Date: 2023-11-30

Time: 14:00-15:00 (UK time)

Strand S5.20

Abstract

Markov chain Monte Carlo (MCMC) methods are the most widely used tools in Bayesian statistics for making inferences from complex posterior distributions. For challenging problems where the posterior is high-dimensional with well-separated modes, MCMC algorithms can get trapped exploring local regions of high probability. Parallel tempering (PT) tackles this problem by delegating the task of global exploration to a tractable reference distribution (e.g. prior) which communicates to the target (e.g. posterior) through a sequence of parallel MCMC algorithms targeting distributions of increasing complexity to the target.

The classical approach to designing PT algorithms relied on a reversibility assumption, making PT challenging to tune and even deteriorating performance when introducing too many parallel chains. This talk will introduce a new non-reversible paradigm for PT (NRPT) that dominates its reversible counterpart while avoiding the performance collapse endemic to reversible PT methods. We will then establish near-optimal tuning guidelines, an efficient black-box methodology scalable to GPUs. Finally, I will discuss the recent application of NRPT in various scientific domains, including cancer genomics, nuclear fusion, political science, and astronomy.

Speaker

Dr. Saifuddin Syed is a Florence Nightingale Bicentennial Fellow in computational statistics and machine learning at the University of Oxford’s Department of Statistics, where he is also a member of the Algorithms and Inference Working Group for the Next Generation Event Horizon Telescope. His research involves designing scalable and robust algorithms for Bayesian inference with scientific applications in mind.