ITE Technical Group Submission System
Conference Paper's Information
Online Proceedings
[Sign in]
 Go Top Page Go Previous   [Japanese] / [English] 

Paper Abstract and Keywords
Presentation 2022-10-14 10:00
Robust Semi-Supervised Learning for Noisy Labels Using Early-learning Regularization and Weighted Loss
Ryota Higashimoto, Soh Yoshida, Mitsuji Muneyasu (Kansai Univ.)
Abstract (in Japanese) (See Japanese page) 
(in English) Training Deep Neural Networks (DNNs) on datasets with incorrect labels (label noise) is an important challenge. In the presence of label noise, DNNs adapt to training samples with correct labels in the early training phase and adapt to samples with label noise in the late training phase. Recently, methods based on semi-supervised learning have shown promise in exploiting this property. In this conventional method, the sample loss distribution is modeled by a Gaussian mixture model, and the DNN is trained by assuming that samples with label noise are unlabeled. On the other hand, the accuracy of the division into labeled and unlabeled samples depends on the performance of the DNN during warm-up. In addition, the conventional method cannot consider the effect of the number of labeled samples, which varies with the percentage of label noise in the training data. In this paper, we introduce Early-learning Regularization (ELR), which suppresses the DNN's adaptation to label noise in the initial learning phase. Furthermore, we propose a learning method that introduces a weighted loss that can dynamically change the loss depending on the fraction of labeled samples. To verify the effectiveness of the proposed method, we conducted experiments using the CIFAR-100 image classification dataset with pseudo-label noise. The results show that the proposed method outperforms the existing methods on CIFAR-100 with 20textasciitilde90% label noise.
Keyword (in Japanese) (See Japanese page) 
(in English) Learning with noisy labels / Semi-supervised learning / Early-learning regularization / weighted loss / / / /  
Reference Info. ITE Tech. Rep.
Paper #  
Date of Issue  
ISSN  
Download PDF

Conference Information
Committee BCT IEICE-SIS  
Conference Date 2022-10-13 - 2022-10-14 
Place (in Japanese) (See Japanese page) 
Place (in English) Hachinohe Institute of Technology 
Topics (in Japanese) (See Japanese page) 
Topics (in English) System Implementation Technology, Short Range Wireless Systems, Smart Multimedia Systems, Broadcasting Technology, etc. 
Paper Information
Registration To IEICE-SIS 
Conference Code 2022-10-SIS-BCT 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Robust Semi-Supervised Learning for Noisy Labels Using Early-learning Regularization and Weighted Loss 
Sub Title (in English)  
Keyword(1) Learning with noisy labels  
Keyword(2) Semi-supervised learning  
Keyword(3) Early-learning regularization  
Keyword(4) weighted loss  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Ryota Higashimoto  
1st Author's Affiliation Kansai University (Kansai Univ.)
2nd Author's Name Soh Yoshida  
2nd Author's Affiliation Kansai University (Kansai Univ.)
3rd Author's Name Mitsuji Muneyasu  
3rd Author's Affiliation Kansai University (Kansai 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 2022-10-14 10:00:00 
Presentation Time 20 minutes 
Registration for IEICE-SIS 
Paper #  
Volume (vol) vol.46 
Number (no)  
Page  
#Pages  
Date of Issue  


[Return to Top Page]

[Return to ITE Web Page]


The Institute of Image Information and Television Engineers (ITE), Japan