Expectation Propagation in Factor Graphs: A Tutorial
Charles A. Sutton · 2005
Expectation propagation is an important variational inference algorithm for graphical models, especially if some of the variables are continuous. This tutorial presents two views EP: as repeatedly projecting into an approximating family, and as a message-passing algorithm. We present EP in terms of factor graphs, which simplifies some of the presentation and provides concreteness, while remaining completely general. We give special emphasis to explaining why belief propagation is a special case of EP, and how EP can be used to approximate a factor graph’s partition function. 1