Learning Gaussian Process Kernels via Hierarchical Bayes

We present a novel method for learning with Gaussian process regression in a hierarchical Bayesian framework. In a first step, kernel matrices on a fixed set of input points are learned from data using a simple and efficient EM algorithm. This step is nonparametric, in that it does not require a parametric form of covariance function. In a second step, kernel functions are fitted to approximate the learned covariance matrix using a generalized Nystr

• Type: Conference
• Year: 2004
• Where: NIPS
• Authors: Anton Schwaighofer, Volker Tresp, Kai Yu
• Keywords: Gaussian Process Regression, Hierarchical Bayesian Framework, Hierarchical Bayesian Method, NIPS 2004, NIPS 2007