A0522
Title: Variational Bayesian procedures with frequentist guarantees
Authors: Dennis Nieman - VU Amsterdam (Netherlands) [presenting]
Abstract: Among the various reasons for the use of variational Bayesian (VB) methods are the possibility of model selection and the reduction of computation time. A theoretical approach to these motivations is taken, studying variational posteriors from a frequentist perspective. Using minimax theory, statistical and computational needs are balanced by finding the minimal dimension of the VB approximation to the posterior required to achieve the optimal convergence rate. The coverage of variational posterior credible regions is also studied. VB model selection entails the tuning of hyperparameters by optimization of the evidence lower bound, which can lead to smoothness-adaptive convergence rates.
