Learning Label Distribution with Subtasks
Lists
Only you see this
Abstract
Label distribution learning (LDL) is a novel learning paradigm that emulates label polysemy by assigning label distributions over the label space. However, recent LDL work seems to exhibit a notable contradiction: 1) some existing LDL methods employ auxiliary tasks to enhance performance, which narrows their focus to specific domains, thereby lacking generalization capability; 2) conversely, LDL methods without auxiliary tasks rely on losses tailored solely to label distributions of the primary task, lacking additional supervised information to guide the learning process. In this paper, we propose $\mathcal{S}$-LDL, a novel and minimalist solution that partitions the label distribution of the primary task into subtask label distributions, i.e., a form of pseudo-supervised information, to reconcile the above contradiction. $\mathcal{S}$-LDL encompasses two key aspects: 1) an algorithm capable of generating subtasks without any extra knowledge, with subtasks deemed valid and reconstructable via our analysis; and 2) a plug-and-play framework seamlessly compatible with existing LDL methods, and even adaptable to derivative tasks of LDL. Experiments demonstrate that $\mathcal{S}$-LDL is effective and efficient. To the best of our knowledge, this represents the first endeavor to address LDL via subtasks. The code will soon be available on GitHub to facilitate reproducible research.
BibTeX
@misc{wu2025learning,
author = {Haitao Wu and Weiwei Li and Xiuyi Jia},
title = {{Learning Label Distribution with Subtasks}},
year = {2025},
url = {https://openreview.net/forum?id=GMu1yJST86},
howpublished = {OpenReview},
note = {Submitted to ICLR 2025},
}Loading related content...