Abstract

Generative Bayesian Computation (GBC) is a framework for posterior simulation that replaces traditional MCMC with deep quantile neural networks trained to transport a base distribution onto the posterior. Once trained, the network provides draws from the posterior in a single forward pass and this scales to high dimensions and likelihood-free inference. Our theory follows that for quantile inference in bayes inference.

A number of applied applications will be discussed including sequential inference, spatial and casual inference problems. I conclude with open problems on architecture selection.