GTR: Semi-supervised Learning with Grouping and Transporting for Robust Thresholding

Weiyang Jin, Siyuan Li, Zicheng Liu, Zedong Wang, Juanxi Tian, Di Wu, Cheng Tan, Chang Yu, Stan Z. Li

Keywords: Semi-supervised Learning, Grouping, Thresholding, Plug-and-play, Pseudo-labeling

TL;DR: This paper designs a Grouping and Transporting method for semi-supervised learning that robustly selects semi-hard samples with test-time augmentations and consistency constraints.

Abstract: Semi-supervised learning (SSL) digs unlabeled data by pseudo-labeling when labeled data is limited. Despite various auxiliary strategies enhancing SSL training, the main challenge is how to determine reliable pseudo labels through a robust thresholding algorithm based on quality indicators (e.g., confidence scores). However, the existing strategies for distinguishing low or high-quality labels through simple grouping indicators remain in trivial design, ignoring the characteristics of the data distribution itself, which cannot guarantee robustness and efficiency. To this end, we group the quality indicators of pseudo labels into three clusters (easy, semi-hard, and hard) and statistically reveal the real bottleneck of threshold selection, i.e., the sensitivity of semi-hard samples, through empirical analysis. We propose an adaptive Grouping and Transporting method that Robustly selects semi-hard samples with test-time augmentations and consistency constraints while saving the selection budgets of easy and hard samples, dubbed as GTR. Our proposed GTR can effectively determine high-quality data when applied to existing SSL methods while reducing redundant costs in the selection. Extensive experiments on 11 SSL benchmarks across three modalities verify that GTR can achieve significant performance gains and speedups over Pseudo Label, FixMatch, and FlexMatch.

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: 4731