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 2022-01-17 11:20
CAMRI Loss: Class-wise Additive Angular Margin Loss for Improving Recall of a Specific Class
Daiki Nishiyama (Univ. Tsukuba), Fukuchi Kazuto, Yohei Akimoto, Jun Sakuma (Univ. Tsukuba/RIKEN) IBISML2021-22
Abstract (in Japanese) (See Japanese page) 
(in English) In real-world applications of multiclass classification models, there is a need to increase the recall of classes where misjudgments lead to serious losses. In this paper, we propose a loss function that can improve the recall of an important class while maintaining accuracy compared to the case using Cross-Entropy Loss. Existing studies have shown good results by giving the loss based on the magnitude of the angle between the weights of the last all-connected layer and the feature vector. In the proposed method, in addition to this, we add a penalty only when we learn important class features for that angle. Experiments showed that the proposed method effectively reduced the variance of the angles between the features and weights of the important classes relative to the other classes. Experiments on a total of 9 classes from CIFAR-10, GTSRB, and AwA2 showed that the proposed method could achieve up to about 9% recall improvement on Cross-Entropy Loss without sacrificing accuracy.
Keyword (in Japanese) (See Japanese page) 
(in English) Machine Learning / CNN / Multiclass Classification / Loss Function / Real-World / ArcFace / Cost-Sensitive Learning /  
Reference Info. IEICE Tech. Rep., vol. 121, no. 321, IBISML2021-22, pp. 29-36, Jan. 2022.
Paper # IBISML2021-22 
Date of Issue 2022-01-10 (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 IBISML2021-22

Conference Information
Committee IBISML  
Conference Date 2022-01-17 - 2022-01-18 
Place (in Japanese) (See Japanese page) 
Place (in English) Online 
Topics (in Japanese) (See Japanese page) 
Topics (in English) Machine Learning, etc. 
Paper Information
Registration To IBISML 
Conference Code 2022-01-IBISML 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) CAMRI Loss: Class-wise Additive Angular Margin Loss for Improving Recall of a Specific Class 
Sub Title (in English)  
Keyword(1) Machine Learning  
Keyword(2) CNN  
Keyword(3) Multiclass Classification  
Keyword(4) Loss Function  
Keyword(5) Real-World  
Keyword(6) ArcFace  
Keyword(7) Cost-Sensitive Learning  
Keyword(8)  
1st Author's Name Daiki Nishiyama  
1st Author's Affiliation University of Tsukuba (Univ. Tsukuba)
2nd Author's Name Fukuchi Kazuto  
2nd Author's Affiliation University of Tsukuba/RIKEN AIP (Univ. Tsukuba/RIKEN)
3rd Author's Name Yohei Akimoto  
3rd Author's Affiliation University of Tsukuba/RIKEN AIP (Univ. Tsukuba/RIKEN)
4th Author's Name Jun Sakuma  
4th Author's Affiliation University of Tsukuba/RIKEN AIP (Univ. Tsukuba/RIKEN)
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 2022-01-17 11:20:00 
Presentation Time 20 minutes 
Registration for IBISML 
Paper # IBISML2021-22 
Volume (vol) vol.121 
Number (no) no.321 
Page pp.29-36 
#Pages
Date of Issue 2022-01-10 (IBISML) 


[Return to Top Page]

[Return to IEICE Web Page]


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