Non-Factorised Variational Inference in Dynamical Systems

Abstract

We introduce and investigate several variations of non-factorising approximate posteriors for Gaussian process state-space models a Bayesian non-parametric model for describing time-series governed by underlying dynamics. We introduce the variations in an order of increasing complexity, paying close attention to what correlations or distributional propertieseach method sacrifices, and connect to existing literature.

Publication
Symposium on Advances in Approximate Bayesian Inference
Date