Optimal Thinning of MCMC Output

Chris Oates (U Newcastle upon Tyne)

Apr 04. 2022, 13:30 — 14:20

Computation can pose a major challenge in Bayesian inverse problems when partial differential equations (PDEs) are involved.  In particular, the effectiveness of Markov chain Monte Carlo (MCMC) can be quite limited.  This talk presents a solution called Stein Thinning, that (1) automatically identifies when MCMC has or has not converged, (2) performs bias-removal for biased MCMC, and (3) provides compression of MCMC output.  The challenge of scaling Stein Thinning to high-dimensional parameter spaces will be highlighted.

Further Information
ESI Boltzmann Lecture Hall
Associated Event:
Adaptivity, High Dimensionality and Randomness (Workshop)
Carsten Carstensen (HU Berlin)
Albert Cohen (Sorbonne U, Paris)
Michael Feischl (TU Vienna)
Christoph Schwab (ETH Zurich)