Presentation 2009-07-14
Statistical Mechanics of Node-perturbation leraning
Kazuyuki HARA, Kentaro KATAHIRA, Kazuo OKANOYA, Masato OKADA,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Node-perturbation learning is a stochastic gradient method, and it can apply to the problem where the objective function is not defined. Werfel gave the theoretical analysis of node-perturbation as a desecrate time learning curve obtained by averaging the error for possible inputs. On the other hand, we derive deterministic equations of the order parameters, which depict behavior of node-perturbation learning by using statistical mechanics method, and treated the generalization error as a function of the order parameters. From our analysis, we found the generalization error becomes large when number of outputs of the network becomes large through the cross-talk noise of the other outputs. We also found that Werfel's learning curve is identical to the generalization error when the thermodynamic limit is assumed.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Node-perturbation learning / generalization error / statistical mechanics method / linear perceptron
Paper # NLP2009-38,NC2009-31
Date of Issue

Conference Information
Committee NC
Conference Date 2009/7/6(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) Statistical Mechanics of Node-perturbation leraning
Sub Title (in English)
Keyword(1) Node-perturbation learning
Keyword(2) generalization error
Keyword(3) statistical mechanics method
Keyword(4) linear perceptron
1st Author's Name Kazuyuki HARA
1st Author's Affiliation Tokyo Metropolitan College of Industrial Engineering()
2nd Author's Name Kentaro KATAHIRA
2nd Author's Affiliation The University of Tokyo:Riken:Japan Science Technology Agency, ERATO Okanoya Emotional Information Project
3rd Author's Name Kazuo OKANOYA
3rd Author's Affiliation Riken:Japan Science Technology Agency, ERATO Okanoya Emotional Information Project
4th Author's Name Masato OKADA
4th Author's Affiliation The University of Tokyo:Riken:Japan Science Technology Agency, ERATO Okanoya Emotional Information Project
Date 2009-07-14
Paper # NLP2009-38,NC2009-31
Volume (vol) vol.109
Number (no) 125
Page pp.pp.-
#Pages 6
Date of Issue