Open Seminar: Scalable, robust, Bayesian inference for modern applications

Professor Chris Holmes -  Professor of Biostatistics at the University of Oxford and
Programme Director for Health at The Alan Turing Institute.

Professor Holmes will present a new approach for Bayesian updating that uses nonparametric learning (NPL) to train parametric models.

He will show that the approach is robust under model misspecification and highly scalable, allowing trivial parallelization to generate 100,000s of Monte Carlo samples from the posterior at the equivalent computational cost of a single sample.

In one sense he replaces MCMC with optimization of randomized (re-weighted) objective functions. Examples are given in logistic regression with sparsity priors, variational Bayes (VB) inference, and Bayesian randomized forests.