Presentation 2014-10-18
Analysis of Learning Characteristics of RBM and Automatic Method for Deciding the Number of Hidden Neurons
Masahiko OSAWA, Masafumi HAGIWARA,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In this paper, we analyze the learning characteristics of Restricted Boltzmann Machine (RBM) by computer simulation. Then, using knowledge from the results of analyses, we propose an automatic method for deciding the number of hidden neurons. The proposed method utilizes two findings obtained from the computer simulation. First one is that, reduction of cross entropy is almost linear with the number of hidden neurons under certain conditions. Another one is that the minimum number of hidden neurons exists for reductions of the cross entropy if we use the dataset with large variance. The proposed method utilizing these findings can estimate the proper number of hidden neurons. We confirmed the effectiveness of the proposed method through computer experiments.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Restricted Boltzmann Machine / Deep Learning
Paper # NC2014-22
Date of Issue

Conference Information
Committee NC
Conference Date 2014/10/11(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 Learning Characteristics of RBM and Automatic Method for Deciding the Number of Hidden Neurons
Sub Title (in English)
Keyword(1) Restricted Boltzmann Machine
Keyword(2) Deep Learning
1st Author's Name Masahiko OSAWA
1st Author's Affiliation Faculty of Science and Technology, Keio University()
2nd Author's Name Masafumi HAGIWARA
2nd Author's Affiliation Faculty of Science and Technology, Keio University
Date 2014-10-18
Paper # NC2014-22
Volume (vol) vol.114
Number (no) 259
Page pp.pp.-
#Pages 6
Date of Issue