Skip to main content Skip to navigation

Carl Rasmussen: Data Analysis using Gaussian Processes

Gaussian processes (GPs) are well known non-parametric Bayesian models, but surprisingly they are not used extensively in practice. In this mainly methodological talk I'll show three very different types of GP models used commonly in the machine learning community: 1) inferring complex structures in regression via hierarchical learning of covariance functions, 2) non-linear dimensionality reduction using the GP Latent Variable Models (GP-LVM) and 3) GP classification using the Expectation Propagation algorithm. These examples highlight that the fundamental ability of GPs to express and manipulate (Bayesian) distributions over functions make them a powerful, practical foundation for numerous types of applications.