Presentation 2009-01-19
On the Effect of Hyperparamater to Generalization Error in Variational Bayes Learning
Shinji OYAMA, Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In variational Bayes learning, the probability distribution of the hidden variable and parameter is made by the mean field approximation. In a mixture of exponential probability distributions such a normal mixture, the mean field approximation is applied to wide practical problems because it provides the fast recursive learning algorithm. It was theoretically clarified the phase transition structure of variational Bayes learning for hyperparameter. In this paper, we study the generalization errors of variational Bayes learning and experimentally show the following facts. (1)The behavior of the generalization error changes between the phase transition point. (2)At an ordinary point, the generalization error strongly depends on the condition that the true distribution is contained in the statistical model or not, whereas, at the critical point, the generalization error is not so depend on the condition.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Singular learning machines / variational Bayes generalization error / Gibbs generalization error / hyperparameter
Paper # NC2008-87
Date of Issue

Conference Information
Committee NC
Conference Date 2009/1/12(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) On the Effect of Hyperparamater to Generalization Error in Variational Bayes Learning
Sub Title (in English)
Keyword(1) Singular learning machines
Keyword(2) variational Bayes generalization error
Keyword(3) Gibbs generalization error
Keyword(4) hyperparameter
1st Author's Name Shinji OYAMA
1st Author's Affiliation Department of Computer Science, Tokyo Institute of Technology()
2nd Author's Name Sumio WATANABE
2nd Author's Affiliation Precision and Intelligence Laboratory, Tokyo Institute of Technology
Date 2009-01-19
Paper # NC2008-87
Volume (vol) vol.108
Number (no) 383
Page pp.pp.-
#Pages 6
Date of Issue