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 2026-01-29 11:00
Inverse Error Ratio: Subband Weighting in Training DNNs
Naoyuki Ichimura (AIST) PRMU2025-29
Abstract (in Japanese) (See Japanese page) 
(in English) In the training of deep neural networks (DNNs), high-frequency components are known to be learned more slowly—a phenomenon referred to as spectral bias. To suppress this bias in image-processing DNNs, this study employs the Spatial Frequency Loss, formulated as a weighted sum of the mean squared errors between the subband features of output and target images. To effectively learn high-frequency components, the weights corresponding to high-frequency subbands need to be set higher. However, empirically determining multiple weights is not straightforward due to the combinatorial complexity and computational cost of training required to validate multiple candidates. To address this issue, a subband weighting method based on the Inverse Error Ratio (IER) is proposed. The IER is defined as the ratio of the mean squared error of the entire image to that of each subband feature. This enables automatic and data-driven subband weighting that simultaneously accounts for the initial frequency response of DNNs and compensates for the amplitude energy imbalance in datasets. Experiments using a Vector-Quantized Variational Auto-Encoder demonstrate the effectiveness of the proposed method in comparison with conventional loss functions.
Keyword (in Japanese) (See Japanese page) 
(in English) Deep neural network / Spectral bias / Spatial frequency loss / Subband weighting / / / /  
Reference Info. IEICE Tech. Rep., vol. 125, no. 348, PRMU2025-29, pp. 25-30, Jan. 2026.
Paper # PRMU2025-29 
Date of Issue 2026-01-22 (PRMU) 
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 PRMU2025-29

Conference Information
Committee PRMU IPSJ-CVIM VRSJ-SIG-MR MVE  
Conference Date 2026-01-29 - 2026-01-30 
Place (in Japanese) (See Japanese page) 
Place (in English)  
Topics (in Japanese) (See Japanese page) 
Topics (in English)  
Paper Information
Registration To PRMU 
Conference Code 2026-01-PRMU-CVIM-SIG-MR-MVE 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Inverse Error Ratio: Subband Weighting in Training DNNs 
Sub Title (in English)  
Keyword(1) Deep neural network  
Keyword(2) Spectral bias  
Keyword(3) Spatial frequency loss  
Keyword(4) Subband weighting  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Naoyuki Ichimura  
1st Author's Affiliation National Institute of Advanced Industrial Science and Technology (AIST)
2nd Author's Name  
2nd Author's Affiliation ()
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 2026-01-29 11:00:00 
Presentation Time 15 minutes 
Registration for PRMU 
Paper # PRMU2025-29 
Volume (vol) vol.125 
Number (no) no.348 
Page pp.25-30 
#Pages 6 
Date of Issue 2026-01-22 (PRMU) 

[Return to Top Page]

[Return to IEICE Web Page]
The Institute of Electronics, Information and Communication Engineers (IEICE), Japan