BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//IDSS - ECPv5.16.3.1//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-ORIGINAL-URL:https://idss.mit.edu
X-WR-CALDESC:Events for IDSS
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:America/New_York
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20210314T070000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20211107T060000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20220313T070000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20221106T060000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220506T110000
DTEND;TZID=America/New_York:20220506T120000
DTSTAMP:20220809T064551
CREATED:20220124T213354Z
LAST-MODIFIED:20220428T191240Z
UID:15698-1651834800-1651838400@idss.mit.edu
SUMMARY:Sampling rare events in Earth and planetary science
DESCRIPTION:Abstract: This talk will cover recent work in our group developing and applying algorithms to simulate rare events in atmospheric science and other areas. I will review a rare event simulation scheme that biases model simulations toward the rare event of interest by preferentially duplicating simulations making progress toward the event and removing others. I will describe applications of this approach to rapid intensification of tropical cyclones and instability of Mercury’s orbit with an emphasis on the elements of algorithm design that most affect performance.\n\nBio: Jonathan Weare is currently an associate professor of mathematics in the Courant Institute of Mathematical Sciences at New York University. Previously he was an associate professor in the statistics department and in the James Franck Institute at the University of Chicago and\, before that\, an assistant professor in the mathematics department there. Before moving to Chicago he was a Courant Instructor of mathematics at NYU and a PhD student in mathematics at the University of California at Berkeley.
URL:https://idss.mit.edu/calendar/weare/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220429T110000
DTEND;TZID=America/New_York:20220429T120000
DTSTAMP:20220809T064551
CREATED:20220124T211441Z
LAST-MODIFIED:20220413T140106Z
UID:15695-1651230000-1651233600@idss.mit.edu
SUMMARY:Is quantile regression a suitable method to understand tax incentives for charitable giving? Case study from the Canton of Geneva\, Switzerland
DESCRIPTION:Abstract: Under the current Swiss law\, taxpayers can deduct charitable donations from their individual’s taxable income subject to a 20%-ceiling. This deductible ceiling was increased at the communal and cantonal level from a previous 5%-ceiling in 2009. The goal of the reform was boosting charitable giving to non-profit entities. However\, the effects of this reform\, and more generally of the existing Swiss system of tax deductions for charitable giving has never been empirically studied. The aim of this work is to provide as many taxation insights and deducters characteristics as possible into both the effects of the 2009 reform\, as well as into the patterns of giving and deducting by different classes of deducters by income and wealth. \nUsing unique panel data\, shared by the Geneva Tax Administration\, for a time framework of 11 years: 2001-2011\, an in-depth statistical analysis was conducted. The overall taxpayers population has been described\, dividing them into six categories according to the income distribution. We studied the changes in the volume of deductions between categories. Quantile regressions models for each year has been fitted to underlying the different income behaviors toward deductions. Moreover\, a specific subset of deducters more sensitive to the deductible ceiling for their donations was identified and studied in detail. The overall net income\, gross wealth\, together with the year of birth\, were the main covariates of interest. Standard linear regression and robust regression models were performed and significant variables\, which help answering the questions of taxpayers’ charitable giving behavior\, were identified. \nIncome has resulted the most significant variable\, driving donations\, and robust regressions the statistical techniques better incorporating the data peculiarity\, without giving too much weight to outliers\, and with an excellent model fitting. This paper seeks to provide both Swiss and foreign academics and policymakers with new research and policy insights. \n– \nBios: Giedre Lideikyte Huber is a Senior lecturer at the Faculty of Law and a Swiss National Science Foundation researcher. She specializes in tax law\, and more specifically in taxation of philanthropy\, corporate taxation and sustainable tax systems (including gender and climate issues in taxation). She has received numerous academic awards and grants\, awarded by the Swiss National Science Foundation (FNS)\, the Fondation Zdenek et Michaela Bakala\, the University of Geneva (Subside Tremplin) and Centre Maurice Chalumeau en sciences des sexualités. (see more) \nMarta Pittavino is a Senior Lecturer and a Senior Research Associate at the Research Center for Statistics of the Geneva School of Economics and Management (GSEM)\, within the University of Geneva. She is scientific coordinator and manager of the Master of Science in Business Analytics. Marta holds a PhD in Biostatistics and Epidemiology from the University of Zurich. Before joining the GSEM\, she was a post-doctoral scientist\, applied statistician\, at the International Agency for Research on Cancer\, part of the World Health Organization\, in Lyon\, France. Her research interests lie in the applied statistics field: data analysis\, forecasting and regression methods\, with a focus on the development of Bayesian hierarchical models applied to epidemiological studies
URL:https://idss.mit.edu/calendar/lideikyte-huber/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220422T110000
DTEND;TZID=America/New_York:20220422T120000
DTSTAMP:20220809T064551
CREATED:20220124T210948Z
LAST-MODIFIED:20220422T092752Z
UID:15693-1650625200-1650628800@idss.mit.edu
SUMMARY:Learning with Random Features and Kernels: Sharp Asymptotics and Universality Laws
DESCRIPTION:Abstract: Many new random matrix ensembles arise in learning and modern signal processing. As shown in recent studies\, the spectral properties of these matrices help answer crucial questions regarding the training and generalization performance of neural networks\, and the fundamental limits of high-dimensional signal recovery. As a result\, there has been growing interest in precisely understanding the spectra and other asymptotic properties of these matrices. Unlike their classical counterparts\, these new random matrices are often highly structured and are the result of nonlinear transformations. This combination of structure and nonlinearity leads to substantial technical challenges when applying existing tools from random matrix theory to these new random matrix ensembles. \nIn this talk\, we will consider learning by random feature models and the related problem of kernel ridge regression. In each case\, a nonlinear random matrix plays a prominent role. We provide an exact characterization of the asymptotic training and generalization errors of these models. These results reveal the important roles played by the regularization\, the loss function and the activation function in the mitigation of the “double descent phenomenon” in learning. The asymptotic analysis is made possible by a general universality theorem\, which establishes the asymptotic equivalence between the nonlinear random matrices and a surrogate linear random matrix ensemble that is much easier to work with. \n– \nBio: Yue M. Lu attended the University of Illinois at Urbana-Champaign\, where he received the M.Sc. degree in mathematics and the Ph.D. degree in electrical engineering\, both in 2007. After his postdoctoral training at the Audiovisual Communications Laboratory at Ecole Polytechnique Fédérale de Lausanne (EPFL)\, Switzerland\, he joined Harvard University\, where he is currently Gordon McKay Professor of Electrical Engineering and of Applied Mathematics at the John A. Paulson School of Engineering and Applied Sciences. He is also fortunate to have held visiting appointments at Duke University in 2016 and at the École Normale Supérieure (ENS) in 2019. His research interests include theoretical and algorithmic aspects of high-dimensional signal and information processing. He is an IEEE Signal Processing Society Distinguished Lecturer and a recipient of the ECE Illinois Young Alumni Achievement Award. \n– \nA full schedule for Spring 2022 Stochastics and Statistics Seminars can be found here:https://stat.mit.edu/seminars/upcoming/
URL:https://idss.mit.edu/calendar/lu2/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220415T110000
DTEND;TZID=America/New_York:20220415T120000
DTSTAMP:20220809T064551
CREATED:20220124T210116Z
LAST-MODIFIED:20220411T135103Z
UID:15691-1650020400-1650024000@idss.mit.edu
SUMMARY:Causal Representation Learning – A Proposal
DESCRIPTION:Abstract: The development of CRISPR-based assays and small molecule screens holds the promise of engineering precise cell state transitions to move cells from one cell type to another or from a diseased state to a healthy state. The main bottleneck is the huge space of possible perturbations/interventions\, where even with the breathtaking technological advances in single-cell biology it will never be possible to experimentally perturb all combinations of thousands of genes or compounds. This important biological problem calls for a framework that can integrate data from different modalities to identify causal representations\, predict the effect of unseen interventions\, and identify the optimal interventions to induce precise cell state transition. Traditional representation learning methods\, although often highly successful in predictive tasks\, do not generally elucidate causal relationships. In this talk\, we will present initial ideas towards building a statistical and computational framework for causal representation learning and its application towards optimal intervention design. \n– \nBio: Caroline Uhler is the Henry L. and Grace Doherty associate professor in EECS and IDSS\, a member of SDSC\, LIDS and the ORC\, Machine Learning at MIT\, and is also core member of the Broad Institute\, where she co-directs the Eric and Wendy Schmidt Center. She is an elected member of the International Statistical Institute and the recipient of a Simons Investigator Award\, a Sloan Research Fellowship\, an NSF Career Award\, a Sofja Kovalevskaja Award from the Humboldt Foundation\, and a START Award from the Austrian Science Fund.
URL:https://idss.mit.edu/calendar/uhler/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220408T110000
DTEND;TZID=America/New_York:20220408T120000
DTSTAMP:20220809T064551
CREATED:20220124T205412Z
LAST-MODIFIED:20220331T125923Z
UID:15688-1649415600-1649419200@idss.mit.edu
SUMMARY:The query complexity of certification
DESCRIPTION:Abstract: We study the problem of certification: given queries to an n-variable boolean function f with certificate complexity k and an input x\, output a size-k certificate for f’s value on x. This abstractly models a problem of interest in explainable machine learning\, where we think of f as a blackbox model that we seek to explain the predictions of. \nFor monotone functions\, classic algorithms of Valiant and Angluin accomplish this task with n queries to f. Our main result is a new algorithm for certifying monotone functions with O(k^8 log(n)) queries\, which comes close to matching the information-theoretic lower bound of Omega(k log(n)). The design and analysis of our algorithm are based on a new connection to threshold phenomena in monotone functions. \nJoint work with Guy Blanc\, Caleb Koch\, and Jane Lange. Available at https://arxiv.org/abs/2201.07736. \n– \nBio: Li-Yang Tan is an assistant professor of computer science at Stanford. He is broadly interested in theoretical computer science\, with an emphasis on computational complexity. A main theme in his work is the development of techniques to understand boolean function complexity\, and the application of these techniques to a range of areas in theoretical computer science. His work has been recognized with best paper awards at FOCS and CCC.
URL:https://idss.mit.edu/calendar/tan/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220318T110000
DTEND;TZID=America/New_York:20220318T120000
DTSTAMP:20220809T064551
CREATED:20220124T205021Z
LAST-MODIFIED:20220314T121813Z
UID:15686-1647601200-1647604800@idss.mit.edu
SUMMARY:Mean-field approximations for high-dimensional Bayesian Regression
DESCRIPTION:Abstract:\nVariational approximations provide an attractive computational alternative to MCMC-based strategies for approximating the posterior distribution in Bayesian inference. Despite their popularity in applications\, supporting theoretical guarantees are limited\, particularly in high-dimensional settings. \nIn the first part of the talk\, we will study bayesian inference in the context of a linear model with product priors\, and derive sufficient conditions for the correctness (to leading order) of the naive mean-field approximation. To this end\, we will utilize recent advances in the theory of non-linear large deviations (Chatterjee and Dembo 2014). Next\, we analyze the naive mean-field variational problem\, and precisely characterize the asymptotic properties of the posterior distribution in this setting. \nIn the second part of the talk\, we will turn to linear regression with iid gaussian design under a proportional asymptotic setting. The naive mean- field approximation is conjectured to be inaccurate in this case|instead\, the Thouless-Anderson-Palmer approximation from statistical physics is expected to provide a tight approximation. We will rigorously establish the TAP formula under a uniform spherical prior on the regression coefficients. This is based on joint work with Sumit Mukherjee (Columbia University) and Jiaze Qiu (Harvard University). \n– \nBio:\nSubhabrata Sen is an assistant professor in the Department of Statistics\, Harvard University. His research interests span Applied Probability\, Statistics\, and Machine Learning. He was a Schramm postdoc at Microsoft Research New England and MIT Mathematics from 2017-2019. He graduated from the Stanford Statistics Department in 2017\, where he was jointly advised by Prof Amir Dembo and Prof Andrea Montanari. Prior to joining Stanford\, he received his undergraduate and Masters degrees in Statistics from the Indian Statistical Institute\, Kolkata.
URL:https://idss.mit.edu/calendar/sen/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220311T110000
DTEND;TZID=America/New_York:20220311T120000
DTSTAMP:20220809T064551
CREATED:20220124T204503Z
LAST-MODIFIED:20220302T145054Z
UID:15683-1646996400-1647000000@idss.mit.edu
SUMMARY:Inference on Winners
DESCRIPTION:Abstract: Many empirical questions concern target parameters selected through optimization. For example\, researchers may be interested in the effectiveness of the best policy found in a randomized trial\, or the best-performing investment strategy based on historical data. Such settings give rise to a winner’s curse\, where conventional estimates are biased and conventional confidence intervals are unreliable. This paper develops optimal confidence intervals and median-unbiased estimators that are valid conditional on the target selected and so overcome this winner’s curse. If one requires validity only on average over targets that might have been selected\, we develop hybrid procedures that combine conditional and projection confidence intervals to offer further performance gains relative to existing alternatives. \nBio: Isaiah Andrews is a Professor of Economics at Harvard University\, a Research Associate at the National Bureau of Economic Research (NBER)\, a fellow of the Econometric Society\, and a co-editor at the American Economic Review. He specializes in econometrics\, and his research focuses on developing methods for inference that are robust to common problems in empirical work\, including insufficiently informative data (weak identification) and model misspecification. He received a MacArthur fellowship in 2020 and the John Bates Clark Medal in 2021.
URL:https://idss.mit.edu/calendar/andrews/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220304T110000
DTEND;TZID=America/New_York:20220304T120000
DTSTAMP:20220809T064551
CREATED:20220304T141324Z
LAST-MODIFIED:20220304T141437Z
UID:16039-1646391600-1646395200@idss.mit.edu
SUMMARY:Optimal testing for calibration of predictive models
DESCRIPTION:Abstract: The prediction accuracy of machine learning methods is steadily increasing\, but the calibration of their uncertainty predictions poses a significant challenge. Numerous works focus on obtaining well-calibrated predictive models\, but less is known about reliably assessing model calibration. This limits our ability to know when algorithms for improving calibration have a real effect\, and when their improvements are merely artifacts due to random noise in finite datasets. In this work\, we consider the problem of detecting mis-calibration of predictive models using a finite validation dataset. Due to the randomness in the data\, plug-in measures of calibration need to be compared against a proper background distribution to reliably assess calibration. Thus\, detecting mis-calibration in a classification setting can be formulated as a statistical hypothesis testing problem. The null hypothesis is that the model is perfectly calibrated\, while the alternative hypothesis is that the deviation from calibration is sufficiently large. \nWe find that detecting mis-calibration is only possible when the conditional probabilities of the classes are sufficiently smooth functions of the predictions. When the conditional class probabilities are H\”older continuous\, we propose a minimax optimal test for calibration based on a debiased plug-in estimator of the $\ell_2$-Expected Calibration Error (ECE). We further propose a version that is adaptive to unknown smoothness. We verify our theoretical findings with a broad range of experiments\, including with several popular deep neural net architectures and several standard post-hoc calibration methods. Our algorithm is a general-purpose tool\, which—combined with classical tests for calibration of discrete-valued predictors—can be used to test the calibration of virtually any classification method. \n— \nBio: Edgar Dobriban is an assistant professor of statistics & computer science at the University of Pennsylvania. He obtained a PhD in statistics from Stanford University in 2017\, and a BA in Mathematics from Princeton University in 2012. His research interests include the statistical analysis of large datasets\, and the theoretical analysis of machine learning. He has received a Theodore W. Anderson award for the best PhD in theoretical statistics from Stanford University\, and an NSF CAREER award. More information is available at his website \nhttps://statistics.wharton.upenn.edu/profile/dobriban/. \n— \nA full schedule for Spring 2022 Stochastics and Statistics Seminars can be found here:https://stat.mit.edu/seminars/upcoming/
URL:https://idss.mit.edu/calendar/optimal-testing-for-calibration-of-predictive-models/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220218T110000
DTEND;TZID=America/New_York:20220218T120000
DTSTAMP:20220809T064551
CREATED:20220124T203258Z
LAST-MODIFIED:20220214T212225Z
UID:15679-1645182000-1645185600@idss.mit.edu
SUMMARY:On the power of Lenstra-Lenstra-Lovasz in noiseless inference
DESCRIPTION:Abstract: In this talk\, we are going to discuss a new polynomial-time algorithmic framework for inference problems\, based on the celebrated Lenstra-Lenstra-Lovasz lattice basis reduction algorithm. Potentially surprisingly\, this algorithmic framework is able to successfully bypass multiple suggested notions of “computational hardness for inference” for various noiseless settings. Such settings include 1) sparse regression\, where there is Overlap Gap Property and low-degree methods fail\, 2) phase retrieval where Approximate Message Passing fails and 3) Gaussian clustering where the SoS hierarchy fails. In particular\, our results\, similar to the folklore but specific Gaussian elimination application in random XORSAT\, highlight the crucial but subtle role of noise in the onset of statistical-to-computational gaps. \nThis is based on joint works with David Gamarnik\, Eren Kizildag\, Min Jae Song\, Joan Bruna and Alex Wein. \n– \nBio: Ilias Zadik is a postdoctoral researcher at the Mathematics Department of MIT\, working with Professor Elchanan Mossel and Professor Nike Sun. Previously\, he spent two years as a postdoctoral fellow at the Center of Data Science of NYU\, and before that he received his PhD in 2019 from the Operations Research Center of MIT under the supervision of Professor David Gamarnik. His research lies broadly on the interface of high-dimensional statistics\, the theory of machine learning and applied probability. He is especially interested in understanding the statistical and computational trade-offs appearing in high dimensional inference.
URL:https://idss.mit.edu/calendar/zadik/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20220204T110000
DTEND;TZID=America/New_York:20220204T120000
DTSTAMP:20220809T064551
CREATED:20220125T135627Z
LAST-MODIFIED:20220202T211029Z
UID:15702-1643972400-1643976000@idss.mit.edu
SUMMARY:The Brownian transport map
DESCRIPTION:Abstract: \nThe existence of a transport map from the standard Gaussian leads to succinctrepresentations for\, potentially complicated\, measures. Inspired by result from optimal transport\, we introduce the Brownian transport map that pushes forward the Wiener measure to a target measure in a finite-dimensional Euclidean space. \nUsing tools from Ito’s and Malliavin’s calculus\, we show that the map is Lipschitz in several cases of interest. Specifically\, our results apply when the target measure satisfies one of the following: \n– More log-concave than the Gaussian\, recovering a result of Caffarelli. \n– Bounded convex support with a semi log-concave density\, providing an affirmative answer to a question first posed for the Brenier map. \n– A mixture of Gaussians\, explaining recent results about dimension-free functional inequalities for such measures. \n– log-concave and isotropic. In this case\, we establish a direct connection between the Poincare constant and the (averaged) Lipschitz constant of the Brownian transport map. Since the Poincare constant is the object of the famous KLS conjecture\, we essentially show that the conjecture is equivalent to the existence of a suitable transportation map. \nJoint work with Yair Shenfeld \nBio: \nDan Mikulincer is an instructor (postdoc) at MIT Mathematics. Previously\, he finished his Ph.D. at the Weizmann Institute\, Faculty of Mathematics\, where he was advised by Ronen Eldan. He completed his B.Sc. in Mathematics and Computer Science at Ben-Gurion University (BGU)\, where he also studied Cognitive Neuroscience. He spent the summer of 2019 at Microsoft Research AI\, hosted by Sébastien Bubeck\, and is also an Azrieli fellow. His research interests broadly lie at the union of high-dimensional geometry\, probability\, statistics\, information theory\, and their relation to data science and learning theory. He is particularly interested in normal approximations and dimension-free phenomena.
URL:https://idss.mit.edu/calendar/tbd-25/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211203T110000
DTEND;TZID=America/New_York:20211203T120000
DTSTAMP:20220809T064551
CREATED:20210818T194203Z
LAST-MODIFIED:20211119T135518Z
UID:14844-1638529200-1638532800@idss.mit.edu
SUMMARY:The Geometry of Particle Collisions: Hidden in Plain Sight
DESCRIPTION:Abstract:\nSince the 1960s\, particle physicists have developed a variety of data analysis strategies for the goal of comparing experimental measurements to theoretical predictions. Despite their numerous successes\, these techniques can seem esoteric and ad hoc\, even to practitioners in the field. In this talk\, I explain how many particle physics analysis tools have a natural geometric interpretation in an emergent “space” of collider events induced by the Wasserstein metric. This in turn suggests new analysis strategies to interpret generic point cloud data sets.\n\n\nBio:\nJesse Thaler is a theoretical particle physicist who fuses techniques from quantum field theory and machine learning to address outstanding questions in fundamental physics. His current research is focused on maximizing the discovery potential of the Large Hadron Collider through new theoretical frameworks and novel data analysis techniques. Prof. Thaler joined the MIT Physics Department in 2010\, and is currently a Professor in the Center for Theoretical Physics. He was a Miller Fellow at U.C. Berkeley from 2006 to 2009\, and he received his Ph.D. in Physics from Harvard. He was awarded a Presidential Early Career Award for Scientists and Engineers in 2012 and a Sloan Research Fellowship in 2013. In 2020\, Prof. Thaler became the inaugural Director of the NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI).
URL:https://stat.mit.edu/calendar/thaler/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211119T110000
DTEND;TZID=America/New_York:20211119T120000
DTSTAMP:20220809T064551
CREATED:20210818T192910Z
LAST-MODIFIED:20211101T143329Z
UID:14842-1637319600-1637323200@idss.mit.edu
SUMMARY:Precise high-dimensional asymptotics for AdaBoost via max-margins & min-norm interpolants
DESCRIPTION:Abstract: This talk will introduce a precise high-dimensional asymptotic theory for AdaBoost on separable data\, taking both statistical and computational perspectives. We will consider the common modern setting where the number of features p and the sample size n are both large and comparable\, and in particular\, look at scenarios where the data is asymptotically separable. Under a class of statistical models\, we will provide an (asymptotically) exact analysis of the max-min-L1-margin and the min-L1-norm interpolant. In turn\, this will characterize the generalization error of AdaBoost\, when the algorithm interpolates the training data and maximizes an empirical L1 margin. On the computational front\, we will provide a sharp analysis of the stopping time when boosting approximately maximizes the empirical L1 margin. Our theory provides several insights into properties of AdaBoost; for instance\, the larger the dimensionality ratio p/n\, the faster the optimization reaches interpolation. Our statistical and computational arguments can handle (1) finite-rank spiked covariance models for the feature distribution and (2) variants of AdaBoost corresponding to general Lq-geometry\, for q in [1\,2]. This is based on joint work with Tengyuan Liang. \n—————– \nBio: \nPragya Sur is an Assistant Professor in the Statistics Department at Harvard University. Her research broadly spans high-dimensional statistics\, statistical machine learning\, robust inference and prediction for multi-study/multi-environment heterogeneous data. She is simultaneously interested in applications of large scale statistical methods to computational neuroscience and genetics. Her research is currently supported by a William F. Milton Fund Award and an NSF DMS award. Previously\, she was a postdoctoral fellow at the Center for Research on Computation and Society\, Harvard John A. Paulson School of Engineering and Applied Sciences. She received a Ph.D. in Statistics from Stanford University in 2019\, where she received the Theodore W. Anderson Theory of Statistics Dissertation Award and a Ric Weiland Graduate Fellowship in the Humanities and Sciences. \n
URL:https://stat.mit.edu/calendar/sur/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211112T110000
DTEND;TZID=America/New_York:20211112T120000
DTSTAMP:20220809T064551
CREATED:20210818T192354Z
LAST-MODIFIED:20211021T182021Z
UID:14840-1636714800-1636718400@idss.mit.edu
SUMMARY:Characterizing the Type 1-Type 2 Error Trade-off for SLOPE
DESCRIPTION:Abstract: Sorted L1 regularization has been incorporated into many methods for solving high-dimensional statistical estimation problems\, including the SLOPE estimator in linear regression. In this talk\, we study how this relatively new regularization technique improves variable selection by characterizing the optimal SLOPE trade-off between the false discovery proportion (FDP) and true positive proportion (TPP) or\, equivalently\, between measures of type I and type II error. Additionally\, we show that on any problem instance\, SLOPE with a certain regularization sequence outperforms the Lasso\, in the sense of having a smaller FDP\, larger TPP and smaller L2 estimation risk simultaneously. Our proofs are based on a novel technique that reduces a variational calculus problem to a class of infinite-dimensional convex optimization problems and a very recent result from approximate message passing (AMP) theory. With SLOPE being a particular example\, we discuss these results in the context of a general program for systematically deriving exact expressions for the asymptotic risk of estimators that are solutions to a broad class of convex optimization problems via AMP. Collaborators on this work include Zhiqi Bu\, Jason Klusowski\, and Weijie Su (https://arxiv.org/abs/1907.07502 and https://arxiv.org/abs/2105.13302) and Oliver Feng\, Ramji Venkataramanan\, and Richard Samworth (https://arxiv.org/abs/2105.02180). \n– \nBio: Cynthia Rush is the Howard Levene Assistant Professor of Statistics in the Department of Statistics at Columbia University. In May\, 2016\, she received a Ph.D. in Statistics from Yale University under the supervision of Andrew Barron and she completed her undergraduate coursework at the University of North Carolina at Chapel Hill where she obtained a B.S. in Mathematics. She received an NSF CRIII award in 2019\, was an NTT Research Fellow at the Simons Institute for the Theory of Computing for the program on Probability\, Computation\, and Geometry in High Dimensions in Fall 2020\, and is currently a Google Research Fellow at the Simons Institute for the Theory of Computing for the program on Computational Complexity of Statistical Inference.
URL:https://stat.mit.edu/calendar/rush/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211105T110000
DTEND;TZID=America/New_York:20211105T120000
DTSTAMP:20220809T064551
CREATED:20210818T192103Z
LAST-MODIFIED:20211102T175713Z
UID:14838-1636110000-1636113600@idss.mit.edu
SUMMARY:Asymptotics of learning on dependent and structured random objects
DESCRIPTION:Abstract: Classical statistical inference relies on numerous tools from probability theory to study the properties of estimators. However\, these same tools are often inadequate to study modern machine problems that frequently involve structured data (e.g networks) or complicated dependence structures (e.g dependent random matrices). In this talk\, we extend universal limit theorems beyond the classical setting. Firstly\, we consider distributionally \structured” and dependent random object i.e random objects whose distribution are invariant under the action of an amenable group. We show\, under mild moment and mixing conditions\, a series of universal second and third order limit theorems: central-limit theorems\, concentration inequalities\, Wigner semi-circular law and Berry-Esseen bounds. The utility of these will be illustrated by a series of examples in machine learning\, network and information theory. Secondly\nby building on these results\, we establish the asymptotic distribution of the cross-\nvalidated risk with the number of folds allowed to grow at an arbitrary rate. Using this\, we study the statistical speed-up of cross validation compared to a train-test split procedure\, which reveals surprising results even when used on simple estimators. \nBio: Morgane Austern is an assistant professor at Harvard University in the statistics department. Her research focuses on problems in probability and statistics that are motivated by machine learning. She graduated with a PhD in statistics from Columbia University in 2019 where she worked in collaboration with Peter Orbanz and Arian Maleki on limit theorems for dependent and structured data. Previously\, she was also a postdoctoral researcher at Microsoft Research New England.
URL:https://stat.mit.edu/calendar/austern/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211029T110000
DTEND;TZID=America/New_York:20211029T120000
DTSTAMP:20220809T064551
CREATED:20210818T191208Z
LAST-MODIFIED:20211020T145253Z
UID:14836-1635505200-1635508800@idss.mit.edu
SUMMARY:Revealing the simplicity of high-dimensional objects via pathwise analysis
DESCRIPTION:Abstract: One of the main reasons behind the success of high-dimensional statistics and modern machine learning in taming the curse of dimensionality is that many classes of high-dimensional distributions are surprisingly well-behaved and\, when viewed correctly\, exhibit a simple structure. This emergent simplicity is in the center of the theory of “high-dimensional phenomena”\, and is manifested in principles such as “Gaussian-like behavior” (objects of interest often inherit the properties of the Gaussian measure)\, “dimension-free behavior” (expressed in inequalities which do not depend on the dimension) and “mean-field behavior” (where the behavior of a system having many degrees of freedom can be compared to a “limit object” having a small number thereof). \nIn this talk\, we present a new analytic approach that helps reveal phenomenona of this nature. The approach is based on pathwise analysis: We construct a sampling procedure associated with the high-dimensional object\, which uses the randomness coming from a Brownian motion. This gives rise to a stochastic process which allows us to make the object tractable by the analysis of the process\, via Ito calculus (the theory of diffusing particles) and relate quantities of interest of the object with the behavior of the process\, for example\, through differentiation with respect to time. \nI will try to explain how this approach works and will briefly discuss several results\, of relevance to high dimensional statistics and machine learning\, that stem from it. These results include concentration inequalities\, central limit theorems and “mean-field” structure theorems. \n\n\n–\n\nBio: Ronen Eldan works at the Weizmann Institute of Science\, and spends the current year at the Institute for Advanced Studies in Princeton. He studies phenomena that arise in high-dimensional settings in probability\, analysis\, mathematical physics and combinatorics\, as well as the application of these phenomena to high dimensional statistics\, machine learning and optimization and theory of computer science. One of his main projects in recent years has been to develop methods which help understand the behavior of high dimensional objects by establishing new connections with the field of stochastic calculus.
URL:https://stat.mit.edu/calendar/eldan/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211022T110000
DTEND;TZID=America/New_York:20211022T120000
DTSTAMP:20220809T064551
CREATED:20210818T190944Z
LAST-MODIFIED:20211018T122124Z
UID:14834-1634900400-1634904000@idss.mit.edu
SUMMARY:Instance Dependent PAC Bounds for Bandits and Reinforcement Learning
DESCRIPTION:Abstract: The sample complexity of an interactive learning problem\, such as multi-armed bandits or reinforcement learning\, is the number of interactions with nature required to output an answer (e.g.\, a recommended arm or policy) that is approximately close to optimal with high probability. While minimax guarantees can be useful rules of thumb to gauge the difficulty of a problem class\, algorithms optimized for this worst-case metric often fail to adapt to “easy” instances where fewer samples suffice. In this talk\, I will highlight some my group’s work on algorithms that obtain optimal\, finite time\, instance dependent sample complexities that scale with the true difficult of the particular instance\, versus just the worst-case. In particular\, I will describe a unifying experimental design based approach used to obtain such algorithms for best-arm identification for linear bandits\, contextual bandits with arbitrary policy classes\, and smooth losses for linear dynamical systems.\n–\nBio: Kevin Jamieson is an Assistant Professor in the Paul G. Allen School of Computer Science & Engineering at the University of Washington and is the Guestrin Endowed Professor in Artificial Intelligence and Machine Learning. He received his B.S. in 2009 from the University of Washington\, his M.S. in 2010 from Columbia University\, and his Ph.D. in 2015 from the University of Wisconsin – Madison under the advisement of Robert Nowak\, all in electrical engineering. He returned to the University of Washington as faculty in 2017 after a postdoc with Benjamin Recht at the University of California\, Berkeley. Jamieson’s research explores how to leverage already-collected data to inform what future measurements to make next\, in a closed loop. His work ranges from theory to practical algorithms with guarantees to open-source machine learning systems and has been adopted in a range of applications\, including measuring human perception in psychology studies\, adaptive A/B/n testing in dynamic web-environments\, numerical optimization\, and efficient tuning of hyperparameters for deep neural networks.
URL:https://stat.mit.edu/calendar/jamieson/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211015T110000
DTEND;TZID=America/New_York:20211015T120000
DTSTAMP:20220809T064551
CREATED:20210818T185538Z
LAST-MODIFIED:20211007T193028Z
UID:14832-1634295600-1634299200@idss.mit.edu
SUMMARY:Breaking the Sample Size Barrier in Reinforcement Learning
DESCRIPTION:Abstract:\nReinforcement learning (RL)\, which is frequently modeled as sequential learning and decision making in the face of uncertainty\, is garnering growing interest in recent years due to its remarkable success in practice. In contemporary RL applications\, it is increasingly more common to encounter environments with prohibitively large state and action space\, thus imposing stringent requirements on the sample efficiency of the RL algorithms in use. Despite the empirical success\, however\, the theoretical underpinnings for many popular RL algorithms remain highly inadequate even for the tabular setting.\nIn this talk\, we present two vignettes regarding the sample efficiency of RL algorithms. The first vignette demonstrates that a perturbed model-based RL approach is minimax optimal under a generative model\, without suffering from a sample size barrier that was present in all past work. In the second vignette\, we pin down the sample complexity of Q-learning on Markovian samples\, which substantially improves upon prior results by a factor at least as large as the dimension of the state-action space. These results cover two distinctive RL paradigms and might shed light on the efficacy of these algorithms in more complicated scenarios. \nBio:\nYuting Wei is currently an assistant professor in the Statistics and Data Science Department at the Wharton School\, University of Pennsylvania. Prior to that\, Yuting spent two years at Carnegie Mellon University as an assistant professor of statistics\, and one year at Stanford University as a Stein Fellow. She received her Ph.D. in statistics at University of California\, Berkeley\, working with Martin Wainwright and Aditya Guntuboyina. She was the recipient of the 2018 Erich L. Lehmann Citation from the Berkeley statistics department for her Ph.D. dissertation in theoretical statistics. Her research interests include high-dimensional and non-parametric statistics\, statistical machine learning\, and reinforcement learning.
URL:https://stat.mit.edu/calendar/wei/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211008T110000
DTEND;TZID=America/New_York:20211008T120000
DTSTAMP:20220809T064551
CREATED:20210818T185330Z
LAST-MODIFIED:20210924T191943Z
UID:14830-1633690800-1633694400@idss.mit.edu
SUMMARY:Recent results in planted assignment problems
DESCRIPTION:Abstract: Motivated by applications such as particle tracking\, network de-anonymization\, and computer vision\, a recent thread of research is devoted to statistical models of assignment problems\, in which the data are random weight graphs correlated with the latent permutation. In contrast to problems such as planted clique or stochastic block model\, the major difference here is the lack of low-rank structures\, which brings forth new challenges in both statistical analysis and algorithm design.\n\n \nIn the first half of the talk\, we discuss the linear assignment problem\, where the goal is to reconstruct a perfect matching planted in a randomly weighted bipartite graph\, whose planted and unplanted edge weights are independently drawn from two different distributions. We determine the sharp threshold at which the optimal reconstruction error (fraction of misclassified edges) exhibits a phase transition from imperfect to perfect. Furthermore\, for exponential weight distributions\, this phase transition is shown to be of infinite order\, confirming the conjecture of Semerjian\, Sicuro\, and Zdeborova. The negative result is shown by proving that\, below the threshold\, the posterior distribution is concentrated away from the hidden matching by constructing exponentially many long augmenting cycles. \nIn the second half of the talk\, we discuss the quadratic assignment problem (graph matching)\, where the goal is to recover the hidden vertex correspondence between two edge-correlated Erdos-Renyi graphs. We prove that there exists a sharp threshold\, above which one can correctly match all but a vanishing fraction of the vertices and below which matching any positive fraction is impossible\, a phenomenon known as the “all-or-nothing” phase transition. The proof builds upon a tight characterization of the mutual information via the truncated second-moment method and an appropriate “area theorem”. Achieving these thresholds with efficient algorithms remains open. \nThis talk is based on joint work with Jian Ding\, Jiaming Xu\, Dana Yang and Sophie Yu. Preprints available at:https://arxiv.org/abs/2103.09383\, https://arxiv.org/abs/2008.10097\, https://arxiv.org/abs/2102.00082. \n– \n\nBio: Yihong Wu is an associate professor in the Department of Statistics and Data Science at Yale University. He received his B.E. degree from Tsinghua University in 2006 and Ph.D. from Princeton University in 2011. He is a recipient of the NSF CAREER award in 2017 and Sloan fellowship in 2018. He is broadly interested in the theoretical and algorithmic aspects of high-dimensional statistics\, information theory\, and optimization.
URL:https://stat.mit.edu/calendar/wu/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20211001T110000
DTEND;TZID=America/New_York:20211001T120000
DTSTAMP:20220809T064551
CREATED:20210818T184600Z
LAST-MODIFIED:20210922T121320Z
UID:14828-1633086000-1633089600@idss.mit.edu
SUMMARY:Causal Matrix Completion
DESCRIPTION:Matrix completion is the study of recovering an underlying matrix from a sparse subset of noisy observations. Traditionally\, it is assumed that the entries of the matrix are “missing completely atrandom” (MCAR)\, i.e.\, each entry is revealed at random\, independent of everything else\, with uniform probability. This is likely unrealistic due to the presence of “latent confounders”\, i.e.\, unobserved factors that determine both the entries of the underlying matrix and the missingness pattern in the observed matrix. \n\nIn general\, these confounders yield “missing not at random” (MNAR) data\, which can severely impact any inference procedure that does not correct for this bias. We develop a formal causal model for matrix completion with MNAR data through the language of potential outcomes\, and provide identification arguments for causal estimand of interest. We design a procedure\, which we call “synthetic nearest neighbors” (SNN)\, to estimate these causal estimands. We prove finite-sample consistency and asymptotic normality of our estimator. Our analysis also leads to new theoretical results for the matrix completion literature. In particular\, we establish entry-wise\, i.e.\, max-norm\, finite-sample consistency and asymptotic normality results for matrix completion with MNAR data. As a special case\, this also provides entry-wise bounds for matrix completion with MCAR data. Across simulated and real data\, we demonstrate the efficacy of our proposed estimator.\n\n\n\nThis is based on joint works with Anish Agarwal (MIT)\, Munzer Dahleh (MIT) and Dennis Shen (UC Berkeley).\n\n\nBio: Devavrat Shah is the Andrew (1956) and Erna Viterbi Professor with the department of electrical engineering and computer science\, MIT. He is a member of LIDS and the ORC\, and the Faculty Director of the MicroMasters in Statistics and Data Science program at IDSS. His research focus is on theory of large complex networks\, which includes network algorithms\, stochastic networks\, network information theory and large-scale statistical inference.
URL:https://stat.mit.edu/calendar/shah/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210924T110000
DTEND;TZID=America/New_York:20210924T120000
DTSTAMP:20220809T064551
CREATED:20210818T182019Z
LAST-MODIFIED:20210920T132357Z
UID:14826-1632481200-1632484800@idss.mit.edu
SUMMARY:Representation and generalization
DESCRIPTION:Abstract: Self-supervised learning is an increasingly popular approach for learning representations of data that can be used for downstream representation tasks. A practical advantage of self-supervised learning is that it can be used on unlabeled data. However\, even when labels are available\, self-supervised learning can be competitive with the more “traditional” approach of supervised learning. \n \nIn this talk we consider “self supervised + simple classifier (SSS)” algorithms\, which are obtained by first learning a self-supervised classifier on data\, and then reusing the same data and its labels to fit a simple (e.g.\, linear) classifier on these data. We show that: \n \n1) Unlike traditional end-to-end supervised learning algorithms\, SSS algorithm have small generalization gaps in practice\, even when the final classifier is highly over-parameterized. \n \n2) Under natural assumptions we can prove that the generalization gap will tend to zero if the number of samples is sufficiently larger than the complexity of the simple classifier\, independently of the complexity of the self-supervised model. We show that the bound proven yields non-vacuous guarantees for many popular representation-learning based classifiers on CIFAR-10 and ImageNet\, including SimCLR\, AMDIM and BigBiGAN. \n \n3) We give evidence that self-supervised and fully supervised models learn similar representations\, by showing that the self-supervised layers can be “stitched” to the bottom of a fully-supervised model\, and vice versa\, without a significant loss of performance. \n \nBased on joint works with Bansal and Kaplun ( https://arxiv.org/abs/2010.08508 )\, and Bansal and Nakkiran ( https://arxiv.org/abs/2106.07682 ). \n \nBio: Boaz Barak is the Gordon McKay professor of Computer Science at Harvard University’s John A. Paulson school of Engineering and Applied Sciences. His research interests include all areas of theoretical computer science and in particular cryptography and computational complexity. Previously\, he was a principal researcher at Microsoft Research New England\, and before that an associate professor (with tenure) at Princeton University’s computer science department. Barak has won the ACM dissertation award\, the Packard and Sloan fellowships\, and was also selected for Foreign Policy magazine’s list of 100 leading global thinkers for 2014 and chosen as a Simons investigator in 2017 .
URL:https://stat.mit.edu/calendar/barak/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210917T110000
DTEND;TZID=America/New_York:20210917T120000
DTSTAMP:20220809T064551
CREATED:20210818T175733Z
LAST-MODIFIED:20210913T190400Z
UID:14822-1631876400-1631880000@idss.mit.edu
SUMMARY:Interpolation and learning with scale dependent kernels
DESCRIPTION:Speaker: Lorenzo Rosasco (MIT/Universita’ di Genova) \nTitle: Interpolation and learning with scale dependent kernels \nAbstract: We study the learning properties of nonparametric ridge-less least squares. In particular\, we consider the common case of estimators defined by scale dependent (Matern) kernels\, and focus on the role scale and smoothness. These estimators interpolate the data and the scale can be shown to control their stability to noise and sampling. Larger scales\, corresponding to smoother functions\, improve stability with respect to sampling. However\, smaller scales\, corresponding to more complex functions\, improve stability to noise. We will discuss to which extent these results can explain the learning curves observed for large overparameterized models. Our analysis combines\, probabilistic results with analytic techniques from interpolation theory. \nBio: Lorenzo Rosasco is an assistant professor at the University of Genova\, Italy. He is also affiliated with the Massachusetts Institute of Technology(MIT)\, where is a visiting professor\, and with the Istituto Italiano di Tecnologia (IIT)\, where he is an external collaborator. He is leading the efforts to establish the Laboratory for Computational and Statistical Learning (LCSL)\, born from a collaborative agreement between IIT and MIT. He received his PhD from the University of Genova in 2006. Dr. Rosasco has developed and analyzed methods to learn from small as well as large samples of high dimensional data\, using analytical and probabilistic tools\, within a multidisciplinary approach drawing concepts and techniques primarily from computer science but also from statistics\, engineering and applied mathematics. \n
URL:https://stat.mit.edu/calendar/rosasco/
LOCATION:E18-304\, United States
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210514T110000
DTEND;TZID=America/New_York:20210514T120000
DTSTAMP:20220809T064551
CREATED:20210113T154925Z
LAST-MODIFIED:20210505T153802Z
UID:13878-1620990000-1620993600@idss.mit.edu
SUMMARY:Likelihood-Free Frequentist Inference
DESCRIPTION:Abstract: Many areas of the physical\, engineering and biological sciences make extensive use of computer simulators to model complex systems. Confidence sets and hypothesis testing are the hallmarks of statistical inference\, but classical methods are poorly suited for scientific applications involving complex simulators without a tractable likelihood. Recently\, many techniques have been introduced that learn a surrogate likelihood using forward-simulated data\, but these methods do not guarantee frequentist confidence sets and tests with nominal coverage and Type I error control\, respectively. \nIn this talk\, I will describe our recent and ongoing research on developing scalable and modular tools for constructing frequentist confidence sets with finite-sample validity. These tools apply to settings where we have access to a high-fidelity simulator but the likelihood cannot be evaluated and observed data are limited. Rather than relying on large-sample (asymptotic) theory or costly Monte Carlo samples at fixed parameter values\, we leverage machine learning tools and simulated data in the neighborhood of a parameter to estimate critical values\, p-values\, and conditional coverage of confidence sets. We refer to our general machinery as “likelihood-free frequentist inference”. Any method that estimates a test statistic\, such as the likelihood ratio\, can be plugged into our framework to efficiently compute valid hypothesis tests and confidence sets\, and run diagnostics. \nPart of this work is joint with Niccolo Dalmasso\, Rafael Izbicki and David Zhao. An earlier version of this work can be found in PMLR: \nhttp://proceedings.mlr.press/v119/dalmasso20a.html\n\n\n\n–\n\n\n\nBiography: Dr Lee is a professor in the Department of Statistics & Data Science at Carnegie Mellon University\, with a joint appointment in the Machine Learning Department. Prior to joining CMU in 2005\, she was the J.W. Gibbs Assistant Professor in the department of mathematics at Yale University\, and before that she served a year as a visiting research associate at the department of applied mathematics at Brown University. Dr Lee’s interests are in developing statistical methodology for the type of complex data and problems often encountered in the physical sciences. She co-directs the Statistical Methods for the Physical Sciences (STAMPS) research group at CMU\, and is key personnel in the recently founded NSF AI Planning Institute for Data-Driven Discovery in Physics.
URL:https://stat.mit.edu/calendar/lee/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210423T110000
DTEND;TZID=America/New_York:20210423T120000
DTSTAMP:20220809T064551
CREATED:20210113T155740Z
LAST-MODIFIED:20210415T144430Z
UID:13887-1619175600-1619179200@idss.mit.edu
SUMMARY:Prioritizing genes from genome-wide association studies
DESCRIPTION:Abstract: Prioritizing likely causal genes from genome-wide association studies (GWAS) is a fundamental problem. There are many methods for GWAS gene prioritization\, including methods that map candidate SNPs to their target genes and methods that leverage patterns of enrichment from across the genome. In this talk\, I will introduce a new method for leveraging genome-wide patterns of enrichment to prioritize genes at GWAS loci\, incorporating information about genes from many sources. I will then discuss the problem of benchmarking gene prioritization methods\, and I will describe a large-scale analysis to benchmark many different methods and combinations of methods on data from the UK Biobank. Our analyses show that the highest confidence can be achieved by combining multiple lines of evidence. I will conclude by giving examples of prioritized genes.\n\n–\n\nBio: Hilary Finucane is a faculty member at Harvard Medical School and Massachusetts General Hospital and the co-director of the Medical and Population Genetics Program at the Broad Institute of MIT and Harvard. She holds a BA from Harvard College in mathematics\, an MSc from the Weizmann Institute of Science in theoretical computer science\, and a PhD from MIT in applied mathematics. She did her PhD research in statistical genetics in the lab of Alkes Price\, funded by a Hertz Foundation Fellowship\, after which she started a research group at the Broad Institute as a Schmidt Fellow. She is also a recipient of the NIH Director’s Early Independence Award.
URL:https://stat.mit.edu/calendar/finucane/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210416T110000
DTEND;TZID=America/New_York:20210416T120000
DTSTAMP:20220809T064551
CREATED:20210113T155334Z
LAST-MODIFIED:20210412T172143Z
UID:13882-1618570800-1618574400@idss.mit.edu
SUMMARY:Sample Size Considerations in Precision Medicine
DESCRIPTION:Abstract: Sequential Multiple Assignment Randomized Trials (SMARTs) are considered the gold standard for estimation and evaluation of treatment regimes. SMARTs are typically sized to ensure sufficient power for a simple comparison\, e.g.\, the comparison of two fixed treatment sequences. Estimation of an optimal treatment regime is conducted as part of a secondary and hypothesis-generating analysis with formal evaluation of the estimated optimal regime deferred to a follow-up trial. However\, running a follow-up trial to evaluate an estimated optimal treatment regime is costly and time-consuming; furthermore\, the estimated optimal regime that is to be evaluated in such a follow-up trial may be far from optimal if the original trial was underpowered for estimation of an optimal regime. We derive sample size procedures for a SMART that ensure: (i) sufficient power for comparing the optimal treatment regime with standard of care; and (ii) the estimated optimal regime is within a given tolerance of the true optimal regime with high-probability. We establish asymptotic validity of the proposed procedures and demonstrate their finite sample performance in a series of simulation experiments. \nBio: Eric Laber is Professor of Statistical Science and Biostatistics and Bioinformatics at Duke University. His research focuses on data-driven decision making in health\, defense\, intelligence\, and retail. He is also passionate about K-12 data science outreach (more information is available at laber-labs.com).
URL:https://stat.mit.edu/calendar/laber/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210409T110000
DTEND;TZID=America/New_York:20210409T120000
DTSTAMP:20220809T064551
CREATED:20210113T155126Z
LAST-MODIFIED:20210409T121045Z
UID:13880-1617966000-1617969600@idss.mit.edu
SUMMARY:Functions space view of linear multi-channel convolution networks with bounded weight norm
DESCRIPTION:Abstract: The magnitude of the weights of a neural network is a fundamental measure of complexity that plays a crucial role in the study of implicit and explicit regularization. For example\, in recent work\, gradient descent updates in overparameterized models asymptotically lead to solutions that implicitly minimize the ell_2 norm of the parameters of the model\, resulting in an inductive bias that is highly architecture dependent. To investigate the properties of learned functions\, it is natural to consider a function space view given by the minimum ell_2 norm of weights required to realize a given function with a given network. We call this the “induced regularizer” of the network. Building on a line of recent work\, we study the induced regularizer of linear convolutional neural networks with a focus on the role of kernel size and the number of channels. We introduce an SDP relaxation of the induced regularizer\, that we show is tight for networks with a single input channel. Using this SDP formulation\, we show that the induced regularizer is independent of the number of the output channels for single-input channel networks\, and for multi-input channel networks\, we show independence given sufficiently many output channels. Moreover\, we show that as the kernel size increases\, the induced regularizer interpolates between a basis-invariant norm and a basis-dependent norm that promotes sparse structures in Fourier space.\nBased on joint work with Meena Jagadeesan and Ilya Razenshteyn.\n\n–\n\nFor more on our speaker\, please visit: http://sgunasekar.github.io/
URL:https://stat.mit.edu/calendar/gunasekar/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210402T110000
DTEND;TZID=America/New_York:20210402T120000
DTSTAMP:20220809T064551
CREATED:20210326T164528Z
LAST-MODIFIED:20210326T164528Z
UID:14434-1617361200-1617364800@idss.mit.edu
SUMMARY:Sampler for the Wasserstein barycenter
DESCRIPTION:Abstract: Wasserstein barycenters have become a central object in applied optimal transport as a tool to summarize complex objects that can be represented as distributions. Such objects include posterior distributions in Bayesian statistics\, functions in functional data analysis and images in graphics. In a nutshell a Wasserstein barycenter is a probability distribution that provides a compelling summary of a finite set of input distributions. While the question of computing Wasserstein barycenters has received significant attention\, this talk focuses on a new and important question: sampling from a barycenter given a natural query access to the input distribution. We describe a new methodology built on the theory of Gradient flows over Wasserstein space together with convergence guarantees.\nThis is joint work with Chiheb Daaloul\, Magali Tournus and Jacques Liandrat.\n\n–\n\nBio: Thibaut Le Gouic is an associate professor at the Ecole Centrale de Marseille and the Institut de Mathématiques de Marseille in France. Since 2019\, he is a visiting professor at the Department of Mathematics at MIT. He received his PhD at the Université Paul Sabatier at Toulouse\, France. His work lies at the interaction between geometry\, statistics and probability theory\, in particular via optimal transport.
URL:https://stat.mit.edu/calendar/sampler-for-the-wasserstein-barycenter/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series,Online Events
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210326T110000
DTEND;TZID=America/New_York:20210326T120000
DTSTAMP:20220809T064551
CREATED:20210113T154348Z
LAST-MODIFIED:20210317T152007Z
UID:13876-1616756400-1616760000@idss.mit.edu
SUMMARY:Testing the I.I.D. assumption online
DESCRIPTION:Abstract: Mainstream machine learning\, despite its recent successes\, has a serious drawback: while its state-of-the-art algorithms often produce excellent predictions\, they do not provide measures of their accuracy and reliability that would be both practically useful and provably valid. Conformal prediction adapts rank tests\, popular in nonparametric statistics\, to testing the IID assumption (the observations being independent and identically distributed). This gives us practical measures\, provably valid under the IID assumption\, of the accuracy and reliability of predictions produced by traditional and recent machine-learning algorithms. An interesting application of conformal prediction is the existence of _exchangeability martingales_\, i.e.\, random processes that are martingales under any exchangeable probability measure. In particular\, they are martingales whenever the observations are IID. In this talk I will discuss the construction of exchangeability martingales and their use for different kinds of change detection\, including detecting a point at which the IID assumption becomes violated and detecting concept shift. This may be useful for deciding when a prediction algorithm should be retrained. \nBio: Vladimir Vovk is Professor of Computer Science at Royal Holloway\, University of London. His research interests include machine learning and the foundations of probability and statistics. He was one of the founders of prediction with expert advice\, an area of machine learning avoiding making any statistical assumptions about the data. In 2001 he and Glenn Shafer published a book (“Probability and Finance: It’s Only a Game”\, New York: Wiley) on new game-theoretic foundations of probability; the sequel (“Game-theoretic Foundations for Probability and Finance”\, Hoboken\, NJ: Wiley) appeared in 2019. His second book (“Algorithmic Learning in a Random World”\, New York: Springer\, 2005)\, co-authored with Alex Gammerman and Glenn Shafer\, is the first monograph on conformal prediction\, method of machine learning that provides provably valid measures of confidence for their predictions. His current research centres on the theory and applications of conformal prediction and applications of game-theoretic probability to statistics\, machine learning\, and finance.
URL:https://stat.mit.edu/calendar/vovk/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210319T110000
DTEND;TZID=America/New_York:20210319T120000
DTSTAMP:20220809T064551
CREATED:20210113T153812Z
LAST-MODIFIED:20210305T192730Z
UID:13874-1616151600-1616155200@idss.mit.edu
SUMMARY:Relaxing the I.I.D. Assumption: Adaptively Minimax Optimal Regret via Root-Entropic Regularization
DESCRIPTION:Abstract: We consider sequential prediction with expert advice when data are generated from distributions varying arbitrarily within an unknown constraint set. We quantify relaxations of the classical i.i.d. assumption in terms of these constraint sets\, with i.i.d. sequences at one extreme and adversarial mechanisms at the other. The Hedge algorithm\, long known to be minimax optimal in the adversarial regime\, was recently shown to be minimax optimal for i.i.d. data. We show that Hedge with deterministic learning rates is suboptimal between these extremes\, and present a new algorithm that adaptively achieves the minimax optimal rate of regret with respect to our relaxations of the i.i.d. assumption\, and does so without knowledge of the underlying constraint set. We analyze our algorithm using the follow-the-regularized-leader framework\, and prove it corresponds to Hedge with an adaptive learning rate that implicitly scales as the square root of the entropy of the current predictive distribution\, rather than the entropy of the initial predictive distribution. \nBio: Daniel Roy is an Associate Professor in the Department of Statistical Sciences at the University of Toronto\, with cross appointments in Computer Science and Electrical and Computer Engineering. He is also a CIFAR Canada AI Chair and founding member of the Vector Institute. Prior to joining Toronto\, he was a Research Fellow of Emmanuel College and Newton International Fellow of the Royal Academy of Engineering\, hosted by the University of Cambridge. Roy completed his doctorate in Computer Science at the Massachusetts Institute of Technology\, where his dissertation was awarded an MIT EECS Sprowls Award. \nBio: Blair Bilodeau is a third-year PhD candidate in the Department of Statistical Sciences at the University of Toronto\, supported by an NSERC Doctoral Canada Graduate Scholarship and the Vector Institute. His research focuses on combining techniques from statistics and computer science to obtain theoretical performance guarantees for decision making. His work emphasizes guarantees that are sensitive to model structure\, data assumptions\, and model uncertainty.\n \n
URL:https://stat.mit.edu/calendar/roy/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210312T110000
DTEND;TZID=America/New_York:20210312T120000
DTSTAMP:20220809T064551
CREATED:20210113T153607Z
LAST-MODIFIED:20210308T212813Z
UID:13872-1615546800-1615550400@idss.mit.edu
SUMMARY:On nearly assumption-free tests of nominal confidence interval coverage for causal parameters estimated by machine learning
DESCRIPTION:Abstract: For many causal effect parameters of interest\, doubly robust machine learning (DRML) estimators ψ̂ 1 are the state-of-the-art\, incorporating the good prediction performance of machine learning; the decreased bias of doubly robust estimators; and the analytic tractability and bias reduction of sample splitting with cross fitting. Nonetheless\, even in the absence of confounding by unmeasured factors\, the nominal (1−α) Wald confidence interval ψ̂ 1±zα/2ˆ[ψ̂ 1] may still undercover even in large samples\, because the bias of ψ̂ 1 may be of the same or even larger order than its standard error of order n−1/2. \nIn this paper\, we introduce essentially assumption-free tests that (i) can falsify the null hypothesis that the bias of ψ̂ 1 is of smaller order than its standard error\, (ii) can provide an upper confidence bound on the true coverage of the Wald interval\, and (iii) are valid under the null under no smoothness/sparsity assumptions on the nuisance parameters. The tests\, which we refer to as \underline{A}ssumption \underline{F}ree \underline{E}mpirical \underline{C}overage \underline{T}ests (AFECTs)\, are based on a U-statistic that estimates part of the bias of ψ̂ 1. \nOur claims need to be tempered in several important ways. First no test\, including ours\, of the null hypothesis that the ratio of the bias to its standard error is smaller than some threshold δ can be consistent [with- out additional assumptions (e.g. smoothness or sparsity) that may be in- correct]. Second the above claims only apply to certain parameters in a particular class. For most of the others\, our results are unavoidably less sharp. \nWork with Lin Liu and Rajarshi Mukherjee \n– \nBio: \nThe principal focus of Dr. Robins’ research has been the development of analytic methods appropriate for drawing causal inferences from complex observational and randomized studies with time-varying exposures or treatments. The new methods are to a large extent based on the estimation of the parameters of a new class of causal models – the structural nested models – using a new class of estimators – the G estimators. The usual approach to the estimation of the effect of a time-varying treatment or exposure on time to disease is to model the hazard incidence of failure at time t as a function of past treatment history using a time-dependent Cox proportional hazards model. \nMore information available here.
URL:https://stat.mit.edu/calendar/robins/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/New_York:20210305T110000
DTEND;TZID=America/New_York:20210305T120000
DTSTAMP:20220809T064551
CREATED:20210113T153123Z
LAST-MODIFIED:20210217T145510Z
UID:13870-1614942000-1614945600@idss.mit.edu
SUMMARY:Detection Thresholds for Distribution-Free Non-Parametric Tests: The Curious Case of Dimension 8
DESCRIPTION:Abstract: Two of the fundamental problems in non-parametric statistical inference are goodness-of-fit and two-sample testing. These two problems have been extensively studied and several multivariate tests have been proposed over the last thirty years\, many of which are based on geometric graphs. These include\, among several others\, the celebrated Friedman-Rafsky two-sample test based on the minimal spanning tree and the K-nearest neighbor graphs\, and the Bickel-Breiman spacings tests for goodness-of-fit. These tests are asymptotically distribution-free\, universally consistent\, and computationally efficient (both in sample size and in dimension)\, making them particularly attractive for modern statistical applications. \nIn this talk\, we will derive the detection thresholds and limiting local power of these tests\, thus providing a way to compare and justify the performance of these tests in various applications. Several interesting properties emerge\, such as a curious phase transition in dimension 8\, and a remarkable blessing of dimensionality in detecting scale changes. I will also discuss the emerging theory of multivariate ranks based on optimal transport and how they can be used to construct efficient distribution-free two-sample tests. \n– \nBio: Bhaswar B. Bhattacharya is an Assistant Professor in the Department of Statistics at the Wharton School\, University of Pennsylvania. He received his Ph.D. from the Department of Statistics at Stanford University in 2016. Prior to that\, he obtained his Bachelor and Master degrees in Statistics from the Indian Statistical Institute\, Kolkata in 2009 and 2011\, respectively. His research interests include non-parametric statistics\, combinatorial probability\, and discrete and computational geometry.
URL:https://stat.mit.edu/calendar/bhattacharya/
LOCATION:online
CATEGORIES:Stochastics and Statistics Seminar Series
END:VEVENT
END:VCALENDAR