Information: Join today and make your research activities more affordable! Technical workshop participation fees and annual registration fees are available at member rates.
Notice: [Important] Announcement of Changes to Registration Fee Payment and Manuscript Upload Procedures for IEICE Technical Meetings
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 13:55
Additive or Concatenating Skip-connections Overcome the Degradation Problem of the Classic Feedforward Neural Network
Yasutaka Furusho, Kazushi Ikeda (NAIST) NC2019-17 IBISML2019-15
Abstract (in Japanese) (See Japanese page) 
(in English) The classic feedforward neural networks like the multilayer perceptron (MLP) degrades its empirical risk by training even though it stacks more layers. To overcome this problem, the ResNet which has additive skip-connections and the DenseNet which has concatenating skip-connections were proposed. These skip-connections enable an extreme deep neural network (DNN) to be trained with high performance. However, the reasons for these successes and when to prefer the one skip-connection to the other are unclear. A large ratio of the between-class distance to the within-class distance of feature vectors at the last hidden layer induces a high performance. Thus, we analyzed the change of these distances through hidden layers of the randomly initialized MLP, the ResNet, and the DenseNet. Our results show that the MLP strongly decreases the between-class distance compared with the within-class distance and that both skip-connections relax this decrease of the between-class angle and improve the ratio of the distances. In particular, the concatenating skip-connection is more preferable to the additive skip-connection if a DNN is extremely deep. Moreover, the additive skip-connection relax the exponential decrease of the angle into the sub-exponential decrease and the concatenating skip-connection relax this decrease into the reciprocal decrease. We also analyzed the effects of training on the distances and show that the preservation of the angle through layers at initialization encourages trained neural networks to increase the ratio of the distances. Therefore, both skip-connections induce high performance.
Keyword (in Japanese) (See Japanese page) 
(in English) Deep neural network / ResNet / DenseNet / Skip-connection / / / /  
Reference Info. IEICE Tech. Rep., vol. 119, no. 89, IBISML2019-15, pp. 97-102, June 2019.
Paper # IBISML2019-15 
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 NC2019-17 IBISML2019-15

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) Additive or Concatenating Skip-connections Overcome the Degradation Problem of the Classic Feedforward Neural Network 
Sub Title (in English)  
Keyword(1) Deep neural network  
Keyword(2) ResNet  
Keyword(3) DenseNet  
Keyword(4) Skip-connection  
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 ()
21st Author's Name  
21st Author's Affiliation ()
22nd Author's Name  
22nd Author's Affiliation ()
23rd Author's Name  
23rd Author's Affiliation ()
24th Author's Name  
24th Author's Affiliation ()
25th Author's Name  
25th Author's Affiliation ()
26th Author's Name / /
26th Author's Affiliation ()
()
27th Author's Name / /
27th Author's Affiliation ()
()
28th Author's Name / /
28th Author's Affiliation ()
()
29th Author's Name / /
29th Author's Affiliation ()
()
30th Author's Name / /
30th Author's Affiliation ()
()
31st Author's Name / /
31st Author's Affiliation ()
()
32nd Author's Name / /
32nd Author's Affiliation ()
()
33rd Author's Name / /
33rd Author's Affiliation ()
()
34th Author's Name / /
34th Author's Affiliation ()
()
35th Author's Name / /
35th Author's Affiliation ()
()
36th Author's Name / /
36th Author's Affiliation ()
()
Speaker Author-1 
Date Time 2019-06-18 13:55:00 
Presentation Time 25 minutes 
Registration for IBISML 
Paper # NC2019-17, IBISML2019-15 
Volume (vol) vol.119 
Number (no) no.88(NC), no.89(IBISML) 
Page pp.75-80(NC), pp.97-102(IBISML) 
#Pages 6 
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