Help
RSS
API
Feed
Maltego
Contact
Domain > bamos.github.io
×
Welcome!
Right click nodes and scroll the mouse to navigate the graph.
×
More information on this domain is in
AlienVault OTX
Is this malicious?
Yes
No
DNS Resolutions
Date
IP Address
2014-10-06
23.235.40.133
(
ClassC
)
2014-12-08
23.235.39.133
(
ClassC
)
2015-02-24
199.27.76.133
(
ClassC
)
2015-04-25
-
2016-07-29
151.101.48.133
(
ClassC
)
2024-10-06
185.199.108.153
(
ClassC
)
Port 80
HTTP/1.1 200 OKConnection: keep-aliveContent-Length: 189520Server: GitHub.comContent-Type: text/html; charsetutf-8permissions-policy: interest-cohort()Last-Modified: Wed, 25 Sep 2024 17:48:57 GMTAccess-Control-Allow-Origin: *ETag: 66f44d09-2e450expires: Sun, 06 Oct 2024 01:31:26 GMTCache-Control: max-age600x-proxy-cache: MISSX-GitHub-Request-Id: 50D4:3E3C20:4486007:46721D1:6701E615Accept-Ranges: bytesAge: 0Date: Sun, 06 Oct 2024 01:21:26 GMTVia: 1.1 varnishX-Served-By: cache-bfi-krnt7300045-BFIX-Cache: MISSX-Cache-Hits: 0X-Timer: S1728177686.188615,VS0,VE118Vary: Accept-EncodingX-Fastly-Request-ID: 50dd5dbc38cbfc4834a658162b611ecc80558a9e !DOCTYPE html>html xmlnshttp://www.w3.org/1999/xhtml xml:langen langen>head> meta charsetutf-8> title>Brandon Amos/title> meta nameauthor contentBrandon Amos /> meta namedescription content /> meta nameviewport contentwidthdevice-width, initial-scale1, maximum-scale1> link relalternate typeapplication/rss+xml href/atom.xml /> link href/vendor/css/bootstrap.min.css relstylesheet> link relstylesheet hrefhttps://cdn.jsdelivr.net/npm/fork-awesome@1.2.0/css/fork-awesome.min.css integritysha256-XoaMnoYC5TH6/+ihMEnospgm0J1PM/nioxbOUdnM8HY crossoriginanonymous> link href/vendor/css/academicons.min.css relstylesheet> link href/vendor/pygments/default.css relstylesheet> link href/css/bamos.css relstylesheet> link href/css/sharingbuttons.css relstylesheet> script src/vendor/js/jquery.min.js>/script> meta nameviewport contentwidthdevice-width, initial-scale1>/head>body> div classnavbar navbar-default navbar-fixed-top> div classcontainer> div classrow> div classcol-md-10 col-md-offset-1> div classnavbar-header> div classnavbar-brand> a href/images/me.jpg>img src/images/me-face.jpg classimg-circle>/img>/a> a href/>Brandon Amos/a> /div> button classnavbar-toggle typebutton data-togglecollapse data-target#navbar-main> span classicon-bar>/span> span classicon-bar>/span> span classicon-bar>/span> /button> /div> div classnavbar-collapse collapse idnavbar-main> ul classnav navbar-nav> li> a href/>About/a> /li> li> a href/blog/>Blog/a> /li> /ul> !-- ul classnav navbar-nav navbar-right stylefont-size: 1.5em> --> !-- li> --> !-- a hrefhttp://github.com/bamos target_blank> --> !-- i classfa fa-lg fa-github>/i>/a> --> !-- /li> --> !-- li> --> !-- a hrefhttp://twitter.com/brandondamos target_blank> --> !-- i classfa fa-lg fa-twitter>/i>/a> --> !-- /li> --> !-- li> --> !-- a hrefhttps://scholar.google.com/citations?userd8gdZR4AAAAJ target_blank> --> !-- i classai ai-google-scholar>/i>/a> --> !-- /li> --> !-- li> --> !-- a href/atom.xml target_blank> --> !-- i classfa fa-rss>/i>/a> --> !-- /li> --> !-- /ul> --> /div> /div> /div> /div> /div>div classcontainer> div classrow> div classcol-md-6 col-md-offset-1 vcenter idxHdr> div stylefont-size: 2em; color: #4582ec; font-weight: bold; padding-bottom: 0.3em;>Brandon Amos/div> div stylefont-size: 1.2em;> Research Scientist /div> div stylefont-size: 1.2em> a hrefhttps://ai.facebook.com/>Meta (FAIR)/a> /div> div stylefont-size: 1.2em> a hrefmailto:bda@meta.com>bda@meta.com/a> /div> br/> div stylepadding: 0.3em; background-color: #4582ec; display: inline-block; border-radius: 4px; font-size: 1.2em;> a hrefdata/cv.pdf target_blank styletext-decoration: none;> i stylecolor: white classfa fa-download>/i> /a> a hrefhttps://github.com/bamos/cv target_blank styletext-decoration: none;> i stylecolor: white classfa fa-code-fork>/i> /a> a hrefdata/cv.pdf target_blank stylecolor: white; text-decoration: none;>CV/a> /div> ul classlist-inline idxIcons stylefont-size: 1.9em; margin-top: 0.5em;> li> a hrefhttp://github.com/bamos target_blank> i classfa fa-fw fa-github>/i>/a> /li> li> a hrefhttp://twitter.com/brandondamos target_blank> i classfa fa-fw>đ/i>/a> /li> !-- li> a relme hrefhttps://sigmoid.social/@bamos target_blank> i classfa fa-fw fa-mastodon>/i>/a> /li> --> li> a hrefhttps://scholar.google.com/citations?userd8gdZR4AAAAJ target_blank> i classai ai-google-scholar>/i>/a> /li> li> a hrefhttp://www.facebook.com/bdamos target_blank> i classfa fa-fw fa-facebook>/i>/a> /li> li> a hrefhttp://www.linkedin.com/in/bdamos target_blank> i classfa fa-fw fa-linkedin>/i>/a> /li> li> a href/atom.xml target_blank> i classfa fa-fw fa-rss>/i>/a> /li> /ul> /div> div classcol-md-2 vcenter idxHdr> img src/images/me/2021-wave.gif styleborder-radius: 20px; margin: 10px; max-width: none; altMe./> /div> /div> div classrow> div classcol-md-12>hr />p alignjustify>I am a research scientist in theb>Fundamental AI Research (FAIR)/b>group atb>Meta/b> in NYC and alsoteach machine learning at b>Cornell Tech/b>.I study foundational topics in b>machine learning/b> andb>optimization/b>, recently involvingreinforcement learning, control, optimal transport, and geometry.My research is on learning systems that understand and interact with our worldand focuses on integrating structural information and domain knowledge intothese systems to represent non-trivial reasoning operations./p>p>br />/p>h2 id-education>i classfa fa-chevron-right>/i> Education/h2>table classtable table-hover> tr> td> span classcvdate>2014 - 2019/span> strong>Ph.D. in Computer Science/strong>, em>Carnegie Mellon University/em> (0.00/0.00) br /> p stylemargin-top:-1em;margin-bottom:0em> br /> Thesis: a hrefhttps://github.com/bamos/thesis target_blank>i>Differentiable Optimization-Based Modeling for Machine Learning/i>/a> br /> Advisor: a hrefhttps://zicokolter.com target_blank>J. Zico Kolter/a> /p> /td> /tr> tr> td> span classcvdate>2011 - 2014/span> strong>B.S. in Computer Science/strong>, em>Virginia Tech/em> (3.99/4.00) br /> /td> /tr>/table>h2 id-previous-positions>i classfa fa-chevron-right>/i> Previous Positions/h2>table classtable table-hover>tr> td stylepadding-right:0;>span classcvdate>2016 - 2019/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Carnegie Mellon University/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://zicokolter.com target_blank>J. Zico Kolter/a> on ML and optimization)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2018/span>p stylemargin: 0>strong>Research Intern/strong>, em>Intel Labs/em>, Santa Claraspan stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttp://vladlen.info/ target_blank>Vladlen Koltun/a> on computer vision)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2017/span>p stylemargin: 0>strong>Research Intern/strong>, em>Google DeepMind/em>, Londonspan stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://scholar.google.com/citations?usernzEluBwAAAAJ target_blank>Nando de Freitas/a> and a hrefhttp://mdenil.com/ target_blank>Misha Denil/a> on RL)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2014 - 2016/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Carnegie Mellon University/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a> on mobile systems)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2014/span>p stylemargin: 0>strong>Research Intern/strong>, em>Adobe Research/em>, San Josespan stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://research.adobe.com/person/david-tompkins/ target_blank>David Tompkins/a> on distributed systems)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2013 - 2014/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Virginia Tech/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://people.cs.vt.edu/~ltw/shortvita.html target_blank>Layne Watson/a> and a hrefhttps://dblp.org/pid/75/8682.html target_blank>David Easterling/a> on optimization)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2012 - 2014/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Virginia Tech/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://www.magnum.io/people/jules.html target_blank>Jules White/a> and a hrefhttps://scholar.google.com/citations?userMRKab9cAAAAJ target_blank>Hamilton Turner/a> on mobile systems)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2012 - 2014/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Virginia Tech/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://www.ssrg.ece.vt.edu/ target_blank>Binoy Ravindran/a> and a hrefhttps://scholar.google.com/citations?userUG5yHRIAAAAJ target_blank>Alastair Murray/a> on compilers)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2013 - 2014/span>p stylemargin: 0>strong>Software Intern/strong>, em>Snowplow/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(Scala development)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2013/span>p stylemargin: 0>strong>Software Intern/strong>, em>Qualcomm/em>, San Diegospan stylecolor:grey;font-size:1.3rem;margin: 0>(Python and C++ development)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2012/span>p stylemargin: 0>strong>Software Intern/strong>, em>Phoenix Integration/em>, Virginiaspan stylecolor:grey;font-size:1.3rem;margin: 0>(C++, C#, and Java development)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2011/span>p stylemargin: 0>strong>Network Administrator Intern/strong>, em>Sunapsys/em>, Virginia/p> /td>/tr>/table>h2 id-honors--awards>i classfa fa-chevron-right>/i> Honors & Awards/h2>table classtable table-hover>tr> td> div stylefloat: right>2022/div> div> a hrefhttps://neurips.cc/Conferences/2022/ProgramCommittee>NeurIPS Top Reviewer/a> /div> /td> !-- td classcol-md-2 styletext-align:right;>2022/td> -->/tr>tr> td> div stylefloat: right>2022/div> div> a hrefhttps://icml.cc/Conferences/2022/Reviewers>ICML Outstanding Reviewer/a> /div> /td> !-- td classcol-md-2 styletext-align:right;>2022/td> -->/tr>tr> td> div stylefloat: right>2019/div> div> a hrefhttps://iclr.cc/Conferences/2019/Awards>ICLR Outstanding Reviewer/a> /div> /td> !-- td classcol-md-2 styletext-align:right;>2019/td> -->/tr>tr> td> div stylefloat: right>2016 - 2019/div> div> NSF Graduate Research Fellowship /div> /td> !-- td classcol-md-2 styletext-align:right;>2016 - 2019/td> -->/tr>tr> td> div stylefloat: right>2011 - 2014/div> div> Nine undergraduate scholarships br />p stylecolor:grey;font-size:1.2rem>Roanoke County Public Schools Engineering,Salem-Roanoke County Chamber of Commerce,Papa Johns,Scottish Rite of Freemasonry,VT Intelligence Community Conter for Academic Excellence,VT Pamplin Leader,VT Benjamin F. Bock, VT Gay B. Shober, VT I. Luck Gravett/p> /div> /td> !-- td classcol-md-2 styletext-align:right;>2011 - 2014/td> -->/tr>/table>h2 id-publications>i classfa fa-chevron-right>/i> Publications/h2>p>a hrefhttps://scholar.google.com/citations?userd8gdZR4AAAAJ>Google Scholar/a>: 8.9k+ citations and an h-index of 37 br />Selected publications I am a primary author on are span stylebackground-color: #ffffd0>highlighted./span>/p>p>br />/p>h2>2024/h2>table classtable table-hover>tr idtr-paulus2024advprompter stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>1./td>td>a hrefhttps://arxiv.org/abs/2404.16873 target_blank>img srcimages/publications/paulus2024advprompter.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2404.16873 target_blank>AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs/a> /em> a hrefjavascript:; onclick$("#abs_paulus2024advprompter").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/advprompter target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?usernjZL5CQAAAAJ target_blank>Anselm Paulus*/a>, a hrefhttps://arman-z.github.io/ target_blank>Arman Zharmagambetov*/a>, a hrefhttps://sites.google.com/view/chuanguo target_blank>Chuan Guo/a>, strong>Brandon Amossup>†/sup>/strong>, and a hrefhttps://yuandong-tian.com/ target_blank>Yuandong Tiansup>†/sup>/a>br />arXiv 2024 br />div idabs_paulus2024advprompter styletext-align: justify; display: none> p>While recently Large Language Models (LLMs) have achieved remarkable successes, they are vulnerable to certain jailbreaking attacks that lead to generation of inappropriate or harmful content. Manual red-teaming requires finding adversarial prompts that cause such jailbreaking, e.g. by appending a suffix to a given instruction, which is inefficient and time-consuming.On the other hand, automatic adversarial prompt generation often leads to semantically meaningless attacks that can easily be detected by perplexity-based filters, may require gradient information from the target LLM, or do not scale well due to time-consuming discrete optimization processes over the token space. In this paper, we present a novel method that uses another LLM, called the AdvPrompter, to generate human-readable adversarial prompts in seconds, approximately 800 times faster than existing optimization-based approaches.We train the AdvPrompter using a novel algorithm that does not require access to the gradients of the target LLM. This process alternates between two steps: (1) generating high-quality target adversarial suffixes by optimizing the AdvPrompter predictions, and (2) low-rank fine-tuning of the AdvPrompter with the generated adversarial suffixes. The trained AdvPrompter generates suffixes that veil the input instruction without changing its meaning, such that the target LLM is lured to give a harmful response. Experimental results on popular open source target LLMs show state-of-the-art results on the AdvBench dataset, that also transfer to closed-source black-box LLM APIs. Further, we demonstrate that by fine-tuning on a synthetic dataset generated by Advprompter, LLMs can be made more robust against jailbreaking attacks while maintaining performance, i.e. high MMLU scores./p> /div>/td>/tr>tr idtr-pooladian2024neural stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>2./td>td>a hrefhttps://arxiv.org/abs/2406.00288 target_blank>img srcimages/publications/pooladian2024neural.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2406.00288 target_blank>Neural Optimal Transport with Lagrangian Costs/a> /em> a hrefjavascript:; onclick$("#abs_pooladian2024neural").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/lagrangian-ot target_blank>code/a> br />a hrefhttps://arampooladian.com/ target_blank>Aram-Alexandre Pooladian/a>, a hrefhttps://cdenrich.github.io/ target_blank>Carles Domingo-Enrich/a>, a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, and strong>Brandon Amos/strong>br />UAI 2024 br />div idabs_pooladian2024neural styletext-align: justify; display: none> p>We investigate the optimal transport problem between probability measures when the underlying cost function is understood to satisfy a least action principle, also known as a Lagrangian cost. These generalizations are useful when connecting observations from a physical system, where the transport dynamics are influenced by the geometry of the system, such as obstacles, (e.g., incorporating barrier functions in the Lagrangian) and allows practitioners to incorporate a priori knowledge of the underlying system such as non-Euclidean geometries (e.g., paths must be circular). Our contributions are of computational interest, where we demonstrate the ability to efficiently compute geodesics and amortize spline-based paths, which has not been done before, even in low dimensional problems. Unlike prior work, we also output the resulting Lagrangian optimal transport map without requiring an ODE solver. We demonstrate the effectiveness of our formulation on low-dimensional examples taken from prior work./p> /div>/td>/tr>tr idtr-sambharya2024learning>td alignright stylepadding-left:0;padding-right:0;>3./td>td>a hrefhttps://arxiv.org/abs/2309.07835 target_blank>img srcimages/publications/sambharya2024learning.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2309.07835 target_blank>Learning to Warm-Start Fixed-Point Optimization Algorithms/a> /em> a hrefjavascript:; onclick$("#abs_sambharya2024learning").toggle()>abs/a> a hrefhttps://github.com/stellatogrp/l2ws target_blank>code/a> br />a hrefhttps://rajivsambharya.github.io/ target_blank>Rajiv Sambharya/a>, a hrefhttps://sites.google.com/view/georgina-hall target_blank>Georgina Hall/a>, strong>Brandon Amos/strong>, and a hrefhttps://stellato.io/ target_blank>Bartolomeo Stellato/a>br />JMLR 2024 br />div idabs_sambharya2024learning styletext-align: justify; display: none> p>We introduce a machine-learning framework to warm-start fixed-point optimization algorithms. Our architecture consists of a neural network mapping problem parameters to warm starts, followed by a predefined number of fixed-point iterations. We propose two loss functions designed to either minimize the fixed-point residual or the distance to a ground truth solution. In this way, the neural network predicts warm starts with the end-to-end goal of minimizing the downstream loss. An important feature of our architecture is its flexibility, in that it can predict a warm start for fixed-point algorithms run for any number of steps, without being limited to the number of steps it has been trained on. We provide PAC-Bayes generalization bounds on unseen data for common classes of fixed-point operators: contractive, linearly convergent, and averaged. Applying this framework to well-known applications in control, statistics, and signal processing, we observe a significant reduction in the number of iterations and solution time required to solve these problems, through learned warm starts./p> /div>/td>/tr>tr idtr-lotfi2024unlocking>td alignright stylepadding-left:0;padding-right:0;>4./td>td>a hrefhttps://arxiv.org/abs/2407.18158 target_blank>img srcimages/publications/lotfi2024unlocking.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2407.18158 target_blank>Unlocking Tokens as Data Points for Generalization Bounds on Larger Language Models/a> /em> a hrefjavascript:; onclick$("#abs_lotfi2024unlocking").toggle()>abs/a>br />a hrefhttps://sanaelotfi.github.io/ target_blank>Sanae Lotfi/a>, a hrefhttps://yilunkuang.github.io/ target_blank>Yilun Kuang/a>, a hrefhttps://mfinzi.github.io/ target_blank>Marc Anton Finzi/a>, strong>Brandon Amos/strong>, a hrefhttps://goldblum.github.io/ target_blank>Micah Goldblum/a>, and a hrefhttps://cims.nyu.edu/~andrewgw/ target_blank>Andrew Gordon Wilson/a>br />NeurIPS 2024 br />div idabs_lotfi2024unlocking styletext-align: justify; display: none> p>Large language models (LLMs) with billions of parameters excel atpredicting the next token in a sequence. Recent workcomputes non-vacuous compression-basedgeneralization bounds for LLMs, but these bounds arevacuous for large models at the billion-parameterscale. Moreover, these bounds are obtained throughrestrictive compression techniques, boundingcompressed models that generate low-qualitytext. Additionally, the tightness of these existingbounds depends on the number of IID documents in atraining set rather than the much larger number ofnon-IID constituent tokens, leaving untappedpotential for tighter bounds. In this work, weinstead use properties of martingales to derivegeneralization bounds that benefit from the vastnumber of tokens in LLM training sets. Since adataset contains far more tokens than documents, ourgeneralization bounds not only tolerate but actuallybenefit from far less restrictive compressionschemes. With Monarch matrices, Kroneckerfactorizations, and post-training quantization, weachieve non-vacuous generalization bounds for LLMsas large as LLaMA2-70B. Unlike previous approaches, our work achieves the first non-vacuous bounds formodels that are deployed in practice and generatehigh-quality text./p> /div>/td>/tr>tr idtr-domingoenrich2024stochastic>td alignright stylepadding-left:0;padding-right:0;>5./td>td>a hrefhttps://arxiv.org/abs/2312.02027 target_blank>img srcimages/publications/domingoenrich2024stochastic.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2312.02027 target_blank>Stochastic Optimal Control Matching/a> /em> a hrefjavascript:; onclick$("#abs_domingoenrich2024stochastic").toggle()>abs/a>br />a hrefhttps://cdenrich.github.io/ target_blank>Carles Domingo-Enrich/a>, a hrefhttps://scholar.google.com/citations?userel5gT4AAAAAJ target_blank>Jiequn Han/a>, strong>Brandon Amos/strong>, a hrefhttps://cims.nyu.edu/~bruna/ target_blank>Joan Bruna/a>, and a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>br />NeurIPS 2024 br />div idabs_domingoenrich2024stochastic styletext-align: justify; display: none> p>Stochastic optimal control, which has the goal of driving the behavior of noisy systems, is broadly applicable in science, engineering and artificial intelligence. Our work introduces Stochastic Optimal Control Matching (SOCM), a novel Iterative Diffusion Optimization (IDO) technique for stochastic optimal control that stems from the same philosophy as the conditional score matching loss for diffusion models. That is, the control is learned via a least squares problem by trying to fit a matching vector field. The training loss, which is closely connected to the cross-entropy loss, is optimized with respect to both the control function and a family of reparameterization matrices which appear in the matching vector field. The optimization with respect to the reparameterization matrices aims at minimizing the variance of the matching vector field. Experimentally, our algorithm achieves lower error than all the existing IDO techniques for stochastic optimal control for four different control settings. The key idea underlying SOCM is the path-wise reparameterization trick, a novel technique that is of independent interest, e.g., for generative modeling./p> /div>/td>/tr>tr idtr-atanackovic2024meta>td alignright stylepadding-left:0;padding-right:0;>6./td>td>a hrefhttps://arxiv.org/abs/2408.14608 target_blank>img srcimages/publications/atanackovic2024meta.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2408.14608 target_blank>Meta Flow Matching: Integrating Vector Fields on the Wasserstein Manifold/a> /em> a hrefjavascript:; onclick$("#abs_atanackovic2024meta").toggle()>abs/a>br />a hrefhttps://lazaratan.github.io/ target_blank>Lazar Atanackovic/a>, a hrefhttps://scholar.google.com/citations?userCblgXekAAAAJ target_blank>Xi Zhang/a>, strong>Brandon Amos/strong>, a hrefhttps://www.cs.mcgill.ca/~blanchem/ target_blank>Mathieu Blanchette/a>, a hrefhttps://scholar.google.ca/citations?userDN3LoTEAAAAJ target_blank>Leo J Lee/a>, a hrefhttps://yoshuabengio.org/profile/ target_blank>Yoshua Bengio/a>, a hrefhttps://www.alextong.net/ target_blank>Alexander Tong/a>, and a hrefhttps://necludov.github.io/ target_blank>Kirill Neklyudov/a>br />ICML GRaM Workshop 2024 br />div idabs_atanackovic2024meta styletext-align: justify; display: none> p>Numerous biological and physical processes can be modeled as systems of interacting entities evolving continuously over time, e.g. the dynamics of communicating cells or physical particles. Learning the dynamics of such systems is essential for predicting the temporal evolution of populations across novel samples and unseen environments. Flow-based models allow for learning these dynamics at the population level - they model the evolution of the entire distribution of samples. However, current flow-based models are limited to a single initial population and a set of predefined conditions which describe different dynamics. We argue that multiple processes in natural sciences have to be represented as vector fields on the Wasserstein manifold of probability densities. That is, the change of the population at any moment in time depends on the population itself due to the interactions between samples. In particular, this is crucial for personalized medicine where the development of diseases and their respective treatment response depends on the microenvironment of cells specific to each patient. We propose Meta Flow Matching (MFM), a practical approach to integrating along these vector fields on the Wasserstein manifold by amortizing the flow model over the initial populations. Namely, we embed the population of samples using a Graph Neural Network (GNN) and use these embeddings to train a Flow Matching model. This gives MFM the ability to generalize over the initial distributions unlike previously proposed methods. We demonstrate the ability of MFM to improve prediction of individual treatment responses on a large scale multi-patient single-cell drug screen dataset./p> /div>/td>/tr>tr idtr-silvestri2024score>td alignright stylepadding-left:0;padding-right:0;>7./td>td>a hrefhttps://arxiv.org/abs/2307.05213 target_blank>img srcimages/publications/silvestri2024score.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2307.05213 target_blank>Score Function Gradient Estimation to Widen the Applicability of Decision-Focused Learning/a> /em> a hrefjavascript:; onclick$("#abs_silvestri2024score").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?useryHEb8eAAAAAJ target_blank>Mattia Silvestri/a>, a hrefhttps://scholar.google.com/citations?usersMtjmx4AAAAJ target_blank>Senne Berden/a>, a hrefhttps://jayantamandi.com/ target_blank>Jayanta Mandi/a>, a hrefhttps://scholar.google.com/citations?usermuyZLrYAAAAJ target_blank>Ali Ä°rfan MahmutoÄulları/a>, strong>Brandon Amos/strong>, a hrefhttps://people.cs.kuleuven.be/~tias.guns/ target_blank>Tias Guns/a>, and a hrefhttps://scholar.google.com/citations?userlJJ6EOMAAAAJ target_blank>Michele Lombardi/a>br />arXiv 2024 br />div idabs_silvestri2024score styletext-align: justify; display: none> p>Many real-world optimization problems contain parameters that are unknown before deployment time, either due to stochasticity or to lack of information (e.g., demand or travel times in delivery problems). A common strategy in such cases is to estimate said parameters via machine learning (ML) models trained to minimize the prediction error, which however is not necessarily aligned with the downstream task-level error. The decision-focused learning (DFL) paradigm overcomes this limitation by training to directly minimize a task loss, e.g. regret. Since the latter has non-informative gradients for combinatorial problems, state-of-the-art DFL methods introduce surrogates and approximations that enable training. But these methods exploit specific assumptions about the problem structures (e.g., convex or linear problems, unknown parameters only in the objective function). We propose an alternative method that makes no such assumptions, it combines stochastic smoothing with score function gradient estimation which works on any task loss. This opens up the use of DFL methods to nonlinear objectives, uncertain parameters in the problem constraints, and even two-stage stochastic optimization. Experiments show that it typically requires more epochs, but that it is on par with specialized methods and performs especially well for the difficult case of problems with uncertainty in the constraints, in terms of solution quality, scalability, or both./p> /div>/td>/tr>/table>h2>2023/h2>table classtable table-hover>tr idtr-amos2023tutorial stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>8./td>td>a hrefhttps://arxiv.org/abs/2202.00665 target_blank>img srcimages/publications/amos2023tutorial.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2202.00665 target_blank>Tutorial on amortized optimization/a> /em> a hrefjavascript:; onclick$("#abs_amos2023tutorial").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/amortized-optimization-tutorial target_blank>code/a> br />strong>Brandon Amos/strong>br />Foundations and Trends in Machine Learning 2023 br />div idabs_amos2023tutorial styletext-align: justify; display: none> p>Optimization is a ubiquitous modeling tool and is often deployedin settings which repeatedly solve similar instancesof the same problem. Amortized optimization methodsuse learning to predict the solutions to problems inthese settings, exploiting the shared structurebetween similar problem instances. These methodshave been crucial in variational inference andreinforcement learning and are capable of solvingoptimization problems many orders of magnitudestimes faster than traditional optimization methodsthat do not use amortization. This tutorial presentsan introduction to the amortized optimizationfoundations behind these advancements and overviewstheir applications in variational inference, sparsecoding, gradient-based meta-learning, control, reinforcement learning, convex optimization, optimaltransport, and deep equilibrium networks./p> /div>/td>/tr>tr idtr-amos2023amortizing stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>9./td>td>a hrefhttps://arxiv.org/abs/2210.12153 target_blank>img srcimages/publications/amos2023amortizing.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2210.12153 target_blank>On amortizing convex conjugates for optimal transport/a> /em> a hrefjavascript:; onclick$("#abs_amos2023amortizing").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/w2ot target_blank>code/a> br />strong>Brandon Amos/strong>br />ICLR 2023 br />div idabs_amos2023amortizing styletext-align: justify; display: none> p>This paper focuses on computing the convex conjugate operation thatarises when solving Euclidean Wasserstein-2 optimaltransport problems. This conjugation, which is alsoreferred to as the Legendre-Fenchel conjugate orc-transform, is considered difficult to compute andin practice, Wasserstein-2 methods are limited bynot being able to exactly conjugate the dualpotentials in continuous space. I show thatcombining amortized approximations to the conjugatewith a solver for fine-tuning is computationallyeasy. This combination significantly improves thequality of transport maps learned for theWasserstein-2 benchmark by Korotin et al. (2021) andis able to model many 2-dimensional couplings andflows considered in the literature./p> /div>/td>/tr>tr idtr-sambharya2023l2a>td alignright stylepadding-left:0;padding-right:0;>10./td>td>a hrefhttps://arxiv.org/abs/2212.08260 target_blank>img srcimages/publications/sambharya2023l2a.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2212.08260 target_blank>End-to-End Learning to Warm-Start for Real-Time Quadratic Optimization/a> /em> a hrefjavascript:; onclick$("#abs_sambharya2023l2a").toggle()>abs/a> a hrefhttps://github.com/stellatogrp/l2ws target_blank>code/a> br />a hrefhttps://rajivsambharya.github.io/ target_blank>Rajiv Sambharya/a>, a hrefhttps://sites.google.com/view/georgina-hall target_blank>Georgina Hall/a>, strong>Brandon Amos/strong>, and a hrefhttps://stellato.io/ target_blank>Bartolomeo Stellato/a>br />L4DC 2023 br />div idabs_sambharya2023l2a styletext-align: justify; display: none> p>First-order methods are widely used to solve convex quadratic programs(QPs) in real-time applications because of their lowper-iteration cost. However, they can suffer fromslow convergence to accurate solutions. In thispaper, we present a framework which learns aneffective warm-start for a popular first-ordermethod in real-time applications, Douglas-Rachford(DR) splitting, across a family of parametricQPs. This framework consists of two modules: afeedforward neural network block, which takes asinput the parameters of the QP and outputs awarm-start, and a block which performs a fixednumber of iterations of DR splitting from thiswarm-start and outputs a candidate solution. A keyfeature of our framework is its ability to doend-to-end learning as we differentiate through theDR iterations. To illustrate the effectiveness ofour method, we provide generalization bounds (basedon Rademacher complexity) that improve with thenumber of training problems and number of iterationssimultaneously. We further apply our method to threereal-time applications and observe that, by learninggood warm-starts, we are able to significantlyreduce the number of iterations required to obtainhigh-quality solutions./p> /div>/td>/tr>tr idtr-amos2023meta stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>11./td>td>a hrefhttps://arxiv.org/abs/2206.05262 target_blank>img srcimages/publications/amos2023meta.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2206.05262 target_blank>Meta Optimal Transport/a> /em> a hrefjavascript:; onclick$("#abs_amos2023meta").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/meta-ot target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, a hrefhttps://giulslu.github.io/ target_blank>Giulia Luise/a>, and a hrefhttps://ievred.github.io target_blank>Ievgen Redko/a>br />ICML 2023 br />div idabs_amos2023meta styletext-align: justify; display: none> p>We study the use of amortized optimization to predict optimaltransport (OT) maps from the input measures, whichwe call Meta OT. This helps repeatedly solve similarOT problems between different measures by leveragingthe knowledge and information present from pastproblems to rapidly predict and solve newproblems. Otherwise, standard methods ignore theknowledge of the past solutions and suboptimallyre-solve each problem from scratch. Meta OT modelssurpass the standard convergence rates oflog-Sinkhorn solvers in the discrete setting andconvex potentials in the continuous setting. Weimprove the computational time of standard OTsolvers by multiple orders of magnitude in discreteand continuous transport settings between images, spherical data, and color palettes./p> /div>/td>/tr>tr idtr-pooladian2023multisample>td alignright stylepadding-left:0;padding-right:0;>12./td>td>a hrefhttps://arxiv.org/abs/2304.14772 target_blank>img srcimages/publications/pooladian2023multisample.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2304.14772 target_blank>Multisample Flow Matching: Straightening Flows with Minibatch Couplings/a> /em> a hrefjavascript:; onclick$("#abs_pooladian2023multisample").toggle()>abs/a>br />a hrefhttps://arampooladian.com/ target_blank>Aram-Alexandre Pooladian/a>, a hrefhttps://helibenhamu.github.io/ target_blank>Heli Ben-Hamu/a>, a hrefhttps://cdenrich.github.io/ target_blank>Carles Domingo-Enrich/a>, strong>Brandon Amos/strong>, a hrefhttps://www.wisdom.weizmann.ac.il/~ylipman/ target_blank>Yaron Lipman/a>, and a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>br />ICML 2023 br />div idabs_pooladian2023multisample styletext-align: justify; display: none> p>Simulation-free methods for training continuous-time generative modelsconstruct probability paths that go between noisedistributions and individual data samples. Recentworks, such as Flow Matching, derived paths that areoptimal for each data sample. However, thesealgorithms rely on independent data and noisesamples, and do not exploit underlying structure inthe data distribution for constructing probabilitypaths. We propose Multisample Flow Matching, a moregeneral framework that uses non-trivial couplingsbetween data and noise samples while satisfying thecorrect marginal constraints. At very small overheadcosts, this generalization allows us to (i) reducegradient variance during training, (ii) obtainstraighter flows for the learned vector field, whichallows us to generate high-quality samples usingfewer function evaluations, and (iii) obtaintransport maps with lower cost in high dimensions, which has applications beyond generativemodeling. Importantly, we do so in a completelysimulation-free manner with a simple minimizationobjective. We show that our proposed methods improvesample consistency on downsampled ImageNet datasets, and lead to better low-cost sample generation./p> /div>/td>/tr>tr idtr-zheng2023semi>td alignright stylepadding-left:0;padding-right:0;>13./td>td>a hrefhttps://arxiv.org/abs/2210.06518 target_blank>img srcimages/publications/zheng2023semi.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2210.06518 target_blank>Semi-Supervised Offline Reinforcement Learning with Action-Free Trajectories/a> /em> a hrefjavascript:; onclick$("#abs_zheng2023semi").toggle()>abs/a>br />a hrefhttps://enosair.github.io/ target_blank>Qinqing Zheng/a>, a hrefhttps://www.mikaelhenaff.com/ target_blank>Mikael Henaff/a>, strong>Brandon Amos/strong>, and a hrefhttps://aditya-grover.github.io/ target_blank>Aditya Grover/a>br />ICML 2023 br />div idabs_zheng2023semi styletext-align: justify; display: none> p>Natural agents can effectively learn from multiple data sources thatdiffer in size, quality, and types ofmeasurements. We study this heterogeneity in thecontext of offline reinforcement learning (RL) byintroducing a new, practically motivatedsemi-supervised setting. Here, an agent has accessto two sets of trajectories: labelled trajectoriescontaining state, action, reward triplets at everytimestep, along with unlabelled trajectories thatcontain only state and reward information. For thissetting, we develop a simple meta-algorithmicpipeline that learns an inverse-dynamics model onthe labelled data to obtain proxy-labels for theunlabelled data, followed by the use of any offlineRL algorithm on the true and proxy-labelledtrajectories. Empirically, we find this simplepipeline to be highly successful - on several D4RLbenchmarks, certain offline RLalgorithms can match the performance of variantstrained on a fully labeled dataset even when welabel only 10% trajectories from the low returnregime. Finally, we perform a large-scale controlledempirical study investigating the interplay ofdata-centric properties of the labelled andunlabelled datasets, with algorithmic design choices(e.g., inverse dynamics, offline RL algorithm) toidentify general trends and best practices fortraining RL agents on semi-supervised offlinedatasets./p> /div>/td>/tr>tr idtr-bansal2023taskmet stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>14./td>td>a hrefhttps://arxiv.org/abs/2312.05250 target_blank>img srcimages/publications/bansal2023taskmet.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2312.05250 target_blank>TaskMet: Task-Driven Metric Learning for Model Learning/a> /em> a hrefjavascript:; onclick$("#abs_bansal2023taskmet").toggle()>abs/a>br />a hrefhttps://dishank-b.github.io/ target_blank>Dishank Bansal/a>, a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, a hrefhttps://www.mustafamukadam.com/ target_blank>Mustafa Mukadam/a>, and strong>Brandon Amos/strong>br />NeurIPS 2023 br />div idabs_bansal2023taskmet styletext-align: justify; display: none> p>Deep learning models are often used with some downstreamtask. Models solely trained to achieve accuratepredictions may struggle to perform well onthe desired downstream tasks. We propose using thetaskâs loss to learn a metric which parameterizes aloss to train the model.This approach does not alterthe optimal prediction model itself, but ratherchanges the model learning to emphasize theinformation important for the downstream task.Thisenables us to achieve the best of both worlds:aprediction model trained in the original predictionspace while also being valuable for the desireddownstream task.We validate our approach throughexperiments conducted in two main settings: 1)decision-focused model learning scenarios involvingportfolio optimization and budget allocation, and2)reinforcement learning in noisy environments withdistracting states./p> /div>/td>/tr>tr idtr-zharmagambetov2023landscape>td alignright stylepadding-left:0;padding-right:0;>15./td>td>a hrefhttps://arxiv.org/abs/2307.08964 target_blank>img srcimages/publications/zharmagambetov2023landscape.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2307.08964 target_blank>Landscape Surrogate: Learning Decision Losses for Mathematical Optimization Under Partial Information/a> /em> a hrefjavascript:; onclick$("#abs_zharmagambetov2023landscape").toggle()>abs/a>br />a hrefhttps://arman-z.github.io/ target_blank>Arman Zharmagambetov/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userTuVq07oAAAAJ target_blank>Aaron Ferber/a>, a hrefhttps://taoanhuang.github.io/ target_blank>Taoan Huang/a>, a hrefhttps://scholar.google.com/citations?user1jjyaBYAAAAJ target_blank>Bistra Dilkina/a>, and a hrefhttps://yuandong-tian.com/ target_blank>Yuandong Tian/a>br />NeurIPS 2023 br />div idabs_zharmagambetov2023landscape styletext-align: justify; display: none> p>Recent works in learning-integrated optimization have shown promise insettings where the optimization problem is onlypartially observed or where general-purposeoptimizers perform poorly without expert tuning. Bylearning an optimizer g to tackle these challengingproblems with f as the objective, the optimizationprocess can be substantially accelerated byleveraging past experience. Training the optimizercan be done with supervision from known optimalsolutions (not always available) or implicitly byoptimizing the compound function f â g , but theimplicit approach is slow and challenging due tofrequent calls to the optimizer and sparsegradients, particularly for combinatorialsolvers. To address these challenges, we proposeusing a smooth and learnable Landscape SurrogateM instead of composing f with g . This surrogate can be computedfaster than g, provides dense and smooth gradientsduring training, can generalize to unseenoptimization problems, and is efficiently learnedvia alternating optimization. We test our approachon both synthetic problems and real-world problems, achieving comparable or superior objective valuescompared to state-of-the-art baselines whilereducing the number of calls to g . Notably, ourapproach outperforms existing methods forcomputationally expensive high-dimensional problems./p> /div>/td>/tr>tr idtr-retchin2023koopman>td alignright stylepadding-left:0;padding-right:0;>16./td>td>a hrefhttps://differentiable.xyz/papers/paper_45.pdf target_blank>img srcimages/publications/retchin2023koopman.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://differentiable.xyz/papers/paper_45.pdf target_blank>Koopman Constrained Policy Optimization: A Koopman operator theoretic method for differentiable optimal control in robotics/a> /em> a hrefjavascript:; onclick$("#abs_retchin2023koopman").toggle()>abs/a>br />a hrefhttps://www.linkedin.com/in/matthew-retchin/ target_blank>Matthew Retchin/a>, strong>Brandon Amos/strong>, a hrefhttps://www.eigensteve.com/ target_blank>Steven Brunton/a>, and a hrefhttps://shurans.github.io/ target_blank>Shuran Song/a>br />ICML Differentiable Almost Everything Workshop 2023 br />div idabs_retchin2023koopman styletext-align: justify; display: none> p>We introduce Koopman Constrained Policy Optimization (KCPO), combining implicitly differentiable model predictivecontrol with a deep Koopman autoencoder for robotlearning in unknown and nonlinear dynamicalsystems. KCPO is a new policy optimization algorithmthat trains neural policies end-to-end with hard boxconstraints on controls. Guaranteed satisfaction ofhard constraints helps ensure the performance andsafety of robots. We perform imitation learning withKCPO to recover expert policies on the SimplePendulum, Cartpole Swing-Up, Reacher, andDifferential Drive environments, outperformingbaseline methods in generalizing toout-of-distribution constraints in most environmentsafter training./p> /div>/td>/tr>/table>h2>2022/h2>table classtable table-hover>tr idtr-fickinger2021crossdomain stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>17./td>td>a hrefhttps://arxiv.org/abs/2110.03684 target_blank>img srcimages/publications/fickinger2021crossdomain.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2110.03684 target_blank>Cross-Domain Imitation Learning via Optimal Transport/a> /em> a hrefjavascript:; onclick$("#abs_fickinger2021crossdomain").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/gwil target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?userbBFN_qwAAAAJ target_blank>Arnaud Fickinger/a>, a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, a hrefhttp://people.eecs.berkeley.edu/~russell/ target_blank>Stuart Russell/a>, and strong>Brandon Amos/strong>br />ICLR 2022 br />div idabs_fickinger2021crossdomain styletext-align: justify; display: none> p>Cross-domain imitation learning studies how to leverage expertdemonstrations of one agent to train an imitationagent with a different embodiment ormorphology. Comparing trajectories and stationarydistributions between the expert and imitationagents is challenging because they live on differentsystems that may not even have the samedimensionality. We propose Gromov-WassersteinImitation Learning (GWIL), a method for cross-domainimitation that uses the Gromov-Wasserstein distanceto align and compare states between the differentspaces of the agents. Our theory formallycharacterizes the scenarios where GWIL preservesoptimality, revealing its possibilities andlimitations. We demonstrate the effectiveness ofGWIL in non-trivial continuous control domainsranging from simple rigid transformation of theexpert domain to arbitrary transformation of thestate-action space./p> /div>/td>/tr>tr idtr-benhamu2022matching>td alignright stylepadding-left:0;padding-right:0;>18./td>td>a hrefhttps://arxiv.org/abs/2207.04711 target_blank>img srcimages/publications/benhamu2022matching.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2207.04711 target_blank>Matching Normalizing Flows and Probability Paths on Manifolds/a> /em> a hrefjavascript:; onclick$("#abs_benhamu2022matching").toggle()>abs/a>br />a hrefhttps://helibenhamu.github.io/ target_blank>Heli Ben-Hamu*/a>, a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen*/a>, a hrefhttps://joeybose.github.io/ target_blank>Joey Bose/a>, strong>Brandon Amos/strong>, a hrefhttps://aditya-grover.github.io/ target_blank>Aditya Grover/a>, a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>, a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, and a hrefhttps://www.wisdom.weizmann.ac.il/~ylipman/ target_blank>Yaron Lipman/a>br />ICML 2022 br />div idabs_benhamu2022matching styletext-align: justify; display: none> p>Continuous Normalizing Flows (CNFs) are a class of generative modelsthat transform a prior distribution to a modeldistribution by solving an ordinary differentialequation (ODE). We propose to train CNFs onmanifolds by minimizing probability path divergence(PPD), a novel family of divergences between theprobability density path generated by the CNF and atarget probability density path. PPD is formulatedusing a logarithmic mass conservation formula whichis a linear first order partial differentialequation relating the log target probabilities andthe CNFâs defining vector field. PPD has several keybenefits over existing methods: it sidesteps theneed to solve an ODE per iteration, readily appliesto manifold data, scales to high dimensions, and iscompatible with a large family of target pathsinterpolating pure noise and data in finitetime. Theoretically, PPD is shown to bound classicalprobability divergences. Empirically, we show thatCNFs learned by minimizing PPD achievestate-of-the-art results in likelihoods and samplequality on existing low-dimensional manifoldbenchmarks, and is the first example of a generativemodel to scale to moderately high dimensionalmanifolds./p> /div>/td>/tr>tr idtr-chen2022semi>td alignright stylepadding-left:0;padding-right:0;>19./td>td>a hrefhttps://arxiv.org/abs/2203.06832 target_blank>img srcimages/publications/chen2022semi.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2203.06832 target_blank>Semi-Discrete Normalizing Flows through Differentiable Tessellation/a> /em> a hrefjavascript:; onclick$("#abs_chen2022semi").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, strong>Brandon Amos/strong>, and a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>br />NeurIPS 2022 br />div idabs_chen2022semi styletext-align: justify; display: none> p>Mapping between discrete and continuous distributions is a difficulttask and many have had to resort to approximate orheuristical approaches. We propose atessellation-based approach that directly learnsquantization boundaries on a continuous space, complete with exact likelihood evaluations. This isdone through constructing normalizing flows onconvex polytopes parameterized through adifferentiable Voronoi tessellation. Using a simplehomeomorphism with an efficient log determinantJacobian, we can then cheaply parameterizedistributions on convex polytopes./p> p>We explore this approach in two application settings, mapping fromdiscrete to continuous and vice versa. Firstly, aVoronoi dequantization allows automatically learningquantization boundaries in a multidimensionalspace. The location of boundaries and distancesbetween regions can encode useful structuralrelations between the quantized discretevalues. Secondly, a Voronoi mixture model hasconstant computation cost for likelihood evaluationregardless of the number of mixturecomponents. Empirically, we show improvements overexisting methods across a range of structured datamodalities, and find that we can achieve asignificant gain from just adding Voronoi mixturesto a baseline model./p> /div>/td>/tr>tr idtr-pineda2022theseus stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>20./td>td>a hrefhttps://arxiv.org/abs/2207.09442 target_blank>img srcimages/publications/pineda2022theseus.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2207.09442 target_blank>Theseus: A Library for Differentiable Nonlinear Optimization/a> /em> a hrefjavascript:; onclick$("#abs_pineda2022theseus").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/theseus target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?userrebEn8oAAAAJ target_blank>Luis Pineda/a>, a hrefhttps://scholar.google.com/citations?user3PJeg1wAAAAJ target_blank>Taosha Fan/a>, a hrefhttps://scholar.google.com/citations?usergpgb4LgAAAAJ target_blank>Maurizio Monge/a>, a hrefhttps://scholar.google.com/citations?userBFWurDEAAAAJ target_blank>Shobha Venkataraman/a>, a hrefhttps://psodhi.github.io/ target_blank>Paloma Sodhi/a>, a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky Chen/a>, a hrefhttps://joeaortiz.github.io/ target_blank>Joseph Ortiz/a>, a hrefhttps://danieldetone.com/ target_blank>Daniel DeTone/a>, a hrefhttps://scholar.google.com/citations?userkeDqjK0AAAAJ target_blank>Austin Wang/a>, a hrefhttps://scholar.google.com/citations?user8orqBsYAAAAJ target_blank>Stuart Anderson/a>, a hrefhttps://www.linkedin.com/in/jing-dong-24b26ab3/ target_blank>Jing Dong/a>, strong>Brandon Amos/strong>, and a hrefhttps://www.mustafamukadam.com/ target_blank>Mustafa Mukadam/a>br />NeurIPS 2022 br />div idabs_pineda2022theseus styletext-align: justify; display: none> p>We present Theseus, an efficient application-agnostic open sourcelibrary for differentiable nonlinear least squares(DNLS) optimization built on PyTorch, providing acommon framework for end-to-end structured learningin robotics and vision. Existing DNLSimplementations are application specific and do notalways incorporate many ingredients important forefficiency. Theseus is application-agnostic, as weillustrate with several example applications thatare built using the same underlying differentiablecomponents, such as second-order optimizers, standard costs functions, and Lie groups. Forefficiency, Theseus incorporates support for sparsesolvers, automatic vectorization, batching, GPUacceleration, and gradient computation with implicitdifferentiation and direct loss minimization. We doextensive performance evaluation in a set ofapplications, demonstrating significant efficiencygains and better scalability when these features areincorporated./p> /div>/td>/tr>tr idtr-vinitsky2022nocturne>td alignright stylepadding-left:0;padding-right:0;>21./td>td>a hrefhttps://arxiv.org/abs/2206.09889 target_blank>img srcimages/publications/vinitsky2022nocturne.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2206.09889 target_blank>Nocturne: a driving benchmark for multi-agent learning/a> /em> a hrefjavascript:; onclick$("#abs_vinitsky2022nocturne").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/nocturne target_blank>code/a> br />a hrefhttps://www.eugenevinitsky.com target_blank>Eugene Vinitsky/a>, a hrefhttps://www.nathanlct.com/about target_blank>Nathan LichtlĂ©/a>, a hrefhttps://www.linkedin.com/in/xiaomeng-yang-356a976b/ target_blank>Xiaomeng Yang/a>, strong>Brandon Amos/strong>, and a hrefhttps://www.jakobfoerster.com/ target_blank>Jakob Foerster/a>br />NeurIPS Datasets and Benchmarks Track 2022 br />div idabs_vinitsky2022nocturne styletext-align: justify; display: none> p>We introduce Nocturne, a new 2D driving simulator forinvestigating multi-agent coordination under partialobservability. The focus of Nocturne is to enableresearch into inference and theory of mind inreal-world multi-agent settings without thecomputational overhead of computer vision andfeature extraction from images. Agents in thissimulator only observe an obstructed view of thescene, mimicking human visual sensingconstraints. Unlike existing benchmarks that arebottlenecked by rendering human-like observationsdirectly using a camera input, Nocturne usesefficient intersection methods to compute avectorized set of visible features in a C++back-end, allowing the simulator to run at 2000+steps-per-second. Using open-source trajectory andmap data, we construct a simulator to load andreplay arbitrary trajectories and scenes fromreal-world driving data. Using this environment, webenchmark reinforcement-learning andimitation-learning agents and demonstrate that theagents are quite far from human-level coordinationability and deviate significantly from the experttrajectories./p> /div>/td>/tr>/table>h2>2021/h2>table classtable table-hover>tr idtr-amos2021modelbased stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>22./td>td>a hrefhttps://arxiv.org/abs/2008.12775 target_blank>img srcimages/publications/amos2021modelbased.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2008.12775 target_blank>On the model-based stochastic value gradient for continuous reinforcement learning/a> /em> a hrefjavascript:; onclick$("#abs_amos2021modelbased").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/svg target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2021.svg.pdf target_blank>slides/a> br />strong>Brandon Amos/strong>, a hrefhttps://samuelstanton.github.io/ target_blank>Samuel Stanton/a>, a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>, and a hrefhttps://cims.nyu.edu/~andrewgw/ target_blank>Andrew Gordon Wilson/a>br />L4DC 2021 (Oral) br />div idabs_amos2021modelbased styletext-align: justify; display: none> p>Model-based reinforcement learning approaches add explicit domainknowledge to agents in hopes of improving thesample-efficiency in comparison to model-freeagents. However, in practice model-based methods areunable to achieve the same asymptotic performance onchallenging continuous control tasks due to thecomplexity of learning and controlling an explicitworld model. In this paper we investigate thestochastic value gradient (SVG), which is awell-known family of methods for controllingcontinuous systems which includes model-basedapproaches that distill a model-based valueexpansion into a model-free policy. We consider avariant of the model-based SVG that scales to largersystems and uses 1) an entropy regularization tohelp with exploration, 2) a learned deterministicworld model to improve the short-horizon valueestimate, and 3) a learned model-free value estimateafter the modelâs rollout. This SVG variationcaptures the model-free soft actor-critic method asan instance when the model rollout horizon is zero, and otherwise uses short-horizon model rollouts toimprove the value estimate for the policy update. Wesurpass the asymptotic performance of othermodel-based methods on the proprioceptive MuJoColocomotion tasks from the OpenAI gym, including ahumanoid. We notably achieve these results with asimple deterministic world model without requiringan ensemble./p> /div>/td>/tr>tr idtr-cohen2021riemannian stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>23./td>td>a hrefhttps://arxiv.org/abs/2106.10272 target_blank>img srcimages/publications/cohen2021riemannian.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2106.10272 target_blank>Riemannian Convex Potential Maps/a> /em> a hrefjavascript:; onclick$("#abs_cohen2021riemannian").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/rcpm target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2021.rcpm.pdf target_blank>slides/a> br />a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen*/a>, strong>Brandon Amos*/strong>, and a hrefhttps://www.wisdom.weizmann.ac.il/~ylipman/ target_blank>Yaron Lipman/a>br />ICML 2021 br />div idabs_cohen2021riemannian styletext-align: justify; display: none> p>Modeling distributions on Riemannian manifolds is a crucialcomponent in understanding non-Euclidean data thatarises, e.g., in physics and geology. The buddingapproaches in this space are limited byrepresentational and computational tradeoffs. Wepropose and study a class of flows that uses convexpotentials from Riemannian optimal transport. Theseare universal and can model distributions on anycompact Riemannian manifold without requiring domainknowledge of the manifold to be integrated into thearchitecture. We demonstrate that these flows canmodel standard distributions on spheres, and tori, on synthetic and geological data./p> /div>/td>/tr>tr idtr-paulus2021comboptnet>td alignright stylepadding-left:0;padding-right:0;>24./td>td>a hrefhttps://arxiv.org/abs/2105.02343 target_blank>img srcimages/publications/paulus2021comboptnet.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2105.02343 target_blank>CombOptNet: Fit the Right NP-Hard Problem by Learning Integer Programming Constraints/a> /em> a hrefjavascript:; onclick$("#abs_paulus2021comboptnet").toggle()>abs/a> a hrefhttps://github.com/martius-lab/CombOptNet target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?usernjZL5CQAAAAJ target_blank>Anselm Paulus/a>, a hrefhttps://mrolinek.github.io/ target_blank>Michal Rolínek/a>, a hrefhttps://scholar.google.com/citations?userhA1rlU4AAAAJ target_blank>Vít Musil/a>, strong>Brandon Amos/strong>, and a hrefhttps://al.is.mpg.de/person/gmartius target_blank>Georg Martius/a>br />ICML 2021 br />div idabs_paulus2021comboptnet styletext-align: justify; display: none> p>Bridging logical and algorithmic reasoning with modern machinelearning techniques is a fundamental challenge withpotentially transformative impact. On thealgorithmic side, many NP-hard problems can beexpressed as integer programs, in which theconstraints play the role of their âcombinatorialspecificationâ. In this work, we aim to integrateinteger programming solvers into neural networkarchitectures as layers capable of learning both thecost terms and the constraints. The resultingend-to-end trainable architectures jointly extractfeatures from raw data and solve a suitable(learned) combinatorial problem withstate-of-the-art integer programming solvers. Wedemonstrate the potential of such layers with anextensive performance analysis on synthetic data andwith a demonstration on a competitive computervision keypoint matching benchmark./p> /div>/td>/tr>tr idtr-fickinger2021scalable>td alignright stylepadding-left:0;padding-right:0;>25./td>td>a hrefhttps://arxiv.org/abs/2109.15316 target_blank>img srcimages/publications/fickinger2021scalable.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2109.15316 target_blank>Scalable Online Planning via Reinforcement Learning Fine-Tuning/a> /em> a hrefjavascript:; onclick$("#abs_fickinger2021scalable").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userbBFN_qwAAAAJ target_blank>Arnaud Fickinger/a>, a hrefhttps://scholar.google.com/citations?usersJwwn54AAAAJ target_blank>Hengyuan Hu/a>, strong>Brandon Amos/strong>, a hrefhttp://people.eecs.berkeley.edu/~russell/ target_blank>Stuart Russell/a>, and a hrefhttps://noambrown.github.io/ target_blank>Noam Brown/a>br />NeurIPS 2021 br />div idabs_fickinger2021scalable styletext-align: justify; display: none> p>Lookahead search has been a critical component of recent AI successes, such as in the games of chess, go, andpoker. However, the search methods used in thesegames, and in many other settings, aretabular. Tabular search methods do not scale wellwith the size of the search space, and this problemis exacerbated by stochasticity and partialobservability. In this work we replace tabularsearch with online model-based fine-tuning of apolicy neural network via reinforcement learning, and show that this approach outperformsstate-of-the-art search algorithms in benchmarksettings. In particular, we use our search algorithmto achieve a new state-of-the-art result inself-play Hanabi, and show the generality of ouralgorithm by also showing that it outperformstabular search in the Atari game Ms. Pacman./p> /div>/td>/tr>tr idtr-cohen2020aligning>td alignright stylepadding-left:0;padding-right:0;>26./td>td>a hrefhttps://arxiv.org/abs/2006.12648 target_blank>img srcimages/publications/cohen2020aligning.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2006.12648 target_blank>Aligning Time Series on Incomparable Spaces/a> /em> a hrefjavascript:; onclick$("#abs_cohen2020aligning").toggle()>abs/a> a hrefhttps://github.com/samcohen16/Aligning-Time-Series target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2021.gdtw.pdf target_blank>slides/a> br />a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, a hrefhttps://giulslu.github.io/ target_blank>Giulia Luise/a>, a hrefhttps://avt.im/ target_blank>Alexander Terenin/a>, strong>Brandon Amos/strong>, and a hrefhttps://www.deisenroth.cc/ target_blank>Marc Peter Deisenroth/a>br />AISTATS 2021 br />div idabs_cohen2020aligning styletext-align: justify; display: none> p>Dynamic time warping (DTW) is a useful method for aligning, comparingand combining time series, but it requires them tolive in comparable spaces. In this work, we considera setting in which time series live on differentspaces without a sensible ground metric, causing DTWto become ill-defined. To alleviate this, we proposeGromov dynamic time warping (GDTW), a distancebetween time series on potentially incomparablespaces that avoids the comparability requirement byinstead considering intra-relational geometry. Wederive a Frank-Wolfe algorithm for computing it anddemonstrate its effectiveness at aligning, combiningand comparing time series living on incomparablespaces. We further propose a smoothed version ofGDTW as a differentiable loss and assess itsproperties in a variety of settings, includingbarycentric averaging, generative modeling andimitation learning./p> /div>/td>/tr>tr idtr-chen2021learning>td alignright stylepadding-left:0;padding-right:0;>27./td>td>a hrefhttps://arxiv.org/abs/2011.03902 target_blank>img srcimages/publications/chen2021learning.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2011.03902 target_blank>Learning Neural Event Functions for Ordinary Differential Equations/a> /em> a hrefjavascript:; onclick$("#abs_chen2021learning").toggle()>abs/a> a hrefhttps://github.com/rtqichen/torchdiffeq target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, strong>Brandon Amos/strong>, and a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>br />ICLR 2021 br />div idabs_chen2021learning styletext-align: justify; display: none> p>The existing Neural ODE formulation relies on an explicitknowledge of the termination time. We extend NeuralODEs to implicitly defined termination criteriamodeled by neural event functions, which can bechained together and differentiated through. NeuralEvent ODEs are capable of modeling discrete(instantaneous) changes in a continuous-time system, without prior knowledge of when these changes shouldoccur or how many such changes should exist. We testour approach in modeling hybrid discrete- andcontinuous- systems such as switching dynamicalsystems and collision in multi-body systems, and wepropose simulation-based training of point processeswith applications in discrete control./p> /div>/td>/tr>tr idtr-chen2021neural>td alignright stylepadding-left:0;padding-right:0;>28./td>td>a hrefhttps://arxiv.org/abs/2011.04583 target_blank>img srcimages/publications/chen2021neural.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2011.04583 target_blank>Neural Spatio-Temporal Point Processes/a> /em> a hrefjavascript:; onclick$("#abs_chen2021neural").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/neural_stpp target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, strong>Brandon Amos/strong>, and a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>br />ICLR 2021 br />div idabs_chen2021neural styletext-align: justify; display: none> p>We propose a new class of parameterizations for spatio-temporalpoint processes which leverage Neural ODEs as acomputational method and enable flexible, high-fidelity models of discrete events that arelocalized in continuous time and space. Central toour approach is a combination of recurrentcontinuous-time neural networks with two novelneural architectures, i.e., Jump and AttentiveContinuous-time Normalizing Flows. This approachallows us to learn complex distributions for boththe spatial and temporal domain and to conditionnon-trivially on the observed event history. Wevalidate our models on data sets from a wide varietyof contexts such as seismology, epidemiology, urbanmobility, and neuroscience./p> /div>/td>/tr>tr idtr-yarats2021improving>td alignright stylepadding-left:0;padding-right:0;>29./td>td>a hrefhttps://arxiv.org/abs/1910.01741 target_blank>img srcimages/publications/yarats2021improving.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1910.01741 target_blank>Improving Sample Efficiency in Model-Free Reinforcement Learning from Images/a> /em> a hrefjavascript:; onclick$("#abs_yarats2021improving").toggle()>abs/a> a hrefhttps://sites.google.com/view/sac-ae target_blank>code/a> br />a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>, a hrefhttps://amyzhang.github.io/ target_blank>Amy Zhang/a>, a hrefhttps://scholar.google.com/citations?userPTS2AOgAAAAJ target_blank>Ilya Kostrikov/a>, strong>Brandon Amos/strong>, a hrefhttps://www.cs.mcgill.ca/~jpineau/ target_blank>Joelle Pineau/a>, and a hrefhttps://scholar.google.com/citations?userGgQ9GEkAAAAJ&h target_blank>Rob Fergus/a>br />AAAI 2021 br />div idabs_yarats2021improving styletext-align: justify; display: none> p>Training an agent to solve control tasks directly fromhigh-dimensional images with model-freereinforcement learning (RL) has provendifficult. The agent needs to learn a latentrepresentation together with a control policy toperform the task. Fitting a high-capacity encoderusing a scarce reward signal is not only sampleinefficient, but also prone to suboptimalconvergence. Two ways to improve sample efficiencyare to extract relevant features for the task anduse off-policy algorithms. We dissect variousapproaches of learning good latent features, andconclude that the image reconstruction loss is theessential ingredient that enables efficient andstable representation learning in image-basedRL. Following these findings, we devise anoff-policy actor-critic algorithm with an auxiliarydecoder that trains end-to-end and matchesstate-of-the-art performance across both model-freeand model-based algorithms on many challengingcontrol tasks. We release our code to encouragefuture research on image-based RL./p> /div>/td>/tr>tr idtr-venkataraman2021neural>td alignright stylepadding-left:0;padding-right:0;>30./td>td>a hrefhttps://arxiv.org/abs/2107.10254 target_blank>img srcimages/publications/venkataraman2021neural.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2107.10254 target_blank>Neural Fixed-Point Acceleration for Convex Optimization/a> /em> a hrefjavascript:; onclick$("#abs_venkataraman2021neural").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/neural-scs target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?userBFWurDEAAAAJ target_blank>Shobha Venkataraman*/a> and strong>Brandon Amos*/strong>br />ICML AutoML Workshop 2021 br />div idabs_venkataraman2021neural styletext-align: justify; display: none> p>Fixed-point iterations are at the heart of numerical computing andare often a computational bottleneck in real-timeapplications that typically need a fast solution ofmoderate accuracy. We present neural fixed-pointacceleration which combines ideas from meta-learningand classical acceleration methods to automaticallylearn to accelerate fixed-point problems that aredrawn from a distribution. We apply our framework toSCS, the state-of-the-art solver for convex coneprogramming, and design models and loss functions toovercome the challenges of learning over unrolledoptimization and acceleration instabilities. Ourwork brings neural acceleration into anyoptimization problem expressible with CVXPY./p> /div>/td>/tr>tr idtr-cohen2021sliced>td alignright stylepadding-left:0;padding-right:0;>31./td>td>a hrefhttps://arxiv.org/abs/2102.07115 target_blank>img srcimages/publications/cohen2021sliced.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2102.07115 target_blank>Sliced Multi-Marginal Optimal Transport/a> /em> a hrefjavascript:; onclick$("#abs_cohen2021sliced").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, a hrefhttps://avt.im/ target_blank>Alexander Terenin/a>, a hrefhttps://scholar.google.com/citations?userjmM-JlIAAAAJ target_blank>Yannik Pitcan/a>, strong>Brandon Amos/strong>, a hrefhttps://www.deisenroth.cc/ target_blank>Marc Peter Deisenroth/a>, and a hrefhttps://scholar.google.co.in/citations?userFPVUA-YAAAAJ target_blank>K S Sesh Kumar/a>br />NeurIPS OTML Workshop 2021 br />div idabs_cohen2021sliced styletext-align: justify; display: none> p>Multi-marginal optimal transport enables one to compare multipleprobability measures, which increasingly findsapplication in multi-task learning problems. Onepractical limitation of multi-marginal transport iscomputational scalability in the number of measures, samples and dimensionality. In this work, we proposea multi-marginal optimal transport paradigm based onrandom one-dimensional projections, whose(generalized) distance we term the slicedmulti-marginal Wasserstein distance. To constructthis distance, we introduce a characterization ofthe one-dimensional multi-marginal Kantorovichproblem and use it to highlight a number ofproperties of the sliced multi-marginal Wassersteindistance. In particular, we show that (i) the slicedmulti-marginal Wasserstein distance is a(generalized) metric that induces the same topologyas the standard Wasserstein distance, (ii) it admitsa dimension-free sample complexity, (iii) it istightly connected with the problem of barycentricaveraging under the sliced-Wasserstein metric. Weconclude by illustrating the sliced multi-marginalWasserstein on multi-task density estimation andmulti-dynamics reinforcement learning problems./p> /div>/td>/tr>tr idtr-richterpowell2021input>td alignright stylepadding-left:0;padding-right:0;>32./td>td>a hrefhttps://arxiv.org/abs/2111.12187 target_blank>img srcimages/publications/richterpowell2021input.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2111.12187 target_blank>Input Convex Gradient Networks/a> /em> a hrefjavascript:; onclick$("#abs_richterpowell2021input").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?hles&userL78pVMMAAAAJ target_blank>Jack Richter-Powell/a>, a hrefhttps://scholar.google.com/citations?userHzf8bu0AAAAJ target_blank>Jonathan Lorraine/a>, and strong>Brandon Amos/strong>br />NeurIPS OTML Workshop 2021 br />div idabs_richterpowell2021input styletext-align: justify; display: none> p>The gradients of convex functions are expressive models of non-trivialvector fields. For example, Brenierâs theorem yieldsthat the optimal transport map between any twomeasures on Euclidean space under the squareddistance is realized as a convex gradient, which isa key insight used in recent generative flowmodels. In this paper, we study how to model convexgradients by integrating a Jacobian-vector productparameterized by a neural network, which we call theInput Convex Gradient Network (ICGN). Wetheoretically study ICGNs and compare them to takingthe gradient of an Input-Convex Neural Network(ICNN), empirically demonstrating that a singlelayer ICGN can fit a toy example better than asingle layer ICNN. Lastly, we explore extensions todeeper networks and connections to constructionsfrom Riemannian geometry./p> /div>/td>/tr>tr idtr-cohen2021imitation>td alignright stylepadding-left:0;padding-right:0;>33./td>td>a hrefhttps://openreview.net/pdf?idXe5MFhFvYGX target_blank>img srcimages/publications/cohen2021imitation.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://openreview.net/pdf?idXe5MFhFvYGX target_blank>Imitation Learning from Pixel Observations for Continuous Control/a> /em> a hrefjavascript:; onclick$("#abs_cohen2021imitation").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, strong>Brandon Amos/strong>, a hrefhttps://www.deisenroth.cc/ target_blank>Marc Peter Deisenroth/a>, a hrefhttps://www.mikaelhenaff.com/ target_blank>Mikael Henaff/a>, a hrefhttps://www.eugenevinitsky.com target_blank>Eugene Vinitsky/a>, and a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>br />NeurIPS DeepRL Workshop 2021 br />div idabs_cohen2021imitation styletext-align: justify; display: none> p>We study imitation learning from visual observations only forcontrolling dynamical systems with continuous statesand actions. This setting is attractive due to thelarge amount of video data available from whichagents could learn from. However, it is challengingdue to i) not observing the actions and ii) thehigh-dimensional visual space. In this setting, weexplore recipes for imitation learning based onadversarial learning and optimal transport. Theserecipes enable us to scale these methods to attainexpert-level performance on visual continuouscontrol tasks in the DeepMind control suite. Weinvestigate the tradeoffs of these approaches andpresent a comprehensive evaluation of the key designchoices. To encourage reproducible research in thisarea, we provide an easy-to-use implementation forbenchmarking visual imitation learning, includingour methods and expert demonstrations./p> /div>/td>/tr>tr idtr-pineda2021mbrl>td alignright stylepadding-left:0;padding-right:0;>34./td>td>a hrefhttps://arxiv.org/abs/2104.10159 target_blank>img srcimages/publications/pineda2021mbrl.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2104.10159 target_blank>MBRL-Lib: A Modular Library for Model-based Reinforcement Learning/a> /em> a hrefjavascript:; onclick$("#abs_pineda2021mbrl").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/mbrl-lib target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?userrebEn8oAAAAJ target_blank>Luis Pineda/a>, strong>Brandon Amos/strong>, a hrefhttps://amyzhang.github.io/ target_blank>Amy Zhang/a>, a hrefhttps://www.natolambert.com/ target_blank>Nathan Lambert/a>, and a hrefhttps://www.robertocalandra.com/about/ target_blank>Roberto Calandra/a>br />arXiv 2021 br />div idabs_pineda2021mbrl styletext-align: justify; display: none> p>Model-based reinforcement learning is a compelling framework fordata-efficient learning of agents that interact withthe world. This family of algorithms has manysubcomponents that need to be carefully selected andtuned. As a result the entry-bar for researchers toapproach the field and to deploy it in real-worldtasks can be daunting. In this paper, we presentMBRL-Lib - a machine learning library formodel-based reinforcement learning in continuousstate-action spaces based on PyTorch. MBRL-Lib isdesigned as a platform for both researchers, toeasily develop, debug and compare new algorithms, and non-expert user, to lower the entry-bar ofdeploying state-of-the-art algorithms./p> /div>/td>/tr>/table>h2>2020/h2>table classtable table-hover>tr idtr-amos2020differentiable stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>35./td>td>a hrefhttps://arxiv.org/abs/1909.12830 target_blank>img srcimages/publications/amos2020differentiable.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1909.12830 target_blank>The Differentiable Cross-Entropy Method/a> /em> a hrefjavascript:; onclick$("#abs_amos2020differentiable").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/dcem target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2020.dcem.pdf target_blank>slides/a> br />strong>Brandon Amos/strong> and a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>br />ICML 2020 br />div idabs_amos2020differentiable styletext-align: justify; display: none> p>We study the Cross-Entropy Method (CEM) for the non-convexoptimization of a continuous and parameterizedobjective function and introduce a differentiablevariant (DCEM) that enables us to differentiate theoutput of CEM with respect to the objectivefunctionâs parameters. In the machine learningsetting this brings CEM inside of the end-to-endlearning pipeline where this has otherwise beenimpossible. We show applications in a syntheticenergy-based structured prediction task and innon-convex continuous control. In the controlsetting we show on the simulated cheetah and walkertasks that we can embed their optimal actionsequences with DCEM and then use policy optimizationto fine-tune components of the controller as a steptowards combining model-based and model-free RL./p> /div>/td>/tr>tr idtr-lambert2020objective>td alignright stylepadding-left:0;padding-right:0;>36./td>td>a hrefhttps://arxiv.org/abs/2002.04523 target_blank>img srcimages/publications/lambert2020objective.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2002.04523 target_blank>Objective Mismatch in Model-based Reinforcement Learning/a> /em> a hrefjavascript:; onclick$("#abs_lambert2020objective").toggle()>abs/a>br />a hrefhttps://www.natolambert.com/ target_blank>Nathan Lambert/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userzSsW478AAAAJ target_blank>Omry Yadan/a>, and a hrefhttps://www.robertocalandra.com/about/ target_blank>Roberto Calandra/a>br />L4DC 2020 br />div idabs_lambert2020objective styletext-align: justify; display: none> p>Model-based reinforcement learning (MBRL) has been shown to be a powerful framework for data-efficiently learning control of continuous tasks. Recent work in MBRL has mostly focused on using more advanced function approximators and planning schemes, with little development of the general framework. In this paper, we identify a fundamental issue of the standard MBRL framework-what we call the objective mismatch issue. Objective mismatch arises when one objective is optimized in the hope that a second, often uncorrelated, metric will also be optimized. In the context of MBRL, we characterize the objective mismatch between training the forward dynamics model with respect to the likelihood of the one-step ahead prediction, and the overall goal of improving performance on a downstream control task. For example, this issue can emerge with the realization that dynamics models effective for a specific task do not necessarily need to be globally accurate, and vice versa globally accurate models might not be sufficiently accurate locally to obtain good control performance on a specific task. In our experiments, we study this objective mismatch issue and demonstrate that the likelihood of one-step ahead predictions is not always correlated with control performance. This observation highlights a critical limitation in the MBRL framework which will require further research to be fully understood and addressed. We propose an initial method to mitigate the mismatch issue by re-weighting dynamics model training. Building on it, we conclude with a discussion about other potential directions of research for addressing this issue./p> /div>/td>/tr>tr idtr-amos2020QNSTOP>td alignright stylepadding-left:0;padding-right:0;>37./td>td>a hrefhttps://vtechworks.lib.vt.edu/bitstream/handle/10919/49672/qnTOMS14.pdf target_blank>img srcimages/publications/amos2020QNSTOP.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://vtechworks.lib.vt.edu/bitstream/handle/10919/49672/qnTOMS14.pdf target_blank>QNSTOP: Quasi-Newton Algorithm for Stochastic Optimization/a> /em> a hrefjavascript:; onclick$("#abs_amos2020QNSTOP").toggle()>abs/a> a hrefhttps://github.com/vtopt/qnstop target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://dblp.org/pid/75/8682.html target_blank>David Easterling/a>, a hrefhttps://people.cs.vt.edu/~ltw/shortvita.html target_blank>Layne T. Watson/a>, a hrefhttps://scholar.google.com/citations?user2I6IgikAAAAJ target_blank>William Thacker/a>, a hrefhttps://dblp.org/pid/142/1258.html target_blank>Brent Castle/a>, and a hrefhttps://mtrosset.pages.iu.edu/ target_blank>Michael Trosset/a>br />ACM TOMS 2020 br />div idabs_amos2020QNSTOP styletext-align: justify; display: none> p>QNSTOP consists of serial and parallel (OpenMP) Fortran 2003 codes for thequasi-Newton stochastic optimization method of Castle and Trosset. Forstochastic problems, convergence theory exists for the particularalgorithmic choices and parameter values used in QNSTOP. Both the paralleldriver subroutine, which offers several parallel decomposition strategies, and the serial driver subroutine can be used for stochastic optimization ordeterministic global optimization, based on an input switch. QNSTOP isparticularly effective for ânoisyâ deterministic problems, using onlyobjective function values. Some performance data for computational systemsbiology problems is given./p> /div>/td>/tr>tr idtr-sercu2020neural>td alignright stylepadding-left:0;padding-right:0;>38./td>td>a hrefhttps://www.biorxiv.org/content/10.1101/2021.04.08.439084v1.abstract target_blank>img srcimages/publications/sercu2020neural.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://www.biorxiv.org/content/10.1101/2021.04.08.439084v1.abstract target_blank>Neural Potts Model/a> /em> a hrefjavascript:; onclick$("#abs_sercu2020neural").toggle()>abs/a>br />a hrefhttps://tom.sercu.me/ target_blank>Tom Sercu/a>, a hrefhttps://dblp.org/pid/296/8930.html target_blank>Robert Verkuil/a>, a hrefhttps://scholar.google.com/citations?user2M0OltAAAAAJ target_blank>Joshua Meier/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userZDjmMuwAAAAJ target_blank>Zeming Lin/a>, a hrefhttps://www.linkedin.com/in/caroline-chen/ target_blank>Caroline Chen/a>, a hrefhttps://www.linkedin.com/in/jasonliu6/ target_blank>Jason Liu/a>, a hrefhttp://yann.lecun.com/ target_blank>Yann LeCun/a>, and a hrefhttps://scholar.google.com/citations?uservqb78-gAAAAJ target_blank>Alexander Rives/a>br />MLCB 2020 br />div idabs_sercu2020neural styletext-align: justify; display: none> p>We propose the Neural Potts Model objective as an amortizedoptimization problem. The objective enables traininga single model with shared parameters to explicitlymodel energy landscapes across multiple proteinfamilies. Given a protein sequence as input, themodel is trained to predict a pairwise couplingmatrix for a Potts model energy function describingthe local evolutionary landscape of thesequence. Couplings can be predicted for novelsequences. A controlled ablation experimentassessing unsupervised contact prediction on sets ofrelated protein families finds a gain fromamortization for low-depth multiple sequencealignments; the result is then confirmed on adatabase with broad coverage of protein sequences./p> /div>/td>/tr>tr idtr-lou2020riemannian>td alignright stylepadding-left:0;padding-right:0;>39./td>td>a hrefhttps://drive.google.com/file/d/1Ewro0Ne1tvK15nHyYopY4wZ59QTVB-1c/view target_blank>img srcimages/publications/lou2020riemannian.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://drive.google.com/file/d/1Ewro0Ne1tvK15nHyYopY4wZ59QTVB-1c/view target_blank>Deep Riemannian Manifold Learning/a> /em> a hrefjavascript:; onclick$("#abs_lou2020riemannian").toggle()>abs/a>br />a hrefhttps://aaronlou.com/ target_blank>Aaron Lou/a>, a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>, and strong>Brandon Amos/strong>br />NeurIPS Geo4dl Workshop 2020 br />div idabs_lou2020riemannian styletext-align: justify; display: none> p>We present a new class of learnable Riemannian manifolds with a metricparameterized by a deep neural network. The core manifold operationsâspecificallythe Riemannian exponential and logarithmic mapsâare solved using approximatenumerical techniques. Input and parameter gradients are computed with anadjoint sensitivity analysis. This enables us to fit geodesics and distances withgradient-based optimization of both on-manifold values and the manifold itself.We demonstrate our methodâs capability to model smooth, flexible metric structuresin graph embedding tasks./p> /div>/td>/tr>/table>h2>2019/h2>table classtable table-hover>tr idtr-amos2019differentiable stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>40./td>td>a hrefhttps://github.com/bamos/thesis/raw/master/bamos_thesis.pdf target_blank>img srcimages/publications/amos2019differentiable.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://github.com/bamos/thesis/raw/master/bamos_thesis.pdf target_blank>Differentiable Optimization-Based Modeling for Machine Learning/a> /em> a hrefjavascript:; onclick$("#abs_amos2019differentiable").toggle()>abs/a> a hrefhttps://github.com/bamos/thesis target_blank>code/a> br />strong>Brandon Amos/strong>br />Ph.D. Thesis 2019 br />div idabs_amos2019differentiable styletext-align: justify; display: none> p>Domain-specific modeling priors and specialized components are becomingincreasingly important to the machine learning field. These components integrate specialized knowledge that we have as humans into model. We argue inthis thesis that optimization methods provide an expressive set of operationsthat should be part of the machine learning practitionerâs modeling toolbox.We present two foundational approaches for optimization-based modeling:1) the OptNet architecture that integrates optimization problems as individuallayers in larger end-to-end trainable deep networks, and 2) the input-convexneural network (ICNN) architecture that helps make inference and learning indeep energy-based models and structured prediction more tractable.We then show how to use the OptNet approach 1) as a way of combiningmodel-free and model-based reinforcement learning and 2) for top-k learningproblems. We conclude by showing how to differentiate cone programs and turnthe cvxpy domain specific language into a differentiable optimization layer thatenables rapid prototyping of the approaches in this thesis./p> /div>/td>/tr>tr idtr-amos2019differentiable3 stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>41./td>td>a hrefhttp://web.stanford.edu/~boyd/papers/pdf/diff_cvxpy.pdf target_blank>img srcimages/publications/amos2019differentiable3.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://web.stanford.edu/~boyd/papers/pdf/diff_cvxpy.pdf target_blank>Differentiable Convex Optimization Layers/a> /em> a hrefjavascript:; onclick$("#abs_amos2019differentiable3").toggle()>abs/a> a hrefhttps://github.com/cvxgrp/cvxpylayers target_blank>code/a> br />a hrefhttps://www.akshayagrawal.com/ target_blank>Akshay Agrawal*/a>, strong>Brandon Amos*/strong>, a hrefhttps://scholar.google.com/citations?userHmCZLyoAAAAJ target_blank>Shane Barratt*/a>, a hrefhttps://web.stanford.edu/~boyd/ target_blank>Stephen Boyd*/a>, a hrefhttps://stevendiamond.me/ target_blank>Steven Diamond*/a>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter*/a>br />NeurIPS 2019 br />div idabs_amos2019differentiable3 styletext-align: justify; display: none> p>Recent work has shown how to embed differentiable optimization problems (that is, problems whose solutions can be backpropagated through) as layers within deep learning architectures. This method provides a useful inductive bias for certain problems, but existing software for differentiable optimization layers is rigid and difficult to apply to new settings. In this paper, we propose an approach to differentiating through disciplined convex programs, a subclass of convex optimization problems used by domain-specific languages (DSLs) for convex optimization. We introduce disciplined parametrized programming, a subset of disciplined convex programming, and we show that every disciplined parametrized program can be represented as the composition of an affine map from parameters to problem data, a solver, and an affine map from the solverâs solution to a solution of the original problem (a new form we refer to as affine-solver-affine form). We then demonstrate how to efficiently differentiate through each of these components, allowing for end-to-end analytical differentiation through the entire convex program. We implement our methodology in version 1.1 of CVXPY, a popular Python-embedded DSL for convex optimization, and additionally implement differentiable layers for disciplined convex programs in PyTorch and TensorFlow 2.0. Our implementation significantly lowers the barrier to using convex optimization problems in differentiable programs. We present applications in linear machine learning models and in stochastic control, and we show that our layer is competitive (in execution time) compared to specialized differentiable solvers from past work./p> /div>/td>/tr>tr idtr-amos2019limited>td alignright stylepadding-left:0;padding-right:0;>42./td>td>a hrefhttps://arxiv.org/abs/1906.08707 target_blank>img srcimages/publications/amos2019limited.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1906.08707 target_blank>The Limited Multi-Label Projection Layer/a> /em> a hrefjavascript:; onclick$("#abs_amos2019limited").toggle()>abs/a> a hrefhttps://github.com/locuslab/lml target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttp://vladlen.info/ target_blank>Vladlen Koltun/a>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />arXiv 2019 br />div idabs_amos2019limited styletext-align: justify; display: none> p>We propose the Limited Multi-Label (LML) projection layer as a newprimitive operation for end-to-end learning systems. The LML layerprovides a probabilistic way of modeling multi-label predictionslimited to having exactly k labels. We derive efficient forward andbackward passes for this layer and show how the layer can be used tooptimize the top-k recall for multi-label tasks with incomplete labelinformation. We evaluate LML layers on top-k CIFAR-100 classificationand scene graph generation. We show that LML layers add a negligibleamount of computational overhead, strictly improve the modelâsrepresentational capacity, and improve accuracy. We also revisit thetruncated top-k entropy method as a competitive baseline for top-kclassification./p> /div>/td>/tr>tr idtr-grefenstette2019generalized>td alignright stylepadding-left:0;padding-right:0;>43./td>td>a hrefhttps://arxiv.org/abs/1910.01727 target_blank>img srcimages/publications/grefenstette2019generalized.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1910.01727 target_blank>Generalized Inner Loop Meta-Learning/a> /em> a hrefjavascript:; onclick$("#abs_grefenstette2019generalized").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/higher target_blank>code/a> br />a hrefhttps://www.egrefen.com/ target_blank>Edward Grefenstette/a>, strong>Brandon Amos/strong>, a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>, a hrefhttps://phumonhtut.me/ target_blank>Phu Mon Htut/a>, a hrefhttps://amolchanov86.github.io/ target_blank>Artem Molchanov/a>, a hrefhttps://fmeier.github.io/ target_blank>Franziska Meier/a>, a hrefhttps://douwekiela.github.io/ target_blank>Douwe Kiela/a>, a hrefhttps://kyunghyuncho.me/ target_blank>Kyunghyun Cho/a>, and a hrefhttps://soumith.ch/ target_blank>Soumith Chintala/a>br />arXiv 2019 br />div idabs_grefenstette2019generalized styletext-align: justify; display: none> p>Many (but not all) approaches self-qualifying as âmeta-learningâ indeep learning and reinforcement learning fit acommon pattern of approximating the solution to anested optimization problem. In this paper, we givea formalization of this shared pattern, which wecall GIMLI, prove its general requirements, andderive a general-purpose algorithm for implementingsimilar approaches. Based on this analysis andalgorithm, we describe a library of our design, higher, which we share with the community to assistand enable future research into these kinds ofmeta-learning approaches. We end the paper byshowcasing the practical applications of thisframework and library through illustrativeexperiments and ablation studies which theyfacilitate./p> /div>/td>/tr>/table>h2>2018/h2>table classtable table-hover>tr idtr-amos2018learning stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>44./td>td>a hrefhttps://arxiv.org/abs/1804.06318 target_blank>img srcimages/publications/amos2018learning.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1804.06318 target_blank>Learning Awareness Models/a> /em> a hrefjavascript:; onclick$("#abs_amos2018learning").toggle()>abs/a>br />strong>Brandon Amos/strong>, a hrefhttps://laurent-dinh.github.io/ target_blank>Laurent Dinh/a>, a hrefhttps://scholar.google.com/citations?userl-HhJaUAAAAJ target_blank>Serkan Cabi/a>, a hrefhttps://dblp.org/pid/188/6045.html target_blank>Thomas Rothörl/a>, a hrefhttps://scholar.google.com/citations?user0Dkf68EAAAAJ target_blank>Sergio Gómez Colmenarejo/a>, a hrefhttps://scholar.google.com/citations?userYfgdfyYAAAAJ target_blank>Alistair Muldal/a>, a hrefhttps://scholar.google.com/citations?usergVFnjOcAAAAJ target_blank>Tom Erez/a>, a hrefhttps://scholar.google.com/citations?userCjOTm_4AAAAJ target_blank>Yuval Tassa/a>, a hrefhttps://scholar.google.com/citations?usernzEluBwAAAAJ target_blank>Nando de Freitas/a>, and a hrefhttps://mdenil.com/ target_blank>Misha Denil/a>br />ICLR 2018 br />div idabs_amos2018learning styletext-align: justify; display: none> p>We consider the setting of an agent with a fixed body interacting with anunknown and uncertain external world. We show that modelstrained to predict proprioceptive information about theagentâs body come to represent objects in the external world.In spite of being trained with only internally availablesignals, these dynamic body models come to represent externalobjects through the necessity of predicting their effects onthe agentâs own body. That is, the model learns holisticpersistent representations of objects in the world, eventhough the only training signals are body signals. Ourdynamics model is able to successfully predict distributionsover 132 sensor readings over 100 steps into the future and wedemonstrate that even when the body is no longer in contactwith an object, the latent variables of the dynamics modelcontinue to represent its shape. We show that active datacollection by maximizing the entropy of predictions about thebody-touch sensors, proprioception and vestibularinformation-leads to learning of dynamic models that showsuperior performance when used for control. We also collectdata from a real robotic hand and show that the same modelscan be used to answer questions about properties of objects inthe real world. Videos with qualitative results of our modelsare available a hrefhttps://goo.gl/mZuqAV>here/a>./p> /div>/td>/tr>tr idtr-amos2018end stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>45./td>td>a hrefhttps://arxiv.org/abs/1810.13400 target_blank>img srcimages/publications/amos2018end.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1810.13400 target_blank>Differentiable MPC for End-to-end Planning and Control/a> /em> a hrefjavascript:; onclick$("#abs_amos2018end").toggle()>abs/a> a hrefhttps://locuslab.github.io/mpc.pytorch/ target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://ivandariojr.io/ target_blank>Ivan Dario Jimenez Rodriguez/a>, a hrefhttps://scholar.google.com/citations?userTh4PuGkAAAAJ target_blank>Jacob Sacks/a>, a hrefhttps://homes.cs.washington.edu/~bboots/ target_blank>Byron Boots/a>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />NeurIPS 2018 br />div idabs_amos2018end styletext-align: justify; display: none> p>In this paper we present foundations for using model predictive control (MPC) as a differentiable policy class in reinforcement learning. Specifically, we differentiate through MPC by using the KKT conditions of the convex approximation at a fixed point of the solver. Using this strategy, we are able to learn the cost and dynamics of a controller via end-to-end learning in a larger system. We empirically show results in an imitation learning setting, demonstrating that we can recover the underlying dynamics and cost more efficiently and reliably than with a generic neural network policy class/p> /div>/td>/tr>tr idtr-brown2018depth>td alignright stylepadding-left:0;padding-right:0;>46./td>td>a hrefhttp://arxiv.org/abs/1805.08195 target_blank>img srcimages/publications/brown2018depth.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://arxiv.org/abs/1805.08195 target_blank>Depth-Limited Solving for Imperfect-Information Games/a> /em> a hrefjavascript:; onclick$("#abs_brown2018depth").toggle()>abs/a>br />a hrefhttps://noambrown.github.io/ target_blank>Noam Brown/a>, a hrefhttp://www.cs.cmu.edu/~sandholm/ target_blank>Tuomas Sandholm/a>, and strong>Brandon Amos/strong>br />NeurIPS 2018 br />div idabs_brown2018depth styletext-align: justify; display: none> p>A fundamental challenge in imperfect-information games is that states do not have well-defined values. As a result, depth-limited search algorithms used in single-agent settings and perfect-information games do not apply. This paper introduces a principled way to conduct depth-limited solving in imperfect-information games by allowing the opponent to choose among a number of strategies for the remainder of the game at the depth limit. Each one of these strategies results in a different set of values for leaf nodes. This forces an agent to be robust to the different strategies an opponent may employ. We demonstrate the effectiveness of this approach by building a master-level heads-up no-limit Texas holdâem poker AI that defeats two prior top agents using only a 4-core CPU and 16 GB of memory. Developing such a powerful agent would have previously required a supercomputer./p> /div>/td>/tr>tr idtr-wang2018enabling>td alignright stylepadding-left:0;padding-right:0;>47./td>td>a hrefhttps://dl.acm.org/citation.cfm?id3209659 target_blank>img srcimages/publications/wang2018enabling.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://dl.acm.org/citation.cfm?id3209659 target_blank>Enabling Live Video Analytics with a Scalable and Privacy-Aware Framework/a> /em> a hrefjavascript:; onclick$("#abs_wang2018enabling").toggle()>abs/a>br />a hrefhttps://www.junjuewang.com target_blank>Junjue Wang/a>, strong>Brandon Amos/strong>, a hrefhttps://anupamdas.org/ target_blank>Anupam Das/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://www.normsadeh.org/ target_blank>Norman Sadeh/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />ACM TOMM 2018 br />div idabs_wang2018enabling styletext-align: justify; display: none> p>We show how to build the components of a privacy-aware, live videoanalytics ecosystem from the bottom up, startingwith OpenFace, our new open-source face recognitionsystem that approaches state-of-the-artaccuracy. Integrating OpenFace with interframetracking, we build RTFace, a mechanism fordenaturing video streams that selectively blursfaces according to specified policies at full framerates. This enables privacy management for livevideo analytics while providing a secure approachfor handling retrospective policyexceptions. Finally, we present a scalable, privacy-aware architecture for large camera networksusing RTFace and show how it can be an enabler for avibrant ecosystem and marketplace of privacy-awarevideo streams and analytics services./p> /div>/td>/tr>/table>h2>2017/h2>table classtable table-hover>tr idtr-amos2017optnet stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>48./td>td>a hrefhttp://arxiv.org/abs/1703.00443 target_blank>img srcimages/publications/amos2017optnet.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://arxiv.org/abs/1703.00443 target_blank>OptNet: Differentiable Optimization as a Layer in Neural Networks/a> /em> a hrefjavascript:; onclick$("#abs_amos2017optnet").toggle()>abs/a> a hrefhttps://github.com/locuslab/optnet target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2017.optnet.pdf target_blank>slides/a> br />strong>Brandon Amos/strong> and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />ICML 2017 br />div idabs_amos2017optnet styletext-align: justify; display: none> p>This paper presents OptNet, a network architecture that integratesoptimization problems (here, specifically in the form of quadratic programs)as individual layers in larger end-to-end trainable deep networks.These layers encode constraints and complex dependenciesbetween the hidden states that traditional convolutional andfully-connected layers often cannot capture.In this paper, we explore the foundations for such an architecture:we show how techniques from sensitivity analysis, bileveloptimization, and implicit differentiation can be used toexactly differentiate through these layers and with respectto layer parameters;we develop a highly efficient solver for these layers that exploits fastGPU-based batch solves within a primal-dual interior point method, and whichprovides backpropagation gradients with virtually no additional cost on top ofthe solve;and we highlight the application of these approaches in several problems.In one notable example, we show that the method iscapable of learning to play mini-Sudoku (4x4) given just input and output games, with no a priori information about the rules of the game;this highlights the ability of our architecture to learn hardconstraints better than other neural architectures./p> /div>/td>/tr>tr idtr-amos2017input stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>49./td>td>a hrefhttp://arxiv.org/abs/1609.07152 target_blank>img srcimages/publications/amos2017input.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://arxiv.org/abs/1609.07152 target_blank>Input Convex Neural Networks/a> /em> a hrefjavascript:; onclick$("#abs_amos2017input").toggle()>abs/a> a hrefhttps://github.com/locuslab/icnn target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2017.icnn.pdf target_blank>slides/a> br />strong>Brandon Amos/strong>, a hrefhttps://leixx.io/ target_blank>Lei Xu/a>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />ICML 2017 br />div idabs_amos2017input styletext-align: justify; display: none> p>This paper presents the input convex neural networkarchitecture. These are scalar-valued (potentially deep) neuralnetworks with constraints on the network parameters such that theoutput of the network is a convex function of (some of) the inputs.The networks allow for efficient inference via optimization over someinputs to the network given others, and can be applied to settingsincluding structured prediction, data imputation, reinforcementlearning, and others. In this paper we lay the basic groundwork forthese models, proposing methods for inference, optimization andlearning, and analyze their representational power. We show that manyexisting neural network architectures can be made input-convex witha minor modification, and develop specialized optimizationalgorithms tailored to this setting. Finally, we highlight theperformance of the methods on multi-label prediction, imagecompletion, and reinforcement learning problems, where we showimprovement over the existing state of the art in many cases./p> /div>/td>/tr>tr idtr-donti2017task>td alignright stylepadding-left:0;padding-right:0;>50./td>td>a hrefhttp://arxiv.org/abs/1703.04529 target_blank>img srcimages/publications/donti2017task.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://arxiv.org/abs/1703.04529 target_blank>Task-based End-to-end Model Learning/a> /em> a hrefjavascript:; onclick$("#abs_donti2017task").toggle()>abs/a> a hrefhttps://github.com/locuslab/e2e-model-learning target_blank>code/a> br />a hrefhttps://priyadonti.com/ target_blank>Priya L. Donti/a>, strong>Brandon Amos/strong>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />NeurIPS 2017 br />div idabs_donti2017task styletext-align: justify; display: none> p>As machine learning techniques have become more ubiquitous, it hasbecome common to see machine learning prediction algorithms operatingwithin some larger process. However, the criteria by which we trainmachine learning algorithms often differ from the ultimate criteria onwhich we evaluate them. This paper proposes an end-to-end approach forlearning probabilistic machine learning models within the context ofstochastic programming, in a manner that directly captures theultimate task-based objective for which they will be used. We thenpresent two experimental evaluations of the proposed approach, one asapplied to a generic inventory stock problem and the second to areal-world electrical grid scheduling task. In both cases, we showthat the proposed approach can outperform both a traditional modelingapproach and a purely black-box policy optimization approach./p> /div>/td>/tr>tr idtr-chen2017quasi>td alignright stylepadding-left:0;padding-right:0;>51./td>td>a hrefhttps://par.nsf.gov/servlets/purl/10111392 target_blank>img srcimages/publications/chen2017quasi.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://par.nsf.gov/servlets/purl/10111392 target_blank>Quasi-Newton Stochastic Optimization Algorithm for Parameter Estimation of a Stochastic Model of the Budding Yeast Cell Cycle/a> /em> a hrefjavascript:; onclick$("#abs_chen2017quasi").toggle()>abs/a>br />a hrefhttps://chenm.sites.wfu.edu/publications/ target_blank>Minghan Chen/a>, strong>Brandon Amos/strong>, a hrefhttps://people.cs.vt.edu/~ltw/shortvita.html target_blank>Layne T. Watson/a>, a hrefhttps://scholar.google.com/citations?usersyETjMMAAAAJ target_blank>John Tyson/a>, a hrefhttps://people.cs.vt.edu/~ycao/ target_blank>Yang Cao/a>, a hrefhttps://people.cs.vt.edu/shaffer/ target_blank>Cliff Shaffer/a>, a hrefhttps://mtrosset.pages.iu.edu/ target_blank>Michael Trosset/a>, a hrefhttps://scholar.google.com/citations?userZ4534DUAAAAJ target_blank>Cihan Oguz/a>, and a hrefhttps://dblp.org/pid/235/5473.html target_blank>Gisella Kakoti/a>br />IEEE/ACM TCBB 2017 br />div idabs_chen2017quasi styletext-align: justify; display: none> p>Parameter estimation in discrete or continuous deterministic cellcycle models is challenging for several reasons, including the nature of what can be observed, andthe accuracy and quantity of those observations. Thechallenge is even greater for stochastic models, where the number of simulations and amount ofempirical data must be even larger to obtainstatistically valid parameter estimates. The twomain contributions of this work are (1) stochasticmodel parameter estimation based on directlymatching multivariate probability distributions, and(2) a new quasi-Newton algorithm class QNSTOP forstochastic optimization problems. QNSTOP directlyuses the random objective function value samplesrather than creating ensemble statistics. QNSTOP isused here to directly match empirical and simulatedjoint probability distributions rather than matchingsummary statistics. Results are given for a currentstate-of-the-art stochastic cell cycle model ofbudding yeast, whose predictions match well somesummary statistics and one-dimensional distributionsfrom empirical data, but do not match well theempirical joint distributions. The nature of themismatch provides insight into the weakness in thestochastic model./p> /div>/td>/tr>tr idtr-ha2017you>td alignright stylepadding-left:0;padding-right:0;>52./td>td>a hrefhttps://dl.acm.org/doi/10.1145/3132211.3134453 target_blank>img srcimages/publications/ha2017you.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://dl.acm.org/doi/10.1145/3132211.3134453 target_blank>You can teach elephants to dance: agile VM handoff for edge computing/a> /em> a hrefjavascript:; onclick$("#abs_ha2017you").toggle()>abs/a>br />a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://dblp.org/pid/18/1620.html target_blank>Yoshihisa Abe/a>, a hrefhttps://dblp.org/pid/207/9122.html target_blank>Thomas Eiszler/a>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, strong>Brandon Amos/strong>, a hrefhttps://dblp.org/pid/207/9123.html target_blank>Rohit Upadhyaya/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />SEC 2017 br />div idabs_ha2017you styletext-align: justify; display: none> p>VM handoff enables rapid and transparent placement changes toexecuting code in edge computing use cases where thesafety and management attributes of VM encapsulationare important. This versatile primitive offers thefunctionality of classic live migration but ishighly optimized for the edge. Over WAN bandwidthsranging from 5 to 25 Mbps, VM handoff migrates arunning 8 GB VM in about a minute, with a downtimeof a few tens of seconds. By dynamically adapting tovarying network bandwidth and processing load, VMhandoff is more than an order of magnitude fasterthan live migration at those bandwidths./p> /div>/td>/tr>tr idtr-chen2017empirical>td alignright stylepadding-left:0;padding-right:0;>53./td>td>a hrefhttps://www.cs.cmu.edu/~zhuoc/papers/latency2017.pdf target_blank>img srcimages/publications/chen2017empirical.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://www.cs.cmu.edu/~zhuoc/papers/latency2017.pdf target_blank>An Empirical Study of Latency in an Emerging Class of Edge Computing Applications for Wearable Cognitive Assistance/a> /em> a hrefjavascript:; onclick$("#abs_chen2017empirical").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, a hrefhttps://www.junjuewang.com target_blank>Junjue Wang/a>, a hrefhttps://scholar.google.com/citations?user0OpjwCMAAAAJ target_blank>Siyan Zhao/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?user0pF6i38AAAAJ target_blank>Guanhang Wu/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://scholar.google.com/citations?userR9r5_GIAAAAJ target_blank>Khalid Elgazzar/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://scholar.google.com/citations?userwUPKh58AAAAJ target_blank>Roberta Klatzky/a>, a hrefhttp://www.cs.cmu.edu/~dps/ target_blank>Daniel Siewiorek/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />SEC 2017 br />div idabs_chen2017empirical styletext-align: justify; display: none> p>An emerging class of interactive wearable cognitive assistanceapplications is poised to become one of the keydemonstrators of edge computing infrastructure. Inthis paper, we design seven such applications andevaluate their performance in terms of latencyacross a range of edge computing configurations, mobile hardware, and wireless networks, including 4GLTE. We also devise a novel multi-algorithm approachthat leverages temporal locality to reduceend-to-end latency by 60% to 70%, withoutsacrificing accuracy. Finally, we derive targetlatencies for our applications, and show that edgecomputing is crucial to meeting these targets./p> /div>/td>/tr>tr idtr-wang2017scalable>td alignright stylepadding-left:0;padding-right:0;>54./td>td>a hrefhttp://elijah.cs.cmu.edu/DOCS/wang-mmsys2017.pdf target_blank>img srcimages/publications/wang2017scalable.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://elijah.cs.cmu.edu/DOCS/wang-mmsys2017.pdf target_blank>A Scalable and Privacy-Aware IoT Service for Live Video Analytics/a> /em> a hrefjavascript:; onclick$("#abs_wang2017scalable").toggle()>abs/a> a hrefhttp://cmusatyalab.github.io/openface/ target_blank>code/a> br />a hrefhttps://www.junjuewang.com target_blank>Junjue Wang/a>, strong>Brandon Amos/strong>, a hrefhttps://anupamdas.org/ target_blank>Anupam Das/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://www.normsadeh.org/ target_blank>Norman Sadeh/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />ACM MMSys 2017 (Best Paper) br />div idabs_wang2017scalable styletext-align: justify; display: none> p>We present OpenFace, our new open-source face recognition systemthat approaches state-of-the-art accuracy. Integrating OpenFace withinter-frame tracking, we build RTFace, a mechanism for denaturing videostreams that selectively blurs faces according to specifiedpolicies at full frame rates. This enables privacy management forlive video analytics while providing a secure approach for handlingretrospective policy exceptions. Finally, we present a scalable, privacy-aware architecture for large camera networks using RTFace./p> /div>/td>/tr>/table>h2>2016/h2>table classtable table-hover>tr idtr-amos2016openface stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>55./td>td>a hrefhttp://reports-archive.adm.cs.cmu.edu/anon/anon/2016/CMU-CS-16-118.pdf target_blank>img srcimages/publications/amos2016openface.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://reports-archive.adm.cs.cmu.edu/anon/anon/2016/CMU-CS-16-118.pdf target_blank>OpenFace: A general-purpose face recognition library with mobile applications/a> /em> a hrefjavascript:; onclick$("#abs_amos2016openface").toggle()>abs/a> a hrefhttps://cmusatyalab.github.io/openface target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://www.linkedin.com/in/bartosz-ludwiczuk-a677a760 target_blank>Bartosz Ludwiczuk/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />CMU 2016 br />div idabs_amos2016openface styletext-align: justify; display: none> p>Cameras are becoming ubiquitous in the Internet of Things (IoT) andcan use face recognition technology to improve context. There is alarge accuracy gap between todayâs publicly available face recognitionsystems and the state-of-the-art private face recognitionsystems. This paper presents our OpenFace face recognition librarythat bridges this accuracy gap. We show that OpenFace providesnear-human accuracy on the LFW benchmark and present a newclassification benchmark for mobile scenarios. This paper is intendedfor non-experts interested in using OpenFace and provides a lightintroduction to the deep neural network techniques we use./p> p>We released OpenFace in October 2015 as an open source library underthe Apache 2.0 license. It is available at:a hrefhttp://cmusatyalab.github.io/openface/>http://cmusatyalab.github.io/openface//a>/p> /div>/td>/tr>tr idtr-zhao2016collapsed>td alignright stylepadding-left:0;padding-right:0;>56./td>td>a hrefhttp://proceedings.mlr.press/v48/zhaoa16.html target_blank>img srcimages/publications/zhao2016collapsed.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://proceedings.mlr.press/v48/zhaoa16.html target_blank>Collapsed Variational Inference for Sum-Product Networks/a> /em> a hrefjavascript:; onclick$("#abs_zhao2016collapsed").toggle()>abs/a>br />a hrefhttps://hanzhaoml.github.io/ target_blank>Han Zhao/a>, a hrefhttps://tameemadel.wordpress.com/ target_blank>Tameem Adel/a>, a hrefhttp://www.cs.cmu.edu/~ggordon/ target_blank>Geoff Gordon/a>, and strong>Brandon Amos/strong>br />ICML 2016 br />div idabs_zhao2016collapsed styletext-align: justify; display: none> p>Sum-Product Networks (SPNs) are probabilistic inference machines that admitexact inference in linear time in the size of the network. Existingparameter learning approaches for SPNs are largely based on the maximumlikelihood principle and hence are subject to overfitting compared tomore Bayesian approaches. Exact Bayesian posterior inference for SPNs iscomputationally intractable. Both standard variational inference andposterior sampling for SPNs are computationally infeasible even fornetworks of moderate size due to the large number of local latentvariables per instance. In this work, we propose a novel deterministiccollapsed variational inference algorithm for SPNs that iscomputationally efficient, easy to implement and at the same time allowsus to incorporate prior information into the optimization formulation.Extensive experiments show a significant improvement in accuracy comparedwith a maximum likelihood based approach./p> /div>/td>/tr>tr idtr-hu2016quantifying>td alignright stylepadding-left:0;padding-right:0;>57./td>td>a hrefhttps://dl.acm.org/doi/10.1145/2967360.2967369 target_blank>img srcimages/publications/hu2016quantifying.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://dl.acm.org/doi/10.1145/2967360.2967369 target_blank>Quantifying the impact of edge computing on mobile applications/a> /em> a hrefjavascript:; onclick$("#abs_hu2016quantifying").toggle()>abs/a>br />a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, a hrefhttps://www.linkedin.com/in/joelyinggao/ target_blank>Ying Gao/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://www.junjuewang.com target_blank>Junjue Wang/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />ACM SIGOPS 2016 br />div idabs_hu2016quantifying styletext-align: justify; display: none> p>Computational offloading services at the edge of the Internet formobile devices are becoming a reality. Using a widerange of mobile applications, we explore how suchinfrastructure improves latency and energyconsumption relative to the cloud. We presentexperimental results from WiFi and 4G LTE networksthat confirm substantial wins from edge computingfor highly interactive mobile applications./p> /div>/td>/tr>tr idtr-davies2016privacy>td alignright stylepadding-left:0;padding-right:0;>58./td>td>a hrefhttp://eprints.lancs.ac.uk/78255/1/44691.pdf target_blank>img srcimages/publications/davies2016privacy.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://eprints.lancs.ac.uk/78255/1/44691.pdf target_blank>Privacy mediators: helping IoT cross the chasm/a> /em> a hrefjavascript:; onclick$("#abs_davies2016privacy").toggle()>abs/a>br />a hrefhttps://www.lancaster.ac.uk/sci-tech/about-us/people/nigel-davies target_blank>Nigel Davies/a>, a hrefhttps://scholar.google.com/citations?userBItCgjYAAAAJ target_blank>Nina Taft/a>, a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>, a hrefhttp://www.sclinch.com/ target_blank>Sarah Clinch/a>, and strong>Brandon Amos/strong>br />HotMobile 2016 br />div idabs_davies2016privacy styletext-align: justify; display: none> p>Unease over data privacy will retard consumer acceptance of IoTdeployments. The primary source of discomfort is a lack of usercontrol over raw data that is streamed directly from sensors to thecloud. This is a direct consequence of the over-centralization oftodayâs cloud-based IoT hub designs. We propose a solution thatinterposes a locally-controlled software component called a privacymediator on every raw sensor stream. Each mediator is in the sameadministrative domain as the sensors whose data is being collected, and dynamically enforces the current privacy policies of the ownersof the sensors or mobile users within the domain. This solution necessitatesa logical point of presence for mediators within the administrativeboundaries of each organization. Such points of presenceare provided by cloudlets, which are small locally-administered datacenters at the edge of the Internet that can support code mobility.The use of cloudlet-based mediators aligns well with natural personaland organizational boundaries of trust and responsibility./p> /div>/td>/tr>/table>h2>2015 and earlier/h2>table classtable table-hover>tr idtr-satyanarayanan2015edge>td alignright stylepadding-left:0;padding-right:0;>59./td>td>a hrefhttps://www.cs.cmu.edu/~satya/docdir/satya-edge2015.pdf target_blank>img srcimages/publications/satyanarayanan2015edge.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://www.cs.cmu.edu/~satya/docdir/satya-edge2015.pdf target_blank>Edge Analytics in the Internet of Things/a> /em> a hrefjavascript:; onclick$("#abs_satyanarayanan2015edge").toggle()>abs/a>br />a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>, a hrefhttps://www.ugent.be/ea/idlab/en/members/pieter-simoens.htm target_blank>Pieter Simoens/a>, a hrefhttps://scholar.google.com/citations?userZeRhyWsAAAAJ target_blank>Yu Xiao/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, and strong>Brandon Amos/strong>br />IEEE Pervasive Computing 2015 br />div idabs_satyanarayanan2015edge styletext-align: justify; display: none> p>High-data-rate sensors, such as video cameras, are becoming ubiquitous in theInternet of Things. This article describes GigaSight, an Internet-scalerepository of crowd-sourced video content, with strong enforcement of privacypreferences and access controls. The GigaSight architecture is a federatedsystem of VM-based cloudlets that perform video analytics at the edge of theInternet, thus reducing the demand for ingress bandwidth into the cloud.Denaturing, which is an owner-specific reduction in fidelity of video contentto preserve privacy, is one form of analytics on cloudlets. Content-basedindexing for search is another form of cloudlet-based analytics. This articleis part of a special issue on smart spaces./p> /div>/td>/tr>tr idtr-turner2015bad>td alignright stylepadding-left:0;padding-right:0;>60./td>td>a hrefhttp://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber7118094 target_blank>img srcimages/publications/turner2015bad.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber7118094 target_blank>Bad Parts: Are Our Manufacturing Systems at Risk of Silent Cyberattacks?/a> /em> a hrefjavascript:; onclick$("#abs_turner2015bad").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userMRKab9cAAAAJ target_blank>Hamilton Turner/a>, a hrefhttps://scholar.google.com/citations?user10HSX90AAAAJ target_blank>Jules White/a>, a hrefhttps://scholar.google.com/citations?usertWmVBNwAAAAJ target_blank>Jaime A. Camelio/a>, a hrefhttps://scholar.google.com/citations?userAW81mosAAAAJ target_blank>Christopher Williams/a>, strong>Brandon Amos/strong>, and a hrefhttps://ieeexplore.ieee.org/author/37085729541 target_blank>Robert Parker/a>br />IEEE Security & Privacy 2015 br />div idabs_turner2015bad styletext-align: justify; display: none> p>Recent cyberattacks have highlighted the risk of physical equipment operatingoutside designed tolerances to produce catastrophic failures. A relatedthreat is cyberattacks that change the design and manufacturing of amachineâs part, such as an automobile brake component, so it no longerfunctions properly. These risks stem from the lack of cyber-physical modelsto identify ongoing attacks as well as the lack of rigorous application ofknown cybersecurity best practices. To protect manufacturing processes in thefuture, research will be needed on a number of critical cyber-physicalmanufacturing security topics./p> /div>/td>/tr>tr idtr-chen2015early>td alignright stylepadding-left:0;padding-right:0;>61./td>td>a hrefhttp://www.cs.cmu.edu/~satya/docdir/chen-wearsys2015.pdf target_blank>img srcimages/publications/chen2015early.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://www.cs.cmu.edu/~satya/docdir/chen-wearsys2015.pdf target_blank>Early Implementation Experience with Wearable Cognitive Assistance Applications/a> /em> a hrefjavascript:; onclick$("#abs_chen2015early").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://www.lujiang.info/ target_blank>Lu Jiang/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, strong>Brandon Amos/strong>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttp://www.cs.cmu.edu/~alex/ target_blank>Alex Hauptmann/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />WearSys 2015 br />div idabs_chen2015early styletext-align: justify; display: none> p>A cognitive assistance application combines a wearable device suchas Google Glass with cloudlet processing to provide step-by-stepguidance on a complex task. In this paper, we focus on user assistancefor narrow and well-defined tasks that require specializedknowledge and/or skills. We describe proof-of-concept implementationsfor four different tasks: assembling 2D Lego models, freehandsketching, playing ping-pong, and recommending context-relevantYouTube tutorials. We then reflect on the difficulties we faced inbuilding these applications, and suggest future research that couldsimplify the creation of similar applications./p> /div>/td>/tr>tr idtr-hu2014case>td alignright stylepadding-left:0;padding-right:0;>62./td>td>a hrefhttp://www.cs.cmu.edu/~satya/docdir/hu-hotmobile2015.pdf target_blank>img srcimages/publications/hu2014case.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://www.cs.cmu.edu/~satya/docdir/hu-hotmobile2015.pdf target_blank>The Case for Offload Shaping/a> /em> a hrefjavascript:; onclick$("#abs_hu2014case").toggle()>abs/a>br />a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://scholar.google.com/citations?uservU6bKxEAAAAJ target_blank>Wolfgang Richter/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://github.com/bgilbert target_blank>Benjamin Gilbert/a>, a hrefhttps://scholar.google.com/citations?userjj5tN8sAAAAJ target_blank>Jan Harkes/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />HotMobile 2015 br />div idabs_hu2014case styletext-align: justify; display: none> p>When offloading computation from a mobile device, we showthat it can pay to perform additional on-device work in orderto reduce the offloading workload. We call this offload shaping, and demonstrate its application at many different levelsof abstraction using a variety of techniques. We show thatoffload shaping can produce significant reduction in resourcedemand, with little loss of application-level fidelity/p> /div>/td>/tr>tr idtr-gao2015cloudlets>td alignright stylepadding-left:0;padding-right:0;>63./td>td>a hrefhttp://reports-archive.adm.cs.cmu.edu/anon/anon/2015/CMU-CS-15-139.pdf target_blank>img srcimages/publications/gao2015cloudlets.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://reports-archive.adm.cs.cmu.edu/anon/anon/2015/CMU-CS-15-139.pdf target_blank>Are Cloudlets Necessary?/a> /em> a hrefjavascript:; onclick$("#abs_gao2015cloudlets").toggle()>abs/a>br />a hrefhttps://www.linkedin.com/in/joelyinggao/ target_blank>Ying Gao/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, strong>Brandon Amos/strong>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />CMU 2015 br />div idabs_gao2015cloudlets styletext-align: justify; display: none> p>We present experimental results from Wi-Fi and 4G LTE networks to validate theintuition that low end-to-end latency of cloud services improves applicationresponse time and reduces energy consumption on mobile devices. We focusspecifically on computational offloading as a cloud service. Using a widerange of applications, and exploring both pre-partitioned and dynamicallypartitioned approaches, we demonstrate the importance of low latency forcloud offload services. We show the best performance is achieved byoffloading to cloudlets, which are small-scale edge-located data centers. Ourresults show that cloudlets can improve response times 51% and reduce energyconsumption in a mobile device by up to 42% compared to cloud offload./p> /div>/td>/tr>tr idtr-ha2015adaptive>td alignright stylepadding-left:0;padding-right:0;>64./td>td>a hrefhttp://ra.adm.cs.cmu.edu/anon/2015/CMU-CS-15-113.pdf target_blank>img srcimages/publications/ha2015adaptive.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://ra.adm.cs.cmu.edu/anon/2015/CMU-CS-15-113.pdf target_blank>Adaptive VM handoff across cloudlets/a> /em> a hrefjavascript:; onclick$("#abs_ha2015adaptive").toggle()>abs/a>br />a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://dblp.org/pid/18/1620.html target_blank>Yoshihisa Abe/a>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, strong>Brandon Amos/strong>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />CMU 2015 br />div idabs_ha2015adaptive styletext-align: justify; display: none> p>Cloudlet offload is a valuable technique for ensuring low end-to-end latency ofresource-intensive cloud processing for many emerging mobile applications.This paper examines the impact of user mobility on cloudlet offload, andshows that even modest user mobility can result in significant networkdegradation. We propose VM handoff as a technique for seamlessly transferringVM-encapsulated execution to a more optimal offload site as users move. Ourapproach can perform handoff in roughly a minute even over limited WANs byadaptively reducing data transferred. We present experimental results tovalidate our implementation and to demonstrate effectiveness of adaptation tochanging network conditions and processing capacity/p> /div>/td>/tr>tr idtr-andrew2014global>td alignright stylepadding-left:0;padding-right:0;>65./td>td>a hrefhttp://dl.acm.org/citation.cfm?id2685662 target_blank>img srcimages/publications/andrew2014global.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://dl.acm.org/citation.cfm?id2685662 target_blank>Global Parameter Estimation for a Eukaryotic Cell Cycle Model in Systems Biology/a> /em> a hrefjavascript:; onclick$("#abs_andrew2014global").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userqpRt_KYAAAAJ target_blank>Tricity Andrew/a>, strong>Brandon Amos/strong>, a hrefhttps://dblp.org/pid/75/8682.html target_blank>David Easterling/a>, a hrefhttps://scholar.google.com/citations?userZ4534DUAAAAJ target_blank>Cihan Oguz/a>, a hrefhttps://scholar.google.com/citations?userfAmU38gAAAAJ target_blank>William Baumann/a>, a hrefhttps://scholar.google.com/citations?usersyETjMMAAAAJ target_blank>John Tyson/a>, and a hrefhttps://people.cs.vt.edu/~ltw/shortvita.html target_blank>Layne T. Watson/a>br />SummerSim 2014 br />div idabs_andrew2014global styletext-align: justify; display: none> p>The complicated process by which a yeast cell divides, known as the cellcycle, has been modeled by a system of 26 nonlinear ordinary differentialequations (ODEs) with 149 parameters. This model captures the chemicalkinetics of the regulatory networks controlling the cell division processin budding yeast cells. Empirical data is discrete and matched againstdiscrete inferences (e.g., whether a particular mutant cell lives or dies)computed from the ODE solution trajectories. The problem ofestimating the ODE parameters to best fit the model to the data is a149-dimensional global optimization problem attacked by the deterministicalgorithm VTDIRECT95 and by the nondeterministic algorithms differentialevolution, QNSTOP, and simulated annealing, whose performances arecompared./p> /div>/td>/tr>tr idtr-amos2013applying>td alignright stylepadding-left:0;padding-right:0;>66./td>td>a hrefhttp://bamos.github.io/data/papers/amos-iwcmc2013.pdf target_blank>img srcimages/publications/amos2013applying.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://bamos.github.io/data/papers/amos-iwcmc2013.pdf target_blank>Applying machine learning classifiers to dynamic Android malware detection at scale/a> /em> a hrefjavascript:; onclick$("#abs_amos2013applying").toggle()>abs/a> a hrefhttps://github.com/VT-Magnum-Research/antimalware target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userMRKab9cAAAAJ target_blank>Hamilton Turner/a>, and a hrefhttps://scholar.google.com/citations?user10HSX90AAAAJ target_blank>Jules White/a>br />IWCMC 2013 br />div idabs_amos2013applying styletext-align: justify; display: none> p>The widespread adoption and contextually sensitivenature of smartphone devices has increased concerns over smartphonemalware. Machine learning classifiers are a current methodfor detecting malicious applications on smartphone systems. Thispaper presents the evaluation of a number of existing classifiers, using a dataset containing thousands of real (i.e. not synthetic)applications. We also present our STREAM framework, whichwas developed to enable rapid large-scale validation of mobilemalware machine learning classifiers./p> /div>/td>/tr>/table>h2 id-open-source-repositories>i classfa fa-chevron-right>/i> Open Source Repositories/h2>p>29.6k+ GitHub stars across all repositories./p>table classtable table-hover>tr> td alignright stylepadding-right:0;padding-left:0;>1./td> td> span classcvdate>2024/span> a hrefhttps://github.com/facebookresearch/advprompter>facebookresearch/advprompter/a> | i classfa fas fa-star>/i> 112 | em>Fast Adaptive Adversarial Prompting for LLMs/em> !-- --> !-- facebookresearch/advprompter --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>2./td> td> span classcvdate>2024/span> a hrefhttps://github.com/facebookresearch/lagrangian-ot>facebookresearch/lagrangian-ot/a> | i classfa fas fa-star>/i> 39 | em>Lagrangian OT/em> !-- --> !-- facebookresearch/lagrangian-ot --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>3./td> td> span classcvdate>2023/span> a hrefhttps://github.com/facebookresearch/amortized-optimization-tutorial>facebookresearch/amortized-optimization-tutorial/a> | i classfa fas fa-star>/i> 236 | em>Tutorial on amortized optimization/em> !-- --> !-- facebookresearch/amortized-optimization-tutorial --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>4./td> td> span classcvdate>2023/span> a hrefhttps://github.com/facebookresearch/taskmet>facebookresearch/taskmet/a> | i classfa fas fa-star>/i> 18 | em>TaskMet: Task-Driven Metric Learning for Model Learning/em> !-- --> !-- facebookresearch/taskmet --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>5./td> td> span classcvdate>2023/span> a hrefhttps://github.com/facebookresearch/w2ot>facebookresearch/w2ot/a> | i classfa fas fa-star>/i> 43 | em>Wasserstein-2 optimal transport in JAX/em> !-- --> !-- facebookresearch/w2ot --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>6./td> td> span classcvdate>2022/span> a hrefhttps://github.com/facebookresearch/theseus>facebookresearch/theseus/a> | i classfa fas fa-star>/i> 1.7k | em>Differentiable non-linear optimization library/em> !-- --> !-- facebookresearch/theseus --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>7./td> td> span classcvdate>2022/span> a hrefhttps://github.com/facebookresearch/meta-ot>facebookresearch/meta-ot/a> | i classfa fas fa-star>/i> 94 | em>Meta Optimal Transport/em> !-- --> !-- facebookresearch/meta-ot --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>8./td> td> span classcvdate>2022/span> a hrefhttps://github.com/bamos/presentations>bamos/presentations/a> | i classfa fas fa-star>/i> 141 | em>Source for my major presentations/em> !-- --> !-- bamos/presentations --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>9./td> td> span classcvdate>2021/span> a hrefhttps://github.com/facebookresearch/rcpm>facebookresearch/rcpm/a> | i classfa fas fa-star>/i> 68 | em>Riemannian Convex Potential Maps/em> !-- --> !-- facebookresearch/rcpm --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>10./td> td> span classcvdate>2021/span> a hrefhttps://github.com/facebookresearch/svg>facebookresearch/svg/a> | i classfa fas fa-star>/i> 54 | em>Model-based stochastic value gradient/em> !-- --> !-- facebookresearch/svg --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>11./td> td> span classcvdate>2021/span> a hrefhttps://github.com/facebookresearch/mbrl-lib>facebookresearch/mbrl-lib/a> | i classfa fas fa-star>/i> 954 | em>Model-based reinforcement learning library/em> !-- --> !-- facebookresearch/mbrl-lib --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>12./td> td> span classcvdate>2020/span> a hrefhttps://github.com/facebookresearch/dcem>facebookresearch/dcem/a> | i classfa fas fa-star>/i> 122 | em>The Differentiable Cross-Entropy Method/em> !-- --> !-- facebookresearch/dcem --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>13./td> td> span classcvdate>2019/span> a hrefhttps://github.com/facebookresearch/higher>facebookresearch/higher/a> | i classfa fas fa-star>/i> 1.6k | em>PyTorch higher-order gradient and optimization library/em> !-- --> !-- facebookresearch/higher --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>14./td> td> span classcvdate>2019/span> a hrefhttps://github.com/bamos/thesis>bamos/thesis/a> | i classfa fas fa-star>/i> 318 | em>Ph.D. Thesis LaTeX source code/em> !-- --> !-- bamos/thesis --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>15./td> td> span classcvdate>2019/span> a hrefhttps://github.com/cvxgrp/cvxpylayers>cvxgrp/cvxpylayers/a> | i classfa fas fa-star>/i> 1.8k | em>Differentiable Convex Optimization Layers/em> !-- --> !-- cvxgrp/cvxpylayers --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>16./td> td> span classcvdate>2019/span> a hrefhttps://github.com/locuslab/lml>locuslab/lml/a> | i classfa fas fa-star>/i> 58 | em>The Limited Multi-Label Projection Layer/em> !-- --> !-- locuslab/lml --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>17./td> td> span classcvdate>2018/span> a hrefhttps://github.com/locuslab/mpc.pytorch>locuslab/mpc.pytorch/a> | i classfa fas fa-star>/i> 865 | em>Differentiable PyTorch Model Predictive Control library/em> !-- --> !-- locuslab/mpc.pytorch --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>18./td> td> span classcvdate>2018/span> a hrefhttps://github.com/locuslab/differentiable-mpc>locuslab/differentiable-mpc/a> | i classfa fas fa-star>/i> 239 | em>Differentiable MPC experiments/em> !-- --> !-- locuslab/differentiable-mpc --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>19./td> td> span classcvdate>2017/span> a hrefhttps://github.com/locuslab/icnn>locuslab/icnn/a> | i classfa fas fa-star>/i> 274 | em>Input Convex Neural Network experiments/em> !-- --> !-- locuslab/icnn --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>20./td> td> span classcvdate>2017/span> a hrefhttps://github.com/locuslab/optnet>locuslab/optnet/a> | i classfa fas fa-star>/i> 507 | em>OptNet experiments/em> !-- --> !-- locuslab/optnet --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>21./td> td> span classcvdate>2017/span> a hrefhttps://github.com/locuslab/qpth>locuslab/qpth/a> | i classfa fas fa-star>/i> 673 | em>Differentiable PyTorch QP solver/em> !-- --> !-- locuslab/qpth --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>22./td> td> span classcvdate>2017/span> a hrefhttps://github.com/bamos/densenet.pytorch>bamos/densenet.pytorch/a> | i classfa fas fa-star>/i> 823 | em>PyTorch DenseNet implementation/em> !-- --> !-- bamos/densenet.pytorch --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>23./td> td> span classcvdate>2017/span> a hrefhttps://github.com/bamos/block>bamos/block/a> | i classfa fas fa-star>/i> 297 | em>Intelligent block matrix constructions/em> !-- --> !-- bamos/block --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>24./td> td> span classcvdate>2017/span> a hrefhttps://github.com/bamos/setGPU>bamos/setGPU/a> | i classfa fas fa-star>/i> 106 | em>Automatically use the least-loaded GPU/em> !-- --> !-- bamos/setGPU --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>25./td> td> span classcvdate>2016/span> a hrefhttps://github.com/bamos/dcgan-completion.tensorflow>bamos/dcgan-completion.tensorflow/a> | i classfa fas fa-star>/i> 1.3k | em>Image completion with GANs/em> !-- --> !-- bamos/dcgan-completion.tensorflow --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>26./td> td> span classcvdate>2015/span> a hrefhttps://github.com/cmusatyalab/openface>cmusatyalab/openface/a> | i classfa fas fa-star>/i> 15.1k | em>Face recognition with deep neural networks/em> !-- --> !-- cmusatyalab/openface --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>27./td> td> span classcvdate>2014/span> a hrefhttps://github.com/vtopt/qnstop>vtopt/qnstop/a> | i classfa fas fa-star>/i> 10 | em>Fortran quasi-Newton stochastic optimization library/em> !-- --> !-- vtopt/qnstop --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>28./td> td> span classcvdate>2014/span> a hrefhttps://github.com/bamos/snowglobe>bamos/snowglobe/a> | i classfa fas fa-star>/i> 27 | em>Haskell-driven, self-hosted web analytics with minimal configuration/em> !-- --> !-- bamos/snowglobe --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>29./td> td> span classcvdate>2014/span> a hrefhttps://github.com/bamos/zsh-history-analysis>bamos/zsh-history-analysis/a> | i classfa fas fa-star>/i> 224 | em>Analyze and plot your zsh history/em> !-- --> !-- bamos/zsh-history-analysis --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>30./td> td> span classcvdate>2014/span> a hrefhttps://github.com/bamos/beamer-snippets>bamos/beamer-snippets/a> | i classfa fas fa-star>/i> 109 | em>Beamer and TikZ snippets/em> !-- --> !-- bamos/beamer-snippets --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>31./td> td> span classcvdate>2013/span> a hrefhttps://github.com/bamos/latex-templates>bamos/latex-templates/a> | i classfa fas fa-star>/i> 366 | em>LaTeX templates/em> !-- --> !-- bamos/latex-templates --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>32./td> td> span classcvdate>2013/span> a hrefhttps://github.com/cparse/cparse>cparse/cparse/a> | i classfa fas fa-star>/i> 336 | em>C++ expression parser using Dijkstras shunting-yard algorithm/em> !-- --> !-- cparse/cparse --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>33./td> td> span classcvdate>2013/span> a hrefhttps://github.com/bamos/cv>bamos/cv/a> | i classfa fas fa-star>/i> 398 | em>Source for this CV: Creates LaTeX/Markdown from YAML/BibTeX/em> !-- --> !-- bamos/cv --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>34./td> td> span classcvdate>2013/span> a hrefhttps://github.com/bamos/python-scripts>bamos/python-scripts/a> | i classfa fas fa-star>/i> 197 | em>Short and fun Python scripts/em> !-- --> !-- bamos/python-scripts --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>35./td> td> span classcvdate>2013/span> a hrefhttps://github.com/bamos/reading-list>bamos/reading-list/a> | i classfa fas fa-star>/i> 185 | em>YAML reading list and notes system/em> !-- --> !-- bamos/reading-list --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>36./td> td> span classcvdate>2012/span> a hrefhttps://github.com/bamos/dotfiles>bamos/dotfiles/a> | i classfa fas fa-star>/i> 238 | em>i classfa fas fa-heart>/i> Linux, xmonad, emacs, vim, zsh, tmux/em> !-- --> !-- bamos/dotfiles --> !-- --> /td>/tr>/table>h2 id-invited-talks>i classfa fa-chevron-right>/i> Invited Talks/h2>p>Slides for my major presentations are open-sourced with a CC-BY license ata hrefhttps://github.com/bamos/presentations>bamos/presentations/a>./p>table classtable table-hover>tr> td alignright stylepadding-right:0;padding-left:0;>1./td> td stylepadding-right:0;> span classcvdate>2024/span> em>Amortized optimization for optimal transport and LLM attacks/em>, ISMP /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>2./td> td stylepadding-right:0;> span classcvdate>2024/span> em>Differentiable optimization for robotics/em>, a hrefhttps://sites.google.com/robotics.utias.utoronto.ca/frontiers-optimization-rss24/schedule>RSS Optimization for Robotics Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>3./td> td stylepadding-right:0;> span classcvdate>2024/span> em>Amortized optimization-based reasoning for AI/em>, University of Amsterdam /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>4./td> td stylepadding-right:0;> span classcvdate>2024/span> em>End-to-end learning geometries for graphs, dynamical systems, and regression/em>, a hrefhttps://logmeetupnyc.github.io/>LoG New York/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>5./td> td stylepadding-right:0;> span classcvdate>2023/span> em>Amortized optimization for optimal transport/em>, a hrefhttps://otmlworkshop.github.io/schedule/>NeurIPS Optimal Transport and ML Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>6./td> td stylepadding-right:0;> span classcvdate>2023/span> em>On optimal control and machine learning/em>, a hrefhttps://frontiers4lcd.github.io/>ICML Learning, Control, and Dynamical Systems Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>7./td> td stylepadding-right:0;> span classcvdate>2023/span> em>Tutorial on amortized optimization/em>, Brown University /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>8./td> td stylepadding-right:0;> span classcvdate>2023/span> em>Learning with differentiable and amortized optimization/em>, NYU AI Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>9./td> td stylepadding-right:0;> span classcvdate>2023/span> em>Learning with differentiable and amortized optimization/em>, Vanderbilt ML Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>10./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Learning with differentiable and amortized optimization/em>, Microsoft Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>11./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Amortized optimization for computing optimal transport maps/em>, a hrefhttps://sites.google.com/view/sampling-transport-diffusions/home>Flatiron Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>12./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Learning with differentiable and amortized optimization/em>, Cornell AI Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>13./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Learning with differentiable and amortized optimization/em>, Cornell Tech Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>14./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Learning with differentiable and amortized optimization/em>, Argonne National Laboratory /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>15./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Theseus: A library for differentiable nonlinear optimization/em>, NYU /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>16./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Theseus: A library for differentiable nonlinear optimization/em>, University of Zurich /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>17./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization-based modeling for machine learning/em>, Colorado Mines AMS Colloquium /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>18./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization/em>, a hrefhttps://guaguakai.github.io/IJCAI22-differentiable-optimization/>IJCAI Tutorial/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>19./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization for control and RL/em>, a hrefhttps://darl-workshop.github.io/>ICML Workshop on Decision Awareness in RL/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>20./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization-based modeling for machine learning/em>, a hrefhttps://sites.google.com/usc.edu/cpaior-2022/master_class>CPAIOR Master Class/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>21./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Tutorial on amortized optimization/em>, ICCOPT /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>22./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization for control and RL/em>, Gridmatic /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>23./td> td stylepadding-right:0;> span classcvdate>2021/span> em>Learning for control with differentiable optimization and ODEs/em>, Columbia University /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>24./td> td stylepadding-right:0;> span classcvdate>2021/span> em>Differentiable optimization-based modeling for machine learning/em>, IBM Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>25./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Differentiable optimization for control/em>, Max Planck Institute (TĂŒbingen) /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>26./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Differentiable optimization-based modeling for machine learning/em>, Mila Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>27./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Deep Declarative Networks/em>, a hrefhttps://anucvml.github.io/ddn-eccvt2020/>ECCV Tutorial/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>28./td> td stylepadding-right:0;> span classcvdate>2020/span> em>On differentiable optimization for control and vision/em>, a hrefhttps://anucvml.github.io/ddn-cvprw2020/>CVPR Deep Declarative Networks Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>29./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Differentiable optimization-based modeling for machine learning/em>, a hrefhttps://sites.google.com/view/cs-159-spring-2020/lectures>Caltech CS 159 (Guest Lecture)/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>30./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Unrolled optimization for learning deep energy models/em>, a hrefhttps://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE67922>SIAM MDS Minisymposium/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>31./td> td stylepadding-right:0;> span classcvdate>2019/span> em>Differentiable optimization-based modeling for machine learning/em>, NYU CILVR Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>32./td> td stylepadding-right:0;> span classcvdate>2019/span> em>Differentiable optimization-based modeling for machine learning/em>, INFORMS /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>33./td> td stylepadding-right:0;> span classcvdate>2019/span> em>Differentiable optimization-based modeling for machine learning/em>, Facebook AI Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>34./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, ISMP /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>35./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Google Brain /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>36./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Bosch Center for AI /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>37./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Waymo Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>38./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Tesla AI /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>39./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, NVIDIA Robotics /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>40./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Salesforce Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>41./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, OpenAI /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>42./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, NNAISENSE /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>43./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization and control/em>, UC Berkeley /td>/tr>/table>h2 id-interns-and-students>i classfa fa-chevron-right>/i> Interns and Students/h2>table classtable table-hover>tr> td stylepadding-right:0;> span classcvdate>2024 - present/span> a hrefhttps://scholar.google.com/citations?userHBAXF6YAAAAJ>Aaron Havens/a> (visiting FAIR from UIUC) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2022 - 2024/span> a hrefhttps://arampooladian.com/>Aram-Alexandre Pooladian/a> (visiting FAIR from NYU) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2022 - 2024/span> a hrefhttps://cdenrich.github.io/>Carles Domingo-Enrich/a> (visiting FAIR from NYU, now at MSR) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023 - 2024/span> a hrefhttps://scholar.google.com/citations?usernjZL5CQAAAAJ>Anselm Paulus/a> (visiting FAIR from Max Planck Institute, TĂŒbingen) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023/span> a hrefhttps://www.mhr.ai>Matthew Retchin/a> (Columbia MS thesis committee, now at Harvard) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2022 - 2023/span> a hrefhttps://sanaelotfi.github.io/>Sanae Lotfi/a> (visiting FAIR from NYU) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2022 - 2023/span> a hrefhttps://dishank-b.github.io>Dishank Bansal/a> (AI resident at FAIR) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2021 - 2022/span> a hrefhttps://www.linkedin.com/in/arnaudfickinger/>Arnaud Fickinger/a> (visiting FAIR from Berkeley) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020 - 2022/span> a hrefhttps://aaronlou.com/>Aaron Lou/a> (visiting FAIR from Cornell and Stanford, now scientist at OpenAI) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2021 - 2022/span> a hrefhttps://www.eugenevinitsky.com>Eugene Vinitsky/a> (visiting FAIR from Berkeley, now professor at NYU) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2021 - 2022/span> a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ>Samuel Cohen/a> (visiting FAIR from UCL, now CEO at FairGen) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttp://www.cs.toronto.edu/~rtqichen/>Ricky Chen/a> (visiting FAIR from Toronto, now scientist at FAIR) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttp://www.cs.cmu.edu/~pliang/>Paul Liang/a> (visiting FAIR from CMU, now professor at MIT) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2018/span> a hrefhttps://phillipkwang.com/>Phillip Wang/a> (at CMU, now CEO at a hrefhttps://gather.town/ target_blank>Gather/a>) /td>/tr>/table>h2 id-professional-activities>i classfa fa-chevron-right>/i> Professional Activities/h2>table classtable table-hover>tr> td stylepadding-right:0;> span classcvdate>2025/span> AAAI Senior Program Committee /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2024/span> NeurIPS Area Chair /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2024/span> NeurIPS Datasets and Benchmarks Area Chair /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2024/span> AAAI Senior Program Committee /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023/span> NeurIPS Area Chair /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023/span> NeurIPS Datasets and Benchmarks Area Chair /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023/span> AAAI Senior Program Committee /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttps://sites.google.com/view/lmca2020/home>NeurIPS Learning Meets Combinatorial Optimization Workshop Organizer/a> /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttps://anucvml.github.io/ddn-cvprw2020/>CVPR Deep Declarative Networks Workshop Organizer/a> /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttps://anucvml.github.io/ddn-eccvt2020/>ECCV Deep Declarative Networks Tutorial Organizer/a> /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2014 - 2015/span> CMU CSD MS Admissions /td>/tr>/table>h3 idreviewing>Reviewing/h3>table classtable table-hover>tr> td stylepadding-right:0;>AAAI Conference on Artificial Intelligence/td>/tr>tr> td stylepadding-right:0;>American Controls Conference (ACC)/td>/tr>tr> td stylepadding-right:0;>IEEE Conference on Computer Vision and Pattern Recognition (CVPR)/td>/tr>tr> td stylepadding-right:0;>IEEE Conference on Decision and Control (CDC)/td>/tr>tr> td stylepadding-right:0;>IEEE Control Systems Letters (L-CSS)/td>/tr>tr> td stylepadding-right:0;>IEEE International Conference on Computer Vision (ICCV)/td>/tr>tr> td stylepadding-right:0;>IEEE International Conference on Intelligent Robots and Systems (IROS)/td>/tr>tr> td stylepadding-right:0;>IEEE International Conference on Robotics and Automation (ICRA)/td>/tr>tr> td stylepadding-right:0;>International Conference on the Constraint Programming, AI, and Operations Research (CPAIOR)/td>/tr>tr> td stylepadding-right:0;>International Conference on Learning Representations (ICLR)/td>/tr>tr> td stylepadding-right:0;>International Conference on Machine Learning (ICML)/td>/tr>tr> td stylepadding-right:0;>International Conference on Machine Learning (ICML) SODS Workshop/td>/tr>tr> td stylepadding-right:0;>Journal of Machine Learning Research (JMLR)/td>/tr>tr> td stylepadding-right:0;>Learning for Dynamics and Control (L4DC)/td>/tr>tr> td stylepadding-right:0;>Mathematical Programming Computation (MPC)/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS)/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS) OPT Workshop/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS) DiffCVGP Workshop/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS) Deep RL Workshop/td>/tr>tr> td stylepadding-right:0;>Optimization Letters/td>/tr>tr> td stylepadding-right:0;>Transactions on Machine Learning Research (TMLR)/td>/tr>/table>h2 id-teaching>i classfa fa-chevron-right>/i> Teaching/h2>table classtable table-hover>tr> td stylepadding-right:0>strong>Applied Machine Learning/strong> (Cornell Tech CS5785), Co-instructor/td> td classcol-md-2 styletext-align:right; padding-left:0;>F2024/td>/tr>tr> td stylepadding-right:0>strong>Graduate AI/strong> (CMU 15-780), TA/td> td classcol-md-2 styletext-align:right; padding-left:0;>S2017/td>/tr>tr> td stylepadding-right:0>strong>Distributed Systems/strong> (CMU 15-440/640), TA/td> td classcol-md-2 styletext-align:right; padding-left:0;>S2016/td>/tr>tr> td stylepadding-right:0>strong>Software Design and Data Structures/strong> (VT CS2114), TA/td> td classcol-md-2 styletext-align:right; padding-left:0;>S2013/td>/tr>/table>h2 id-skills>i classfa fa-chevron-right>/i> Skills/h2>table classtable table-hover>tr> td classcol-md-2>Programming/td> td>C, C++, Fortran, Haskell, Java, Lua, Make, Mathematica, Python, R, Scala /td>/tr>tr> td classcol-md-2>Frameworks/td> td>JAX, NumPy, Pandas, PyTorch, SciPy, TensorFlow, Torch7 /td>/tr>tr> td classcol-md-2>Toolbox/td> td>Linux, emacs, vim, evil, org, mu4e, xmonad, git, tmux, zsh /td>/tr>/table>!--## i classfa fa-chevron-right>/i> Recent Blog Poststable classtable table-hover> tr> td>a href/2016/08/09/deep-completion/>Image Completion with Deep Learning in TensorFlow/a>/td> td classcol-md-3 styletext-align: right;>August 9, 2016/td> /tr> /table>h4>a href/blog>View all/a>/h4>## i classfa fa-chevron-right>/i> Fun Side Projects+ CS conference tracker(https://github.com/bamos/conference-tracker).+ SnowGlobe(https://github.com/bamos/snowglobe): Haskell-driven, small-scale web analytics with minimal configuration.+ My reading list(http://bamos.github.io/reading-list/): YAML data and hosted on GitHub pages.+ dotfiles(https://github.com/bamos/dotfiles): ♥ Arch Linux(https://www.archlinux.org/), OSX, mutt(http://www.mutt.org/), xmonad(http://xmonad.org/), i3(https://i3wm.org/), vim(http://www.vim.org/), emacs(https://www.gnu.org/software/emacs/), zsh(http://www.zsh.org/), mpv(http://mpv.io/), cmus(https://cmus.github.io/).+ girl(https://github.com/bamos/girl): Scala program to find broken links in GitHub projects.+ zsh-history-analysis(https://github.com/bamos/zsh-history-analysis): Analyze shell usage patterns with Python and R.+ python-scripts(https://github.com/bamos/python-scripts): Short and fun Python scripts.+ This website(https://github.com/bamos/bamos.github.io): Built with Jekyll and hosted on GitHub pages.+ cv(https://github.com/bamos/cv): Python-driven resume-curriculum vitae with Jinja templates.+ yaml-mailer(https://github.com/bamos/yaml-mailer): Email many people different messages.+ latex-templates(https://github.com/bamos/latex-templates) and beamer-snippets(https://github.com/bamos/beamer-snippets): Personal collection and previewing of LaTeX and Beamer snippets. Admittedly, I now use Keynote for presentations.-->hr />p>Last updated on 2024-09-25/p> /div> /div>/div> script src/js/sp.js>/script> script src/vendor/js/bootstrap.min.js>/script> script src/vendor/js/anchor.min.js>/script> script src/vendor/js/jquery.toc.js>/script> script typetext/javascript> (function(i,s,o,g,r,a,m){iGoogleAnalyticsObjectr;irir||function(){ (ir.qir.q||).push(arguments)},ir.l1*new Date();as.createElement(o), ms.getElementsByTagName(o)0;a.async1;a.srcg;m.parentNode.insertBefore(a,m) })(window,document,script,https://www.google-analytics.com/analytics.js,ga); ga(create, UA-102191838-1, auto); ga(send, pageview); // $(#toc).toc({ // headings: h2,h3 // }); // anchors.add(h2,h3); /script>/body>/html>
Port 443
HTTP/1.1 200 OKConnection: keep-aliveContent-Length: 189520Server: GitHub.comContent-Type: text/html; charsetutf-8permissions-policy: interest-cohort()Last-Modified: Wed, 25 Sep 2024 17:48:57 GMTAccess-Control-Allow-Origin: *ETag: 66f44d09-2e450expires: Sun, 06 Oct 2024 00:32:01 GMTCache-Control: max-age600x-proxy-cache: MISSX-GitHub-Request-Id: DDEE:3FD743:213EB1F:22421B5:6701D829Accept-Ranges: bytesAge: 0Date: Sun, 06 Oct 2024 01:21:26 GMTVia: 1.1 varnishX-Served-By: cache-bfi-kbfi7400055-BFIX-Cache: HITX-Cache-Hits: 0X-Timer: S1728177686.372498,VS0,VE87Vary: Accept-EncodingX-Fastly-Request-ID: 916dde0612b28f61253b2d9f98fdd20938dc38ed !DOCTYPE html>html xmlnshttp://www.w3.org/1999/xhtml xml:langen langen>head> meta charsetutf-8> title>Brandon Amos/title> meta nameauthor contentBrandon Amos /> meta namedescription content /> meta nameviewport contentwidthdevice-width, initial-scale1, maximum-scale1> link relalternate typeapplication/rss+xml href/atom.xml /> link href/vendor/css/bootstrap.min.css relstylesheet> link relstylesheet hrefhttps://cdn.jsdelivr.net/npm/fork-awesome@1.2.0/css/fork-awesome.min.css integritysha256-XoaMnoYC5TH6/+ihMEnospgm0J1PM/nioxbOUdnM8HY crossoriginanonymous> link href/vendor/css/academicons.min.css relstylesheet> link href/vendor/pygments/default.css relstylesheet> link href/css/bamos.css relstylesheet> link href/css/sharingbuttons.css relstylesheet> script src/vendor/js/jquery.min.js>/script> meta nameviewport contentwidthdevice-width, initial-scale1>/head>body> div classnavbar navbar-default navbar-fixed-top> div classcontainer> div classrow> div classcol-md-10 col-md-offset-1> div classnavbar-header> div classnavbar-brand> a href/images/me.jpg>img src/images/me-face.jpg classimg-circle>/img>/a> a href/>Brandon Amos/a> /div> button classnavbar-toggle typebutton data-togglecollapse data-target#navbar-main> span classicon-bar>/span> span classicon-bar>/span> span classicon-bar>/span> /button> /div> div classnavbar-collapse collapse idnavbar-main> ul classnav navbar-nav> li> a href/>About/a> /li> li> a href/blog/>Blog/a> /li> /ul> !-- ul classnav navbar-nav navbar-right stylefont-size: 1.5em> --> !-- li> --> !-- a hrefhttp://github.com/bamos target_blank> --> !-- i classfa fa-lg fa-github>/i>/a> --> !-- /li> --> !-- li> --> !-- a hrefhttp://twitter.com/brandondamos target_blank> --> !-- i classfa fa-lg fa-twitter>/i>/a> --> !-- /li> --> !-- li> --> !-- a hrefhttps://scholar.google.com/citations?userd8gdZR4AAAAJ target_blank> --> !-- i classai ai-google-scholar>/i>/a> --> !-- /li> --> !-- li> --> !-- a href/atom.xml target_blank> --> !-- i classfa fa-rss>/i>/a> --> !-- /li> --> !-- /ul> --> /div> /div> /div> /div> /div>div classcontainer> div classrow> div classcol-md-6 col-md-offset-1 vcenter idxHdr> div stylefont-size: 2em; color: #4582ec; font-weight: bold; padding-bottom: 0.3em;>Brandon Amos/div> div stylefont-size: 1.2em;> Research Scientist /div> div stylefont-size: 1.2em> a hrefhttps://ai.facebook.com/>Meta (FAIR)/a> /div> div stylefont-size: 1.2em> a hrefmailto:bda@meta.com>bda@meta.com/a> /div> br/> div stylepadding: 0.3em; background-color: #4582ec; display: inline-block; border-radius: 4px; font-size: 1.2em;> a hrefdata/cv.pdf target_blank styletext-decoration: none;> i stylecolor: white classfa fa-download>/i> /a> a hrefhttps://github.com/bamos/cv target_blank styletext-decoration: none;> i stylecolor: white classfa fa-code-fork>/i> /a> a hrefdata/cv.pdf target_blank stylecolor: white; text-decoration: none;>CV/a> /div> ul classlist-inline idxIcons stylefont-size: 1.9em; margin-top: 0.5em;> li> a hrefhttp://github.com/bamos target_blank> i classfa fa-fw fa-github>/i>/a> /li> li> a hrefhttp://twitter.com/brandondamos target_blank> i classfa fa-fw>đ/i>/a> /li> !-- li> a relme hrefhttps://sigmoid.social/@bamos target_blank> i classfa fa-fw fa-mastodon>/i>/a> /li> --> li> a hrefhttps://scholar.google.com/citations?userd8gdZR4AAAAJ target_blank> i classai ai-google-scholar>/i>/a> /li> li> a hrefhttp://www.facebook.com/bdamos target_blank> i classfa fa-fw fa-facebook>/i>/a> /li> li> a hrefhttp://www.linkedin.com/in/bdamos target_blank> i classfa fa-fw fa-linkedin>/i>/a> /li> li> a href/atom.xml target_blank> i classfa fa-fw fa-rss>/i>/a> /li> /ul> /div> div classcol-md-2 vcenter idxHdr> img src/images/me/2021-wave.gif styleborder-radius: 20px; margin: 10px; max-width: none; altMe./> /div> /div> div classrow> div classcol-md-12>hr />p alignjustify>I am a research scientist in theb>Fundamental AI Research (FAIR)/b>group atb>Meta/b> in NYC and alsoteach machine learning at b>Cornell Tech/b>.I study foundational topics in b>machine learning/b> andb>optimization/b>, recently involvingreinforcement learning, control, optimal transport, and geometry.My research is on learning systems that understand and interact with our worldand focuses on integrating structural information and domain knowledge intothese systems to represent non-trivial reasoning operations./p>p>br />/p>h2 id-education>i classfa fa-chevron-right>/i> Education/h2>table classtable table-hover> tr> td> span classcvdate>2014 - 2019/span> strong>Ph.D. in Computer Science/strong>, em>Carnegie Mellon University/em> (0.00/0.00) br /> p stylemargin-top:-1em;margin-bottom:0em> br /> Thesis: a hrefhttps://github.com/bamos/thesis target_blank>i>Differentiable Optimization-Based Modeling for Machine Learning/i>/a> br /> Advisor: a hrefhttps://zicokolter.com target_blank>J. Zico Kolter/a> /p> /td> /tr> tr> td> span classcvdate>2011 - 2014/span> strong>B.S. in Computer Science/strong>, em>Virginia Tech/em> (3.99/4.00) br /> /td> /tr>/table>h2 id-previous-positions>i classfa fa-chevron-right>/i> Previous Positions/h2>table classtable table-hover>tr> td stylepadding-right:0;>span classcvdate>2016 - 2019/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Carnegie Mellon University/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://zicokolter.com target_blank>J. Zico Kolter/a> on ML and optimization)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2018/span>p stylemargin: 0>strong>Research Intern/strong>, em>Intel Labs/em>, Santa Claraspan stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttp://vladlen.info/ target_blank>Vladlen Koltun/a> on computer vision)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2017/span>p stylemargin: 0>strong>Research Intern/strong>, em>Google DeepMind/em>, Londonspan stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://scholar.google.com/citations?usernzEluBwAAAAJ target_blank>Nando de Freitas/a> and a hrefhttp://mdenil.com/ target_blank>Misha Denil/a> on RL)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2014 - 2016/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Carnegie Mellon University/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a> on mobile systems)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2014/span>p stylemargin: 0>strong>Research Intern/strong>, em>Adobe Research/em>, San Josespan stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://research.adobe.com/person/david-tompkins/ target_blank>David Tompkins/a> on distributed systems)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2013 - 2014/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Virginia Tech/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://people.cs.vt.edu/~ltw/shortvita.html target_blank>Layne Watson/a> and a hrefhttps://dblp.org/pid/75/8682.html target_blank>David Easterling/a> on optimization)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2012 - 2014/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Virginia Tech/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://www.magnum.io/people/jules.html target_blank>Jules White/a> and a hrefhttps://scholar.google.com/citations?userMRKab9cAAAAJ target_blank>Hamilton Turner/a> on mobile systems)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2012 - 2014/span>p stylemargin: 0>strong>Research Assistant/strong>, em>Virginia Tech/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(with a hrefhttps://www.ssrg.ece.vt.edu/ target_blank>Binoy Ravindran/a> and a hrefhttps://scholar.google.com/citations?userUG5yHRIAAAAJ target_blank>Alastair Murray/a> on compilers)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2013 - 2014/span>p stylemargin: 0>strong>Software Intern/strong>, em>Snowplow/em>span stylecolor:grey;font-size:1.3rem;margin: 0>(Scala development)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2013/span>p stylemargin: 0>strong>Software Intern/strong>, em>Qualcomm/em>, San Diegospan stylecolor:grey;font-size:1.3rem;margin: 0>(Python and C++ development)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2012/span>p stylemargin: 0>strong>Software Intern/strong>, em>Phoenix Integration/em>, Virginiaspan stylecolor:grey;font-size:1.3rem;margin: 0>(C++, C#, and Java development)/span>/p> /td>/tr>tr> td stylepadding-right:0;>span classcvdate>2011/span>p stylemargin: 0>strong>Network Administrator Intern/strong>, em>Sunapsys/em>, Virginia/p> /td>/tr>/table>h2 id-honors--awards>i classfa fa-chevron-right>/i> Honors & Awards/h2>table classtable table-hover>tr> td> div stylefloat: right>2022/div> div> a hrefhttps://neurips.cc/Conferences/2022/ProgramCommittee>NeurIPS Top Reviewer/a> /div> /td> !-- td classcol-md-2 styletext-align:right;>2022/td> -->/tr>tr> td> div stylefloat: right>2022/div> div> a hrefhttps://icml.cc/Conferences/2022/Reviewers>ICML Outstanding Reviewer/a> /div> /td> !-- td classcol-md-2 styletext-align:right;>2022/td> -->/tr>tr> td> div stylefloat: right>2019/div> div> a hrefhttps://iclr.cc/Conferences/2019/Awards>ICLR Outstanding Reviewer/a> /div> /td> !-- td classcol-md-2 styletext-align:right;>2019/td> -->/tr>tr> td> div stylefloat: right>2016 - 2019/div> div> NSF Graduate Research Fellowship /div> /td> !-- td classcol-md-2 styletext-align:right;>2016 - 2019/td> -->/tr>tr> td> div stylefloat: right>2011 - 2014/div> div> Nine undergraduate scholarships br />p stylecolor:grey;font-size:1.2rem>Roanoke County Public Schools Engineering,Salem-Roanoke County Chamber of Commerce,Papa Johns,Scottish Rite of Freemasonry,VT Intelligence Community Conter for Academic Excellence,VT Pamplin Leader,VT Benjamin F. Bock, VT Gay B. Shober, VT I. Luck Gravett/p> /div> /td> !-- td classcol-md-2 styletext-align:right;>2011 - 2014/td> -->/tr>/table>h2 id-publications>i classfa fa-chevron-right>/i> Publications/h2>p>a hrefhttps://scholar.google.com/citations?userd8gdZR4AAAAJ>Google Scholar/a>: 8.9k+ citations and an h-index of 37 br />Selected publications I am a primary author on are span stylebackground-color: #ffffd0>highlighted./span>/p>p>br />/p>h2>2024/h2>table classtable table-hover>tr idtr-paulus2024advprompter stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>1./td>td>a hrefhttps://arxiv.org/abs/2404.16873 target_blank>img srcimages/publications/paulus2024advprompter.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2404.16873 target_blank>AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs/a> /em> a hrefjavascript:; onclick$("#abs_paulus2024advprompter").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/advprompter target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?usernjZL5CQAAAAJ target_blank>Anselm Paulus*/a>, a hrefhttps://arman-z.github.io/ target_blank>Arman Zharmagambetov*/a>, a hrefhttps://sites.google.com/view/chuanguo target_blank>Chuan Guo/a>, strong>Brandon Amossup>†/sup>/strong>, and a hrefhttps://yuandong-tian.com/ target_blank>Yuandong Tiansup>†/sup>/a>br />arXiv 2024 br />div idabs_paulus2024advprompter styletext-align: justify; display: none> p>While recently Large Language Models (LLMs) have achieved remarkable successes, they are vulnerable to certain jailbreaking attacks that lead to generation of inappropriate or harmful content. Manual red-teaming requires finding adversarial prompts that cause such jailbreaking, e.g. by appending a suffix to a given instruction, which is inefficient and time-consuming.On the other hand, automatic adversarial prompt generation often leads to semantically meaningless attacks that can easily be detected by perplexity-based filters, may require gradient information from the target LLM, or do not scale well due to time-consuming discrete optimization processes over the token space. In this paper, we present a novel method that uses another LLM, called the AdvPrompter, to generate human-readable adversarial prompts in seconds, approximately 800 times faster than existing optimization-based approaches.We train the AdvPrompter using a novel algorithm that does not require access to the gradients of the target LLM. This process alternates between two steps: (1) generating high-quality target adversarial suffixes by optimizing the AdvPrompter predictions, and (2) low-rank fine-tuning of the AdvPrompter with the generated adversarial suffixes. The trained AdvPrompter generates suffixes that veil the input instruction without changing its meaning, such that the target LLM is lured to give a harmful response. Experimental results on popular open source target LLMs show state-of-the-art results on the AdvBench dataset, that also transfer to closed-source black-box LLM APIs. Further, we demonstrate that by fine-tuning on a synthetic dataset generated by Advprompter, LLMs can be made more robust against jailbreaking attacks while maintaining performance, i.e. high MMLU scores./p> /div>/td>/tr>tr idtr-pooladian2024neural stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>2./td>td>a hrefhttps://arxiv.org/abs/2406.00288 target_blank>img srcimages/publications/pooladian2024neural.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2406.00288 target_blank>Neural Optimal Transport with Lagrangian Costs/a> /em> a hrefjavascript:; onclick$("#abs_pooladian2024neural").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/lagrangian-ot target_blank>code/a> br />a hrefhttps://arampooladian.com/ target_blank>Aram-Alexandre Pooladian/a>, a hrefhttps://cdenrich.github.io/ target_blank>Carles Domingo-Enrich/a>, a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, and strong>Brandon Amos/strong>br />UAI 2024 br />div idabs_pooladian2024neural styletext-align: justify; display: none> p>We investigate the optimal transport problem between probability measures when the underlying cost function is understood to satisfy a least action principle, also known as a Lagrangian cost. These generalizations are useful when connecting observations from a physical system, where the transport dynamics are influenced by the geometry of the system, such as obstacles, (e.g., incorporating barrier functions in the Lagrangian) and allows practitioners to incorporate a priori knowledge of the underlying system such as non-Euclidean geometries (e.g., paths must be circular). Our contributions are of computational interest, where we demonstrate the ability to efficiently compute geodesics and amortize spline-based paths, which has not been done before, even in low dimensional problems. Unlike prior work, we also output the resulting Lagrangian optimal transport map without requiring an ODE solver. We demonstrate the effectiveness of our formulation on low-dimensional examples taken from prior work./p> /div>/td>/tr>tr idtr-sambharya2024learning>td alignright stylepadding-left:0;padding-right:0;>3./td>td>a hrefhttps://arxiv.org/abs/2309.07835 target_blank>img srcimages/publications/sambharya2024learning.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2309.07835 target_blank>Learning to Warm-Start Fixed-Point Optimization Algorithms/a> /em> a hrefjavascript:; onclick$("#abs_sambharya2024learning").toggle()>abs/a> a hrefhttps://github.com/stellatogrp/l2ws target_blank>code/a> br />a hrefhttps://rajivsambharya.github.io/ target_blank>Rajiv Sambharya/a>, a hrefhttps://sites.google.com/view/georgina-hall target_blank>Georgina Hall/a>, strong>Brandon Amos/strong>, and a hrefhttps://stellato.io/ target_blank>Bartolomeo Stellato/a>br />JMLR 2024 br />div idabs_sambharya2024learning styletext-align: justify; display: none> p>We introduce a machine-learning framework to warm-start fixed-point optimization algorithms. Our architecture consists of a neural network mapping problem parameters to warm starts, followed by a predefined number of fixed-point iterations. We propose two loss functions designed to either minimize the fixed-point residual or the distance to a ground truth solution. In this way, the neural network predicts warm starts with the end-to-end goal of minimizing the downstream loss. An important feature of our architecture is its flexibility, in that it can predict a warm start for fixed-point algorithms run for any number of steps, without being limited to the number of steps it has been trained on. We provide PAC-Bayes generalization bounds on unseen data for common classes of fixed-point operators: contractive, linearly convergent, and averaged. Applying this framework to well-known applications in control, statistics, and signal processing, we observe a significant reduction in the number of iterations and solution time required to solve these problems, through learned warm starts./p> /div>/td>/tr>tr idtr-lotfi2024unlocking>td alignright stylepadding-left:0;padding-right:0;>4./td>td>a hrefhttps://arxiv.org/abs/2407.18158 target_blank>img srcimages/publications/lotfi2024unlocking.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2407.18158 target_blank>Unlocking Tokens as Data Points for Generalization Bounds on Larger Language Models/a> /em> a hrefjavascript:; onclick$("#abs_lotfi2024unlocking").toggle()>abs/a>br />a hrefhttps://sanaelotfi.github.io/ target_blank>Sanae Lotfi/a>, a hrefhttps://yilunkuang.github.io/ target_blank>Yilun Kuang/a>, a hrefhttps://mfinzi.github.io/ target_blank>Marc Anton Finzi/a>, strong>Brandon Amos/strong>, a hrefhttps://goldblum.github.io/ target_blank>Micah Goldblum/a>, and a hrefhttps://cims.nyu.edu/~andrewgw/ target_blank>Andrew Gordon Wilson/a>br />NeurIPS 2024 br />div idabs_lotfi2024unlocking styletext-align: justify; display: none> p>Large language models (LLMs) with billions of parameters excel atpredicting the next token in a sequence. Recent workcomputes non-vacuous compression-basedgeneralization bounds for LLMs, but these bounds arevacuous for large models at the billion-parameterscale. Moreover, these bounds are obtained throughrestrictive compression techniques, boundingcompressed models that generate low-qualitytext. Additionally, the tightness of these existingbounds depends on the number of IID documents in atraining set rather than the much larger number ofnon-IID constituent tokens, leaving untappedpotential for tighter bounds. In this work, weinstead use properties of martingales to derivegeneralization bounds that benefit from the vastnumber of tokens in LLM training sets. Since adataset contains far more tokens than documents, ourgeneralization bounds not only tolerate but actuallybenefit from far less restrictive compressionschemes. With Monarch matrices, Kroneckerfactorizations, and post-training quantization, weachieve non-vacuous generalization bounds for LLMsas large as LLaMA2-70B. Unlike previous approaches, our work achieves the first non-vacuous bounds formodels that are deployed in practice and generatehigh-quality text./p> /div>/td>/tr>tr idtr-domingoenrich2024stochastic>td alignright stylepadding-left:0;padding-right:0;>5./td>td>a hrefhttps://arxiv.org/abs/2312.02027 target_blank>img srcimages/publications/domingoenrich2024stochastic.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2312.02027 target_blank>Stochastic Optimal Control Matching/a> /em> a hrefjavascript:; onclick$("#abs_domingoenrich2024stochastic").toggle()>abs/a>br />a hrefhttps://cdenrich.github.io/ target_blank>Carles Domingo-Enrich/a>, a hrefhttps://scholar.google.com/citations?userel5gT4AAAAAJ target_blank>Jiequn Han/a>, strong>Brandon Amos/strong>, a hrefhttps://cims.nyu.edu/~bruna/ target_blank>Joan Bruna/a>, and a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>br />NeurIPS 2024 br />div idabs_domingoenrich2024stochastic styletext-align: justify; display: none> p>Stochastic optimal control, which has the goal of driving the behavior of noisy systems, is broadly applicable in science, engineering and artificial intelligence. Our work introduces Stochastic Optimal Control Matching (SOCM), a novel Iterative Diffusion Optimization (IDO) technique for stochastic optimal control that stems from the same philosophy as the conditional score matching loss for diffusion models. That is, the control is learned via a least squares problem by trying to fit a matching vector field. The training loss, which is closely connected to the cross-entropy loss, is optimized with respect to both the control function and a family of reparameterization matrices which appear in the matching vector field. The optimization with respect to the reparameterization matrices aims at minimizing the variance of the matching vector field. Experimentally, our algorithm achieves lower error than all the existing IDO techniques for stochastic optimal control for four different control settings. The key idea underlying SOCM is the path-wise reparameterization trick, a novel technique that is of independent interest, e.g., for generative modeling./p> /div>/td>/tr>tr idtr-atanackovic2024meta>td alignright stylepadding-left:0;padding-right:0;>6./td>td>a hrefhttps://arxiv.org/abs/2408.14608 target_blank>img srcimages/publications/atanackovic2024meta.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2408.14608 target_blank>Meta Flow Matching: Integrating Vector Fields on the Wasserstein Manifold/a> /em> a hrefjavascript:; onclick$("#abs_atanackovic2024meta").toggle()>abs/a>br />a hrefhttps://lazaratan.github.io/ target_blank>Lazar Atanackovic/a>, a hrefhttps://scholar.google.com/citations?userCblgXekAAAAJ target_blank>Xi Zhang/a>, strong>Brandon Amos/strong>, a hrefhttps://www.cs.mcgill.ca/~blanchem/ target_blank>Mathieu Blanchette/a>, a hrefhttps://scholar.google.ca/citations?userDN3LoTEAAAAJ target_blank>Leo J Lee/a>, a hrefhttps://yoshuabengio.org/profile/ target_blank>Yoshua Bengio/a>, a hrefhttps://www.alextong.net/ target_blank>Alexander Tong/a>, and a hrefhttps://necludov.github.io/ target_blank>Kirill Neklyudov/a>br />ICML GRaM Workshop 2024 br />div idabs_atanackovic2024meta styletext-align: justify; display: none> p>Numerous biological and physical processes can be modeled as systems of interacting entities evolving continuously over time, e.g. the dynamics of communicating cells or physical particles. Learning the dynamics of such systems is essential for predicting the temporal evolution of populations across novel samples and unseen environments. Flow-based models allow for learning these dynamics at the population level - they model the evolution of the entire distribution of samples. However, current flow-based models are limited to a single initial population and a set of predefined conditions which describe different dynamics. We argue that multiple processes in natural sciences have to be represented as vector fields on the Wasserstein manifold of probability densities. That is, the change of the population at any moment in time depends on the population itself due to the interactions between samples. In particular, this is crucial for personalized medicine where the development of diseases and their respective treatment response depends on the microenvironment of cells specific to each patient. We propose Meta Flow Matching (MFM), a practical approach to integrating along these vector fields on the Wasserstein manifold by amortizing the flow model over the initial populations. Namely, we embed the population of samples using a Graph Neural Network (GNN) and use these embeddings to train a Flow Matching model. This gives MFM the ability to generalize over the initial distributions unlike previously proposed methods. We demonstrate the ability of MFM to improve prediction of individual treatment responses on a large scale multi-patient single-cell drug screen dataset./p> /div>/td>/tr>tr idtr-silvestri2024score>td alignright stylepadding-left:0;padding-right:0;>7./td>td>a hrefhttps://arxiv.org/abs/2307.05213 target_blank>img srcimages/publications/silvestri2024score.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2307.05213 target_blank>Score Function Gradient Estimation to Widen the Applicability of Decision-Focused Learning/a> /em> a hrefjavascript:; onclick$("#abs_silvestri2024score").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?useryHEb8eAAAAAJ target_blank>Mattia Silvestri/a>, a hrefhttps://scholar.google.com/citations?usersMtjmx4AAAAJ target_blank>Senne Berden/a>, a hrefhttps://jayantamandi.com/ target_blank>Jayanta Mandi/a>, a hrefhttps://scholar.google.com/citations?usermuyZLrYAAAAJ target_blank>Ali Ä°rfan MahmutoÄulları/a>, strong>Brandon Amos/strong>, a hrefhttps://people.cs.kuleuven.be/~tias.guns/ target_blank>Tias Guns/a>, and a hrefhttps://scholar.google.com/citations?userlJJ6EOMAAAAJ target_blank>Michele Lombardi/a>br />arXiv 2024 br />div idabs_silvestri2024score styletext-align: justify; display: none> p>Many real-world optimization problems contain parameters that are unknown before deployment time, either due to stochasticity or to lack of information (e.g., demand or travel times in delivery problems). A common strategy in such cases is to estimate said parameters via machine learning (ML) models trained to minimize the prediction error, which however is not necessarily aligned with the downstream task-level error. The decision-focused learning (DFL) paradigm overcomes this limitation by training to directly minimize a task loss, e.g. regret. Since the latter has non-informative gradients for combinatorial problems, state-of-the-art DFL methods introduce surrogates and approximations that enable training. But these methods exploit specific assumptions about the problem structures (e.g., convex or linear problems, unknown parameters only in the objective function). We propose an alternative method that makes no such assumptions, it combines stochastic smoothing with score function gradient estimation which works on any task loss. This opens up the use of DFL methods to nonlinear objectives, uncertain parameters in the problem constraints, and even two-stage stochastic optimization. Experiments show that it typically requires more epochs, but that it is on par with specialized methods and performs especially well for the difficult case of problems with uncertainty in the constraints, in terms of solution quality, scalability, or both./p> /div>/td>/tr>/table>h2>2023/h2>table classtable table-hover>tr idtr-amos2023tutorial stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>8./td>td>a hrefhttps://arxiv.org/abs/2202.00665 target_blank>img srcimages/publications/amos2023tutorial.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2202.00665 target_blank>Tutorial on amortized optimization/a> /em> a hrefjavascript:; onclick$("#abs_amos2023tutorial").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/amortized-optimization-tutorial target_blank>code/a> br />strong>Brandon Amos/strong>br />Foundations and Trends in Machine Learning 2023 br />div idabs_amos2023tutorial styletext-align: justify; display: none> p>Optimization is a ubiquitous modeling tool and is often deployedin settings which repeatedly solve similar instancesof the same problem. Amortized optimization methodsuse learning to predict the solutions to problems inthese settings, exploiting the shared structurebetween similar problem instances. These methodshave been crucial in variational inference andreinforcement learning and are capable of solvingoptimization problems many orders of magnitudestimes faster than traditional optimization methodsthat do not use amortization. This tutorial presentsan introduction to the amortized optimizationfoundations behind these advancements and overviewstheir applications in variational inference, sparsecoding, gradient-based meta-learning, control, reinforcement learning, convex optimization, optimaltransport, and deep equilibrium networks./p> /div>/td>/tr>tr idtr-amos2023amortizing stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>9./td>td>a hrefhttps://arxiv.org/abs/2210.12153 target_blank>img srcimages/publications/amos2023amortizing.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2210.12153 target_blank>On amortizing convex conjugates for optimal transport/a> /em> a hrefjavascript:; onclick$("#abs_amos2023amortizing").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/w2ot target_blank>code/a> br />strong>Brandon Amos/strong>br />ICLR 2023 br />div idabs_amos2023amortizing styletext-align: justify; display: none> p>This paper focuses on computing the convex conjugate operation thatarises when solving Euclidean Wasserstein-2 optimaltransport problems. This conjugation, which is alsoreferred to as the Legendre-Fenchel conjugate orc-transform, is considered difficult to compute andin practice, Wasserstein-2 methods are limited bynot being able to exactly conjugate the dualpotentials in continuous space. I show thatcombining amortized approximations to the conjugatewith a solver for fine-tuning is computationallyeasy. This combination significantly improves thequality of transport maps learned for theWasserstein-2 benchmark by Korotin et al. (2021) andis able to model many 2-dimensional couplings andflows considered in the literature./p> /div>/td>/tr>tr idtr-sambharya2023l2a>td alignright stylepadding-left:0;padding-right:0;>10./td>td>a hrefhttps://arxiv.org/abs/2212.08260 target_blank>img srcimages/publications/sambharya2023l2a.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2212.08260 target_blank>End-to-End Learning to Warm-Start for Real-Time Quadratic Optimization/a> /em> a hrefjavascript:; onclick$("#abs_sambharya2023l2a").toggle()>abs/a> a hrefhttps://github.com/stellatogrp/l2ws target_blank>code/a> br />a hrefhttps://rajivsambharya.github.io/ target_blank>Rajiv Sambharya/a>, a hrefhttps://sites.google.com/view/georgina-hall target_blank>Georgina Hall/a>, strong>Brandon Amos/strong>, and a hrefhttps://stellato.io/ target_blank>Bartolomeo Stellato/a>br />L4DC 2023 br />div idabs_sambharya2023l2a styletext-align: justify; display: none> p>First-order methods are widely used to solve convex quadratic programs(QPs) in real-time applications because of their lowper-iteration cost. However, they can suffer fromslow convergence to accurate solutions. In thispaper, we present a framework which learns aneffective warm-start for a popular first-ordermethod in real-time applications, Douglas-Rachford(DR) splitting, across a family of parametricQPs. This framework consists of two modules: afeedforward neural network block, which takes asinput the parameters of the QP and outputs awarm-start, and a block which performs a fixednumber of iterations of DR splitting from thiswarm-start and outputs a candidate solution. A keyfeature of our framework is its ability to doend-to-end learning as we differentiate through theDR iterations. To illustrate the effectiveness ofour method, we provide generalization bounds (basedon Rademacher complexity) that improve with thenumber of training problems and number of iterationssimultaneously. We further apply our method to threereal-time applications and observe that, by learninggood warm-starts, we are able to significantlyreduce the number of iterations required to obtainhigh-quality solutions./p> /div>/td>/tr>tr idtr-amos2023meta stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>11./td>td>a hrefhttps://arxiv.org/abs/2206.05262 target_blank>img srcimages/publications/amos2023meta.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2206.05262 target_blank>Meta Optimal Transport/a> /em> a hrefjavascript:; onclick$("#abs_amos2023meta").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/meta-ot target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, a hrefhttps://giulslu.github.io/ target_blank>Giulia Luise/a>, and a hrefhttps://ievred.github.io target_blank>Ievgen Redko/a>br />ICML 2023 br />div idabs_amos2023meta styletext-align: justify; display: none> p>We study the use of amortized optimization to predict optimaltransport (OT) maps from the input measures, whichwe call Meta OT. This helps repeatedly solve similarOT problems between different measures by leveragingthe knowledge and information present from pastproblems to rapidly predict and solve newproblems. Otherwise, standard methods ignore theknowledge of the past solutions and suboptimallyre-solve each problem from scratch. Meta OT modelssurpass the standard convergence rates oflog-Sinkhorn solvers in the discrete setting andconvex potentials in the continuous setting. Weimprove the computational time of standard OTsolvers by multiple orders of magnitude in discreteand continuous transport settings between images, spherical data, and color palettes./p> /div>/td>/tr>tr idtr-pooladian2023multisample>td alignright stylepadding-left:0;padding-right:0;>12./td>td>a hrefhttps://arxiv.org/abs/2304.14772 target_blank>img srcimages/publications/pooladian2023multisample.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2304.14772 target_blank>Multisample Flow Matching: Straightening Flows with Minibatch Couplings/a> /em> a hrefjavascript:; onclick$("#abs_pooladian2023multisample").toggle()>abs/a>br />a hrefhttps://arampooladian.com/ target_blank>Aram-Alexandre Pooladian/a>, a hrefhttps://helibenhamu.github.io/ target_blank>Heli Ben-Hamu/a>, a hrefhttps://cdenrich.github.io/ target_blank>Carles Domingo-Enrich/a>, strong>Brandon Amos/strong>, a hrefhttps://www.wisdom.weizmann.ac.il/~ylipman/ target_blank>Yaron Lipman/a>, and a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>br />ICML 2023 br />div idabs_pooladian2023multisample styletext-align: justify; display: none> p>Simulation-free methods for training continuous-time generative modelsconstruct probability paths that go between noisedistributions and individual data samples. Recentworks, such as Flow Matching, derived paths that areoptimal for each data sample. However, thesealgorithms rely on independent data and noisesamples, and do not exploit underlying structure inthe data distribution for constructing probabilitypaths. We propose Multisample Flow Matching, a moregeneral framework that uses non-trivial couplingsbetween data and noise samples while satisfying thecorrect marginal constraints. At very small overheadcosts, this generalization allows us to (i) reducegradient variance during training, (ii) obtainstraighter flows for the learned vector field, whichallows us to generate high-quality samples usingfewer function evaluations, and (iii) obtaintransport maps with lower cost in high dimensions, which has applications beyond generativemodeling. Importantly, we do so in a completelysimulation-free manner with a simple minimizationobjective. We show that our proposed methods improvesample consistency on downsampled ImageNet datasets, and lead to better low-cost sample generation./p> /div>/td>/tr>tr idtr-zheng2023semi>td alignright stylepadding-left:0;padding-right:0;>13./td>td>a hrefhttps://arxiv.org/abs/2210.06518 target_blank>img srcimages/publications/zheng2023semi.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2210.06518 target_blank>Semi-Supervised Offline Reinforcement Learning with Action-Free Trajectories/a> /em> a hrefjavascript:; onclick$("#abs_zheng2023semi").toggle()>abs/a>br />a hrefhttps://enosair.github.io/ target_blank>Qinqing Zheng/a>, a hrefhttps://www.mikaelhenaff.com/ target_blank>Mikael Henaff/a>, strong>Brandon Amos/strong>, and a hrefhttps://aditya-grover.github.io/ target_blank>Aditya Grover/a>br />ICML 2023 br />div idabs_zheng2023semi styletext-align: justify; display: none> p>Natural agents can effectively learn from multiple data sources thatdiffer in size, quality, and types ofmeasurements. We study this heterogeneity in thecontext of offline reinforcement learning (RL) byintroducing a new, practically motivatedsemi-supervised setting. Here, an agent has accessto two sets of trajectories: labelled trajectoriescontaining state, action, reward triplets at everytimestep, along with unlabelled trajectories thatcontain only state and reward information. For thissetting, we develop a simple meta-algorithmicpipeline that learns an inverse-dynamics model onthe labelled data to obtain proxy-labels for theunlabelled data, followed by the use of any offlineRL algorithm on the true and proxy-labelledtrajectories. Empirically, we find this simplepipeline to be highly successful - on several D4RLbenchmarks, certain offline RLalgorithms can match the performance of variantstrained on a fully labeled dataset even when welabel only 10% trajectories from the low returnregime. Finally, we perform a large-scale controlledempirical study investigating the interplay ofdata-centric properties of the labelled andunlabelled datasets, with algorithmic design choices(e.g., inverse dynamics, offline RL algorithm) toidentify general trends and best practices fortraining RL agents on semi-supervised offlinedatasets./p> /div>/td>/tr>tr idtr-bansal2023taskmet stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>14./td>td>a hrefhttps://arxiv.org/abs/2312.05250 target_blank>img srcimages/publications/bansal2023taskmet.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2312.05250 target_blank>TaskMet: Task-Driven Metric Learning for Model Learning/a> /em> a hrefjavascript:; onclick$("#abs_bansal2023taskmet").toggle()>abs/a>br />a hrefhttps://dishank-b.github.io/ target_blank>Dishank Bansal/a>, a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, a hrefhttps://www.mustafamukadam.com/ target_blank>Mustafa Mukadam/a>, and strong>Brandon Amos/strong>br />NeurIPS 2023 br />div idabs_bansal2023taskmet styletext-align: justify; display: none> p>Deep learning models are often used with some downstreamtask. Models solely trained to achieve accuratepredictions may struggle to perform well onthe desired downstream tasks. We propose using thetaskâs loss to learn a metric which parameterizes aloss to train the model.This approach does not alterthe optimal prediction model itself, but ratherchanges the model learning to emphasize theinformation important for the downstream task.Thisenables us to achieve the best of both worlds:aprediction model trained in the original predictionspace while also being valuable for the desireddownstream task.We validate our approach throughexperiments conducted in two main settings: 1)decision-focused model learning scenarios involvingportfolio optimization and budget allocation, and2)reinforcement learning in noisy environments withdistracting states./p> /div>/td>/tr>tr idtr-zharmagambetov2023landscape>td alignright stylepadding-left:0;padding-right:0;>15./td>td>a hrefhttps://arxiv.org/abs/2307.08964 target_blank>img srcimages/publications/zharmagambetov2023landscape.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2307.08964 target_blank>Landscape Surrogate: Learning Decision Losses for Mathematical Optimization Under Partial Information/a> /em> a hrefjavascript:; onclick$("#abs_zharmagambetov2023landscape").toggle()>abs/a>br />a hrefhttps://arman-z.github.io/ target_blank>Arman Zharmagambetov/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userTuVq07oAAAAJ target_blank>Aaron Ferber/a>, a hrefhttps://taoanhuang.github.io/ target_blank>Taoan Huang/a>, a hrefhttps://scholar.google.com/citations?user1jjyaBYAAAAJ target_blank>Bistra Dilkina/a>, and a hrefhttps://yuandong-tian.com/ target_blank>Yuandong Tian/a>br />NeurIPS 2023 br />div idabs_zharmagambetov2023landscape styletext-align: justify; display: none> p>Recent works in learning-integrated optimization have shown promise insettings where the optimization problem is onlypartially observed or where general-purposeoptimizers perform poorly without expert tuning. Bylearning an optimizer g to tackle these challengingproblems with f as the objective, the optimizationprocess can be substantially accelerated byleveraging past experience. Training the optimizercan be done with supervision from known optimalsolutions (not always available) or implicitly byoptimizing the compound function f â g , but theimplicit approach is slow and challenging due tofrequent calls to the optimizer and sparsegradients, particularly for combinatorialsolvers. To address these challenges, we proposeusing a smooth and learnable Landscape SurrogateM instead of composing f with g . This surrogate can be computedfaster than g, provides dense and smooth gradientsduring training, can generalize to unseenoptimization problems, and is efficiently learnedvia alternating optimization. We test our approachon both synthetic problems and real-world problems, achieving comparable or superior objective valuescompared to state-of-the-art baselines whilereducing the number of calls to g . Notably, ourapproach outperforms existing methods forcomputationally expensive high-dimensional problems./p> /div>/td>/tr>tr idtr-retchin2023koopman>td alignright stylepadding-left:0;padding-right:0;>16./td>td>a hrefhttps://differentiable.xyz/papers/paper_45.pdf target_blank>img srcimages/publications/retchin2023koopman.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://differentiable.xyz/papers/paper_45.pdf target_blank>Koopman Constrained Policy Optimization: A Koopman operator theoretic method for differentiable optimal control in robotics/a> /em> a hrefjavascript:; onclick$("#abs_retchin2023koopman").toggle()>abs/a>br />a hrefhttps://www.linkedin.com/in/matthew-retchin/ target_blank>Matthew Retchin/a>, strong>Brandon Amos/strong>, a hrefhttps://www.eigensteve.com/ target_blank>Steven Brunton/a>, and a hrefhttps://shurans.github.io/ target_blank>Shuran Song/a>br />ICML Differentiable Almost Everything Workshop 2023 br />div idabs_retchin2023koopman styletext-align: justify; display: none> p>We introduce Koopman Constrained Policy Optimization (KCPO), combining implicitly differentiable model predictivecontrol with a deep Koopman autoencoder for robotlearning in unknown and nonlinear dynamicalsystems. KCPO is a new policy optimization algorithmthat trains neural policies end-to-end with hard boxconstraints on controls. Guaranteed satisfaction ofhard constraints helps ensure the performance andsafety of robots. We perform imitation learning withKCPO to recover expert policies on the SimplePendulum, Cartpole Swing-Up, Reacher, andDifferential Drive environments, outperformingbaseline methods in generalizing toout-of-distribution constraints in most environmentsafter training./p> /div>/td>/tr>/table>h2>2022/h2>table classtable table-hover>tr idtr-fickinger2021crossdomain stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>17./td>td>a hrefhttps://arxiv.org/abs/2110.03684 target_blank>img srcimages/publications/fickinger2021crossdomain.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2110.03684 target_blank>Cross-Domain Imitation Learning via Optimal Transport/a> /em> a hrefjavascript:; onclick$("#abs_fickinger2021crossdomain").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/gwil target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?userbBFN_qwAAAAJ target_blank>Arnaud Fickinger/a>, a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, a hrefhttp://people.eecs.berkeley.edu/~russell/ target_blank>Stuart Russell/a>, and strong>Brandon Amos/strong>br />ICLR 2022 br />div idabs_fickinger2021crossdomain styletext-align: justify; display: none> p>Cross-domain imitation learning studies how to leverage expertdemonstrations of one agent to train an imitationagent with a different embodiment ormorphology. Comparing trajectories and stationarydistributions between the expert and imitationagents is challenging because they live on differentsystems that may not even have the samedimensionality. We propose Gromov-WassersteinImitation Learning (GWIL), a method for cross-domainimitation that uses the Gromov-Wasserstein distanceto align and compare states between the differentspaces of the agents. Our theory formallycharacterizes the scenarios where GWIL preservesoptimality, revealing its possibilities andlimitations. We demonstrate the effectiveness ofGWIL in non-trivial continuous control domainsranging from simple rigid transformation of theexpert domain to arbitrary transformation of thestate-action space./p> /div>/td>/tr>tr idtr-benhamu2022matching>td alignright stylepadding-left:0;padding-right:0;>18./td>td>a hrefhttps://arxiv.org/abs/2207.04711 target_blank>img srcimages/publications/benhamu2022matching.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2207.04711 target_blank>Matching Normalizing Flows and Probability Paths on Manifolds/a> /em> a hrefjavascript:; onclick$("#abs_benhamu2022matching").toggle()>abs/a>br />a hrefhttps://helibenhamu.github.io/ target_blank>Heli Ben-Hamu*/a>, a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen*/a>, a hrefhttps://joeybose.github.io/ target_blank>Joey Bose/a>, strong>Brandon Amos/strong>, a hrefhttps://aditya-grover.github.io/ target_blank>Aditya Grover/a>, a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>, a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, and a hrefhttps://www.wisdom.weizmann.ac.il/~ylipman/ target_blank>Yaron Lipman/a>br />ICML 2022 br />div idabs_benhamu2022matching styletext-align: justify; display: none> p>Continuous Normalizing Flows (CNFs) are a class of generative modelsthat transform a prior distribution to a modeldistribution by solving an ordinary differentialequation (ODE). We propose to train CNFs onmanifolds by minimizing probability path divergence(PPD), a novel family of divergences between theprobability density path generated by the CNF and atarget probability density path. PPD is formulatedusing a logarithmic mass conservation formula whichis a linear first order partial differentialequation relating the log target probabilities andthe CNFâs defining vector field. PPD has several keybenefits over existing methods: it sidesteps theneed to solve an ODE per iteration, readily appliesto manifold data, scales to high dimensions, and iscompatible with a large family of target pathsinterpolating pure noise and data in finitetime. Theoretically, PPD is shown to bound classicalprobability divergences. Empirically, we show thatCNFs learned by minimizing PPD achievestate-of-the-art results in likelihoods and samplequality on existing low-dimensional manifoldbenchmarks, and is the first example of a generativemodel to scale to moderately high dimensionalmanifolds./p> /div>/td>/tr>tr idtr-chen2022semi>td alignright stylepadding-left:0;padding-right:0;>19./td>td>a hrefhttps://arxiv.org/abs/2203.06832 target_blank>img srcimages/publications/chen2022semi.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2203.06832 target_blank>Semi-Discrete Normalizing Flows through Differentiable Tessellation/a> /em> a hrefjavascript:; onclick$("#abs_chen2022semi").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, strong>Brandon Amos/strong>, and a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>br />NeurIPS 2022 br />div idabs_chen2022semi styletext-align: justify; display: none> p>Mapping between discrete and continuous distributions is a difficulttask and many have had to resort to approximate orheuristical approaches. We propose atessellation-based approach that directly learnsquantization boundaries on a continuous space, complete with exact likelihood evaluations. This isdone through constructing normalizing flows onconvex polytopes parameterized through adifferentiable Voronoi tessellation. Using a simplehomeomorphism with an efficient log determinantJacobian, we can then cheaply parameterizedistributions on convex polytopes./p> p>We explore this approach in two application settings, mapping fromdiscrete to continuous and vice versa. Firstly, aVoronoi dequantization allows automatically learningquantization boundaries in a multidimensionalspace. The location of boundaries and distancesbetween regions can encode useful structuralrelations between the quantized discretevalues. Secondly, a Voronoi mixture model hasconstant computation cost for likelihood evaluationregardless of the number of mixturecomponents. Empirically, we show improvements overexisting methods across a range of structured datamodalities, and find that we can achieve asignificant gain from just adding Voronoi mixturesto a baseline model./p> /div>/td>/tr>tr idtr-pineda2022theseus stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>20./td>td>a hrefhttps://arxiv.org/abs/2207.09442 target_blank>img srcimages/publications/pineda2022theseus.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2207.09442 target_blank>Theseus: A Library for Differentiable Nonlinear Optimization/a> /em> a hrefjavascript:; onclick$("#abs_pineda2022theseus").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/theseus target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?userrebEn8oAAAAJ target_blank>Luis Pineda/a>, a hrefhttps://scholar.google.com/citations?user3PJeg1wAAAAJ target_blank>Taosha Fan/a>, a hrefhttps://scholar.google.com/citations?usergpgb4LgAAAAJ target_blank>Maurizio Monge/a>, a hrefhttps://scholar.google.com/citations?userBFWurDEAAAAJ target_blank>Shobha Venkataraman/a>, a hrefhttps://psodhi.github.io/ target_blank>Paloma Sodhi/a>, a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky Chen/a>, a hrefhttps://joeaortiz.github.io/ target_blank>Joseph Ortiz/a>, a hrefhttps://danieldetone.com/ target_blank>Daniel DeTone/a>, a hrefhttps://scholar.google.com/citations?userkeDqjK0AAAAJ target_blank>Austin Wang/a>, a hrefhttps://scholar.google.com/citations?user8orqBsYAAAAJ target_blank>Stuart Anderson/a>, a hrefhttps://www.linkedin.com/in/jing-dong-24b26ab3/ target_blank>Jing Dong/a>, strong>Brandon Amos/strong>, and a hrefhttps://www.mustafamukadam.com/ target_blank>Mustafa Mukadam/a>br />NeurIPS 2022 br />div idabs_pineda2022theseus styletext-align: justify; display: none> p>We present Theseus, an efficient application-agnostic open sourcelibrary for differentiable nonlinear least squares(DNLS) optimization built on PyTorch, providing acommon framework for end-to-end structured learningin robotics and vision. Existing DNLSimplementations are application specific and do notalways incorporate many ingredients important forefficiency. Theseus is application-agnostic, as weillustrate with several example applications thatare built using the same underlying differentiablecomponents, such as second-order optimizers, standard costs functions, and Lie groups. Forefficiency, Theseus incorporates support for sparsesolvers, automatic vectorization, batching, GPUacceleration, and gradient computation with implicitdifferentiation and direct loss minimization. We doextensive performance evaluation in a set ofapplications, demonstrating significant efficiencygains and better scalability when these features areincorporated./p> /div>/td>/tr>tr idtr-vinitsky2022nocturne>td alignright stylepadding-left:0;padding-right:0;>21./td>td>a hrefhttps://arxiv.org/abs/2206.09889 target_blank>img srcimages/publications/vinitsky2022nocturne.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2206.09889 target_blank>Nocturne: a driving benchmark for multi-agent learning/a> /em> a hrefjavascript:; onclick$("#abs_vinitsky2022nocturne").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/nocturne target_blank>code/a> br />a hrefhttps://www.eugenevinitsky.com target_blank>Eugene Vinitsky/a>, a hrefhttps://www.nathanlct.com/about target_blank>Nathan LichtlĂ©/a>, a hrefhttps://www.linkedin.com/in/xiaomeng-yang-356a976b/ target_blank>Xiaomeng Yang/a>, strong>Brandon Amos/strong>, and a hrefhttps://www.jakobfoerster.com/ target_blank>Jakob Foerster/a>br />NeurIPS Datasets and Benchmarks Track 2022 br />div idabs_vinitsky2022nocturne styletext-align: justify; display: none> p>We introduce Nocturne, a new 2D driving simulator forinvestigating multi-agent coordination under partialobservability. The focus of Nocturne is to enableresearch into inference and theory of mind inreal-world multi-agent settings without thecomputational overhead of computer vision andfeature extraction from images. Agents in thissimulator only observe an obstructed view of thescene, mimicking human visual sensingconstraints. Unlike existing benchmarks that arebottlenecked by rendering human-like observationsdirectly using a camera input, Nocturne usesefficient intersection methods to compute avectorized set of visible features in a C++back-end, allowing the simulator to run at 2000+steps-per-second. Using open-source trajectory andmap data, we construct a simulator to load andreplay arbitrary trajectories and scenes fromreal-world driving data. Using this environment, webenchmark reinforcement-learning andimitation-learning agents and demonstrate that theagents are quite far from human-level coordinationability and deviate significantly from the experttrajectories./p> /div>/td>/tr>/table>h2>2021/h2>table classtable table-hover>tr idtr-amos2021modelbased stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>22./td>td>a hrefhttps://arxiv.org/abs/2008.12775 target_blank>img srcimages/publications/amos2021modelbased.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2008.12775 target_blank>On the model-based stochastic value gradient for continuous reinforcement learning/a> /em> a hrefjavascript:; onclick$("#abs_amos2021modelbased").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/svg target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2021.svg.pdf target_blank>slides/a> br />strong>Brandon Amos/strong>, a hrefhttps://samuelstanton.github.io/ target_blank>Samuel Stanton/a>, a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>, and a hrefhttps://cims.nyu.edu/~andrewgw/ target_blank>Andrew Gordon Wilson/a>br />L4DC 2021 (Oral) br />div idabs_amos2021modelbased styletext-align: justify; display: none> p>Model-based reinforcement learning approaches add explicit domainknowledge to agents in hopes of improving thesample-efficiency in comparison to model-freeagents. However, in practice model-based methods areunable to achieve the same asymptotic performance onchallenging continuous control tasks due to thecomplexity of learning and controlling an explicitworld model. In this paper we investigate thestochastic value gradient (SVG), which is awell-known family of methods for controllingcontinuous systems which includes model-basedapproaches that distill a model-based valueexpansion into a model-free policy. We consider avariant of the model-based SVG that scales to largersystems and uses 1) an entropy regularization tohelp with exploration, 2) a learned deterministicworld model to improve the short-horizon valueestimate, and 3) a learned model-free value estimateafter the modelâs rollout. This SVG variationcaptures the model-free soft actor-critic method asan instance when the model rollout horizon is zero, and otherwise uses short-horizon model rollouts toimprove the value estimate for the policy update. Wesurpass the asymptotic performance of othermodel-based methods on the proprioceptive MuJoColocomotion tasks from the OpenAI gym, including ahumanoid. We notably achieve these results with asimple deterministic world model without requiringan ensemble./p> /div>/td>/tr>tr idtr-cohen2021riemannian stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>23./td>td>a hrefhttps://arxiv.org/abs/2106.10272 target_blank>img srcimages/publications/cohen2021riemannian.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2106.10272 target_blank>Riemannian Convex Potential Maps/a> /em> a hrefjavascript:; onclick$("#abs_cohen2021riemannian").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/rcpm target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2021.rcpm.pdf target_blank>slides/a> br />a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen*/a>, strong>Brandon Amos*/strong>, and a hrefhttps://www.wisdom.weizmann.ac.il/~ylipman/ target_blank>Yaron Lipman/a>br />ICML 2021 br />div idabs_cohen2021riemannian styletext-align: justify; display: none> p>Modeling distributions on Riemannian manifolds is a crucialcomponent in understanding non-Euclidean data thatarises, e.g., in physics and geology. The buddingapproaches in this space are limited byrepresentational and computational tradeoffs. Wepropose and study a class of flows that uses convexpotentials from Riemannian optimal transport. Theseare universal and can model distributions on anycompact Riemannian manifold without requiring domainknowledge of the manifold to be integrated into thearchitecture. We demonstrate that these flows canmodel standard distributions on spheres, and tori, on synthetic and geological data./p> /div>/td>/tr>tr idtr-paulus2021comboptnet>td alignright stylepadding-left:0;padding-right:0;>24./td>td>a hrefhttps://arxiv.org/abs/2105.02343 target_blank>img srcimages/publications/paulus2021comboptnet.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2105.02343 target_blank>CombOptNet: Fit the Right NP-Hard Problem by Learning Integer Programming Constraints/a> /em> a hrefjavascript:; onclick$("#abs_paulus2021comboptnet").toggle()>abs/a> a hrefhttps://github.com/martius-lab/CombOptNet target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?usernjZL5CQAAAAJ target_blank>Anselm Paulus/a>, a hrefhttps://mrolinek.github.io/ target_blank>Michal Rolínek/a>, a hrefhttps://scholar.google.com/citations?userhA1rlU4AAAAJ target_blank>Vít Musil/a>, strong>Brandon Amos/strong>, and a hrefhttps://al.is.mpg.de/person/gmartius target_blank>Georg Martius/a>br />ICML 2021 br />div idabs_paulus2021comboptnet styletext-align: justify; display: none> p>Bridging logical and algorithmic reasoning with modern machinelearning techniques is a fundamental challenge withpotentially transformative impact. On thealgorithmic side, many NP-hard problems can beexpressed as integer programs, in which theconstraints play the role of their âcombinatorialspecificationâ. In this work, we aim to integrateinteger programming solvers into neural networkarchitectures as layers capable of learning both thecost terms and the constraints. The resultingend-to-end trainable architectures jointly extractfeatures from raw data and solve a suitable(learned) combinatorial problem withstate-of-the-art integer programming solvers. Wedemonstrate the potential of such layers with anextensive performance analysis on synthetic data andwith a demonstration on a competitive computervision keypoint matching benchmark./p> /div>/td>/tr>tr idtr-fickinger2021scalable>td alignright stylepadding-left:0;padding-right:0;>25./td>td>a hrefhttps://arxiv.org/abs/2109.15316 target_blank>img srcimages/publications/fickinger2021scalable.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2109.15316 target_blank>Scalable Online Planning via Reinforcement Learning Fine-Tuning/a> /em> a hrefjavascript:; onclick$("#abs_fickinger2021scalable").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userbBFN_qwAAAAJ target_blank>Arnaud Fickinger/a>, a hrefhttps://scholar.google.com/citations?usersJwwn54AAAAJ target_blank>Hengyuan Hu/a>, strong>Brandon Amos/strong>, a hrefhttp://people.eecs.berkeley.edu/~russell/ target_blank>Stuart Russell/a>, and a hrefhttps://noambrown.github.io/ target_blank>Noam Brown/a>br />NeurIPS 2021 br />div idabs_fickinger2021scalable styletext-align: justify; display: none> p>Lookahead search has been a critical component of recent AI successes, such as in the games of chess, go, andpoker. However, the search methods used in thesegames, and in many other settings, aretabular. Tabular search methods do not scale wellwith the size of the search space, and this problemis exacerbated by stochasticity and partialobservability. In this work we replace tabularsearch with online model-based fine-tuning of apolicy neural network via reinforcement learning, and show that this approach outperformsstate-of-the-art search algorithms in benchmarksettings. In particular, we use our search algorithmto achieve a new state-of-the-art result inself-play Hanabi, and show the generality of ouralgorithm by also showing that it outperformstabular search in the Atari game Ms. Pacman./p> /div>/td>/tr>tr idtr-cohen2020aligning>td alignright stylepadding-left:0;padding-right:0;>26./td>td>a hrefhttps://arxiv.org/abs/2006.12648 target_blank>img srcimages/publications/cohen2020aligning.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2006.12648 target_blank>Aligning Time Series on Incomparable Spaces/a> /em> a hrefjavascript:; onclick$("#abs_cohen2020aligning").toggle()>abs/a> a hrefhttps://github.com/samcohen16/Aligning-Time-Series target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2021.gdtw.pdf target_blank>slides/a> br />a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, a hrefhttps://giulslu.github.io/ target_blank>Giulia Luise/a>, a hrefhttps://avt.im/ target_blank>Alexander Terenin/a>, strong>Brandon Amos/strong>, and a hrefhttps://www.deisenroth.cc/ target_blank>Marc Peter Deisenroth/a>br />AISTATS 2021 br />div idabs_cohen2020aligning styletext-align: justify; display: none> p>Dynamic time warping (DTW) is a useful method for aligning, comparingand combining time series, but it requires them tolive in comparable spaces. In this work, we considera setting in which time series live on differentspaces without a sensible ground metric, causing DTWto become ill-defined. To alleviate this, we proposeGromov dynamic time warping (GDTW), a distancebetween time series on potentially incomparablespaces that avoids the comparability requirement byinstead considering intra-relational geometry. Wederive a Frank-Wolfe algorithm for computing it anddemonstrate its effectiveness at aligning, combiningand comparing time series living on incomparablespaces. We further propose a smoothed version ofGDTW as a differentiable loss and assess itsproperties in a variety of settings, includingbarycentric averaging, generative modeling andimitation learning./p> /div>/td>/tr>tr idtr-chen2021learning>td alignright stylepadding-left:0;padding-right:0;>27./td>td>a hrefhttps://arxiv.org/abs/2011.03902 target_blank>img srcimages/publications/chen2021learning.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2011.03902 target_blank>Learning Neural Event Functions for Ordinary Differential Equations/a> /em> a hrefjavascript:; onclick$("#abs_chen2021learning").toggle()>abs/a> a hrefhttps://github.com/rtqichen/torchdiffeq target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, strong>Brandon Amos/strong>, and a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>br />ICLR 2021 br />div idabs_chen2021learning styletext-align: justify; display: none> p>The existing Neural ODE formulation relies on an explicitknowledge of the termination time. We extend NeuralODEs to implicitly defined termination criteriamodeled by neural event functions, which can bechained together and differentiated through. NeuralEvent ODEs are capable of modeling discrete(instantaneous) changes in a continuous-time system, without prior knowledge of when these changes shouldoccur or how many such changes should exist. We testour approach in modeling hybrid discrete- andcontinuous- systems such as switching dynamicalsystems and collision in multi-body systems, and wepropose simulation-based training of point processeswith applications in discrete control./p> /div>/td>/tr>tr idtr-chen2021neural>td alignright stylepadding-left:0;padding-right:0;>28./td>td>a hrefhttps://arxiv.org/abs/2011.04583 target_blank>img srcimages/publications/chen2021neural.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2011.04583 target_blank>Neural Spatio-Temporal Point Processes/a> /em> a hrefjavascript:; onclick$("#abs_chen2021neural").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/neural_stpp target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?user7MxQd6UAAAAJ target_blank>Ricky T. Q. Chen/a>, strong>Brandon Amos/strong>, and a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>br />ICLR 2021 br />div idabs_chen2021neural styletext-align: justify; display: none> p>We propose a new class of parameterizations for spatio-temporalpoint processes which leverage Neural ODEs as acomputational method and enable flexible, high-fidelity models of discrete events that arelocalized in continuous time and space. Central toour approach is a combination of recurrentcontinuous-time neural networks with two novelneural architectures, i.e., Jump and AttentiveContinuous-time Normalizing Flows. This approachallows us to learn complex distributions for boththe spatial and temporal domain and to conditionnon-trivially on the observed event history. Wevalidate our models on data sets from a wide varietyof contexts such as seismology, epidemiology, urbanmobility, and neuroscience./p> /div>/td>/tr>tr idtr-yarats2021improving>td alignright stylepadding-left:0;padding-right:0;>29./td>td>a hrefhttps://arxiv.org/abs/1910.01741 target_blank>img srcimages/publications/yarats2021improving.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1910.01741 target_blank>Improving Sample Efficiency in Model-Free Reinforcement Learning from Images/a> /em> a hrefjavascript:; onclick$("#abs_yarats2021improving").toggle()>abs/a> a hrefhttps://sites.google.com/view/sac-ae target_blank>code/a> br />a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>, a hrefhttps://amyzhang.github.io/ target_blank>Amy Zhang/a>, a hrefhttps://scholar.google.com/citations?userPTS2AOgAAAAJ target_blank>Ilya Kostrikov/a>, strong>Brandon Amos/strong>, a hrefhttps://www.cs.mcgill.ca/~jpineau/ target_blank>Joelle Pineau/a>, and a hrefhttps://scholar.google.com/citations?userGgQ9GEkAAAAJ&h target_blank>Rob Fergus/a>br />AAAI 2021 br />div idabs_yarats2021improving styletext-align: justify; display: none> p>Training an agent to solve control tasks directly fromhigh-dimensional images with model-freereinforcement learning (RL) has provendifficult. The agent needs to learn a latentrepresentation together with a control policy toperform the task. Fitting a high-capacity encoderusing a scarce reward signal is not only sampleinefficient, but also prone to suboptimalconvergence. Two ways to improve sample efficiencyare to extract relevant features for the task anduse off-policy algorithms. We dissect variousapproaches of learning good latent features, andconclude that the image reconstruction loss is theessential ingredient that enables efficient andstable representation learning in image-basedRL. Following these findings, we devise anoff-policy actor-critic algorithm with an auxiliarydecoder that trains end-to-end and matchesstate-of-the-art performance across both model-freeand model-based algorithms on many challengingcontrol tasks. We release our code to encouragefuture research on image-based RL./p> /div>/td>/tr>tr idtr-venkataraman2021neural>td alignright stylepadding-left:0;padding-right:0;>30./td>td>a hrefhttps://arxiv.org/abs/2107.10254 target_blank>img srcimages/publications/venkataraman2021neural.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2107.10254 target_blank>Neural Fixed-Point Acceleration for Convex Optimization/a> /em> a hrefjavascript:; onclick$("#abs_venkataraman2021neural").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/neural-scs target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?userBFWurDEAAAAJ target_blank>Shobha Venkataraman*/a> and strong>Brandon Amos*/strong>br />ICML AutoML Workshop 2021 br />div idabs_venkataraman2021neural styletext-align: justify; display: none> p>Fixed-point iterations are at the heart of numerical computing andare often a computational bottleneck in real-timeapplications that typically need a fast solution ofmoderate accuracy. We present neural fixed-pointacceleration which combines ideas from meta-learningand classical acceleration methods to automaticallylearn to accelerate fixed-point problems that aredrawn from a distribution. We apply our framework toSCS, the state-of-the-art solver for convex coneprogramming, and design models and loss functions toovercome the challenges of learning over unrolledoptimization and acceleration instabilities. Ourwork brings neural acceleration into anyoptimization problem expressible with CVXPY./p> /div>/td>/tr>tr idtr-cohen2021sliced>td alignright stylepadding-left:0;padding-right:0;>31./td>td>a hrefhttps://arxiv.org/abs/2102.07115 target_blank>img srcimages/publications/cohen2021sliced.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2102.07115 target_blank>Sliced Multi-Marginal Optimal Transport/a> /em> a hrefjavascript:; onclick$("#abs_cohen2021sliced").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, a hrefhttps://avt.im/ target_blank>Alexander Terenin/a>, a hrefhttps://scholar.google.com/citations?userjmM-JlIAAAAJ target_blank>Yannik Pitcan/a>, strong>Brandon Amos/strong>, a hrefhttps://www.deisenroth.cc/ target_blank>Marc Peter Deisenroth/a>, and a hrefhttps://scholar.google.co.in/citations?userFPVUA-YAAAAJ target_blank>K S Sesh Kumar/a>br />NeurIPS OTML Workshop 2021 br />div idabs_cohen2021sliced styletext-align: justify; display: none> p>Multi-marginal optimal transport enables one to compare multipleprobability measures, which increasingly findsapplication in multi-task learning problems. Onepractical limitation of multi-marginal transport iscomputational scalability in the number of measures, samples and dimensionality. In this work, we proposea multi-marginal optimal transport paradigm based onrandom one-dimensional projections, whose(generalized) distance we term the slicedmulti-marginal Wasserstein distance. To constructthis distance, we introduce a characterization ofthe one-dimensional multi-marginal Kantorovichproblem and use it to highlight a number ofproperties of the sliced multi-marginal Wassersteindistance. In particular, we show that (i) the slicedmulti-marginal Wasserstein distance is a(generalized) metric that induces the same topologyas the standard Wasserstein distance, (ii) it admitsa dimension-free sample complexity, (iii) it istightly connected with the problem of barycentricaveraging under the sliced-Wasserstein metric. Weconclude by illustrating the sliced multi-marginalWasserstein on multi-task density estimation andmulti-dynamics reinforcement learning problems./p> /div>/td>/tr>tr idtr-richterpowell2021input>td alignright stylepadding-left:0;padding-right:0;>32./td>td>a hrefhttps://arxiv.org/abs/2111.12187 target_blank>img srcimages/publications/richterpowell2021input.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2111.12187 target_blank>Input Convex Gradient Networks/a> /em> a hrefjavascript:; onclick$("#abs_richterpowell2021input").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?hles&userL78pVMMAAAAJ target_blank>Jack Richter-Powell/a>, a hrefhttps://scholar.google.com/citations?userHzf8bu0AAAAJ target_blank>Jonathan Lorraine/a>, and strong>Brandon Amos/strong>br />NeurIPS OTML Workshop 2021 br />div idabs_richterpowell2021input styletext-align: justify; display: none> p>The gradients of convex functions are expressive models of non-trivialvector fields. For example, Brenierâs theorem yieldsthat the optimal transport map between any twomeasures on Euclidean space under the squareddistance is realized as a convex gradient, which isa key insight used in recent generative flowmodels. In this paper, we study how to model convexgradients by integrating a Jacobian-vector productparameterized by a neural network, which we call theInput Convex Gradient Network (ICGN). Wetheoretically study ICGNs and compare them to takingthe gradient of an Input-Convex Neural Network(ICNN), empirically demonstrating that a singlelayer ICGN can fit a toy example better than asingle layer ICNN. Lastly, we explore extensions todeeper networks and connections to constructionsfrom Riemannian geometry./p> /div>/td>/tr>tr idtr-cohen2021imitation>td alignright stylepadding-left:0;padding-right:0;>33./td>td>a hrefhttps://openreview.net/pdf?idXe5MFhFvYGX target_blank>img srcimages/publications/cohen2021imitation.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://openreview.net/pdf?idXe5MFhFvYGX target_blank>Imitation Learning from Pixel Observations for Continuous Control/a> /em> a hrefjavascript:; onclick$("#abs_cohen2021imitation").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ target_blank>Samuel Cohen/a>, strong>Brandon Amos/strong>, a hrefhttps://www.deisenroth.cc/ target_blank>Marc Peter Deisenroth/a>, a hrefhttps://www.mikaelhenaff.com/ target_blank>Mikael Henaff/a>, a hrefhttps://www.eugenevinitsky.com target_blank>Eugene Vinitsky/a>, and a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>br />NeurIPS DeepRL Workshop 2021 br />div idabs_cohen2021imitation styletext-align: justify; display: none> p>We study imitation learning from visual observations only forcontrolling dynamical systems with continuous statesand actions. This setting is attractive due to thelarge amount of video data available from whichagents could learn from. However, it is challengingdue to i) not observing the actions and ii) thehigh-dimensional visual space. In this setting, weexplore recipes for imitation learning based onadversarial learning and optimal transport. Theserecipes enable us to scale these methods to attainexpert-level performance on visual continuouscontrol tasks in the DeepMind control suite. Weinvestigate the tradeoffs of these approaches andpresent a comprehensive evaluation of the key designchoices. To encourage reproducible research in thisarea, we provide an easy-to-use implementation forbenchmarking visual imitation learning, includingour methods and expert demonstrations./p> /div>/td>/tr>tr idtr-pineda2021mbrl>td alignright stylepadding-left:0;padding-right:0;>34./td>td>a hrefhttps://arxiv.org/abs/2104.10159 target_blank>img srcimages/publications/pineda2021mbrl.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2104.10159 target_blank>MBRL-Lib: A Modular Library for Model-based Reinforcement Learning/a> /em> a hrefjavascript:; onclick$("#abs_pineda2021mbrl").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/mbrl-lib target_blank>code/a> br />a hrefhttps://scholar.google.com/citations?userrebEn8oAAAAJ target_blank>Luis Pineda/a>, strong>Brandon Amos/strong>, a hrefhttps://amyzhang.github.io/ target_blank>Amy Zhang/a>, a hrefhttps://www.natolambert.com/ target_blank>Nathan Lambert/a>, and a hrefhttps://www.robertocalandra.com/about/ target_blank>Roberto Calandra/a>br />arXiv 2021 br />div idabs_pineda2021mbrl styletext-align: justify; display: none> p>Model-based reinforcement learning is a compelling framework fordata-efficient learning of agents that interact withthe world. This family of algorithms has manysubcomponents that need to be carefully selected andtuned. As a result the entry-bar for researchers toapproach the field and to deploy it in real-worldtasks can be daunting. In this paper, we presentMBRL-Lib - a machine learning library formodel-based reinforcement learning in continuousstate-action spaces based on PyTorch. MBRL-Lib isdesigned as a platform for both researchers, toeasily develop, debug and compare new algorithms, and non-expert user, to lower the entry-bar ofdeploying state-of-the-art algorithms./p> /div>/td>/tr>/table>h2>2020/h2>table classtable table-hover>tr idtr-amos2020differentiable stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>35./td>td>a hrefhttps://arxiv.org/abs/1909.12830 target_blank>img srcimages/publications/amos2020differentiable.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1909.12830 target_blank>The Differentiable Cross-Entropy Method/a> /em> a hrefjavascript:; onclick$("#abs_amos2020differentiable").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/dcem target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2020.dcem.pdf target_blank>slides/a> br />strong>Brandon Amos/strong> and a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>br />ICML 2020 br />div idabs_amos2020differentiable styletext-align: justify; display: none> p>We study the Cross-Entropy Method (CEM) for the non-convexoptimization of a continuous and parameterizedobjective function and introduce a differentiablevariant (DCEM) that enables us to differentiate theoutput of CEM with respect to the objectivefunctionâs parameters. In the machine learningsetting this brings CEM inside of the end-to-endlearning pipeline where this has otherwise beenimpossible. We show applications in a syntheticenergy-based structured prediction task and innon-convex continuous control. In the controlsetting we show on the simulated cheetah and walkertasks that we can embed their optimal actionsequences with DCEM and then use policy optimizationto fine-tune components of the controller as a steptowards combining model-based and model-free RL./p> /div>/td>/tr>tr idtr-lambert2020objective>td alignright stylepadding-left:0;padding-right:0;>36./td>td>a hrefhttps://arxiv.org/abs/2002.04523 target_blank>img srcimages/publications/lambert2020objective.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/2002.04523 target_blank>Objective Mismatch in Model-based Reinforcement Learning/a> /em> a hrefjavascript:; onclick$("#abs_lambert2020objective").toggle()>abs/a>br />a hrefhttps://www.natolambert.com/ target_blank>Nathan Lambert/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userzSsW478AAAAJ target_blank>Omry Yadan/a>, and a hrefhttps://www.robertocalandra.com/about/ target_blank>Roberto Calandra/a>br />L4DC 2020 br />div idabs_lambert2020objective styletext-align: justify; display: none> p>Model-based reinforcement learning (MBRL) has been shown to be a powerful framework for data-efficiently learning control of continuous tasks. Recent work in MBRL has mostly focused on using more advanced function approximators and planning schemes, with little development of the general framework. In this paper, we identify a fundamental issue of the standard MBRL framework-what we call the objective mismatch issue. Objective mismatch arises when one objective is optimized in the hope that a second, often uncorrelated, metric will also be optimized. In the context of MBRL, we characterize the objective mismatch between training the forward dynamics model with respect to the likelihood of the one-step ahead prediction, and the overall goal of improving performance on a downstream control task. For example, this issue can emerge with the realization that dynamics models effective for a specific task do not necessarily need to be globally accurate, and vice versa globally accurate models might not be sufficiently accurate locally to obtain good control performance on a specific task. In our experiments, we study this objective mismatch issue and demonstrate that the likelihood of one-step ahead predictions is not always correlated with control performance. This observation highlights a critical limitation in the MBRL framework which will require further research to be fully understood and addressed. We propose an initial method to mitigate the mismatch issue by re-weighting dynamics model training. Building on it, we conclude with a discussion about other potential directions of research for addressing this issue./p> /div>/td>/tr>tr idtr-amos2020QNSTOP>td alignright stylepadding-left:0;padding-right:0;>37./td>td>a hrefhttps://vtechworks.lib.vt.edu/bitstream/handle/10919/49672/qnTOMS14.pdf target_blank>img srcimages/publications/amos2020QNSTOP.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://vtechworks.lib.vt.edu/bitstream/handle/10919/49672/qnTOMS14.pdf target_blank>QNSTOP: Quasi-Newton Algorithm for Stochastic Optimization/a> /em> a hrefjavascript:; onclick$("#abs_amos2020QNSTOP").toggle()>abs/a> a hrefhttps://github.com/vtopt/qnstop target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://dblp.org/pid/75/8682.html target_blank>David Easterling/a>, a hrefhttps://people.cs.vt.edu/~ltw/shortvita.html target_blank>Layne T. Watson/a>, a hrefhttps://scholar.google.com/citations?user2I6IgikAAAAJ target_blank>William Thacker/a>, a hrefhttps://dblp.org/pid/142/1258.html target_blank>Brent Castle/a>, and a hrefhttps://mtrosset.pages.iu.edu/ target_blank>Michael Trosset/a>br />ACM TOMS 2020 br />div idabs_amos2020QNSTOP styletext-align: justify; display: none> p>QNSTOP consists of serial and parallel (OpenMP) Fortran 2003 codes for thequasi-Newton stochastic optimization method of Castle and Trosset. Forstochastic problems, convergence theory exists for the particularalgorithmic choices and parameter values used in QNSTOP. Both the paralleldriver subroutine, which offers several parallel decomposition strategies, and the serial driver subroutine can be used for stochastic optimization ordeterministic global optimization, based on an input switch. QNSTOP isparticularly effective for ânoisyâ deterministic problems, using onlyobjective function values. Some performance data for computational systemsbiology problems is given./p> /div>/td>/tr>tr idtr-sercu2020neural>td alignright stylepadding-left:0;padding-right:0;>38./td>td>a hrefhttps://www.biorxiv.org/content/10.1101/2021.04.08.439084v1.abstract target_blank>img srcimages/publications/sercu2020neural.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://www.biorxiv.org/content/10.1101/2021.04.08.439084v1.abstract target_blank>Neural Potts Model/a> /em> a hrefjavascript:; onclick$("#abs_sercu2020neural").toggle()>abs/a>br />a hrefhttps://tom.sercu.me/ target_blank>Tom Sercu/a>, a hrefhttps://dblp.org/pid/296/8930.html target_blank>Robert Verkuil/a>, a hrefhttps://scholar.google.com/citations?user2M0OltAAAAAJ target_blank>Joshua Meier/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userZDjmMuwAAAAJ target_blank>Zeming Lin/a>, a hrefhttps://www.linkedin.com/in/caroline-chen/ target_blank>Caroline Chen/a>, a hrefhttps://www.linkedin.com/in/jasonliu6/ target_blank>Jason Liu/a>, a hrefhttp://yann.lecun.com/ target_blank>Yann LeCun/a>, and a hrefhttps://scholar.google.com/citations?uservqb78-gAAAAJ target_blank>Alexander Rives/a>br />MLCB 2020 br />div idabs_sercu2020neural styletext-align: justify; display: none> p>We propose the Neural Potts Model objective as an amortizedoptimization problem. The objective enables traininga single model with shared parameters to explicitlymodel energy landscapes across multiple proteinfamilies. Given a protein sequence as input, themodel is trained to predict a pairwise couplingmatrix for a Potts model energy function describingthe local evolutionary landscape of thesequence. Couplings can be predicted for novelsequences. A controlled ablation experimentassessing unsupervised contact prediction on sets ofrelated protein families finds a gain fromamortization for low-depth multiple sequencealignments; the result is then confirmed on adatabase with broad coverage of protein sequences./p> /div>/td>/tr>tr idtr-lou2020riemannian>td alignright stylepadding-left:0;padding-right:0;>39./td>td>a hrefhttps://drive.google.com/file/d/1Ewro0Ne1tvK15nHyYopY4wZ59QTVB-1c/view target_blank>img srcimages/publications/lou2020riemannian.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://drive.google.com/file/d/1Ewro0Ne1tvK15nHyYopY4wZ59QTVB-1c/view target_blank>Deep Riemannian Manifold Learning/a> /em> a hrefjavascript:; onclick$("#abs_lou2020riemannian").toggle()>abs/a>br />a hrefhttps://aaronlou.com/ target_blank>Aaron Lou/a>, a hrefhttps://maxn.io/ target_blank>Maximilian Nickel/a>, and strong>Brandon Amos/strong>br />NeurIPS Geo4dl Workshop 2020 br />div idabs_lou2020riemannian styletext-align: justify; display: none> p>We present a new class of learnable Riemannian manifolds with a metricparameterized by a deep neural network. The core manifold operationsâspecificallythe Riemannian exponential and logarithmic mapsâare solved using approximatenumerical techniques. Input and parameter gradients are computed with anadjoint sensitivity analysis. This enables us to fit geodesics and distances withgradient-based optimization of both on-manifold values and the manifold itself.We demonstrate our methodâs capability to model smooth, flexible metric structuresin graph embedding tasks./p> /div>/td>/tr>/table>h2>2019/h2>table classtable table-hover>tr idtr-amos2019differentiable stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>40./td>td>a hrefhttps://github.com/bamos/thesis/raw/master/bamos_thesis.pdf target_blank>img srcimages/publications/amos2019differentiable.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://github.com/bamos/thesis/raw/master/bamos_thesis.pdf target_blank>Differentiable Optimization-Based Modeling for Machine Learning/a> /em> a hrefjavascript:; onclick$("#abs_amos2019differentiable").toggle()>abs/a> a hrefhttps://github.com/bamos/thesis target_blank>code/a> br />strong>Brandon Amos/strong>br />Ph.D. Thesis 2019 br />div idabs_amos2019differentiable styletext-align: justify; display: none> p>Domain-specific modeling priors and specialized components are becomingincreasingly important to the machine learning field. These components integrate specialized knowledge that we have as humans into model. We argue inthis thesis that optimization methods provide an expressive set of operationsthat should be part of the machine learning practitionerâs modeling toolbox.We present two foundational approaches for optimization-based modeling:1) the OptNet architecture that integrates optimization problems as individuallayers in larger end-to-end trainable deep networks, and 2) the input-convexneural network (ICNN) architecture that helps make inference and learning indeep energy-based models and structured prediction more tractable.We then show how to use the OptNet approach 1) as a way of combiningmodel-free and model-based reinforcement learning and 2) for top-k learningproblems. We conclude by showing how to differentiate cone programs and turnthe cvxpy domain specific language into a differentiable optimization layer thatenables rapid prototyping of the approaches in this thesis./p> /div>/td>/tr>tr idtr-amos2019differentiable3 stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>41./td>td>a hrefhttp://web.stanford.edu/~boyd/papers/pdf/diff_cvxpy.pdf target_blank>img srcimages/publications/amos2019differentiable3.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://web.stanford.edu/~boyd/papers/pdf/diff_cvxpy.pdf target_blank>Differentiable Convex Optimization Layers/a> /em> a hrefjavascript:; onclick$("#abs_amos2019differentiable3").toggle()>abs/a> a hrefhttps://github.com/cvxgrp/cvxpylayers target_blank>code/a> br />a hrefhttps://www.akshayagrawal.com/ target_blank>Akshay Agrawal*/a>, strong>Brandon Amos*/strong>, a hrefhttps://scholar.google.com/citations?userHmCZLyoAAAAJ target_blank>Shane Barratt*/a>, a hrefhttps://web.stanford.edu/~boyd/ target_blank>Stephen Boyd*/a>, a hrefhttps://stevendiamond.me/ target_blank>Steven Diamond*/a>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter*/a>br />NeurIPS 2019 br />div idabs_amos2019differentiable3 styletext-align: justify; display: none> p>Recent work has shown how to embed differentiable optimization problems (that is, problems whose solutions can be backpropagated through) as layers within deep learning architectures. This method provides a useful inductive bias for certain problems, but existing software for differentiable optimization layers is rigid and difficult to apply to new settings. In this paper, we propose an approach to differentiating through disciplined convex programs, a subclass of convex optimization problems used by domain-specific languages (DSLs) for convex optimization. We introduce disciplined parametrized programming, a subset of disciplined convex programming, and we show that every disciplined parametrized program can be represented as the composition of an affine map from parameters to problem data, a solver, and an affine map from the solverâs solution to a solution of the original problem (a new form we refer to as affine-solver-affine form). We then demonstrate how to efficiently differentiate through each of these components, allowing for end-to-end analytical differentiation through the entire convex program. We implement our methodology in version 1.1 of CVXPY, a popular Python-embedded DSL for convex optimization, and additionally implement differentiable layers for disciplined convex programs in PyTorch and TensorFlow 2.0. Our implementation significantly lowers the barrier to using convex optimization problems in differentiable programs. We present applications in linear machine learning models and in stochastic control, and we show that our layer is competitive (in execution time) compared to specialized differentiable solvers from past work./p> /div>/td>/tr>tr idtr-amos2019limited>td alignright stylepadding-left:0;padding-right:0;>42./td>td>a hrefhttps://arxiv.org/abs/1906.08707 target_blank>img srcimages/publications/amos2019limited.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1906.08707 target_blank>The Limited Multi-Label Projection Layer/a> /em> a hrefjavascript:; onclick$("#abs_amos2019limited").toggle()>abs/a> a hrefhttps://github.com/locuslab/lml target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttp://vladlen.info/ target_blank>Vladlen Koltun/a>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />arXiv 2019 br />div idabs_amos2019limited styletext-align: justify; display: none> p>We propose the Limited Multi-Label (LML) projection layer as a newprimitive operation for end-to-end learning systems. The LML layerprovides a probabilistic way of modeling multi-label predictionslimited to having exactly k labels. We derive efficient forward andbackward passes for this layer and show how the layer can be used tooptimize the top-k recall for multi-label tasks with incomplete labelinformation. We evaluate LML layers on top-k CIFAR-100 classificationand scene graph generation. We show that LML layers add a negligibleamount of computational overhead, strictly improve the modelâsrepresentational capacity, and improve accuracy. We also revisit thetruncated top-k entropy method as a competitive baseline for top-kclassification./p> /div>/td>/tr>tr idtr-grefenstette2019generalized>td alignright stylepadding-left:0;padding-right:0;>43./td>td>a hrefhttps://arxiv.org/abs/1910.01727 target_blank>img srcimages/publications/grefenstette2019generalized.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1910.01727 target_blank>Generalized Inner Loop Meta-Learning/a> /em> a hrefjavascript:; onclick$("#abs_grefenstette2019generalized").toggle()>abs/a> a hrefhttps://github.com/facebookresearch/higher target_blank>code/a> br />a hrefhttps://www.egrefen.com/ target_blank>Edward Grefenstette/a>, strong>Brandon Amos/strong>, a hrefhttps://cs.nyu.edu/~dy1042/ target_blank>Denis Yarats/a>, a hrefhttps://phumonhtut.me/ target_blank>Phu Mon Htut/a>, a hrefhttps://amolchanov86.github.io/ target_blank>Artem Molchanov/a>, a hrefhttps://fmeier.github.io/ target_blank>Franziska Meier/a>, a hrefhttps://douwekiela.github.io/ target_blank>Douwe Kiela/a>, a hrefhttps://kyunghyuncho.me/ target_blank>Kyunghyun Cho/a>, and a hrefhttps://soumith.ch/ target_blank>Soumith Chintala/a>br />arXiv 2019 br />div idabs_grefenstette2019generalized styletext-align: justify; display: none> p>Many (but not all) approaches self-qualifying as âmeta-learningâ indeep learning and reinforcement learning fit acommon pattern of approximating the solution to anested optimization problem. In this paper, we givea formalization of this shared pattern, which wecall GIMLI, prove its general requirements, andderive a general-purpose algorithm for implementingsimilar approaches. Based on this analysis andalgorithm, we describe a library of our design, higher, which we share with the community to assistand enable future research into these kinds ofmeta-learning approaches. We end the paper byshowcasing the practical applications of thisframework and library through illustrativeexperiments and ablation studies which theyfacilitate./p> /div>/td>/tr>/table>h2>2018/h2>table classtable table-hover>tr idtr-amos2018learning stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>44./td>td>a hrefhttps://arxiv.org/abs/1804.06318 target_blank>img srcimages/publications/amos2018learning.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1804.06318 target_blank>Learning Awareness Models/a> /em> a hrefjavascript:; onclick$("#abs_amos2018learning").toggle()>abs/a>br />strong>Brandon Amos/strong>, a hrefhttps://laurent-dinh.github.io/ target_blank>Laurent Dinh/a>, a hrefhttps://scholar.google.com/citations?userl-HhJaUAAAAJ target_blank>Serkan Cabi/a>, a hrefhttps://dblp.org/pid/188/6045.html target_blank>Thomas Rothörl/a>, a hrefhttps://scholar.google.com/citations?user0Dkf68EAAAAJ target_blank>Sergio Gómez Colmenarejo/a>, a hrefhttps://scholar.google.com/citations?userYfgdfyYAAAAJ target_blank>Alistair Muldal/a>, a hrefhttps://scholar.google.com/citations?usergVFnjOcAAAAJ target_blank>Tom Erez/a>, a hrefhttps://scholar.google.com/citations?userCjOTm_4AAAAJ target_blank>Yuval Tassa/a>, a hrefhttps://scholar.google.com/citations?usernzEluBwAAAAJ target_blank>Nando de Freitas/a>, and a hrefhttps://mdenil.com/ target_blank>Misha Denil/a>br />ICLR 2018 br />div idabs_amos2018learning styletext-align: justify; display: none> p>We consider the setting of an agent with a fixed body interacting with anunknown and uncertain external world. We show that modelstrained to predict proprioceptive information about theagentâs body come to represent objects in the external world.In spite of being trained with only internally availablesignals, these dynamic body models come to represent externalobjects through the necessity of predicting their effects onthe agentâs own body. That is, the model learns holisticpersistent representations of objects in the world, eventhough the only training signals are body signals. Ourdynamics model is able to successfully predict distributionsover 132 sensor readings over 100 steps into the future and wedemonstrate that even when the body is no longer in contactwith an object, the latent variables of the dynamics modelcontinue to represent its shape. We show that active datacollection by maximizing the entropy of predictions about thebody-touch sensors, proprioception and vestibularinformation-leads to learning of dynamic models that showsuperior performance when used for control. We also collectdata from a real robotic hand and show that the same modelscan be used to answer questions about properties of objects inthe real world. Videos with qualitative results of our modelsare available a hrefhttps://goo.gl/mZuqAV>here/a>./p> /div>/td>/tr>tr idtr-amos2018end stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>45./td>td>a hrefhttps://arxiv.org/abs/1810.13400 target_blank>img srcimages/publications/amos2018end.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://arxiv.org/abs/1810.13400 target_blank>Differentiable MPC for End-to-end Planning and Control/a> /em> a hrefjavascript:; onclick$("#abs_amos2018end").toggle()>abs/a> a hrefhttps://locuslab.github.io/mpc.pytorch/ target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://ivandariojr.io/ target_blank>Ivan Dario Jimenez Rodriguez/a>, a hrefhttps://scholar.google.com/citations?userTh4PuGkAAAAJ target_blank>Jacob Sacks/a>, a hrefhttps://homes.cs.washington.edu/~bboots/ target_blank>Byron Boots/a>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />NeurIPS 2018 br />div idabs_amos2018end styletext-align: justify; display: none> p>In this paper we present foundations for using model predictive control (MPC) as a differentiable policy class in reinforcement learning. Specifically, we differentiate through MPC by using the KKT conditions of the convex approximation at a fixed point of the solver. Using this strategy, we are able to learn the cost and dynamics of a controller via end-to-end learning in a larger system. We empirically show results in an imitation learning setting, demonstrating that we can recover the underlying dynamics and cost more efficiently and reliably than with a generic neural network policy class/p> /div>/td>/tr>tr idtr-brown2018depth>td alignright stylepadding-left:0;padding-right:0;>46./td>td>a hrefhttp://arxiv.org/abs/1805.08195 target_blank>img srcimages/publications/brown2018depth.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://arxiv.org/abs/1805.08195 target_blank>Depth-Limited Solving for Imperfect-Information Games/a> /em> a hrefjavascript:; onclick$("#abs_brown2018depth").toggle()>abs/a>br />a hrefhttps://noambrown.github.io/ target_blank>Noam Brown/a>, a hrefhttp://www.cs.cmu.edu/~sandholm/ target_blank>Tuomas Sandholm/a>, and strong>Brandon Amos/strong>br />NeurIPS 2018 br />div idabs_brown2018depth styletext-align: justify; display: none> p>A fundamental challenge in imperfect-information games is that states do not have well-defined values. As a result, depth-limited search algorithms used in single-agent settings and perfect-information games do not apply. This paper introduces a principled way to conduct depth-limited solving in imperfect-information games by allowing the opponent to choose among a number of strategies for the remainder of the game at the depth limit. Each one of these strategies results in a different set of values for leaf nodes. This forces an agent to be robust to the different strategies an opponent may employ. We demonstrate the effectiveness of this approach by building a master-level heads-up no-limit Texas holdâem poker AI that defeats two prior top agents using only a 4-core CPU and 16 GB of memory. Developing such a powerful agent would have previously required a supercomputer./p> /div>/td>/tr>tr idtr-wang2018enabling>td alignright stylepadding-left:0;padding-right:0;>47./td>td>a hrefhttps://dl.acm.org/citation.cfm?id3209659 target_blank>img srcimages/publications/wang2018enabling.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://dl.acm.org/citation.cfm?id3209659 target_blank>Enabling Live Video Analytics with a Scalable and Privacy-Aware Framework/a> /em> a hrefjavascript:; onclick$("#abs_wang2018enabling").toggle()>abs/a>br />a hrefhttps://www.junjuewang.com target_blank>Junjue Wang/a>, strong>Brandon Amos/strong>, a hrefhttps://anupamdas.org/ target_blank>Anupam Das/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://www.normsadeh.org/ target_blank>Norman Sadeh/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />ACM TOMM 2018 br />div idabs_wang2018enabling styletext-align: justify; display: none> p>We show how to build the components of a privacy-aware, live videoanalytics ecosystem from the bottom up, startingwith OpenFace, our new open-source face recognitionsystem that approaches state-of-the-artaccuracy. Integrating OpenFace with interframetracking, we build RTFace, a mechanism fordenaturing video streams that selectively blursfaces according to specified policies at full framerates. This enables privacy management for livevideo analytics while providing a secure approachfor handling retrospective policyexceptions. Finally, we present a scalable, privacy-aware architecture for large camera networksusing RTFace and show how it can be an enabler for avibrant ecosystem and marketplace of privacy-awarevideo streams and analytics services./p> /div>/td>/tr>/table>h2>2017/h2>table classtable table-hover>tr idtr-amos2017optnet stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>48./td>td>a hrefhttp://arxiv.org/abs/1703.00443 target_blank>img srcimages/publications/amos2017optnet.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://arxiv.org/abs/1703.00443 target_blank>OptNet: Differentiable Optimization as a Layer in Neural Networks/a> /em> a hrefjavascript:; onclick$("#abs_amos2017optnet").toggle()>abs/a> a hrefhttps://github.com/locuslab/optnet target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2017.optnet.pdf target_blank>slides/a> br />strong>Brandon Amos/strong> and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />ICML 2017 br />div idabs_amos2017optnet styletext-align: justify; display: none> p>This paper presents OptNet, a network architecture that integratesoptimization problems (here, specifically in the form of quadratic programs)as individual layers in larger end-to-end trainable deep networks.These layers encode constraints and complex dependenciesbetween the hidden states that traditional convolutional andfully-connected layers often cannot capture.In this paper, we explore the foundations for such an architecture:we show how techniques from sensitivity analysis, bileveloptimization, and implicit differentiation can be used toexactly differentiate through these layers and with respectto layer parameters;we develop a highly efficient solver for these layers that exploits fastGPU-based batch solves within a primal-dual interior point method, and whichprovides backpropagation gradients with virtually no additional cost on top ofthe solve;and we highlight the application of these approaches in several problems.In one notable example, we show that the method iscapable of learning to play mini-Sudoku (4x4) given just input and output games, with no a priori information about the rules of the game;this highlights the ability of our architecture to learn hardconstraints better than other neural architectures./p> /div>/td>/tr>tr idtr-amos2017input stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>49./td>td>a hrefhttp://arxiv.org/abs/1609.07152 target_blank>img srcimages/publications/amos2017input.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://arxiv.org/abs/1609.07152 target_blank>Input Convex Neural Networks/a> /em> a hrefjavascript:; onclick$("#abs_amos2017input").toggle()>abs/a> a hrefhttps://github.com/locuslab/icnn target_blank>code/a> a hrefhttp://bamos.github.io/data/slides/2017.icnn.pdf target_blank>slides/a> br />strong>Brandon Amos/strong>, a hrefhttps://leixx.io/ target_blank>Lei Xu/a>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />ICML 2017 br />div idabs_amos2017input styletext-align: justify; display: none> p>This paper presents the input convex neural networkarchitecture. These are scalar-valued (potentially deep) neuralnetworks with constraints on the network parameters such that theoutput of the network is a convex function of (some of) the inputs.The networks allow for efficient inference via optimization over someinputs to the network given others, and can be applied to settingsincluding structured prediction, data imputation, reinforcementlearning, and others. In this paper we lay the basic groundwork forthese models, proposing methods for inference, optimization andlearning, and analyze their representational power. We show that manyexisting neural network architectures can be made input-convex witha minor modification, and develop specialized optimizationalgorithms tailored to this setting. Finally, we highlight theperformance of the methods on multi-label prediction, imagecompletion, and reinforcement learning problems, where we showimprovement over the existing state of the art in many cases./p> /div>/td>/tr>tr idtr-donti2017task>td alignright stylepadding-left:0;padding-right:0;>50./td>td>a hrefhttp://arxiv.org/abs/1703.04529 target_blank>img srcimages/publications/donti2017task.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://arxiv.org/abs/1703.04529 target_blank>Task-based End-to-end Model Learning/a> /em> a hrefjavascript:; onclick$("#abs_donti2017task").toggle()>abs/a> a hrefhttps://github.com/locuslab/e2e-model-learning target_blank>code/a> br />a hrefhttps://priyadonti.com/ target_blank>Priya L. Donti/a>, strong>Brandon Amos/strong>, and a hrefhttps://zicokolter.com/ target_blank>J. Zico Kolter/a>br />NeurIPS 2017 br />div idabs_donti2017task styletext-align: justify; display: none> p>As machine learning techniques have become more ubiquitous, it hasbecome common to see machine learning prediction algorithms operatingwithin some larger process. However, the criteria by which we trainmachine learning algorithms often differ from the ultimate criteria onwhich we evaluate them. This paper proposes an end-to-end approach forlearning probabilistic machine learning models within the context ofstochastic programming, in a manner that directly captures theultimate task-based objective for which they will be used. We thenpresent two experimental evaluations of the proposed approach, one asapplied to a generic inventory stock problem and the second to areal-world electrical grid scheduling task. In both cases, we showthat the proposed approach can outperform both a traditional modelingapproach and a purely black-box policy optimization approach./p> /div>/td>/tr>tr idtr-chen2017quasi>td alignright stylepadding-left:0;padding-right:0;>51./td>td>a hrefhttps://par.nsf.gov/servlets/purl/10111392 target_blank>img srcimages/publications/chen2017quasi.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://par.nsf.gov/servlets/purl/10111392 target_blank>Quasi-Newton Stochastic Optimization Algorithm for Parameter Estimation of a Stochastic Model of the Budding Yeast Cell Cycle/a> /em> a hrefjavascript:; onclick$("#abs_chen2017quasi").toggle()>abs/a>br />a hrefhttps://chenm.sites.wfu.edu/publications/ target_blank>Minghan Chen/a>, strong>Brandon Amos/strong>, a hrefhttps://people.cs.vt.edu/~ltw/shortvita.html target_blank>Layne T. Watson/a>, a hrefhttps://scholar.google.com/citations?usersyETjMMAAAAJ target_blank>John Tyson/a>, a hrefhttps://people.cs.vt.edu/~ycao/ target_blank>Yang Cao/a>, a hrefhttps://people.cs.vt.edu/shaffer/ target_blank>Cliff Shaffer/a>, a hrefhttps://mtrosset.pages.iu.edu/ target_blank>Michael Trosset/a>, a hrefhttps://scholar.google.com/citations?userZ4534DUAAAAJ target_blank>Cihan Oguz/a>, and a hrefhttps://dblp.org/pid/235/5473.html target_blank>Gisella Kakoti/a>br />IEEE/ACM TCBB 2017 br />div idabs_chen2017quasi styletext-align: justify; display: none> p>Parameter estimation in discrete or continuous deterministic cellcycle models is challenging for several reasons, including the nature of what can be observed, andthe accuracy and quantity of those observations. Thechallenge is even greater for stochastic models, where the number of simulations and amount ofempirical data must be even larger to obtainstatistically valid parameter estimates. The twomain contributions of this work are (1) stochasticmodel parameter estimation based on directlymatching multivariate probability distributions, and(2) a new quasi-Newton algorithm class QNSTOP forstochastic optimization problems. QNSTOP directlyuses the random objective function value samplesrather than creating ensemble statistics. QNSTOP isused here to directly match empirical and simulatedjoint probability distributions rather than matchingsummary statistics. Results are given for a currentstate-of-the-art stochastic cell cycle model ofbudding yeast, whose predictions match well somesummary statistics and one-dimensional distributionsfrom empirical data, but do not match well theempirical joint distributions. The nature of themismatch provides insight into the weakness in thestochastic model./p> /div>/td>/tr>tr idtr-ha2017you>td alignright stylepadding-left:0;padding-right:0;>52./td>td>a hrefhttps://dl.acm.org/doi/10.1145/3132211.3134453 target_blank>img srcimages/publications/ha2017you.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://dl.acm.org/doi/10.1145/3132211.3134453 target_blank>You can teach elephants to dance: agile VM handoff for edge computing/a> /em> a hrefjavascript:; onclick$("#abs_ha2017you").toggle()>abs/a>br />a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://dblp.org/pid/18/1620.html target_blank>Yoshihisa Abe/a>, a hrefhttps://dblp.org/pid/207/9122.html target_blank>Thomas Eiszler/a>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, strong>Brandon Amos/strong>, a hrefhttps://dblp.org/pid/207/9123.html target_blank>Rohit Upadhyaya/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />SEC 2017 br />div idabs_ha2017you styletext-align: justify; display: none> p>VM handoff enables rapid and transparent placement changes toexecuting code in edge computing use cases where thesafety and management attributes of VM encapsulationare important. This versatile primitive offers thefunctionality of classic live migration but ishighly optimized for the edge. Over WAN bandwidthsranging from 5 to 25 Mbps, VM handoff migrates arunning 8 GB VM in about a minute, with a downtimeof a few tens of seconds. By dynamically adapting tovarying network bandwidth and processing load, VMhandoff is more than an order of magnitude fasterthan live migration at those bandwidths./p> /div>/td>/tr>tr idtr-chen2017empirical>td alignright stylepadding-left:0;padding-right:0;>53./td>td>a hrefhttps://www.cs.cmu.edu/~zhuoc/papers/latency2017.pdf target_blank>img srcimages/publications/chen2017empirical.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://www.cs.cmu.edu/~zhuoc/papers/latency2017.pdf target_blank>An Empirical Study of Latency in an Emerging Class of Edge Computing Applications for Wearable Cognitive Assistance/a> /em> a hrefjavascript:; onclick$("#abs_chen2017empirical").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, a hrefhttps://www.junjuewang.com target_blank>Junjue Wang/a>, a hrefhttps://scholar.google.com/citations?user0OpjwCMAAAAJ target_blank>Siyan Zhao/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?user0pF6i38AAAAJ target_blank>Guanhang Wu/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://scholar.google.com/citations?userR9r5_GIAAAAJ target_blank>Khalid Elgazzar/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://scholar.google.com/citations?userwUPKh58AAAAJ target_blank>Roberta Klatzky/a>, a hrefhttp://www.cs.cmu.edu/~dps/ target_blank>Daniel Siewiorek/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />SEC 2017 br />div idabs_chen2017empirical styletext-align: justify; display: none> p>An emerging class of interactive wearable cognitive assistanceapplications is poised to become one of the keydemonstrators of edge computing infrastructure. Inthis paper, we design seven such applications andevaluate their performance in terms of latencyacross a range of edge computing configurations, mobile hardware, and wireless networks, including 4GLTE. We also devise a novel multi-algorithm approachthat leverages temporal locality to reduceend-to-end latency by 60% to 70%, withoutsacrificing accuracy. Finally, we derive targetlatencies for our applications, and show that edgecomputing is crucial to meeting these targets./p> /div>/td>/tr>tr idtr-wang2017scalable>td alignright stylepadding-left:0;padding-right:0;>54./td>td>a hrefhttp://elijah.cs.cmu.edu/DOCS/wang-mmsys2017.pdf target_blank>img srcimages/publications/wang2017scalable.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://elijah.cs.cmu.edu/DOCS/wang-mmsys2017.pdf target_blank>A Scalable and Privacy-Aware IoT Service for Live Video Analytics/a> /em> a hrefjavascript:; onclick$("#abs_wang2017scalable").toggle()>abs/a> a hrefhttp://cmusatyalab.github.io/openface/ target_blank>code/a> br />a hrefhttps://www.junjuewang.com target_blank>Junjue Wang/a>, strong>Brandon Amos/strong>, a hrefhttps://anupamdas.org/ target_blank>Anupam Das/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://www.normsadeh.org/ target_blank>Norman Sadeh/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />ACM MMSys 2017 (Best Paper) br />div idabs_wang2017scalable styletext-align: justify; display: none> p>We present OpenFace, our new open-source face recognition systemthat approaches state-of-the-art accuracy. Integrating OpenFace withinter-frame tracking, we build RTFace, a mechanism for denaturing videostreams that selectively blurs faces according to specifiedpolicies at full frame rates. This enables privacy management forlive video analytics while providing a secure approach for handlingretrospective policy exceptions. Finally, we present a scalable, privacy-aware architecture for large camera networks using RTFace./p> /div>/td>/tr>/table>h2>2016/h2>table classtable table-hover>tr idtr-amos2016openface stylebackground-color: #ffffd0>td alignright stylepadding-left:0;padding-right:0;>55./td>td>a hrefhttp://reports-archive.adm.cs.cmu.edu/anon/anon/2016/CMU-CS-16-118.pdf target_blank>img srcimages/publications/amos2016openface.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://reports-archive.adm.cs.cmu.edu/anon/anon/2016/CMU-CS-16-118.pdf target_blank>OpenFace: A general-purpose face recognition library with mobile applications/a> /em> a hrefjavascript:; onclick$("#abs_amos2016openface").toggle()>abs/a> a hrefhttps://cmusatyalab.github.io/openface target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://www.linkedin.com/in/bartosz-ludwiczuk-a677a760 target_blank>Bartosz Ludwiczuk/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />CMU 2016 br />div idabs_amos2016openface styletext-align: justify; display: none> p>Cameras are becoming ubiquitous in the Internet of Things (IoT) andcan use face recognition technology to improve context. There is alarge accuracy gap between todayâs publicly available face recognitionsystems and the state-of-the-art private face recognitionsystems. This paper presents our OpenFace face recognition librarythat bridges this accuracy gap. We show that OpenFace providesnear-human accuracy on the LFW benchmark and present a newclassification benchmark for mobile scenarios. This paper is intendedfor non-experts interested in using OpenFace and provides a lightintroduction to the deep neural network techniques we use./p> p>We released OpenFace in October 2015 as an open source library underthe Apache 2.0 license. It is available at:a hrefhttp://cmusatyalab.github.io/openface/>http://cmusatyalab.github.io/openface//a>/p> /div>/td>/tr>tr idtr-zhao2016collapsed>td alignright stylepadding-left:0;padding-right:0;>56./td>td>a hrefhttp://proceedings.mlr.press/v48/zhaoa16.html target_blank>img srcimages/publications/zhao2016collapsed.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://proceedings.mlr.press/v48/zhaoa16.html target_blank>Collapsed Variational Inference for Sum-Product Networks/a> /em> a hrefjavascript:; onclick$("#abs_zhao2016collapsed").toggle()>abs/a>br />a hrefhttps://hanzhaoml.github.io/ target_blank>Han Zhao/a>, a hrefhttps://tameemadel.wordpress.com/ target_blank>Tameem Adel/a>, a hrefhttp://www.cs.cmu.edu/~ggordon/ target_blank>Geoff Gordon/a>, and strong>Brandon Amos/strong>br />ICML 2016 br />div idabs_zhao2016collapsed styletext-align: justify; display: none> p>Sum-Product Networks (SPNs) are probabilistic inference machines that admitexact inference in linear time in the size of the network. Existingparameter learning approaches for SPNs are largely based on the maximumlikelihood principle and hence are subject to overfitting compared tomore Bayesian approaches. Exact Bayesian posterior inference for SPNs iscomputationally intractable. Both standard variational inference andposterior sampling for SPNs are computationally infeasible even fornetworks of moderate size due to the large number of local latentvariables per instance. In this work, we propose a novel deterministiccollapsed variational inference algorithm for SPNs that iscomputationally efficient, easy to implement and at the same time allowsus to incorporate prior information into the optimization formulation.Extensive experiments show a significant improvement in accuracy comparedwith a maximum likelihood based approach./p> /div>/td>/tr>tr idtr-hu2016quantifying>td alignright stylepadding-left:0;padding-right:0;>57./td>td>a hrefhttps://dl.acm.org/doi/10.1145/2967360.2967369 target_blank>img srcimages/publications/hu2016quantifying.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://dl.acm.org/doi/10.1145/2967360.2967369 target_blank>Quantifying the impact of edge computing on mobile applications/a> /em> a hrefjavascript:; onclick$("#abs_hu2016quantifying").toggle()>abs/a>br />a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, a hrefhttps://www.linkedin.com/in/joelyinggao/ target_blank>Ying Gao/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://www.junjuewang.com target_blank>Junjue Wang/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />ACM SIGOPS 2016 br />div idabs_hu2016quantifying styletext-align: justify; display: none> p>Computational offloading services at the edge of the Internet formobile devices are becoming a reality. Using a widerange of mobile applications, we explore how suchinfrastructure improves latency and energyconsumption relative to the cloud. We presentexperimental results from WiFi and 4G LTE networksthat confirm substantial wins from edge computingfor highly interactive mobile applications./p> /div>/td>/tr>tr idtr-davies2016privacy>td alignright stylepadding-left:0;padding-right:0;>58./td>td>a hrefhttp://eprints.lancs.ac.uk/78255/1/44691.pdf target_blank>img srcimages/publications/davies2016privacy.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://eprints.lancs.ac.uk/78255/1/44691.pdf target_blank>Privacy mediators: helping IoT cross the chasm/a> /em> a hrefjavascript:; onclick$("#abs_davies2016privacy").toggle()>abs/a>br />a hrefhttps://www.lancaster.ac.uk/sci-tech/about-us/people/nigel-davies target_blank>Nigel Davies/a>, a hrefhttps://scholar.google.com/citations?userBItCgjYAAAAJ target_blank>Nina Taft/a>, a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>, a hrefhttp://www.sclinch.com/ target_blank>Sarah Clinch/a>, and strong>Brandon Amos/strong>br />HotMobile 2016 br />div idabs_davies2016privacy styletext-align: justify; display: none> p>Unease over data privacy will retard consumer acceptance of IoTdeployments. The primary source of discomfort is a lack of usercontrol over raw data that is streamed directly from sensors to thecloud. This is a direct consequence of the over-centralization oftodayâs cloud-based IoT hub designs. We propose a solution thatinterposes a locally-controlled software component called a privacymediator on every raw sensor stream. Each mediator is in the sameadministrative domain as the sensors whose data is being collected, and dynamically enforces the current privacy policies of the ownersof the sensors or mobile users within the domain. This solution necessitatesa logical point of presence for mediators within the administrativeboundaries of each organization. Such points of presenceare provided by cloudlets, which are small locally-administered datacenters at the edge of the Internet that can support code mobility.The use of cloudlet-based mediators aligns well with natural personaland organizational boundaries of trust and responsibility./p> /div>/td>/tr>/table>h2>2015 and earlier/h2>table classtable table-hover>tr idtr-satyanarayanan2015edge>td alignright stylepadding-left:0;padding-right:0;>59./td>td>a hrefhttps://www.cs.cmu.edu/~satya/docdir/satya-edge2015.pdf target_blank>img srcimages/publications/satyanarayanan2015edge.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttps://www.cs.cmu.edu/~satya/docdir/satya-edge2015.pdf target_blank>Edge Analytics in the Internet of Things/a> /em> a hrefjavascript:; onclick$("#abs_satyanarayanan2015edge").toggle()>abs/a>br />a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>, a hrefhttps://www.ugent.be/ea/idlab/en/members/pieter-simoens.htm target_blank>Pieter Simoens/a>, a hrefhttps://scholar.google.com/citations?userZeRhyWsAAAAJ target_blank>Yu Xiao/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, and strong>Brandon Amos/strong>br />IEEE Pervasive Computing 2015 br />div idabs_satyanarayanan2015edge styletext-align: justify; display: none> p>High-data-rate sensors, such as video cameras, are becoming ubiquitous in theInternet of Things. This article describes GigaSight, an Internet-scalerepository of crowd-sourced video content, with strong enforcement of privacypreferences and access controls. The GigaSight architecture is a federatedsystem of VM-based cloudlets that perform video analytics at the edge of theInternet, thus reducing the demand for ingress bandwidth into the cloud.Denaturing, which is an owner-specific reduction in fidelity of video contentto preserve privacy, is one form of analytics on cloudlets. Content-basedindexing for search is another form of cloudlet-based analytics. This articleis part of a special issue on smart spaces./p> /div>/td>/tr>tr idtr-turner2015bad>td alignright stylepadding-left:0;padding-right:0;>60./td>td>a hrefhttp://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber7118094 target_blank>img srcimages/publications/turner2015bad.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber7118094 target_blank>Bad Parts: Are Our Manufacturing Systems at Risk of Silent Cyberattacks?/a> /em> a hrefjavascript:; onclick$("#abs_turner2015bad").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userMRKab9cAAAAJ target_blank>Hamilton Turner/a>, a hrefhttps://scholar.google.com/citations?user10HSX90AAAAJ target_blank>Jules White/a>, a hrefhttps://scholar.google.com/citations?usertWmVBNwAAAAJ target_blank>Jaime A. Camelio/a>, a hrefhttps://scholar.google.com/citations?userAW81mosAAAAJ target_blank>Christopher Williams/a>, strong>Brandon Amos/strong>, and a hrefhttps://ieeexplore.ieee.org/author/37085729541 target_blank>Robert Parker/a>br />IEEE Security & Privacy 2015 br />div idabs_turner2015bad styletext-align: justify; display: none> p>Recent cyberattacks have highlighted the risk of physical equipment operatingoutside designed tolerances to produce catastrophic failures. A relatedthreat is cyberattacks that change the design and manufacturing of amachineâs part, such as an automobile brake component, so it no longerfunctions properly. These risks stem from the lack of cyber-physical modelsto identify ongoing attacks as well as the lack of rigorous application ofknown cybersecurity best practices. To protect manufacturing processes in thefuture, research will be needed on a number of critical cyber-physicalmanufacturing security topics./p> /div>/td>/tr>tr idtr-chen2015early>td alignright stylepadding-left:0;padding-right:0;>61./td>td>a hrefhttp://www.cs.cmu.edu/~satya/docdir/chen-wearsys2015.pdf target_blank>img srcimages/publications/chen2015early.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://www.cs.cmu.edu/~satya/docdir/chen-wearsys2015.pdf target_blank>Early Implementation Experience with Wearable Cognitive Assistance Applications/a> /em> a hrefjavascript:; onclick$("#abs_chen2015early").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://www.lujiang.info/ target_blank>Lu Jiang/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, strong>Brandon Amos/strong>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttp://www.cs.cmu.edu/~alex/ target_blank>Alex Hauptmann/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />WearSys 2015 br />div idabs_chen2015early styletext-align: justify; display: none> p>A cognitive assistance application combines a wearable device suchas Google Glass with cloudlet processing to provide step-by-stepguidance on a complex task. In this paper, we focus on user assistancefor narrow and well-defined tasks that require specializedknowledge and/or skills. We describe proof-of-concept implementationsfor four different tasks: assembling 2D Lego models, freehandsketching, playing ping-pong, and recommending context-relevantYouTube tutorials. We then reflect on the difficulties we faced inbuilding these applications, and suggest future research that couldsimplify the creation of similar applications./p> /div>/td>/tr>tr idtr-hu2014case>td alignright stylepadding-left:0;padding-right:0;>62./td>td>a hrefhttp://www.cs.cmu.edu/~satya/docdir/hu-hotmobile2015.pdf target_blank>img srcimages/publications/hu2014case.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://www.cs.cmu.edu/~satya/docdir/hu-hotmobile2015.pdf target_blank>The Case for Offload Shaping/a> /em> a hrefjavascript:; onclick$("#abs_hu2014case").toggle()>abs/a>br />a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://scholar.google.com/citations?uservU6bKxEAAAAJ target_blank>Wolfgang Richter/a>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, a hrefhttps://github.com/bgilbert target_blank>Benjamin Gilbert/a>, a hrefhttps://scholar.google.com/citations?userjj5tN8sAAAAJ target_blank>Jan Harkes/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />HotMobile 2015 br />div idabs_hu2014case styletext-align: justify; display: none> p>When offloading computation from a mobile device, we showthat it can pay to perform additional on-device work in orderto reduce the offloading workload. We call this offload shaping, and demonstrate its application at many different levelsof abstraction using a variety of techniques. We show thatoffload shaping can produce significant reduction in resourcedemand, with little loss of application-level fidelity/p> /div>/td>/tr>tr idtr-gao2015cloudlets>td alignright stylepadding-left:0;padding-right:0;>63./td>td>a hrefhttp://reports-archive.adm.cs.cmu.edu/anon/anon/2015/CMU-CS-15-139.pdf target_blank>img srcimages/publications/gao2015cloudlets.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://reports-archive.adm.cs.cmu.edu/anon/anon/2015/CMU-CS-15-139.pdf target_blank>Are Cloudlets Necessary?/a> /em> a hrefjavascript:; onclick$("#abs_gao2015cloudlets").toggle()>abs/a>br />a hrefhttps://www.linkedin.com/in/joelyinggao/ target_blank>Ying Gao/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, strong>Brandon Amos/strong>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />CMU 2015 br />div idabs_gao2015cloudlets styletext-align: justify; display: none> p>We present experimental results from Wi-Fi and 4G LTE networks to validate theintuition that low end-to-end latency of cloud services improves applicationresponse time and reduces energy consumption on mobile devices. We focusspecifically on computational offloading as a cloud service. Using a widerange of applications, and exploring both pre-partitioned and dynamicallypartitioned approaches, we demonstrate the importance of low latency forcloud offload services. We show the best performance is achieved byoffloading to cloudlets, which are small-scale edge-located data centers. Ourresults show that cloudlets can improve response times 51% and reduce energyconsumption in a mobile device by up to 42% compared to cloud offload./p> /div>/td>/tr>tr idtr-ha2015adaptive>td alignright stylepadding-left:0;padding-right:0;>64./td>td>a hrefhttp://ra.adm.cs.cmu.edu/anon/2015/CMU-CS-15-113.pdf target_blank>img srcimages/publications/ha2015adaptive.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://ra.adm.cs.cmu.edu/anon/2015/CMU-CS-15-113.pdf target_blank>Adaptive VM handoff across cloudlets/a> /em> a hrefjavascript:; onclick$("#abs_ha2015adaptive").toggle()>abs/a>br />a hrefhttp://krha.kr/ target_blank>Kiryong Ha/a>, a hrefhttps://dblp.org/pid/18/1620.html target_blank>Yoshihisa Abe/a>, a hrefhttps://scholar.google.com/citations?user-wKpBNkAAAAJ target_blank>Zhuo Chen/a>, a hrefhttp://www.cs.cmu.edu/~wenluh/ target_blank>Wenlu Hu/a>, strong>Brandon Amos/strong>, a hrefhttps://www.andrew.cmu.edu/user/pspillai/ target_blank>Padmanabhan Pillai/a>, and a hrefhttps://www.cs.cmu.edu/~satya/ target_blank>Mahadev Satyanarayanan/a>br />CMU 2015 br />div idabs_ha2015adaptive styletext-align: justify; display: none> p>Cloudlet offload is a valuable technique for ensuring low end-to-end latency ofresource-intensive cloud processing for many emerging mobile applications.This paper examines the impact of user mobility on cloudlet offload, andshows that even modest user mobility can result in significant networkdegradation. We propose VM handoff as a technique for seamlessly transferringVM-encapsulated execution to a more optimal offload site as users move. Ourapproach can perform handoff in roughly a minute even over limited WANs byadaptively reducing data transferred. We present experimental results tovalidate our implementation and to demonstrate effectiveness of adaptation tochanging network conditions and processing capacity/p> /div>/td>/tr>tr idtr-andrew2014global>td alignright stylepadding-left:0;padding-right:0;>65./td>td>a hrefhttp://dl.acm.org/citation.cfm?id2685662 target_blank>img srcimages/publications/andrew2014global.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://dl.acm.org/citation.cfm?id2685662 target_blank>Global Parameter Estimation for a Eukaryotic Cell Cycle Model in Systems Biology/a> /em> a hrefjavascript:; onclick$("#abs_andrew2014global").toggle()>abs/a>br />a hrefhttps://scholar.google.com/citations?userqpRt_KYAAAAJ target_blank>Tricity Andrew/a>, strong>Brandon Amos/strong>, a hrefhttps://dblp.org/pid/75/8682.html target_blank>David Easterling/a>, a hrefhttps://scholar.google.com/citations?userZ4534DUAAAAJ target_blank>Cihan Oguz/a>, a hrefhttps://scholar.google.com/citations?userfAmU38gAAAAJ target_blank>William Baumann/a>, a hrefhttps://scholar.google.com/citations?usersyETjMMAAAAJ target_blank>John Tyson/a>, and a hrefhttps://people.cs.vt.edu/~ltw/shortvita.html target_blank>Layne T. Watson/a>br />SummerSim 2014 br />div idabs_andrew2014global styletext-align: justify; display: none> p>The complicated process by which a yeast cell divides, known as the cellcycle, has been modeled by a system of 26 nonlinear ordinary differentialequations (ODEs) with 149 parameters. This model captures the chemicalkinetics of the regulatory networks controlling the cell division processin budding yeast cells. Empirical data is discrete and matched againstdiscrete inferences (e.g., whether a particular mutant cell lives or dies)computed from the ODE solution trajectories. The problem ofestimating the ODE parameters to best fit the model to the data is a149-dimensional global optimization problem attacked by the deterministicalgorithm VTDIRECT95 and by the nondeterministic algorithms differentialevolution, QNSTOP, and simulated annealing, whose performances arecompared./p> /div>/td>/tr>tr idtr-amos2013applying>td alignright stylepadding-left:0;padding-right:0;>66./td>td>a hrefhttp://bamos.github.io/data/papers/amos-iwcmc2013.pdf target_blank>img srcimages/publications/amos2013applying.png onerrorthis.style.displaynone classpublicationImg />/a> em>a hrefhttp://bamos.github.io/data/papers/amos-iwcmc2013.pdf target_blank>Applying machine learning classifiers to dynamic Android malware detection at scale/a> /em> a hrefjavascript:; onclick$("#abs_amos2013applying").toggle()>abs/a> a hrefhttps://github.com/VT-Magnum-Research/antimalware target_blank>code/a> br />strong>Brandon Amos/strong>, a hrefhttps://scholar.google.com/citations?userMRKab9cAAAAJ target_blank>Hamilton Turner/a>, and a hrefhttps://scholar.google.com/citations?user10HSX90AAAAJ target_blank>Jules White/a>br />IWCMC 2013 br />div idabs_amos2013applying styletext-align: justify; display: none> p>The widespread adoption and contextually sensitivenature of smartphone devices has increased concerns over smartphonemalware. Machine learning classifiers are a current methodfor detecting malicious applications on smartphone systems. Thispaper presents the evaluation of a number of existing classifiers, using a dataset containing thousands of real (i.e. not synthetic)applications. We also present our STREAM framework, whichwas developed to enable rapid large-scale validation of mobilemalware machine learning classifiers./p> /div>/td>/tr>/table>h2 id-open-source-repositories>i classfa fa-chevron-right>/i> Open Source Repositories/h2>p>29.6k+ GitHub stars across all repositories./p>table classtable table-hover>tr> td alignright stylepadding-right:0;padding-left:0;>1./td> td> span classcvdate>2024/span> a hrefhttps://github.com/facebookresearch/advprompter>facebookresearch/advprompter/a> | i classfa fas fa-star>/i> 112 | em>Fast Adaptive Adversarial Prompting for LLMs/em> !-- --> !-- facebookresearch/advprompter --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>2./td> td> span classcvdate>2024/span> a hrefhttps://github.com/facebookresearch/lagrangian-ot>facebookresearch/lagrangian-ot/a> | i classfa fas fa-star>/i> 39 | em>Lagrangian OT/em> !-- --> !-- facebookresearch/lagrangian-ot --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>3./td> td> span classcvdate>2023/span> a hrefhttps://github.com/facebookresearch/amortized-optimization-tutorial>facebookresearch/amortized-optimization-tutorial/a> | i classfa fas fa-star>/i> 236 | em>Tutorial on amortized optimization/em> !-- --> !-- facebookresearch/amortized-optimization-tutorial --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>4./td> td> span classcvdate>2023/span> a hrefhttps://github.com/facebookresearch/taskmet>facebookresearch/taskmet/a> | i classfa fas fa-star>/i> 18 | em>TaskMet: Task-Driven Metric Learning for Model Learning/em> !-- --> !-- facebookresearch/taskmet --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>5./td> td> span classcvdate>2023/span> a hrefhttps://github.com/facebookresearch/w2ot>facebookresearch/w2ot/a> | i classfa fas fa-star>/i> 43 | em>Wasserstein-2 optimal transport in JAX/em> !-- --> !-- facebookresearch/w2ot --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>6./td> td> span classcvdate>2022/span> a hrefhttps://github.com/facebookresearch/theseus>facebookresearch/theseus/a> | i classfa fas fa-star>/i> 1.7k | em>Differentiable non-linear optimization library/em> !-- --> !-- facebookresearch/theseus --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>7./td> td> span classcvdate>2022/span> a hrefhttps://github.com/facebookresearch/meta-ot>facebookresearch/meta-ot/a> | i classfa fas fa-star>/i> 94 | em>Meta Optimal Transport/em> !-- --> !-- facebookresearch/meta-ot --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>8./td> td> span classcvdate>2022/span> a hrefhttps://github.com/bamos/presentations>bamos/presentations/a> | i classfa fas fa-star>/i> 141 | em>Source for my major presentations/em> !-- --> !-- bamos/presentations --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>9./td> td> span classcvdate>2021/span> a hrefhttps://github.com/facebookresearch/rcpm>facebookresearch/rcpm/a> | i classfa fas fa-star>/i> 68 | em>Riemannian Convex Potential Maps/em> !-- --> !-- facebookresearch/rcpm --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>10./td> td> span classcvdate>2021/span> a hrefhttps://github.com/facebookresearch/svg>facebookresearch/svg/a> | i classfa fas fa-star>/i> 54 | em>Model-based stochastic value gradient/em> !-- --> !-- facebookresearch/svg --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>11./td> td> span classcvdate>2021/span> a hrefhttps://github.com/facebookresearch/mbrl-lib>facebookresearch/mbrl-lib/a> | i classfa fas fa-star>/i> 954 | em>Model-based reinforcement learning library/em> !-- --> !-- facebookresearch/mbrl-lib --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>12./td> td> span classcvdate>2020/span> a hrefhttps://github.com/facebookresearch/dcem>facebookresearch/dcem/a> | i classfa fas fa-star>/i> 122 | em>The Differentiable Cross-Entropy Method/em> !-- --> !-- facebookresearch/dcem --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>13./td> td> span classcvdate>2019/span> a hrefhttps://github.com/facebookresearch/higher>facebookresearch/higher/a> | i classfa fas fa-star>/i> 1.6k | em>PyTorch higher-order gradient and optimization library/em> !-- --> !-- facebookresearch/higher --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>14./td> td> span classcvdate>2019/span> a hrefhttps://github.com/bamos/thesis>bamos/thesis/a> | i classfa fas fa-star>/i> 318 | em>Ph.D. Thesis LaTeX source code/em> !-- --> !-- bamos/thesis --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>15./td> td> span classcvdate>2019/span> a hrefhttps://github.com/cvxgrp/cvxpylayers>cvxgrp/cvxpylayers/a> | i classfa fas fa-star>/i> 1.8k | em>Differentiable Convex Optimization Layers/em> !-- --> !-- cvxgrp/cvxpylayers --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>16./td> td> span classcvdate>2019/span> a hrefhttps://github.com/locuslab/lml>locuslab/lml/a> | i classfa fas fa-star>/i> 58 | em>The Limited Multi-Label Projection Layer/em> !-- --> !-- locuslab/lml --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>17./td> td> span classcvdate>2018/span> a hrefhttps://github.com/locuslab/mpc.pytorch>locuslab/mpc.pytorch/a> | i classfa fas fa-star>/i> 865 | em>Differentiable PyTorch Model Predictive Control library/em> !-- --> !-- locuslab/mpc.pytorch --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>18./td> td> span classcvdate>2018/span> a hrefhttps://github.com/locuslab/differentiable-mpc>locuslab/differentiable-mpc/a> | i classfa fas fa-star>/i> 239 | em>Differentiable MPC experiments/em> !-- --> !-- locuslab/differentiable-mpc --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>19./td> td> span classcvdate>2017/span> a hrefhttps://github.com/locuslab/icnn>locuslab/icnn/a> | i classfa fas fa-star>/i> 274 | em>Input Convex Neural Network experiments/em> !-- --> !-- locuslab/icnn --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>20./td> td> span classcvdate>2017/span> a hrefhttps://github.com/locuslab/optnet>locuslab/optnet/a> | i classfa fas fa-star>/i> 507 | em>OptNet experiments/em> !-- --> !-- locuslab/optnet --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>21./td> td> span classcvdate>2017/span> a hrefhttps://github.com/locuslab/qpth>locuslab/qpth/a> | i classfa fas fa-star>/i> 673 | em>Differentiable PyTorch QP solver/em> !-- --> !-- locuslab/qpth --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>22./td> td> span classcvdate>2017/span> a hrefhttps://github.com/bamos/densenet.pytorch>bamos/densenet.pytorch/a> | i classfa fas fa-star>/i> 823 | em>PyTorch DenseNet implementation/em> !-- --> !-- bamos/densenet.pytorch --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>23./td> td> span classcvdate>2017/span> a hrefhttps://github.com/bamos/block>bamos/block/a> | i classfa fas fa-star>/i> 297 | em>Intelligent block matrix constructions/em> !-- --> !-- bamos/block --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>24./td> td> span classcvdate>2017/span> a hrefhttps://github.com/bamos/setGPU>bamos/setGPU/a> | i classfa fas fa-star>/i> 106 | em>Automatically use the least-loaded GPU/em> !-- --> !-- bamos/setGPU --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>25./td> td> span classcvdate>2016/span> a hrefhttps://github.com/bamos/dcgan-completion.tensorflow>bamos/dcgan-completion.tensorflow/a> | i classfa fas fa-star>/i> 1.3k | em>Image completion with GANs/em> !-- --> !-- bamos/dcgan-completion.tensorflow --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>26./td> td> span classcvdate>2015/span> a hrefhttps://github.com/cmusatyalab/openface>cmusatyalab/openface/a> | i classfa fas fa-star>/i> 15.1k | em>Face recognition with deep neural networks/em> !-- --> !-- cmusatyalab/openface --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>27./td> td> span classcvdate>2014/span> a hrefhttps://github.com/vtopt/qnstop>vtopt/qnstop/a> | i classfa fas fa-star>/i> 10 | em>Fortran quasi-Newton stochastic optimization library/em> !-- --> !-- vtopt/qnstop --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>28./td> td> span classcvdate>2014/span> a hrefhttps://github.com/bamos/snowglobe>bamos/snowglobe/a> | i classfa fas fa-star>/i> 27 | em>Haskell-driven, self-hosted web analytics with minimal configuration/em> !-- --> !-- bamos/snowglobe --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>29./td> td> span classcvdate>2014/span> a hrefhttps://github.com/bamos/zsh-history-analysis>bamos/zsh-history-analysis/a> | i classfa fas fa-star>/i> 224 | em>Analyze and plot your zsh history/em> !-- --> !-- bamos/zsh-history-analysis --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>30./td> td> span classcvdate>2014/span> a hrefhttps://github.com/bamos/beamer-snippets>bamos/beamer-snippets/a> | i classfa fas fa-star>/i> 109 | em>Beamer and TikZ snippets/em> !-- --> !-- bamos/beamer-snippets --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>31./td> td> span classcvdate>2013/span> a hrefhttps://github.com/bamos/latex-templates>bamos/latex-templates/a> | i classfa fas fa-star>/i> 366 | em>LaTeX templates/em> !-- --> !-- bamos/latex-templates --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>32./td> td> span classcvdate>2013/span> a hrefhttps://github.com/cparse/cparse>cparse/cparse/a> | i classfa fas fa-star>/i> 336 | em>C++ expression parser using Dijkstras shunting-yard algorithm/em> !-- --> !-- cparse/cparse --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>33./td> td> span classcvdate>2013/span> a hrefhttps://github.com/bamos/cv>bamos/cv/a> | i classfa fas fa-star>/i> 398 | em>Source for this CV: Creates LaTeX/Markdown from YAML/BibTeX/em> !-- --> !-- bamos/cv --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>34./td> td> span classcvdate>2013/span> a hrefhttps://github.com/bamos/python-scripts>bamos/python-scripts/a> | i classfa fas fa-star>/i> 197 | em>Short and fun Python scripts/em> !-- --> !-- bamos/python-scripts --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>35./td> td> span classcvdate>2013/span> a hrefhttps://github.com/bamos/reading-list>bamos/reading-list/a> | i classfa fas fa-star>/i> 185 | em>YAML reading list and notes system/em> !-- --> !-- bamos/reading-list --> !-- --> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>36./td> td> span classcvdate>2012/span> a hrefhttps://github.com/bamos/dotfiles>bamos/dotfiles/a> | i classfa fas fa-star>/i> 238 | em>i classfa fas fa-heart>/i> Linux, xmonad, emacs, vim, zsh, tmux/em> !-- --> !-- bamos/dotfiles --> !-- --> /td>/tr>/table>h2 id-invited-talks>i classfa fa-chevron-right>/i> Invited Talks/h2>p>Slides for my major presentations are open-sourced with a CC-BY license ata hrefhttps://github.com/bamos/presentations>bamos/presentations/a>./p>table classtable table-hover>tr> td alignright stylepadding-right:0;padding-left:0;>1./td> td stylepadding-right:0;> span classcvdate>2024/span> em>Amortized optimization for optimal transport and LLM attacks/em>, ISMP /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>2./td> td stylepadding-right:0;> span classcvdate>2024/span> em>Differentiable optimization for robotics/em>, a hrefhttps://sites.google.com/robotics.utias.utoronto.ca/frontiers-optimization-rss24/schedule>RSS Optimization for Robotics Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>3./td> td stylepadding-right:0;> span classcvdate>2024/span> em>Amortized optimization-based reasoning for AI/em>, University of Amsterdam /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>4./td> td stylepadding-right:0;> span classcvdate>2024/span> em>End-to-end learning geometries for graphs, dynamical systems, and regression/em>, a hrefhttps://logmeetupnyc.github.io/>LoG New York/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>5./td> td stylepadding-right:0;> span classcvdate>2023/span> em>Amortized optimization for optimal transport/em>, a hrefhttps://otmlworkshop.github.io/schedule/>NeurIPS Optimal Transport and ML Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>6./td> td stylepadding-right:0;> span classcvdate>2023/span> em>On optimal control and machine learning/em>, a hrefhttps://frontiers4lcd.github.io/>ICML Learning, Control, and Dynamical Systems Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>7./td> td stylepadding-right:0;> span classcvdate>2023/span> em>Tutorial on amortized optimization/em>, Brown University /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>8./td> td stylepadding-right:0;> span classcvdate>2023/span> em>Learning with differentiable and amortized optimization/em>, NYU AI Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>9./td> td stylepadding-right:0;> span classcvdate>2023/span> em>Learning with differentiable and amortized optimization/em>, Vanderbilt ML Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>10./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Learning with differentiable and amortized optimization/em>, Microsoft Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>11./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Amortized optimization for computing optimal transport maps/em>, a hrefhttps://sites.google.com/view/sampling-transport-diffusions/home>Flatiron Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>12./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Learning with differentiable and amortized optimization/em>, Cornell AI Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>13./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Learning with differentiable and amortized optimization/em>, Cornell Tech Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>14./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Learning with differentiable and amortized optimization/em>, Argonne National Laboratory /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>15./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Theseus: A library for differentiable nonlinear optimization/em>, NYU /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>16./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Theseus: A library for differentiable nonlinear optimization/em>, University of Zurich /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>17./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization-based modeling for machine learning/em>, Colorado Mines AMS Colloquium /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>18./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization/em>, a hrefhttps://guaguakai.github.io/IJCAI22-differentiable-optimization/>IJCAI Tutorial/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>19./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization for control and RL/em>, a hrefhttps://darl-workshop.github.io/>ICML Workshop on Decision Awareness in RL/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>20./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization-based modeling for machine learning/em>, a hrefhttps://sites.google.com/usc.edu/cpaior-2022/master_class>CPAIOR Master Class/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>21./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Tutorial on amortized optimization/em>, ICCOPT /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>22./td> td stylepadding-right:0;> span classcvdate>2022/span> em>Differentiable optimization for control and RL/em>, Gridmatic /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>23./td> td stylepadding-right:0;> span classcvdate>2021/span> em>Learning for control with differentiable optimization and ODEs/em>, Columbia University /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>24./td> td stylepadding-right:0;> span classcvdate>2021/span> em>Differentiable optimization-based modeling for machine learning/em>, IBM Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>25./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Differentiable optimization for control/em>, Max Planck Institute (TĂŒbingen) /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>26./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Differentiable optimization-based modeling for machine learning/em>, Mila Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>27./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Deep Declarative Networks/em>, a hrefhttps://anucvml.github.io/ddn-eccvt2020/>ECCV Tutorial/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>28./td> td stylepadding-right:0;> span classcvdate>2020/span> em>On differentiable optimization for control and vision/em>, a hrefhttps://anucvml.github.io/ddn-cvprw2020/>CVPR Deep Declarative Networks Workshop/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>29./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Differentiable optimization-based modeling for machine learning/em>, a hrefhttps://sites.google.com/view/cs-159-spring-2020/lectures>Caltech CS 159 (Guest Lecture)/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>30./td> td stylepadding-right:0;> span classcvdate>2020/span> em>Unrolled optimization for learning deep energy models/em>, a hrefhttps://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE67922>SIAM MDS Minisymposium/a> /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>31./td> td stylepadding-right:0;> span classcvdate>2019/span> em>Differentiable optimization-based modeling for machine learning/em>, NYU CILVR Seminar /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>32./td> td stylepadding-right:0;> span classcvdate>2019/span> em>Differentiable optimization-based modeling for machine learning/em>, INFORMS /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>33./td> td stylepadding-right:0;> span classcvdate>2019/span> em>Differentiable optimization-based modeling for machine learning/em>, Facebook AI Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>34./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, ISMP /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>35./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Google Brain /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>36./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Bosch Center for AI /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>37./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Waymo Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>38./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Tesla AI /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>39./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, NVIDIA Robotics /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>40./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, Salesforce Research /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>41./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, OpenAI /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>42./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization-based modeling for machine learning/em>, NNAISENSE /td>/tr>tr> td alignright stylepadding-right:0;padding-left:0;>43./td> td stylepadding-right:0;> span classcvdate>2018/span> em>Differentiable optimization and control/em>, UC Berkeley /td>/tr>/table>h2 id-interns-and-students>i classfa fa-chevron-right>/i> Interns and Students/h2>table classtable table-hover>tr> td stylepadding-right:0;> span classcvdate>2024 - present/span> a hrefhttps://scholar.google.com/citations?userHBAXF6YAAAAJ>Aaron Havens/a> (visiting FAIR from UIUC) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2022 - 2024/span> a hrefhttps://arampooladian.com/>Aram-Alexandre Pooladian/a> (visiting FAIR from NYU) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2022 - 2024/span> a hrefhttps://cdenrich.github.io/>Carles Domingo-Enrich/a> (visiting FAIR from NYU, now at MSR) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023 - 2024/span> a hrefhttps://scholar.google.com/citations?usernjZL5CQAAAAJ>Anselm Paulus/a> (visiting FAIR from Max Planck Institute, TĂŒbingen) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023/span> a hrefhttps://www.mhr.ai>Matthew Retchin/a> (Columbia MS thesis committee, now at Harvard) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2022 - 2023/span> a hrefhttps://sanaelotfi.github.io/>Sanae Lotfi/a> (visiting FAIR from NYU) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2022 - 2023/span> a hrefhttps://dishank-b.github.io>Dishank Bansal/a> (AI resident at FAIR) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2021 - 2022/span> a hrefhttps://www.linkedin.com/in/arnaudfickinger/>Arnaud Fickinger/a> (visiting FAIR from Berkeley) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020 - 2022/span> a hrefhttps://aaronlou.com/>Aaron Lou/a> (visiting FAIR from Cornell and Stanford, now scientist at OpenAI) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2021 - 2022/span> a hrefhttps://www.eugenevinitsky.com>Eugene Vinitsky/a> (visiting FAIR from Berkeley, now professor at NYU) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2021 - 2022/span> a hrefhttps://scholar.google.com/citations?userCmdjfTsAAAAJ>Samuel Cohen/a> (visiting FAIR from UCL, now CEO at FairGen) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttp://www.cs.toronto.edu/~rtqichen/>Ricky Chen/a> (visiting FAIR from Toronto, now scientist at FAIR) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttp://www.cs.cmu.edu/~pliang/>Paul Liang/a> (visiting FAIR from CMU, now professor at MIT) /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2018/span> a hrefhttps://phillipkwang.com/>Phillip Wang/a> (at CMU, now CEO at a hrefhttps://gather.town/ target_blank>Gather/a>) /td>/tr>/table>h2 id-professional-activities>i classfa fa-chevron-right>/i> Professional Activities/h2>table classtable table-hover>tr> td stylepadding-right:0;> span classcvdate>2025/span> AAAI Senior Program Committee /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2024/span> NeurIPS Area Chair /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2024/span> NeurIPS Datasets and Benchmarks Area Chair /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2024/span> AAAI Senior Program Committee /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023/span> NeurIPS Area Chair /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023/span> NeurIPS Datasets and Benchmarks Area Chair /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2023/span> AAAI Senior Program Committee /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttps://sites.google.com/view/lmca2020/home>NeurIPS Learning Meets Combinatorial Optimization Workshop Organizer/a> /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttps://anucvml.github.io/ddn-cvprw2020/>CVPR Deep Declarative Networks Workshop Organizer/a> /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2020/span> a hrefhttps://anucvml.github.io/ddn-eccvt2020/>ECCV Deep Declarative Networks Tutorial Organizer/a> /td>/tr>tr> td stylepadding-right:0;> span classcvdate>2014 - 2015/span> CMU CSD MS Admissions /td>/tr>/table>h3 idreviewing>Reviewing/h3>table classtable table-hover>tr> td stylepadding-right:0;>AAAI Conference on Artificial Intelligence/td>/tr>tr> td stylepadding-right:0;>American Controls Conference (ACC)/td>/tr>tr> td stylepadding-right:0;>IEEE Conference on Computer Vision and Pattern Recognition (CVPR)/td>/tr>tr> td stylepadding-right:0;>IEEE Conference on Decision and Control (CDC)/td>/tr>tr> td stylepadding-right:0;>IEEE Control Systems Letters (L-CSS)/td>/tr>tr> td stylepadding-right:0;>IEEE International Conference on Computer Vision (ICCV)/td>/tr>tr> td stylepadding-right:0;>IEEE International Conference on Intelligent Robots and Systems (IROS)/td>/tr>tr> td stylepadding-right:0;>IEEE International Conference on Robotics and Automation (ICRA)/td>/tr>tr> td stylepadding-right:0;>International Conference on the Constraint Programming, AI, and Operations Research (CPAIOR)/td>/tr>tr> td stylepadding-right:0;>International Conference on Learning Representations (ICLR)/td>/tr>tr> td stylepadding-right:0;>International Conference on Machine Learning (ICML)/td>/tr>tr> td stylepadding-right:0;>International Conference on Machine Learning (ICML) SODS Workshop/td>/tr>tr> td stylepadding-right:0;>Journal of Machine Learning Research (JMLR)/td>/tr>tr> td stylepadding-right:0;>Learning for Dynamics and Control (L4DC)/td>/tr>tr> td stylepadding-right:0;>Mathematical Programming Computation (MPC)/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS)/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS) OPT Workshop/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS) DiffCVGP Workshop/td>/tr>tr> td stylepadding-right:0;>Neural Information Processing Systems (NeurIPS) Deep RL Workshop/td>/tr>tr> td stylepadding-right:0;>Optimization Letters/td>/tr>tr> td stylepadding-right:0;>Transactions on Machine Learning Research (TMLR)/td>/tr>/table>h2 id-teaching>i classfa fa-chevron-right>/i> Teaching/h2>table classtable table-hover>tr> td stylepadding-right:0>strong>Applied Machine Learning/strong> (Cornell Tech CS5785), Co-instructor/td> td classcol-md-2 styletext-align:right; padding-left:0;>F2024/td>/tr>tr> td stylepadding-right:0>strong>Graduate AI/strong> (CMU 15-780), TA/td> td classcol-md-2 styletext-align:right; padding-left:0;>S2017/td>/tr>tr> td stylepadding-right:0>strong>Distributed Systems/strong> (CMU 15-440/640), TA/td> td classcol-md-2 styletext-align:right; padding-left:0;>S2016/td>/tr>tr> td stylepadding-right:0>strong>Software Design and Data Structures/strong> (VT CS2114), TA/td> td classcol-md-2 styletext-align:right; padding-left:0;>S2013/td>/tr>/table>h2 id-skills>i classfa fa-chevron-right>/i> Skills/h2>table classtable table-hover>tr> td classcol-md-2>Programming/td> td>C, C++, Fortran, Haskell, Java, Lua, Make, Mathematica, Python, R, Scala /td>/tr>tr> td classcol-md-2>Frameworks/td> td>JAX, NumPy, Pandas, PyTorch, SciPy, TensorFlow, Torch7 /td>/tr>tr> td classcol-md-2>Toolbox/td> td>Linux, emacs, vim, evil, org, mu4e, xmonad, git, tmux, zsh /td>/tr>/table>!--## i classfa fa-chevron-right>/i> Recent Blog Poststable classtable table-hover> tr> td>a href/2016/08/09/deep-completion/>Image Completion with Deep Learning in TensorFlow/a>/td> td classcol-md-3 styletext-align: right;>August 9, 2016/td> /tr> /table>h4>a href/blog>View all/a>/h4>## i classfa fa-chevron-right>/i> Fun Side Projects+ CS conference tracker(https://github.com/bamos/conference-tracker).+ SnowGlobe(https://github.com/bamos/snowglobe): Haskell-driven, small-scale web analytics with minimal configuration.+ My reading list(http://bamos.github.io/reading-list/): YAML data and hosted on GitHub pages.+ dotfiles(https://github.com/bamos/dotfiles): ♥ Arch Linux(https://www.archlinux.org/), OSX, mutt(http://www.mutt.org/), xmonad(http://xmonad.org/), i3(https://i3wm.org/), vim(http://www.vim.org/), emacs(https://www.gnu.org/software/emacs/), zsh(http://www.zsh.org/), mpv(http://mpv.io/), cmus(https://cmus.github.io/).+ girl(https://github.com/bamos/girl): Scala program to find broken links in GitHub projects.+ zsh-history-analysis(https://github.com/bamos/zsh-history-analysis): Analyze shell usage patterns with Python and R.+ python-scripts(https://github.com/bamos/python-scripts): Short and fun Python scripts.+ This website(https://github.com/bamos/bamos.github.io): Built with Jekyll and hosted on GitHub pages.+ cv(https://github.com/bamos/cv): Python-driven resume-curriculum vitae with Jinja templates.+ yaml-mailer(https://github.com/bamos/yaml-mailer): Email many people different messages.+ latex-templates(https://github.com/bamos/latex-templates) and beamer-snippets(https://github.com/bamos/beamer-snippets): Personal collection and previewing of LaTeX and Beamer snippets. Admittedly, I now use Keynote for presentations.-->hr />p>Last updated on 2024-09-25/p> /div> /div>/div> script src/js/sp.js>/script> script src/vendor/js/bootstrap.min.js>/script> script src/vendor/js/anchor.min.js>/script> script src/vendor/js/jquery.toc.js>/script> script typetext/javascript> (function(i,s,o,g,r,a,m){iGoogleAnalyticsObjectr;irir||function(){ (ir.qir.q||).push(arguments)},ir.l1*new Date();as.createElement(o), ms.getElementsByTagName(o)0;a.async1;a.srcg;m.parentNode.insertBefore(a,m) })(window,document,script,https://www.google-analytics.com/analytics.js,ga); ga(create, UA-102191838-1, auto); ga(send, pageview); // $(#toc).toc({ // headings: h2,h3 // }); // anchors.add(h2,h3); /script>/body>/html>
View on OTX
|
View on ThreatMiner
Please enable JavaScript to view the
comments powered by Disqus.
Data with thanks to
AlienVault OTX
,
VirusTotal
,
Malwr
and
others
. [
Sitemap
]