Nikola Surjanovic
Department of Statistics, University of British Columbia
Title: Scalable, Automated, and Reliable: The Next Era of Bayesian Inference
Date: Friday, February 28th, 2025
Time: 1:30PM (PDT)
Location: ASB 10900
Abstract: Computational Bayesian inference offers a flexible approach to answering important scientific questions regarding uncertainty. However, the Bayesian approach can reach its computational limits on modern probabilistic models, which may be high-dimensional, weakly identifiable, or incorporate large amounts of data. To tackle modern problems, future Bayesian methods must: take advantage of modern computational resources and parallelization capabilities (scalability); be friendly to use for non-expert practitioners (automation); and offer performance guarantees (reliability).
In this talk I present some of my work on the path to scalable, automated, and reliable Bayesian inference. In particular, I present new developments on parallel tempering Markov chain Monte Carlo (MCMC) methods, which have empirically shown to perform well on complex sampling problems and are inherently parallelizable. I first present a way to combine variational inference methods and parallel tempering MCMC in an automated way that comes with performance guarantees. Next, I present a new convergence theory of parallel tempering that establishes reliable bounds on MCMC burn-in bias for general (including non-log-concave) distributions. This theory sheds light on the empirical success of annealing-based methods and offers easy-to-use convergence diagnostics. Finally, I discuss how the methods in the talk are implemented in our open-source software, Pigeons, which allows for computation to be scaled on compute clusters with up to thousands of machines.