Smooth Bayesian kernel machines

Abstract

In this paper, we consider the possibility of obtaining a kernel machine that is sparse in feature space and smooth in output space. Smooth in output space implies that the underlying function is supposed to have continuous derivatives up to some order. Smoothness is achieved by applying a roughness penalty, a concept from the area of functional data analysis. Sparseness is taken care of by automatic relevance determination. Both are combined in a Bayesian model, which has been implemented and tested. Test results are presented in the paper. © Springer-Verlag Berlin Heidelberg 2005.