Presentation 1996/5/24
Generalization Error of a Linear Neural Network with a Singular Fisher Information Matrix
Kenji Fukumizu,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) We calculate the generalization error of a special linear neural network whose information matrix is singular at the true parameter. It is well known that the information matrix in certain non-nonlinear models can be singular. Many statistical methods are not applicable to such cases, because they essentially depend on the ordinary asymptotic theory which assumes the regularity of the information matrix. Thus, it is very important to clarify the generalization error in the singular cases. In this paper, we discuss a special model of linear neural networks, and show that the maximum likelihood estimator of this model is rigorously solvable. We calculate the generalization error in some special cases where the information matrix is singular at the true parameter, and show that the generalization error is larger than the error of a network with a regular information matrix.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Linear neural network / Maximum likelihood estimation / Generalization error / Information matrix.
Paper # NC96-3
Date of Issue

Conference Information
Committee NC
Conference Date 1996/5/24(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 ENG
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Generalization Error of a Linear Neural Network with a Singular Fisher Information Matrix
Sub Title (in English)
Keyword(1) Linear neural network
Keyword(2) Maximum likelihood estimation
Keyword(3) Generalization error
Keyword(4) Information matrix.
1st Author's Name Kenji Fukumizu
1st Author's Affiliation Information and Communication R & D Center, RlCOH Co., Ltd.()
Date 1996/5/24
Paper # NC96-3
Volume (vol) vol.96
Number (no) 76
Page pp.pp.-
#Pages 8
Date of Issue