Machine Learning and Friends Lunch

home
past talks
resources
conferences

Variational Approximations for Inference over Graphical Models

Abstract

This is a tutorial talk on the use of variational approximations for inference calculations over graphical models. Variational methods connect "integral" and "differential" viewpoints of physical systems. They also provide a framework for developing approximations to the marginal and posterior probabilities that are needed for inference calculations over graphical models. I'll discuss a variational reworking of the Expectation Maximization (EM) algorithm, "mean field" approximations, and the equivalence between the Belief Propagation (BP) algorithm and the Bethe approximation for the 2-D Ising model.