Challenges in Markov chain Monte Carlo for Bayesian neural networks

Abstract

Markov chain Monte Carlo (MCMC) methods have not been broadly adopted in Bayesian neural networks (BNNs). This talk initially reviews the main challenges in sampling from the parameter posterior of a neural network via MCMC. Such challenges culminate to lack of convergence to the parameter posterior. Nevertheless, this talk shows that a non-converged Markov chain, generated via MCMC sampling from the parameter space of a neural network, can yield via Bayesian marginalization a valuable predictive posterior of the output of the neural network. Classification examples based on multilayer perceptrons showcase highly accurate predictive posteriors. The postulate of limited scope for MCMC developments in BNNs is partially valid; an asymptotically exact parameter posterior seems less plausible, yet an accurate predictive posterior is a tenable research avenue. This is joint work with Jacob Hinkle, M. Todd Young and David Womble.

Date
Apr 23, 2021 3:00 PM — 4:00 PM
Location
University of Glasgow, School of Mathematics and Statistics
University Place, Glasgow, G12 8QQ
Theodore Papamarkou
Theodore Papamarkou
Reader in maths of data science

My research interests include approximate inference and complexity theory.