Presentation 2005/6/16
Analysis of on-line learning when a teacher goes around the true teacher
Seiji MIYOSHI, Masato OKADA,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In the framework of on-line learning, the learning machine might move around the teacher due to the difference of structures or output functions between the teacher and the learning machine or the noises. We analyzed the generalization performance of a new student supervised by the moving machine. Calculating generalization errors of a model composed by the true teacher, the moving teacher and the student that are all linear perceptrons with noises analytically by using statistical mechanics, it is proved that the generalization error between the true teacher and the student can be smaller than that between the true teacher and the moving teacher, though the student uses examples from only the moving teacher.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) on-line learning / generalization error / moving teacher / true teacher / unlearnable case
Paper # NC2005-10
Date of Issue

Conference Information
Committee NC
Conference Date 2005/6/16(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) Analysis of on-line learning when a teacher goes around the true teacher
Sub Title (in English)
Keyword(1) on-line learning
Keyword(2) generalization error
Keyword(3) moving teacher
Keyword(4) true teacher
Keyword(5) unlearnable case
1st Author's Name Seiji MIYOSHI
1st Author's Affiliation Kobe City College of Technology()
2nd Author's Name Masato OKADA
2nd Author's Affiliation Division of Transdisciplinary Sciences, Graduate School of Frontier Sciences, The University of Tokyo:RIKEN Brain Science Institute:JST PRESTO
Date 2005/6/16
Paper # NC2005-10
Volume (vol) vol.105
Number (no) 130
Page pp.pp.-
#Pages 6
Date of Issue