Presentation 2010-12-19
Comparison of DIC and WAIC in Neural Bayes Learning
Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Layered neural networks are singular statistical models, because the map taking parameters to probability distributions is not one-to-one and their Fisher information matrices are not positive definite. Recently, we proposed a widely applicable information criterion (WAIC) which enables us to estimate the average of the generalization error even if the true distribution is singular for and unrealizable by a learning machine. In this paper, we compare the widely applicable information criterion with the deviance information criterion (DIC), and show two results theoretically and experimentally. First, if the true distribution is regular for a learning machine, then WAIC and DIC are asymptotically equivalent to each other. Second, if the true distribution is singular for a learning machine, then the average of DIC is not equal to that of the generalization error, whereas the average of WAIC is equal to that of the generalization error.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Algebraic Geometry / Learning Theory / Singular Learning Machine
Paper # MBE2010-69,NC2010-80
Date of Issue

Conference Information
Committee NC
Conference Date 2010/12/12(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) Comparison of DIC and WAIC in Neural Bayes Learning
Sub Title (in English)
Keyword(1) Algebraic Geometry
Keyword(2) Learning Theory
Keyword(3) Singular Learning Machine
1st Author's Name Sumio WATANABE
1st Author's Affiliation Tokyo Institute of Technology Precision and Intellignce Laboratory()
Date 2010-12-19
Paper # MBE2010-69,NC2010-80
Volume (vol) vol.110
Number (no) 355
Page pp.pp.-
#Pages 6
Date of Issue