Presentation 2008-06-27
Application of Adaptive Natural Gradient Descent to Singular Learning Machines
Yoshiya NISHIMURA, Masato INOUE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Fisher information matrix degenerates at singular points in singular learning machines (SLM). Therefore, the adaptive natural gradient descent (ANGD) breaks down because the estimated inverse Fisher information matrix is diverged. In this research, we propose a general method avoiding the divergence of the Fisher information matrix by focusing on one aspect of this problem, i.e., the divergence is caused by the limit of the numerical precision. We have validated the proposed method using soft committee machines. Compared to the original ANGD, the proposed method successfully avoided the divergence without any significant degradation of both the training speed and the calculation cost.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) adaptive natural gradient descent / singular learning machine / soft committee machine
Paper # NC2008-25
Date of Issue

Conference Information
Committee NC
Conference Date 2008/6/19(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) Application of Adaptive Natural Gradient Descent to Singular Learning Machines
Sub Title (in English)
Keyword(1) adaptive natural gradient descent
Keyword(2) singular learning machine
Keyword(3) soft committee machine
1st Author's Name Yoshiya NISHIMURA
1st Author's Affiliation Dept. of Electrical Engineering and Bioscience, Graduate School of Advanced Science and Engineering, Waseda University()
2nd Author's Name Masato INOUE
2nd Author's Affiliation Dept. of Electrical Engineering and Bioscience, Graduate School of Advanced Science and Engineering, Waseda University
Date 2008-06-27
Paper # NC2008-25
Volume (vol) vol.108
Number (no) 101
Page pp.pp.-
#Pages 5
Date of Issue