Scalable Hierarchical Gaussian Process Models for Regression and Pattern Classification

Gaussian processes, which are distributions over functions, are powerful nonparametric tools for the two major machine learning tasks: regression and classification. Both tasks are concerned with learning input-output mappings from example input-output pairs. In Gaussian process (GP) regression and classification, such mappings are modeled by Gaussian processes. In GP regression, the likelihood is Gaussian for continuous outputs, and hence closed-form solutions for prediction and model selection can be obtained. In GP classification, the likelihood is non-Gaussian for discrete/categorical outputs, and hence closed-form solutions are not available, and approximate inference methods must be resorted.