Neural Computation, vol. 16, no. 9, pp. 1779–1810

著者:

  • Shiro Ikeda
  • Toshiyuki Tanaka
  • Shun-ichi Amari

URL:

Abstract:

Belief propagation is a universal method of stochastic reasoning. It gives a good approximate solution, when it is applied to a stochastic model with loopy interactions. AI, statistical physical, and information geometrical methods have so far been used to analyze its performance separately. The present paper gives a unified framework to understand the relation underlying these concepts. In particular, the free energy and its relation to BP and CCCP is elucidated from the point of view of information geometry. We then propose a family of new algorithms. The stability of the algorithms is also analyzed, and methods of accelerating these algorithms are proposed.