EcoSta 2021: Start Registration
View Submission - EcoSta2021
A0550
Title: Variational Bayes for regression with Gaussian process priors: A frequentist Bayesian analysis Authors:  Dennis Nieman - VU Amsterdam (Netherlands) [presenting]
Harry Zanten - University of Amsterdam (Netherlands)
Botond Szabo - Leiden University (Netherlands)
Abstract: A Bayesian nonparametric regression model with Gaussian process priors is considered. In practice, sampling from the exact posterior distribution is computationally expensive. We study an approximative procedure called the variational method, which reduces computation time. Of particular interest is a variational framework that has gained popularity in the machine learning literature in the last decade. We investigate frequentist properties of the Bayesian approach: the data are assumed to be generated from a distribution with a true functional parameter, and conditions are given under which the contraction rate does not deteriorate under the approximation. The developed theory is applied to several examples of Gaussian process priors.