Clustering Properties of Self-Supervised Learning

Xi Weng, Jianing An, Xudong Ma, Binhang Qi, Jie Luo, Xi Yang, Jin Song Dong, Lei Huang
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:66597-66616, 2025.

Abstract

Self-supervised learning (SSL) methods via joint embedding architectures have proven remarkably effective at capturing semantically rich representations with strong clustering properties, magically in the absence of label supervision. Despite this, few of them have explored leveraging these untapped properties to improve themselves. In this paper, we provide an evidence through various metrics that the encoder’s output encoding exhibits superior and more stable clustering properties compared to other components. Building on this insight, we propose a novel positive-feedback SSL method, termed Representation Self-Assignment (ReSA), which leverages the model’s clustering properties to promote learning in a self-guided manner. Extensive experiments on standard SSL benchmarks reveal that models pretrained with ReSA outperform other state-of-the-art SSL methods by a significant margin. Finally, we analyze how ReSA facilitates better clustering properties, demonstrating that it effectively enhances clustering performance at both fine-grained and coarse-grained levels, shaping representations that are inherently more structured and semantically meaningful.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-weng25a, title = {Clustering Properties of Self-Supervised Learning}, author = {Weng, Xi and An, Jianing and Ma, Xudong and Qi, Binhang and Luo, Jie and Yang, Xi and Dong, Jin Song and Huang, Lei}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {66597--66616}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/weng25a/weng25a.pdf}, url = {https://proceedings.mlr.press/v267/weng25a.html}, abstract = {Self-supervised learning (SSL) methods via joint embedding architectures have proven remarkably effective at capturing semantically rich representations with strong clustering properties, magically in the absence of label supervision. Despite this, few of them have explored leveraging these untapped properties to improve themselves. In this paper, we provide an evidence through various metrics that the encoder’s output encoding exhibits superior and more stable clustering properties compared to other components. Building on this insight, we propose a novel positive-feedback SSL method, termed Representation Self-Assignment (ReSA), which leverages the model’s clustering properties to promote learning in a self-guided manner. Extensive experiments on standard SSL benchmarks reveal that models pretrained with ReSA outperform other state-of-the-art SSL methods by a significant margin. Finally, we analyze how ReSA facilitates better clustering properties, demonstrating that it effectively enhances clustering performance at both fine-grained and coarse-grained levels, shaping representations that are inherently more structured and semantically meaningful.} }
Endnote
%0 Conference Paper %T Clustering Properties of Self-Supervised Learning %A Xi Weng %A Jianing An %A Xudong Ma %A Binhang Qi %A Jie Luo %A Xi Yang %A Jin Song Dong %A Lei Huang %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-weng25a %I PMLR %P 66597--66616 %U https://proceedings.mlr.press/v267/weng25a.html %V 267 %X Self-supervised learning (SSL) methods via joint embedding architectures have proven remarkably effective at capturing semantically rich representations with strong clustering properties, magically in the absence of label supervision. Despite this, few of them have explored leveraging these untapped properties to improve themselves. In this paper, we provide an evidence through various metrics that the encoder’s output encoding exhibits superior and more stable clustering properties compared to other components. Building on this insight, we propose a novel positive-feedback SSL method, termed Representation Self-Assignment (ReSA), which leverages the model’s clustering properties to promote learning in a self-guided manner. Extensive experiments on standard SSL benchmarks reveal that models pretrained with ReSA outperform other state-of-the-art SSL methods by a significant margin. Finally, we analyze how ReSA facilitates better clustering properties, demonstrating that it effectively enhances clustering performance at both fine-grained and coarse-grained levels, shaping representations that are inherently more structured and semantically meaningful.
APA
Weng, X., An, J., Ma, X., Qi, B., Luo, J., Yang, X., Dong, J.S. & Huang, L.. (2025). Clustering Properties of Self-Supervised Learning. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:66597-66616 Available from https://proceedings.mlr.press/v267/weng25a.html.

Related Material