Presentation 2008-03-12
An Extension of CCCP Algorithm for Bethe Free Energy
Yu NISHIYAMA, Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Belief Propagation (BP) is an efficient algorithm for computing marginal probabilities of a high-dimensional probability distribution. The marginals computed by BP are equivalent to the extrema of Bethe free energy. Concave convex procedure (CCCP) has been studied for the optimization of Bethe free energy. In this paper, we extend the CCCP algorithm for Bethe free energy and present a new CCCP (NCCCP) algorithm. We practically apply NCCCP algorithm to multi-dimensional Gaussian distributions. As a result, NCCCP algorithm enables inner loop to converge even if all Lagrange multipliers in the loop are simultaneously updated. Moreover, we find that there exists an optimal point of the parameters introduced to NCCCP that can reduce the expensive computational cost.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Loopy Belief Propagation / Bethe Free Energy / CCCP / Gaussian Distribution
Paper # NC2007-126
Date of Issue

Conference Information
Committee NC
Conference Date 2008/3/5(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Neurocomputing (NC)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) An Extension of CCCP Algorithm for Bethe Free Energy
Sub Title (in English)
Keyword(1) Loopy Belief Propagation
Keyword(2) Bethe Free Energy
Keyword(3) CCCP
Keyword(4) Gaussian Distribution
1st Author's Name Yu NISHIYAMA
1st Author's Affiliation Department of Computational Intelligence and Systems Science, Tokyo Institute of Technology()
2nd Author's Name Sumio WATANABE
2nd Author's Affiliation Precision and Intelligence Laboratory, Tokyo Institute of Technology
Date 2008-03-12
Paper # NC2007-126
Volume (vol) vol.107
Number (no) 542
Page pp.pp.-
#Pages 6
Date of Issue