Improving the Mean Field Approximation Via the Use of Mixture Distributions

Abstract

Mean field methods provide computationally efficient approximationsto posterior probability distributions for graphical models. Simplemean field methods make a completely factorized approximation tothe posterior, which is unlikely to be accurate when the posterioris multimodal. Indeed, if the posterior is multi-modal, only oneof the modes can be captured. To improve the mean field approximationin such cases, we employ mixture models as posterior approximations,where each mixture...