Presentation 2002/3/11
Precise Prediction and Knowledge Discovery in Model Selection Problem of Non-identifiable Learning Machines
Koitiro NISHIUE, Sumio WATANABE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) The model selection problem needs criterion to select the most suitable model among several models. The criterion of consistency is on behalf of knowledge discovery and the criterion of effectiveness is on behalf of precise prediction. These criterions are not yet clarified in non-identifiable learning machines such as neural networks or gaussian mixtures. We study the Bayesian model selection methods based on the minimization of the stochastic complexity. We formulate a hypothesis that the positive prior is appropriate for consistency and Jeffreys'prior is appropriate for effectiveness. We examine the hypothesis for the model selection problem of three layer neural networks.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Bayes estimation / Model selection problem / Stochastic complexity / Jeffreys'Prior / Consistency / Effectiveness
Paper # NC2001-151
Date of Issue

Conference Information
Committee NC
Conference Date 2002/3/11(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) Precise Prediction and Knowledge Discovery in Model Selection Problem of Non-identifiable Learning Machines
Sub Title (in English)
Keyword(1) Bayes estimation
Keyword(2) Model selection problem
Keyword(3) Stochastic complexity
Keyword(4) Jeffreys'Prior
Keyword(5) Consistency
Keyword(6) Effectiveness
1st Author's Name Koitiro NISHIUE
1st Author's Affiliation Department of Advanced Applied Electronics Tokyo Institute of Technology()
2nd Author's Name Sumio WATANABE
2nd Author's Affiliation P&I Lab., Tokyo Institute of Technology
Date 2002/3/11
Paper # NC2001-151
Volume (vol) vol.101
Number (no) 735
Page pp.pp.-
#Pages 8
Date of Issue