May 2020
Naive Feature Selection: Sparsity in Naive Bayes
Alexandre d'Aspremont (ENS, CNRS)
online
Abstract: Due to its linear complexity, naive Bayes classification remains an attractive supervised learning method, especially in very large-scale settings. We propose a sparse version of naive Bayes, which can be used for feature selection. This leads to a combinatorial maximum-likelihood problem, for which we provide an exact solution in the case of binary data, or a bound in the multinomial case. We prove that our bound becomes tight as the marginal contribution of additional features decreases. Both binary and…
Find out more »April 2020
How to Trap a Gradient Flow
Sébastien Bubeck (Microsoft Research)
online
Abstract: In 1993, Stephen A. Vavasis proved that in any finite dimension, there exists a faster method than gradient descent to find stationary points of smooth non-convex functions. In dimension 2 he proved that 1/eps gradient queries are enough, and that 1/sqrt(eps) queries are necessary. We close this gap by providing an algorithm based on a new local-to-global phenomenon for smooth non-convex functions. Some higher dimensional results will also be discussed. I will also present an extension of the 1/sqrt(eps)…
Find out more »On Using Graph Distances to Estimate Euclidean and Related Distances
Ery Arias-Castro (University of California, San Diego)
online
Abstract: Graph distances have proven quite useful in machine learning/statistics, particularly in the estimation of Euclidean or geodesic distances. The talk will include a partial review of the literature, and then present more recent developments on the estimation of curvature-constrained distances on a surface, as well as on the estimation of Euclidean distances based on an unweighted and noisy neighborhood graph. – About the Speaker: Ery Arias-Castro received his Ph.D. in Statistics from Stanford University in 2004. He then took…
Find out more »[POSTPONED] The Blessings of Multiple Causes
David Blei (Columbia University)
E18-304
*Please note: this event has been POSTPONED until Fall 2020* See MIT’s COVID-19 policies for more details. Title: The Blessings of Multiple Causes Abstract: Causal inference from observational data is a vital problem, but it comes with strong assumptions. Most methods require that we observe all confounders, variables that affect both the causal variables and the outcome variables. But whether we have observed all confounders is a famously untestable assumption. We describe the deconfounder, a way to do causal…
Find out more »Matrix Concentration for Products
Jonathan Niles-Weed (NYU)
online
Abstract: We develop nonasymptotic concentration bounds for products of independent random matrices. Such products arise in the study of stochastic algorithms, linear dynamical systems, and random walks on groups. Our bounds exactly match those available for scalar random variables and continue the program, initiated by Ahlswede-Winter and Tropp, of extending familiar concentration bounds to the noncommutative setting. Our proof technique relies on geometric properties of the Schatten trace class. Joint work with D. Huang, J. A. Tropp, and R. Ward.…
Find out more »[POSTPONED] Guido Imbens – The Applied Econometrics Professor and Professor of Economics, Graduate School of Business, Stanford University
E18-304
IDSS will host Prof. Guido Imbens as part of the Distinguished Speaker Seminar series. Prof. Guido Imbens’ primary field of interest is Econometrics. Research topics in which he is interested include: causality, program evaluation, identification, Bayesian methods, semi-parametric methods, instrumental variables.
Find out more »Uncovering atomistic mechanisms of crystallization using Machine Learning
Rodrigo Freitas (Stanford University)
online
Abstract: Solid-liquid interfaces have notoriously haphazard atomic environments. While essentially amorphous, the liquid has short-range order and heterogeneous dynamics. The crystal, albeit ordered, contains a plethora of defects ranging from adatoms to dislocation-created spiral steps. All these elements are of paramount importance in the crystal growth process, which makes the crystallization kinetics challenging to describe concisely in a single framework. In this seminar I will introduce a novel data-driven approach to systematically detect, encode, and classify all atomic-scale mechanisms…
Find out more »March 2020
Physics Guided Neural Networks for the Design and Understanding of Materials
Tian Xie (MIT)
online
Abstract: Climate change demands faster material innovations in multiple domains to reduce the carbon emissions of various industrial processes, but it currently takes 10-20 years to develop a single material with conventional human-driven approaches due to the large amounts of trail and errors needed. Machine learning approaches enable the direct learning of structure-property relations of materials from data, which have the potential to accelerate the search of materials and understand new scientific knowledge. However, one central challenge in developing…
Find out more »[POSTPONED] Large-Scale Cyber-Physical Systems: From Control Theory to Deep Learning
Navid Azizan (California Institute of Technology)
online
Abstract: The expansion of large-scale cyber-physical systems such as electrical grids, transportation networks, IoT, and other societal networks has created enormous challenges for controlling them and, at the same time, tremendous opportunities for utilizing the massive amounts of data generated by them. At the core of these data-driven control problems are distributed and stochastic optimization algorithms, such as the widely used stochastic gradient descent (SGD). While easy to use, these algorithms are not so easy to analyze, and in many…
Find out more »Safe Deep Learning in the Loop: Challenges, Methods, and Future Directions
Mahyar Fazlyab (University of Pennsylvania)
E18-304
Abstract: Despite high-profile advances in various decision-making and classification tasks, Deep Neural Networks (DNNs) have found limited application in safety-critical domains such as self-driving cars and automated healthcare. In particular, DNNs can be vulnerable to adversarial attacks and input uncertainties. This issue becomes even more complicated when DNNs are used in closed-loop systems, where a small perturbation (caused by noisy measurements, uncertain initial conditions, disturbances, etc.) can substantially impact the system being controlled. Therefore, it is of utmost importance…
Find out more »