Theoretical insights into pseudo-label-based semi-supervised learning: Convergence rate and sample complexity analysis

Zeping Min, Qian Ge, Cheng Tai

Keywords: Pseudo-label-based semi-supervised learning, Convergence rate, Sample complexity estimate

TL;DR: We analyze pseudo-label-based semi-supervised learning, showcasing its convergence rate, sample complexity, and utility with vast unlabeled data, particularly for under-parameterized models.

Abstract: Pseudo-label-based semi-supervised learning has recently emerged as an effective technique in various domains. In this paper, we present a comprehensive theoretical analysis of the algorithm, significantly advancing our understanding of its empirical successes. Our analysis demonstrates that the algorithm can achieve a remarkable convergence rate of $\mathcal{O}(N^{-1/2})$ order, and we provide an estimate of the sample complexity. We further investigate the algorithm's performance in scenarios with an infinite number of unlabeled data points, highlighting its effectiveness in leveraging large-scale unlabeled data. A key insight of our study is that incorporating pseudo-labeled data can improve model training when correctly labeled data is more valuable than the interference caused by mislabeled data, particularly for under-parameterized models that tend to ignore the impact of incorrect labels. Experimental findings corroborate the accuracy of our estimations. This study elucidates the strengths and limitations of the pseudo-label-based semi-supervised learning algorithm, paving the way for future research in this field. The code can be found at the anonymous URL https://anonymous.4open.science/r/mycode_1-A2EE

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/2024/AuthorGuide.

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: 4328