IEICE Technical Committee Submission System
Conference Paper's Information
Online Proceedings
[Sign in]
Tech. Rep. Archives
 Go Top Page Go Previous   [Japanese] / [English] 

Paper Abstract and Keywords
Presentation 2009-03-11 16:35
A Learning Algorithm of Helmholtz Machine with Mean Field Approximation
Yuki Aoki (Nara Inst. of Sci and Tech.), Shin-ichi Maeda, Shin Ishii (Kyoto Univ.) NC2008-113
Abstract (in Japanese) (See Japanese page) 
(in English) It is often required to extract a compact feature of an original high-dimensional datum. Such a compactfeature is useful to visualize the original data, or to avoid the over-fitting in the learning which usually requires parameters in proportion to the dimensionality of the data. Recently, Hinton and Salakhutdinov [8] proposed a new dimensionality reduction method, which shows a superior performance to the PCA in their demonstration. They used a hierarchical Boltzmann Machine with millions of parameters. It is well known that the learning of such a nonlinear function approximator is difficult because of local minima and overfitting; neverthless, it seems that they had avoided such possible problems by adopting a Contrastive Divergence (CD) learning in the pretraining. Although the CD learning provides a parameter update rule which realizes efficient learning, several potential disadvantages
remain: there are no theoretically justified choice for the step-size in updating the parameters because the cost function of CD learning is unknown.
In this report, we derive a learning algorithm of Helmholtz Machine
using a mean field approximation, which allows analytical evaluations of both of the cost function and its derivatives with respect to the parameters. Also, it is shown that the learning of the Helmholtz Machine is comparative to the CD learning of the Boltzmann machine under certain conditions.
Keyword (in Japanese) (See Japanese page) 
(in English) Restricted Boltzmann Machine, / Helmholtz Machine, / Contrastive Divergence / Feature Extraction / Mean Field Approximation / / /  
Reference Info. IEICE Tech. Rep., vol. 108, no. 480, NC2008-113, pp. 57-62, March 2009.
Paper # NC2008-113 
Date of Issue 2009-03-04 (NC) 
ISSN Print edition: ISSN 0913-5685    Online edition: ISSN 2432-6380
Copyright
and
reproduction
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034)
Download PDF NC2008-113

Conference Information
Committee NC MBE  
Conference Date 2009-03-11 - 2009-03-13 
Place (in Japanese) (See Japanese page) 
Place (in English) Tamagawa Univ. 
Topics (in Japanese) (See Japanese page) 
Topics (in English)  
Paper Information
Registration To NC 
Conference Code 2009-03-NC-MBE 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) A Learning Algorithm of Helmholtz Machine with Mean Field Approximation 
Sub Title (in English)  
Keyword(1) Restricted Boltzmann Machine,  
Keyword(2) Helmholtz Machine,  
Keyword(3) Contrastive Divergence  
Keyword(4) Feature Extraction  
Keyword(5) Mean Field Approximation  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Yuki Aoki  
1st Author's Affiliation Nara Institute of Science and Technology (Nara Inst. of Sci and Tech.)
2nd Author's Name Shin-ichi Maeda  
2nd Author's Affiliation Kyoto University (Kyoto Univ.)
3rd Author's Name Shin Ishii  
3rd Author's Affiliation Kyoto University (Kyoto Univ.)
4th Author's Name  
4th Author's Affiliation ()
5th Author's Name  
5th Author's Affiliation ()
6th Author's Name  
6th Author's Affiliation ()
7th Author's Name  
7th Author's Affiliation ()
8th Author's Name  
8th Author's Affiliation ()
9th Author's Name  
9th Author's Affiliation ()
10th Author's Name  
10th Author's Affiliation ()
11th Author's Name  
11th Author's Affiliation ()
12th Author's Name  
12th Author's Affiliation ()
13th Author's Name  
13th Author's Affiliation ()
14th Author's Name  
14th Author's Affiliation ()
15th Author's Name  
15th Author's Affiliation ()
16th Author's Name  
16th Author's Affiliation ()
17th Author's Name  
17th Author's Affiliation ()
18th Author's Name  
18th Author's Affiliation ()
19th Author's Name  
19th Author's Affiliation ()
20th Author's Name  
20th Author's Affiliation ()
Speaker Author-1 
Date Time 2009-03-11 16:35:00 
Presentation Time 25 minutes 
Registration for NC 
Paper # NC2008-113 
Volume (vol) vol.108 
Number (no) no.480 
Page pp.57-62 
#Pages
Date of Issue 2009-03-04 (NC) 


[Return to Top Page]

[Return to IEICE Web Page]


The Institute of Electronics, Information and Communication Engineers (IEICE), Japan