Expectation Propagation for Graphical Models

Yuan Qi

Abstract

Many real-world problems can be modeled by probabilistic graphical models that have observed and hidden nodes. Solving these problems amounts to inferring the states of the hidden nodes given observed data. As a special case, many sequential data can be modeled by dynamic graphic models with nonlinear and non-Gaussian observations. For these models, there are no analytic methods to do the exact inference. Monte Carlo methods have been used to numerically approximate the solution. But in general Monte Carlo methods are computationally expensive. For general structural data that contains complicated relational information, the corresponding graphical models will contain loops in their structures. The exact inference on loopy graphical models is often too slow. Therefore, approximate inference on loopy graphical models has become an important topic. In this talk, I will present efficient Bayesian approximate inference algorithms for nonlinear dynamic models and loopy graphical models based on the expectation propagation framework, apply these algorithms to wireless digital communications, and compare their performance with Monte Carlo methods and belief propagation.

Back to the Main Page