Presentation 2012-11-07
Free Energy and Generalization error of Local mimima in Variational Bayes Learning
Fumito NAKAMURA, Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Variational Bayes learning approximates the posterior distribution with small computational costs, however, it has several local minima which depend on initial values. The variational free energy can be calculated using only training samples, whereas the generalization error not. To compare several local minima, the relation between the variational free energy and the generalization error are necessary. In this paper, we experimentally calculate the free energy and the generalization error for each local minimum, and report the following results. If a true distribution is regular for a statistical model, then the local minimum that minimizes the variational free energy also makes the generalization error minimal. If a true distribution is singular for a statistical model, then the local minimum that minimizes the variational free energy does not always make the generalization error minimal.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Variational Bayes learning / Variational free energy / generalization error
Paper # IBISML2012-42
Date of Issue

Conference Information
Committee IBISML
Conference Date 2012/10/31(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 Information-Based Induction Sciences and Machine Learning (IBISML)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Free Energy and Generalization error of Local mimima in Variational Bayes Learning
Sub Title (in English)
Keyword(1) Variational Bayes learning
Keyword(2) Variational free energy
Keyword(3) generalization error
1st Author's Name Fumito NAKAMURA
1st Author's Affiliation Tokyo Institute of Technology, Dept. of Computational Intelligence and Systems Science()
2nd Author's Name Sumio WATANABE
2nd Author's Affiliation Tokyo Institute of Technology, Dept. of Computational Intelligence and Systems Science
Date 2012-11-07
Paper # IBISML2012-42
Volume (vol) vol.112
Number (no) 279
Page pp.pp.-
#Pages 7
Date of Issue