Professor Mark Girolami, University of Glasgow
Learning Kernels with Gaussian Process Priors
		
Title: Learning Kernels with Gaussian Process Priors 

Abstract:

The predictive accuracy of classification methods can be significantly
improved by integrating diverse data sources via classifier combination.
Non-parametric methods for heterogeneous data integration within a
classification context have been previously proposed utilising
Semi-Definite-Programming and non-probabilistic binary kernel-based Support
Vector Machines. However it is unclear how these methods may be extended to
the multi-class classification setting. In this contribution it is shown
that full Bayesian inference can be achieved for integrating multiple
datasets in the multiway classification setting employing Gaussian Process
priors, in addition efficient variational approximations are presented as
computationally economic alternatives to Metropolis-within-Gibbs sampling.
The proposed approach to integrating multiple data sets is demonstrated on a
handwritten digit recognition problem and is also applied to a large scale
protein fold prediction problem where we infer the weighted combinations of
data specific covariance functions and achieve state-of-the-art predictive
performance without resorting to any ad hoc parameter tuning and classifier
combination.