Shiro Ikeda, Institute of statistical mathematics

Title: Belief propagation from information geometrical viewpoint

Abstract: 
Belief propagation (BP) is a technique which is used in many fields,
such as error correction codes, statistical physics, and artificial
intelligence. It is an algorithm to calculate the marginal
distributions approximately when direct calculation of them is
intractable. In this talk, I would like to give short introduction of
BP, and after that, would like to introduce our results. We have
investigated the properties of BP and its variants from information
geometrical viewpoint. The viewpoint gives a new mathematical
framework for analysis. Based on the framework, we derive two
conditions which characterize the stationary point of BP and its
variants. We also show the relation of these two conditions and the
Bethe free energy, and propose a new cost function. Finally, with the
cost function, we propose a new variant of BP.

(This is a joint work with Toshiyuki Tanaka and Shun-ichi Amari)

WWW: http://www.ism.ac.jp/~shiro/