Enhancing Deep Partial Label Learning via Casting it to a Satisfiability Problem

Liwei Che, Efthymia Tsamoura, Vladimir Pavlovic

Keywords: weak supervised learning, partial label learning

Abstract: Partial label learning (PLL) is a challenging real-world problem in the field of weakly supervised learning, in which each data instance contains a set of candidate labels with multiple ambiguous labels and one gold label. Although recent progress in PLL using deep representation learning has led to significant advances, the methods continue to experience significant performance drops on data with high label ambiguity and fine-grained categories. By casting PLL into a satisfiability problem and incorporating a loss based on this reduction, we show that the accuracy of those techniques can be further improved. We establish several key theoretical properties of the proposed SATisfiability-based (SAT) loss and its learning error bound. Our extensive empirical comparison reveals that the proposed loss improves over existing PLL techniques by up to 25.12% on multi-class benchmarks and 12.50% on fine-grained categorized benchmarks.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3952