Presentation 2007-03-14
Progressive Contrastive Divergence Method
Kazuya TAKABATAKE, Shotaro AKAHO,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In cases that a gradient method is used for maximum likelihood estimation of a multivariate probability model, evaluation of the gradient vector often requires massive computation. Contrastive divergence method uses some approximation and a Monte Carlo method to reduce the computation for the evaluation of the gradient vector. Due to this approximation, the results from contrastive divergence method differs from the true maximum likelihood estimation. In this paper, progressive contrastive divergence method that is a modification of contrastive divergence method is shown. The result from progressive divergence method theoretically coincide with the true maximum likelihood estimation. Results of experiments that use Boltzmann machine for the probability model are shown.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) contrastive divergence method / Monte Carlo method / Markov chain Monte Carlo / Maximum likelihood estimation / convergence
Paper # NC2006-143
Date of Issue

Conference Information
Committee NC
Conference Date 2007/3/7(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) Progressive Contrastive Divergence Method
Sub Title (in English)
Keyword(1) contrastive divergence method
Keyword(2) Monte Carlo method
Keyword(3) Markov chain Monte Carlo
Keyword(4) Maximum likelihood estimation
Keyword(5) convergence
1st Author's Name Kazuya TAKABATAKE
1st Author's Affiliation Neuroscience Research Institute, AIST()
2nd Author's Name Shotaro AKAHO
2nd Author's Affiliation Neuroscience Research Institute, AIST
Date 2007-03-14
Paper # NC2006-143
Volume (vol) vol.106
Number (no) 588
Page pp.pp.-
#Pages 4
Date of Issue