Presentation 2001/2/2
A Study on Theoretical Modeling of Ensemble Learning
Masato UCHIDA, Hiroyuki SHIOYA, Tsutomu DA-TE,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) The ensemble learning using the expectation of total outputs from some trained machines is proposed to improve the predictive performance. In this paper, we analyze the algorithmic structure of this learning method using the negative logarithm likelihood function. As a result, we can realize the ensemble learning is a simple three-step-minimizing procedure for Kullback divergence. Taking this into account, the asymptotic statistical property is given to elucidate the effect of the ensemble learning. Our framework extends the theoretical range of the ensemble learning, that is, the ensemble learning using α-divergence can be defined in a similar way. In addition, we mention that the learning with non-Bayesian additive term is related to this framework.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) ensemble learning / expectation of predictive error / α-divergence
Paper # NC2000-95
Date of Issue

Conference Information
Committee NC
Conference Date 2001/2/2(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 Study on Theoretical Modeling of Ensemble Learning
Sub Title (in English)
Keyword(1) ensemble learning
Keyword(2) expectation of predictive error
Keyword(3) α-divergence
1st Author's Name Masato UCHIDA
1st Author's Affiliation Graduate School of Engineering, Hokkaido University()
2nd Author's Name Hiroyuki SHIOYA
2nd Author's Affiliation Graduate School of Engineering, Hokkaido University
3rd Author's Name Tsutomu DA-TE
3rd Author's Affiliation Graduate School of Engineering, Hokkaido University
Date 2001/2/2
Paper # NC2000-95
Volume (vol) vol.100
Number (no) 618
Page pp.pp.-
#Pages 7
Date of Issue