Partial Optimal Transport for Open-set Semi-supervised Learning
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.
Keywords: Open-set problem, Optimal transport, Semi-supervised learning, Out-of-distribution detection
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Semi-supervised learning (SSL) is a machine learning paradigm that leverages both labeled and unlabeled data to improve the performance of learning tasks. However, SSL methods make an assumption that the label spaces of labeled and unlabeled data are identical, which may not hold in open-world applications, where the unlabeled data may contain novel categories that were not present in the labeled training data, essentially outliers. This paper tackles open-set semi-supervised learning (OSSL), where detecting these outliers, or out-of-distribution (OOD) data, is critical. In particular, we model the OOD detection problem in OSSL as a partial optimal transport (POT) problem. With the theory of POT, we devise a mass score function (MSF) to measure the likelihood of a sample being an outlier during training. Then, a novel OOD loss is proposed, which allows to adapt the off-the-shelf SSL methods with POT into OSSL settings in an end-to-end training manner.
Furthermore, we conduct extensive experiments on multiple datasets and OSSL configurations, demonstrating that our method consistently achieves superior or competitive results compared to existing approaches.
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: 1549
Loading