Abstract:

Graphical models are used in many scientific disciplines, including statistical physics, machine learning, and error-correcting coding. One typically seeks the marginal probability of selected variables (nodes of the graph) given observations at other variables. Belief propagation (BP) is a fast marginalization method, based on passing local messages. Designed for singly-connected graphs, BP nonetheless works well in many applications involving graphs with loops, for reasons that were not well understood. We characterize the BP solutions, showing that BP can only converge to a stationary point of an approximate free energy, known as the Bethe free energy in statistical physics. This understanding lets us for construct new message-passing algorithms based on improvements to Bethe's approximation introduced by Kikuchi and others. The new generalized belief propagation (GBP) algorithms are much more accurate than ordinary BP for some problems, and permit solutions to Kikuchi approximations for otherwise intractable inhomogeneous systems. We illustrate GBP with a spin-glass example and an error-correcting code, showing dramatically improved estimates of local magnetizations and decoding performance using GBP.