Approximations for Binary Gaussian Process Classification

摘要: We provide a comprehensive overview of many recent algorithms for approximate inference in Gaussian process models probabilistic binary classification. The relationships between several approaches are elucidated theoretically, and the properties different corroborated by experimental results. examine both 1) quality predictive distributions 2) suitability marginal likelihood approximations model selection (selecting hyperparameters) compare to gold standard based on MCMC. Interestingly, some methods produce good although their poor. Strong conclusions drawn about methods: Expectation Propagation algorithm is almost always method choice unless computational budget very tight. also extend existing various ways, unifying code implementing all approaches.