Gaussian Processes and Extensions

Zoubin Ghahramani

Abstract

Gaussian Processes (GPs) define a highly flexible prior on functions which can be used for nonparametric Bayesian kernel regression. In this talk I will give a brief tutorial on GPs and then describe some of our work on extensions of the basic GP framework. In particular, I'll discuss three extensions: (1) GP classification using the EM-EP algorithm, (2) warped Gaussian processes, and (3) infinite mixtures of GP experts. These allow us to learn the kernel, handle regression "pre-processing" in a principled manner, and vary the kernel and noise-level over the input space. [Joint work with Carl E Rasmussen, Hyun-Chul Kim, Ed Snelson]

Back to the Main Page