Title

MCMC Methods for Sparse Deep Learning

Abstract

This talk describes a very fast cyclical asynchronous Markov Chain Monte Carlo (MCMC) sampling framework for dealing with a class of high-dimensional multimodal posterior distributions that arises in sparse Bayesian modeling. We will present some theoretical results of the algorithm that show that on some classes of linear regression problems, the mixing time of the algorithm is linear in the number of regressors. The analysis also captures well some of the observed shortcoming of the approach. We use the sampler to gain some valuable insights on posterior distributions in deep learning models.

Bio

Yves Atchade is Professor of Statistics in the Math and Stats department with a joint appointment in the Faculty of Computing and Data Science at Boston University. His current research interests deal mainly with Markov Chain Monte Carlo methods, the statistical properties of Bayesian procedures, and the development of statistical methods for improved environmental prediction. He received a PhD from the Universite de Montreal in 2003. He is fellow of the Institute of Mathematical statistical, and associate editor of several leading statistical journals including the Harvard Data Science Review, the Bernoulli Journal, and the Journal of the American Statistical Association -- TM.