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 2026-03-24 16:15
Variational Inference for Spatially Correlated Discrete Distributions using Gaussian Fields with KMS Structure and its Application to Segmentation
Ryu Tadokoro (Tohoku Univ.), Tsukasa Takagi, Shin-ichi Maeda (PFN) IBISML2025-46
Abstract (in Japanese) (See Japanese page) 
(in English) In semantic segmentation, obtaining high-quality annotations is prohibitively expensive, and real-world datasets inevitably contain ``noisy labels'' arising from inter-annotator disagreement or human errors. Notably, in the image domain, these noisy labels are not distributed independently but tend to exhibit spatial correlations among adjacent pixels. However, explicitly treating such spatially correlated discrete variables in a probabilistic model renders the computation of the marginal likelihood intractable due to the summation required over all possible label configurations.

To address this computational challenge, we propose a novel discrete probabilistic distribution termed the ELBO-Computable Correlated Discrete Distribution (ECCD), which explicitly handles spatially correlated label noise. The ECCD models the spatial dependencies of discrete labels through a latent continuous Gaussian field. By incorporating a Kac-Murdock-Szeg"{o} (KMS) structure into the covariance matrix of the Gaussian field, we reduce the computational complexity of determinant and inverse matrix operations to a linear order, enabling efficient computation of the Evidence Lower Bound (ELBO).

We applied the proposed method to semantic segmentation tasks in medical and remote sensing imagery and confirmed that our approach achieves superior robustness and accuracy compared to existing noisy-label learning methods in scenarios characterized by spatially correlated noise.
Keyword (in Japanese) (See Japanese page) 
(in English) Label noise / Variational inference / Discrete distribution / / / / /  
Reference Info. IEICE Tech. Rep., vol. 125, no. 425, IBISML2025-46, pp. 38-44, March 2026.
Paper # IBISML2025-46 
Date of Issue 2026-03-17 (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 IBISML2025-46

Conference Information
Committee PRMU IPSJ-CVIM IBISML ITE-SIP  
Conference Date 2026-03-24 - 2026-03-25 
Place (in Japanese) (See Japanese page) 
Place (in English)  
Topics (in Japanese) (See Japanese page) 
Topics (in English)  
Paper Information
Registration To IBISML 
Conference Code 2026-03-PRMU-CVIM-IBISML-SIP 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Variational Inference for Spatially Correlated Discrete Distributions using Gaussian Fields with KMS Structure and its Application to Segmentation 
Sub Title (in English)  
Keyword(1) Label noise  
Keyword(2) Variational inference  
Keyword(3) Discrete distribution  
Keyword(4)  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Ryu Tadokoro  
1st Author's Affiliation Tohoku University (Tohoku Univ.)
2nd Author's Name Tsukasa Takagi  
2nd Author's Affiliation Preferred Networks (PFN)
3rd Author's Name Shin-ichi Maeda  
3rd Author's Affiliation Preferred Networks (PFN)
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-3 
Date Time 2026-03-24 16:15:00 
Presentation Time 20 minutes 
Registration for IBISML 
Paper # IBISML2025-46 
Volume (vol) vol.125 
Number (no) no.425 
Page pp.38-44 
#Pages 7 
Date of Issue 2026-03-17 (IBISML) 

[Return to Top Page]

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