Neural Methods for Amortised Inference

Abstract

Simulation-based methods for statistical inference have evolved dramatically over the past 50 years, keeping pace with technological advancements. The field is undergoing a new revolution as it embraces the representational capacity of neural networks, optimisation libraries and graphics processing units for learning complex mappings between data and inferential targets. The resulting tools are amortised, in the sense that they allow rapid inference through fast feedforward operations, and they have several compelling advantages over classical methods such as Markov chain Monte Carlo or variational Bayes: they do not require knowledge of the likelihood function, are relatively easy to implement, and facilitate inference at a substantially reduced computational cost. In this talk I present the decision-theoretic foundation of neural inference methods, and detail how these methods can be used for point estimation, and approximate and full Bayesian inference. The presentation concludes with a showcase of environmental applications where we have used these methods for making rapid inference with spatial process models. This is joint work with Matthew Sainsbury-Dale and Raphaël Huser.