 |
Center for Nonlinear Studies |
In this talk I will discuss two distinct methods to improve generalized belief propagation (GBP). The first method provides new guidance as to what clusters of variables (or regions) to choose for executing GBP. I will focus on loop-structured regions and expose an elegant relation between GBP on loop-region graphs and the theory cycle bases. The second method I discuss is a new cluster-cumulant expansion at any fixed point of GBP, which provides an alternative to Cherkov & Chernyak's loop-series. I will show that the cluster-cumulant expansion is defined for arbitrary alphabets, generalizes to arbitrary region graphs, has fewer terms than the loop series (e.g. all disconnected clusters have a zero contribution) and empirically seems to be more accurate than the loop series. I will end with some future directions.
Host: Misha Chertkov, chertkov@lanl.gov, 665-8119