Presentation 2001/3/16
On the training error and generalization error of a special Gaussian unit
Katsuyuki HAGIWARA, Kazuhiro KUNO, Shiro USUI,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In constructing a model selection criterion of layered neural nets and radial basis functions, it has been known the problem that true connection weights are not identifiable. Therefore, it has been making efforts to reveal the statistical properties of the training and generalization error for such situation and it is obtained an upper bound of the expected training error and a lower bound of the expected generalization error. In this article, by Combining the notion of 〓-covering and the extreme value theory, we derived an order of the expected training and generalization error for a special Gaussian unit, provided that the output data is a GaLlssian noise sequence.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Unidentifiability / a Gaussian unit / training error / generalization error
Paper # NC2000-159
Date of Issue

Conference Information
Committee NC
Conference Date 2001/3/16(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 training error and generalization error of a special Gaussian unit
Sub Title (in English)
Keyword(1) Unidentifiability
Keyword(2) a Gaussian unit
Keyword(3) training error
Keyword(4) generalization error
1st Author's Name Katsuyuki HAGIWARA
1st Author's Affiliation Faculty of Physics Engineering, Mie University()
2nd Author's Name Kazuhiro KUNO
2nd Author's Affiliation Faculty of Physics Engineering, Mie University
3rd Author's Name Shiro USUI
3rd Author's Affiliation Dept. of Information and Computer Sciences, Toyohashi University of Technology
Date 2001/3/16
Paper # NC2000-159
Volume (vol) vol.100
Number (no) 688
Page pp.pp.-
#Pages 8
Date of Issue