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 2017-11-10 13:00
Statistical Mechanical Analysis of Learning with Two-Layer Perceptron with Multiple Output Units -- Reconsidering Plateau Phenomenon --
Yuki Yoshida (UTokyo), Ryo Karakida (AIST), Masato Okada (UTokyo/AIST/RIKEN BSI), Shun-ichi Amari (RIKEN BSI) IBISML2017-83
Abstract (in Japanese) (See Japanese page) 
(in English) The plateau phenomenon --- stopping decrease of error in the middle of learning --- is problematic.
Since Amari et al. pointed out that singular regions stemming from structure of neural network cause such plateaus,
dynamics of learning with two-layered networks has been studied intensively. However, all these researches deal
with networks which have one-dimensional output, thus two-or-more dimensional output cases are
overlooked.according to Amari et al., such networks might be unlikely to be trapped in a plateau.
In this article, we analyze the dynamics of learning with two-layer perceptron with
multidimensional output under a statistical mechanical formalization.
We derive order parameters which capture macroscopic characteristics of connection weights and
differential equations which they follow.
We show that singular-region-driven plateaus do diminish or vanish
at least in a simple setting with multidimensional output.
We found that more the model being non-degenerative (i.e. far from one-dimensional output),
the more plateaus are alleviated.
Moreover, we showed theoretically that singular-region-driven plateaus are not shown at all
in the case with orthogonalized initializations.
Keyword (in Japanese) (See Japanese page) 
(in English) Neural network / Perceptron / Plateau / Statistical Mechanics / / / /  
Reference Info. IEICE Tech. Rep., vol. 117, no. 293, IBISML2017-83, pp. 347-354, Nov. 2017.
Paper # IBISML2017-83 
Date of Issue 2017-11-02 (IBISML) 
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 IBISML2017-83

Conference Information
Committee IBISML  
Conference Date 2017-11-08 - 2017-11-10 
Place (in Japanese) (See Japanese page) 
Place (in English) Univ. of Tokyo 
Topics (in Japanese) (See Japanese page) 
Topics (in English) Information-Based Induction Science Workshop (IBIS2017) 
Paper Information
Registration To IBISML 
Conference Code 2017-11-IBISML 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Statistical Mechanical Analysis of Learning with Two-Layer Perceptron with Multiple Output Units 
Sub Title (in English) Reconsidering Plateau Phenomenon 
Keyword(1) Neural network  
Keyword(2) Perceptron  
Keyword(3) Plateau  
Keyword(4) Statistical Mechanics  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Yuki Yoshida  
1st Author's Affiliation The University of Tokyo (UTokyo)
2nd Author's Name Ryo Karakida  
2nd Author's Affiliation National Institute of Advanced Industrial Science and Technology (AIST)
3rd Author's Name Masato Okada  
3rd Author's Affiliation The University of Tokyo/National Institute of Advanced Industrial Science and Technology/RIKEN Brain Science Institute (UTokyo/AIST/RIKEN BSI)
4th Author's Name Shun-ichi Amari  
4th Author's Affiliation RIKEN Brain Science Institute (RIKEN BSI)
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 2017-11-10 13:00:00 
Presentation Time 150 minutes 
Registration for IBISML 
Paper # IBISML2017-83 
Volume (vol) vol.117 
Number (no) no.293 
Page pp.347-354 
#Pages
Date of Issue 2017-11-02 (IBISML) 


[Return to Top Page]

[Return to IEICE Web Page]


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