Regression and Classification Using Gaussian Process Priors

摘要: Gaussian processes are a natural way of specifying prior distributions over functions one or more input variables. When such function defines the mean response in regression model with errors, inference can be done using matrix computations, which feasible for datasets up to about thousand cases. The covariance process given hierarchical prior, allows discover high-level properties data, as inputs relevant predicting response. Inference these hyperparameters Markov chain sampling. Classification models defined underlying latent values, also sampled within chain. my view simplest and most obvious defining flexible Bayesian classification models, but despite some past usage, they appear have been rather neglected general-purpose technique. This may partly due confusion between being modeled best predictor this unknown function.