Inference in the face of intractability: Bayesian applications of continuous-time Markov processes
Forskningsprojekt, 2023 –

Thanks to Monte Carlo methods and modern computing power Bayesian inference is more accessible to practitioners than ever. The ability to sample distributions with intractable normalization constants is crucial in spatial statistics, molecular dynamics, statistical mechanics, and more. At the same time, our samplers are taken from a class of processes that are themselves interesting models; the Bayesian notion of uncertainty for hypotheses still respects the Law of Large Numbers. New sampling methods allow us to explore alternative models for more efficient inference, with one example being the advent of non-reversible Monte Carlo methods such as piecewise deterministic Markov processes (PDMPs). The purpose of this project is to develop new, accessible tools and theory for attacking difficult inference problems with and about continuous-time Markov processes. A particular area of interest is gradient estimation, transforming Bayesian inference into optimization problems, used in e.g. variational inference.

Deltagare

Ruben Seyer (kontakt)

Chalmers, Matematiska vetenskaper, Tillämpad matematik och statistik

Moritz Schauer

Chalmers, Matematiska vetenskaper, Tillämpad matematik och statistik

Aila Särkkä

Chalmers, Matematiska vetenskaper, Tillämpad matematik och statistik

Finansiering

Göteborgs universitet

Finansierar Chalmers deltagande under 2023–

Mer information

Senast uppdaterat

2024-06-21