Seminario Datalab
New methods in Markov chain Monte Carlo
Ponente: Max Hird (University of Waterloo and ICMAT)Fecha: jueves 08 de octubre de 2026 - 12:00Lugar: Aula Gris 2, ICMAT
Resumen:
With the advent of the widespread use of powerful personal computers in the 1990s, canonical Markov chain Monte Carlo (MCMC) sampling methods saw a rapid explosion in popularity. However, modern problems encountered by Machine Learning practitioners have necessitated extensions and reconceptions of these original canonical algorithms.
The work I will present is an assortment of ongoing projects concerning the theoretical analysis and creation of these extensions and reconceptions. These projects include: a high dimensional adaptive sampler with reduced computational cost, a way of boosting an arbitrary MCMC algorithm with a variational distribution, a review of linear and nonlinear based preconditioners for MCMC, a non-asymptotic analysis of adaptive MCMC samplers, and the theoretical analysis and practical implementation of a Riemannian MCMC sampler using a Stein kernel.
Hopefully the talk should elucidate what makes modern sampling problems difficult, and how these difficulties are overcome.
