token solves the reversal curse and can improve planning capabilites.</p></div></a><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://arxiv.org/pdf/2406.05183 target=_blank rel=noopener>ArXiv</a></div></div><div class=card-simple><div class=article-metadata><div><span><a href>Ouail Kitouni</a></span>, <span><a href>Niklas Nolte</a></span>, <span><a href>James Hensman</a></span>, <span><a href>Bhaskar Mitra</a></span></div><span class=article-date>September 2023
18</span><span class=middot-divider></span>
19<span class=pub-publication>ICML2024
20</span><span class=middot-divider></span>
21<a href=/publication/kbgen/#disqus_thread></a></div><a href=/publication/kbgen/><div class=img-hover-zoom><img src=/publication/kbgen/featured_hu9d46a69541736948000adc208bbe8759_1109713_577a96c409f1f8820e4b0e9ed1f7352b.png data-src=/publication/kbgen/featured_hu9d46a69541736948000adc208bbe8759_1109713_808x455_fill_lanczos_center_3.png class="article-banner lazyload" alt="DiSK: Diffusion Model for Structured Knowledge"></div></a><h3 class="article-title mb-1 mt-3"><a href=/publication/kbgen/>DiSK: Diffusion Model for Structured Knowledge</a></h3><a href=/publication/kbgen/ class=summary-link><div class=article-style><p>DiSK is a generative framework for structured (dictionary-like) data that can handle various data types, from numbers to complex hierarchical types. This model excels in tasks like populating missing data and is especially proficient at predicting numerical values. Its potential extends to augmenting language models for better information retrieval and knowledge manipulation.</p></div></a><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://arxiv.org/abs/2312.05253 target=_blank rel=noopener>ArXiv</a></div></div><div class=card-simple><div class=article-metadata><div><span><a href>Ouail Kitouni</a></span>, <span><a href>Niklas Nolte</a></span>, <span><a href>Mike Williams</a></span></div><span class=article-date>August 2022
22</span><span class=middot-divider></span>
23<span class=pub-publication>ICLR2023
24</span><span class=middot-divider></span>
25<a href=/publication/lipnn/#disqus_thread></a></div><a href=/publication/lipnn/><div class=img-hover-zoom><img src=/publication/lipnn/featured_huecc6edbee80d16e39badfbdc5fcf80a5_248906_6fc719a9dbafefebf28eaa54ad0c8b7a.png data-src=/publication/lipnn/featured_huecc6edbee80d16e39badfbdc5fcf80a5_248906_808x455_fill_lanczos_smart1_3.png class="article-banner lazyload" alt="Robust and Provably Monotonic Networks"></div></a><h3 class="article-title mb-1 mt-3"><a href=/publication/lipnn/>Robust and Provably Monotonic Networks</a></h3><a href=/publication/lipnn/ class=summary-link><div class=article-style><p>We develop a novel neural architecture with an exact bound on its Lipschitz constant. The model can be made monotonic in any subset of its features. This inductive bias is especially important for fairness and interpretability considerations.</p></div></a><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://arxiv.org/abs/2112.00038 target=_blank rel=noopener>PDF
26</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://github.com/niklasnolte/MonotOneNorm target=_blank rel=noopener>Code
27</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=/project/monotonic/>Project
28</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://ml4physicalsciences.github.io/2021/files/NeurIPS_ML4PS_2021_86_poster.png target=_blank rel=noopener>Poster
29</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://ml4physicalsciences.github.io/2021/files/NeurIPS_ML4PS_2021_86.pdf target=_blank rel=noopener>NeurIPS Abstract</a></div></div><div class=card-simple><div class=article-metadata><div><span><a href>Ziming Liu</a></span>, <span><a href>Ouail Kitouni</a></span>, <span><a href>Niklas Nolte</a></span>, <span><a href>Eric Michaud</a></span>, <span><a href>Mike Williams</a></span>, <span><a href>Max Tegmark</a></span></div><span class=article-date>June 2022
30</span><span class=middot-divider></span>
31<span class=pub-publication>NeurIPS2022
32</span><span class=middot-divider></span>
33<a href=/publication/grok/#disqus_thread></a></div><a href=/publication/grok/><div class=img-hover-zoom><img src=/publication/grok/featured_hu24213661bb9bf86284a2cfe7b6c6f587_3888027_e46bd311b76dc84682aae3c0673018bb.png data-src=/publication/grok/featured_hu24213661bb9bf86284a2cfe7b6c6f587_3888027_8
3308x455_fill_lanczos_center_3.png class="article-banner lazyload" alt="Towards Understanding Grokking: An Effective Theory of Representation Learning"></div></a><h3 class="article-title mb-1 mt-3"><a href=/publication/grok/>Towards Understanding Grokking: An Effective Theory of Representation Learning</a></h3><a href=/publication/grok/ class=summary-link><div class=article-style><p>This study investigates <em>grokking</em>, a generalization phenomenon first observed in transformer models trained on arithmetic data, using microscopic and macroscopic analyses, revealing four learning phases and a “Goldilocks zone” for optimal representation learning, while emphasizing the value of physics-inspired tools in understanding deep learning.</p></div></a><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://arxiv.org/abs/2205.10343 target=_blank rel=noopener>PDF</a></div></div></div></div></div></section><section id=publications class="home-section wg-pages"><div class=home-section-bg></div><div class=container><div class=row><div class="col-12 col-lg-4 section-heading"><h1>Recent Publications</h1></div><div class="col-12 col-lg-8"><div class="media stream-item"><div class=media-body><h3 class="article-title mb-0 mt-0"><a href=/publication/mode/>Controlling Classifier Bias with Moment Decomposition: A Method to Enhance Searches for Resonances</a></h3><a href=/publication/mode/ class=summary-link><div class=article-style>Moment Decorrelation (MoDe) is a tool designed to ensure that a model&rsquo;s output remains uncorrelated with certain parameters, commonly termed as protected attributes in fairness contexts. Beyond mere decorrelation, MoDe can even shape the output of a model to adopt linear or quadratic relationships with any input/protected attribute.</div></a><div class="stream-meta article-metadata"><div><span><a href>Ouail Kitouni</a></span>, <span><a href>Benjamin Nachman</a></span>, <span><a href>Constantin Weisser</a></span>, <span><a href>Mike Williams</a></span></div></div><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://arxiv.org/abs/2010.09745 target=_blank rel=noopener>PDF
34</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://github.com/okitouni/MoDe target=_blank rel=noopener>Code
35</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=/project/mode/>Project
36</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://ml4physicalsciences.github.io/2020/files/NeurIPS_ML4PS_2020_45_poster.pdf target=_blank rel=noopener>Poster
37</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://youtu.be/ASqP0tcU6Ag target=_blank rel=noopener>Video
38</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href="https://inspirehep.net/literature?sort=mostrecent&amp;size=25&amp;page=1&amp;q=find%20eprint%202010.09745" target=_blank rel=noopener>INSPIRE-HEP
39</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://ml4physicalsciences.github.io/2020/files/NeurIPS_ML4PS_2020_45.pdf target=_blank rel=noopener>NeurIPS Abstract</a></div></div><div class=ml-3><a href=/publication/mode/><img src=/publication/mode/featured_hub2e09730819de9c6d35861ee229dff3f_442346_3f8d51a7bc36ece2ef8d1fc59e022f5e.png data-src=/publication/mode/featured_hub2e09730819de9c6d35861ee229dff3f_442346_150x0_resize_lanczos_3.png alt="Controlling Classifier Bias with Moment Decomposition: A Method to Enhance Searches for Resonances" class=lazyload></a></div></div><div class="media stream-item"><div class=media-body><h3 class="article-title mb-0 mt-0"><a href=/publication/neemo/>NEEMo: Geometric Fitting using Neural Estimation of the Energy Mover's Distance</a></h3><a href=/publication/neemo/ class=summary-link><div class=article-style>NEEMo is a technique to fit arbitrary geometries to an arbitrary collection of points. Much like WGANs, it relies on the neural estimation of the Wasserstein metric through the KR dual formulation.</div></a><div class="stream-meta article-metadata"><div><span><a href>Ouail Kitouni</a></span>, <span><a href>Niklas Nolte</a></span>, <span><a href>Mike Williams</a></span></div></div><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://arxiv.org/abs/2209.15624 target=_blank rel=noopener>PDF
40</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://github.com/okitouni/EnergyMover-Dual/tree/neurips2022 target=_blank rel=noopener>Code
41</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=/files/neemo/NeurIPS2022Poster_NEEMo_green.pdf target=_blank rel=noopener>Poster
42</a><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=/files/neemo/neemoACAT2022.pdf>Slides</a></div></div><div class=ml-3><a href=/publication/neemo/><img src=/publication/neemo/featured_hu985a0df1fc451c90ac39870ef8e1c4c1_698146_c116dcdb37c0a054c2f7e87a96438788.jpg data-src=/publication/neemo/featured_hu985a0df1fc451c90ac39870ef8e1c4c1_698146_150x0_resize_q75_lanczos.jpg alt="NEEMo: Geometric Fitting using Neural Estimation of the Energy Mover's Distance" class=lazyload></a></div></div><div class="media stream-item"><div class=media-body>
42<h3 class="article-title mb-0 mt-0"><a href=/publication/nuclr/>NuCLR: Nuclear Co-Learned Representations</a></h3><a href=/publication/nuclr/ class=summary-link><div class=article-style>What information can we extract from neural networks? Do they tend to learn useful representations that can be interpreted by human experts? In this work, we investigate the learned representations of a model trained on nuclear physics data and show that it captures some of the underlying physical principles.</div></a><div class="stream-meta article-metadata"><div><span><a href>Ouail Kitouni</a></span>, <span><a href>Niklas Nolte</a></span>, <span><a href>Sokratis Trifinopoulos</a></span>, <span><a href>Subhash Kantamneni</a></span>, <span><a href>Mike Williams</a></span></div></div><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://arxiv.org/pdf/2306.06099.pdf target=_blank rel=noopener>PDF</a></div></div><div class=ml-3><a href=/publication/nuclr/><img src=/publication/nuclr/featured_hu94ca90c8879a3886ec7c2e7f30b79dfe_233148_c9c7b75847bc6943827135ebdbb9e233.png data-src=/publication/nuclr/featured_hu94ca90c8879a3886ec7c2e7f30b79dfe_233148_150x0_resize_lanczos_3.png alt="NuCLR: Nuclear Co-Learned Representations" class=lazyload></a></div></div><div class="media stream-item"><div class=media-body><h3 class="article-title mb-0 mt-0"><a href=/publication/csp/>ML Optimization for Concentrated Solar Power Plants</a></h3><a href=/publication/csp/ class=summary-link><div class=article-style>We use Bayesian Optimization to improve dry cooling systems for concentrated solar power plants, making them more cost-competitive.</div></a><div class="stream-meta article-metadata"><div><span><a href>Hansley Narasiah</a></span>, <span><a href>Ouail Kitouni</a></span>, <span><a href>Andrea Scorsoglio</a></span>, <span><a href>Bernd Sturdza</a></span>, <span><a href>Shawn Hatcher</a></span>, <span><a href>Dolores Garcia</a></span>, <span><a href>Matt Kusner</a></span></div></div><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=/files/CSP/CSP-latest.pdf>PDF</a></div></div><div class=ml-3><a href=/publication/csp/><img src=/publication/csp/featured_hu7c54993fc06a3ba4c2bc3e03610748fe_1505297_a1390e9842279bb1b28b617db82c0860.png data-src=/publication/csp/featured_hu7c54993fc06a3ba4c2bc3e03610748fe_1505297_150x0_resize_lanczos_3.png alt="ML Optimization for Concentrated Solar Power Plants" class=lazyload></a></div></div><div class="media stream-item"><div class=media-body><h3 class="article-title mb-0 mt-0"><a href=/publication/nuclr-mi/>From Neurons to Neutrons: A Case Study in Interpretability</a></h3><a href=/publication/nuclr-mi/ class=summary-link><div class=article-style>Transformers trained on nuclear physics data learn representations close to human-derived nuclear theory. Mechanistic interpretability of neural networks can be a path towards new scientific understanding.</div></a><div class="stream-meta article-metadata"><div><span><a href>Ouail Kitouni</a></span>, <span><a href>Niklas Nolte</a></span>, <span><a href>Victor Samuel Perez-Diaz</a></span>, <span><a href>Sokratis Trifinopoulos</a></span>, <span><a href>Mike Williams</a></span></div></div><div class=btn-links><a class="btn btn-outline-primary my-1 mr-1 btn-sm" href=https://arxiv.org/pdf/2405.17425 target=_blank rel=noopener>ArXiv</a></div></div><div class=ml-3><a href=/publication/nuclr-mi/><img src=/publication/nuclr-mi/featured_hud21cbea5bde390e9be37fd30ab5c1412_166402_c7e9ec383d78c45c1faa2698e5e8d1e4.png data-src=/publication/nuclr-mi/featured_hud21cbea5bde390e9be37fd30ab5c1412_166402_150x0_resize_lanczos_3.png alt="From Neurons to Neutrons: A Case Study in Interpretability" class=lazyload></a></div></div></div></div></div></section><section id=projects class="home-section wg-portfolio"><div class=home-section-bg></div><div class=container><div class=row><div class="col-12 col-lg-4 section-heading"><h1>Miscellaneous Projects</h1></div><div class="col-12 col-lg-8"><div class="isotope projects-container js-layout-masonry"><div class="project-card project-item isotope-item js-id-Deep-Learning js-id-Interpretability"><div class=card><a href=/project/monotonic/ class="card-image hover-overlay"><img src=/project/monotonic/featured_hu5637e810043307cf969b29aa983b8ad4_390493_fadd55df5345b660e838728c9a3254ef.png data-src=/project/monotonic/featured_hu5637e810043307cf969b29aa983b8ad4_390493_550x0_resize_lanczos_3.png alt="Monotonic Networks" class="img-responsive lazyload"></a><div class=card-text><h4><a href=/project/monotonic/>Monotonic Networks</a></h4><div class=article-style><p>A small package to make neural networks monotonic in any subset of their inputs (this works for individual neurons, too!).</p></div></div></div></div><div class="project-card project-item isotope-item js-id-Deep-Learning js-id-Interpretability"><div class=card><a href=/project/mode/ class="card-image hover-overlay"><img src=/project/mode/featured_hue0692ddb7b23a88ca258392813010338_318205_32471b6204290f87be050b519e0c856a.png data-src=/project/mode/featured_hue0692ddb7b23a88ca258392813010338_318205_550x0_resize_lanczos_3.png alt="MoDe: Controlling Classifier Bias" class="img-responsive lazyload"></a><div class=card-text><h4><a href=/project/mode/>MoDe: Controlling Classifier Bias</a></h4><div class=article-style><p>A regularization to make neural networks&rsquo; output independent from certain features.</p></div></div></div></div>
42<div class="project-card project-item isotope-item js-id-Quantum-Mechanics"><div class=card><a href=/project/bell/ class="card-image hover-overlay"><img src=/project/bell/featured_hua131cf47b5ad40f7a99be322b1ee2497_825193_9bc8bb4f5b9d161f601327e25d28b23e.png data-src=/project/bell/featured_hua131cf47b5ad40f7a99be322b1ee2497_825193_550x0_resize_lanczos_3.png alt="Bell Inequality Experiment" class="img-responsive lazyload"></a><div class=card-text><h4><a href=/project/bell/>Bell Inequality Experiment</a></h4><div class=article-style><p>An experiment to demonstrate the non-locality of quantum mechanics through the violation of Bell’s Inequality.</p></div></div></div></div></div></div></div></div></section><section id=posts class="home-section wg-pages"><div class=home-section-bg></div><div class=container><div class=row><div class="col-12 col-lg-4 section-heading"><h1>Recent Posts</h1></div><div class="col-12 col-lg-8"><div class="media stream-item"><div class=media-body><h3 class="article-title mb-0 mt-0"><a href=/post/nuclr/>Reverse-engineering a Transformer Trained on Nuclear Physics</a></h3><a href=/post/nuclr/ class=summary-link><div class=article-style>A mechanistic interpretability effort that aims to reverse-engineer NuCLR, which accurately predicts nuclear properties and demonstrates an understanding of basic nuclear theory, to potentially derive new insights in nuclear physics.</div></a><div class="stream-meta article-metadata"><div class=article-metadata><div><span><a href>Ouail Kitouni</a></span></div><span class=article-date>Jun 21, 2023
43</span><span class=middot-divider></span>
44<span class=article-reading-time>1 min read
45</span><span class=middot-divider></span>
46<a href=/post/nuclr/#disqus_thread></a></div></div></div><div class=ml-3><a href=/post/nuclr/><img src=/post/nuclr/featured_hud73dc5689ac9884e625cad2f272d8d5c_3814826_61e103a54c1bf1aeb322593c914e242c.png data-src=/post/nuclr/featured_hud73dc5689ac9884e625cad2f272d8d5c_3814826_150x0_resize_lanczos_3.png alt="Reverse-engineering a Transformer Trained on Nuclear Physics" class=lazyload></a></div></div><div class="media stream-item"><div class=media-body><h3 class="article-title mb-0 mt-0"><a href=/post/physics-of-dl/>The Physics of Deep Learning</a></h3><a href=/post/physics-of-dl/ class=summary-link><div class=article-style>Our theoretical understanding of deep learning is still in its infancy. And as we figure out what questions to even ask, perhaps it would be useful to draw inspiration from the field responsible for some of the most successful theories in science. This is a lazy repository of what we know about deep learning.</div></a><div class="stream-meta article-metadata"><div class=article-metadata><div><span><a href>Ouail Kitouni</a></span></div><span class=article-date>Jun 21, 2023
47</span><span class=middot-divider></span>
48<span class=article-reading-time>2 min read
49</span><span class=middot-divider></span>
50<a href=/post/physics-of-dl/#disqus_thread></a></div></div></div><div class=ml-3><a href=/post/physics-of-dl/><img src=/post/physics-of-dl/featured_huf677ab5284dc3059c291ae31652dfe1d_285849_ddeb672ed67c4bb511eb9a38ff2dd7e7.jpeg data-src=/post/physics-of-dl/featured_huf677ab5284dc3059c291ae31652dfe1d_285849_150x0_resize_q75_lanczos.jpeg alt="The Physics of Deep Learning" class=lazyload></a></div></div><div class="media stream-item"><div class=media-body><h3 class="article-title mb-0 mt-0"><a href=/post/animations/>Misc. Animations and Visualizations</a></h3><a href=/post/animations/ class=summary-link><div class=article-style>A collection of visualization I made in various projects</div></a><div class="stream-meta article-metadata"><div class=article-metadata><div><span><a href>Ouail Kitouni</a></span></div><span class=article-date>Nov 18, 2022
51</span><span class=middot-divider></span>
52<span class=article-reading-time>1 min read
53</span><span class=middot-divider></span>
54<a href=/post/animations/#disqus_thread></a></div></div></div><div class=ml-3><a href=/post/animations/><img src=/post/animations/featured_hu31f009535a5091da19bd7ca03a63d277_1046118_4010c3aef5edc0c00046b1d545188435.png data-src=/post/animations/featured_hu31f009535a5091da19bd7ca03a63d277_1046118_150x0_resize_lanczos_3.png alt="Misc. Animations and Visualizations" class=lazyload></a></div></div><div class="media stream-item"><div class=media-body>
54<h3 class="article-title mb-0 mt-0"><a href=/post/grokking-thoughts/>Some thoughts on Grokking</a></h3><a href=/post/grokking-thoughts/ class=summary-link><div class=article-style>A short blog post on &ldquo;Towards Understanding Grokking: an effective theory of representation learning&rdquo;</div></a><div class="stream-meta article-metadata"><div class=article-metadata><div><span><a href>Ouail Kitouni</a></span></div><span class=article-date>Sep 21, 2022
55</span><span class=middot-divider></span>
56<span class=article-reading-time>2 min read
57</span><span class=middot-divider></span>
58<a href=/post/grokking-thoughts/#disqus_thread></a></div></div></div><div class=ml-3><a href=/post/grokking-thoughts/><img src=/post/grokking-thoughts/featured_hu341cd8cc234cde8e7cdab09f50205554_1363591_21ea0918f28944dd552462c41e4dab36.png data-src=/post/grokking-thoughts/featured_hu341cd8cc234cde8e7cdab09f50205554_1363591_150x0_resize_lanczos_3.png alt="Some thoughts on Grokking" class=lazyload></a></div></div></div></div></div></section><section id=contact class="home-section wg-contact"><div class=home-section-bg></div><div class=container><div class=row><div class="col-12 col-lg-4 section-heading"><h1>Contact</h1></div><div class="col-12 col-lg-8"><div class=mb-3><form name=contact method=POST netlify netlify-honeypot=welcome-bot><div class="form-group form-inline"><label class=sr-only for=inputName>Name</label>
59<input type=text name=name class="form-control w-100" id=inputName placeholder=Name required></div><div class="form-group form-inline"><label class=sr-only for=inputEmail>Email</label>
60<input type=email name=email class="form-control w-100" id=inputEmail placeholder=Email required></div><div class=form-group><label class=sr-only for=inputMessage>Message</label>
61<textarea name=message class=form-control id=inputMessage rows=5 placeholder=Message required></textarea></div><div class=d-none><label>Do not fill this field unless you are a bot: <input name=welcome-bot></label></div><button type=submit class="btn btn-outline-primary px-3 py-2">Send</button></form></div><ul class=fa-ul><li><i class="fa-li fas fa-envelope fa-2x" aria-hidden=true></i>
62<span id=person-email><a href=mailto:[email protected]>[email protected]</a></span></li><li><i class="fa-li fas fa-map-marker fa-2x" aria-hidden=true></i>
63<span id=person-address>MIT, Cambridge, MA 02139</span></li></ul></div></div></div></section></div><div class=page-footer><div class=container><footer class=site-footer><p class=powered-by></p><p class=powered-by>Published with
64<a href=https://wowchemy.com target=_blank rel=noopener>Wowchemy</a> —
65the free, <a href=https://github.com/wowchemy/wowchemy-hugo-modules target=_blank rel=noopener>open source</a> website builder that empowers creators.</p></footer></div></div><div id=modal class="modal fade" role=dialog><div class=modal-dialog><div class=modal-content><div class=modal-header><h5 class=modal-title>Cite</h5><button type=button class=close data-dismiss=modal aria-label=Close>
66<span aria-hidden=true>&#215;</span></button></div><div class=modal-body><pre><code class="tex hljs"></code></pre></div><div class=modal-footer><a class="btn btn-outline-primary my-1 js-copy-cite" href=# target=_blank><i class="fas fa-copy"></i> Copy
67</a><a class="btn btn-outline-primary my-1 js-download-cite" href=# target=_blank><i class="fas fa-download"></i> Download</a><div id=modal-error></div></div></div></div></div>
67<script src=https://cdnjs.cloudflare.com/ajax/libs/jquery/3.5.1/jquery.min.js integrity="sha256-9/aliU8dGd2tb6OSsuzixeV4y/faTqgFtohetphbbj0=" crossorigin=anonymous></script>
67<script src=https://cdnjs.cloudflare.com/ajax/libs/jquery.imagesloaded/4.1.4/imagesloaded.pkgd.min.js integrity="sha256-lqvxZrPLtfffUl2G/e7szqSvPBILGbwmsGE1MKlOi0Q=" crossorigin=anonymous></script>
67<script src=https://cdnjs.cloudflare.com/ajax/libs/jquery.isotope/3.0.6/isotope.pkgd.min.js integrity="sha256-CBrpuqrMhXwcLLUd5tvQ4euBHCdh7wGlDfNz8vbu/iI=" crossorigin=anonymous></script>
67<script src=https://cdnjs.cloudflare.com/ajax/libs/fancybox/3.5.7/jquery.fancybox.min.js integrity="sha256-yt2kYMy0w8AbtF89WXb2P1rfjcP/HTHLT7097U8Y5b8=" crossorigin=anonymous></script>
67<script src=https://cdnjs.cloudflare.com/ajax/libs/instant.page/5.1.0/instantpage.min.js integrity="sha512-1+qUtKoh9XZW7j+6LhRMAyOrgSQKenQ4mluTR+cvxXjP1Z54RxZuzstR/H9kgPXQsVB8IW7DMDFUJpzLjvhGSQ==" crossorigin=anonymous></script>
67<script src=https://cdnjs.cloudflare.com/ajax/libs/highlight.js/10.2.0/highlight.min.js integrity="sha512-TDKKr+IvoqZnPzc3l35hdjpHD0m+b2EC2SrLEgKDRWpxf2rFCxemkgvJ5kfU48ip+Y+m2XVKyOCD85ybtlZDmw==" crossorigin=anonymous></script>
67<script src=https://cdnjs.cloudflare.com/ajax/libs/highlight.js/10.2.0/languages/r.min.js></script>
67<script>const code_highlighting=!0</script>
67<script>const search_config={indexURI:"/index.json",minLength:1,threshold:.3},i18n={no_results:"No results found",placeholder:"Search...",results:"results found"},content_type={post:"Posts",project:"Projects",publication:"Publications",talk:"Talks",slides:"Slides"}</script>
67<script id=search-hit-fuse-template type=text/x-template>
68      <div class="search-hit" id="summary-{{key}}">
69      <div class="search-hit-content">
70        <div class="search-hit-name">
71          <a href="{{relpermalink}}">{{title}}</a>
72          <div class="article-metadata search-hit-type">{{type}}</div>
73          <p class="search-hit-description">{{snippet}}</p>
74        </div>
75      </div>
76      </div>
77    </script>
77<script src=https://cdnjs.cloudflare.com/ajax/libs/fuse.js/3.2.1/fuse.min.js integrity="sha256-VzgmKYmhsGNNN4Ph1kMW+BjoYJM2jV5i4IlFoeZA9XI=" crossorigin=anonymous></script>
77<script src=https://cdnjs.cloudflare.com/ajax/libs/mark.js/8.11.1/jquery.mark.min.js integrity="sha256-4HLtjeVgH0eIB3aZ9mLYF6E8oU5chNdjU6p6rrXpl9U=" crossorigin=anonymous></script>
77<script>window.netlifyIdentity&&window.netlifyIdentity.on("init",e=>{e||window.netlifyIdentity.on("login",()=>{document.location.href="/admin/"})})</script>
77<script id=dsq-count-scr src=https://.disqus.com/count.js async></script>
77<script src=/js/wowchemy.min.c4e8ff713e0be9e3a3dc7a62dd497a3c.js></script>
77</body></html>

Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.