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 2019-06-18 14:45
Theoretical Analysis of the Fixup Initialization for Fast Convergence and High Generalization Ability
Yasutaka Furusho, Kazushi Ikeda (NAIST) NC2019-19 IBISML2019-17
Abstract (in Japanese) (See Japanese page) 
(in English) The Fixup initialization is a new initialization method of ResNet for a fast convergence with a high learning rate of SGD and a high generalization ability. However, the reasons for its high performance are not clear.
Both the learning rate and the generalization ability are affected by the loss landscape at initialization, that is, the maximum eigenvalue of the Hessian matrix. Thus, we calculated the maximum eigenvalue of the ResNet and found that the maximum eigenvalue with the Fixup initialization has at most the square root order with respect to the depth of the ResNet while that with the He initialization has the exponential order. This small eigenvalue leads to a fast convergence with a high learning rate of SGD and a high generalization ability.
Keyword (in Japanese) (See Japanese page) 
(in English) Deep neural network / ResNet / Initialization / Fixup / / / /  
Reference Info. IEICE Tech. Rep., vol. 119, no. 89, IBISML2019-17, pp. 109-114, June 2019.
Paper # IBISML2019-17 
Date of Issue 2019-06-10 (NC, IBISML) 
ISSN 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 NC2019-19 IBISML2019-17

Conference Information
Committee NC IBISML IPSJ-MPS IPSJ-BIO  
Conference Date 2019-06-17 - 2019-06-19 
Place (in Japanese) (See Japanese page) 
Place (in English) Okinawa Institute of Science and Technology 
Topics (in Japanese) (See Japanese page) 
Topics (in English) Neurocomputing, Machine Learning Approach to Biodata Mining, and General 
Paper Information
Registration To IBISML 
Conference Code 2019-06-NC-IBISML-MPS-BIO 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Theoretical Analysis of the Fixup Initialization for Fast Convergence and High Generalization Ability 
Sub Title (in English)  
Keyword(1) Deep neural network  
Keyword(2) ResNet  
Keyword(3) Initialization  
Keyword(4) Fixup  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Yasutaka Furusho  
1st Author's Affiliation Nara Institute of Science and Technology (NAIST)
2nd Author's Name Kazushi Ikeda  
2nd Author's Affiliation Nara Institute of Science and Technology (NAIST)
3rd Author's Name  
3rd Author's Affiliation ()
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 2019-06-18 14:45:00 
Presentation Time 25 minutes 
Registration for IBISML 
Paper # NC2019-19, IBISML2019-17 
Volume (vol) vol.119 
Number (no) no.88(NC), no.89(IBISML) 
Page pp.87-92(NC), pp.109-114(IBISML) 
#Pages
Date of Issue 2019-06-10 (NC, IBISML) 


[Return to Top Page]

[Return to IEICE Web Page]


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