Presentation 1999/7/19
A geometrical consideration on the gradient descent learning of multilayer perceptron
Katsuyuki HAGIWARA, Kazuhiro KUNO,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Although the gradient descent learning is the most basic and the simplest learning rule, it is not well-known the behaviour for multilayer perceptron as indicated by the existence of problems such as the ravine of error surface, premature saturation/flat spot problem and symmetric phase/plateau. In this article, first, we gave a geometrical interpretation of the gradient descent learning of multilayer perceptron whose activation function in the output layer is linear function. Based on the interpretation, we considered the case in which the training is not properly performed because of the unbalance between the adjustments of input weight and output weight. To overcome this problem, we gave a simple modification of the ordinal learning rule. The justification of our consideration was shown in the numerical experiments.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) multilayer perceptron / gradient descent learning / linear parameter / nonlinear parameter / linear least squares estimators
Paper # NC99-38
Date of Issue

Conference Information
Committee NC
Conference Date 1999/7/19(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) A geometrical consideration on the gradient descent learning of multilayer perceptron
Sub Title (in English)
Keyword(1) multilayer perceptron
Keyword(2) gradient descent learning
Keyword(3) linear parameter
Keyword(4) nonlinear parameter
Keyword(5) linear least squares estimators
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
Date 1999/7/19
Paper # NC99-38
Volume (vol) vol.99
Number (no) 193
Page pp.pp.-
#Pages 8
Date of Issue