Presentation 2007-05-21
Theoretical Analysis of Accuracy of Belief Propagation in Gaussian Models
Yu NISHIYAMA, Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Belief propagation (BP) is and algorithm which can compute marginal probability distributions with a tractable computational cost. Loopy belief propagation (LBP) applied to the graphs containing loops is known to provide marginal distributions approximately if LBP converges. In this paper, we apply LBP to a multi-dimensional Gaussian distribution that has loops and analytically show how accurate LBP is in some cases. Specifically, we analytically show messages, approximate marginal densities, and the KL distance at fixed points of LBP when the graph corresponding to a Gaussian distribution has at most a single loop. Basides, for the graphs which have arbitrary structures, we derive the expansions of approximate marginal densities when covariances are small.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Loopy Belief Propagation / Kullback Leibler Distance / Single Loop
Paper # NC2007-6
Date of Issue

Conference Information
Committee NC
Conference Date 2007/5/14(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 ENG
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Theoretical Analysis of Accuracy of Belief Propagation in Gaussian Models
Sub Title (in English)
Keyword(1) Loopy Belief Propagation
Keyword(2) Kullback Leibler Distance
Keyword(3) Single Loop
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 2007-05-21
Paper # NC2007-6
Volume (vol) vol.107
Number (no) 50
Page pp.pp.-
#Pages 6
Date of Issue