Presentation 2008-01-15
Analytical and Numerical Solutions of Bethe Approximation in Normal Distributions
Yu NISHIYAMA, Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) For the calculation of marginal distributions, which require the huge computational cost, the algorithm called belief propagation(BP) has shown the effectiveness. Loopy belief propagation(LBP), which is the BP applied to the distributions that have loops, is not guaranteed to converge in general and, if it converges, it computes approximate marginal probabilities. The LBP fixed-points after the convergence are also given by the extrema of the evaluation function called Bethe free energy. In this paper, we clarify the extrema of Bethe free energy in some particular cases of normal distributions, and after we give the convergence condition of LBP and the accuracy of LBP, we do the numerical experiments and compare the results with analytical solutions.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Bethe Approximation / Bethe Free Energy / Loopy Belief Propagation / Single Cycle
Paper # NC2007-87
Date of Issue

Conference Information
Committee NC
Conference Date 2008/1/8(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) Analytical and Numerical Solutions of Bethe Approximation in Normal Distributions
Sub Title (in English)
Keyword(1) Bethe Approximation
Keyword(2) Bethe Free Energy
Keyword(3) Loopy Belief Propagation
Keyword(4) Single Cycle
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 2008-01-15
Paper # NC2007-87
Volume (vol) vol.107
Number (no) 413
Page pp.pp.-
#Pages 6
Date of Issue