Multimodal MCMC with TF Probability and Replica Exchange
Abstract: Popular Markov Chain Monte Carlo (MCMC) algorithms such as Hamiltonian Monte Carlo (HMC) have become the main computational workhorse for Bayesian inference, where intractable distributions, meaning distributions with unknown normalization constant, are ubiquitous. Yet, they often have difficulties with multimodal distributions, which get only more pronounced with increasing number of parameters. In this talk, I will give an introduction to Replica Exchange (aka Parallel Tempering), a MCMC method popular in biomolecular simulation and computational physics. It improves sampling of multimodal distributions by exchanging samples between increasingly "flatter" versions of the target distribution. While Replica Exchange is easy enough to implement from scratch, I will show a few examples using TensorFlow Probability's Replica Exchange implementation, but familiarity with TensorFlow Probability is not required to follow this presentation.
Presenter: Simeon Carstens
Bio: Originally a physicist, Simeon did a PhD and postdoc research in computational biology, focusing on Bayesian determination of three-dimensional chromosome structures. Since May 2019, he is a Data Scientist at Tweag I/O, a Paris-based software innovation lab and consultancy.