![[2602.14272] Radial-VCReg: More Informative Representation Learning Through Radial Gaussianization](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.14272] Radial-VCReg: More Informative Representation Learning Through Radial Gaussianization
Summary
The paper presents Radial-VCReg, a novel approach to self-supervised learning that enhances representation learning by addressing the limitations of existing methods through radial Gaussianization.
Why It Matters
This research is significant as it tackles the challenge of maximizing informative representations in high-dimensional spaces, which is crucial for improving machine learning models. By introducing Radial-VCReg, the authors provide a method that can lead to better performance in various applications, making it relevant for researchers and practitioners in machine learning.
Key Takeaways
- Radial-VCReg improves upon existing methods by incorporating a radial Gaussianization loss.
- The method aligns feature norms with the Chi distribution, enhancing representation learning.
- Empirical results show consistent performance improvements on both synthetic and real-world datasets.
- Radial-VCReg reduces higher-order dependencies, promoting diverse representations.
- The approach broadens the class of distributions that can be transformed towards normality.
Computer Science > Machine Learning arXiv:2602.14272 (cs) [Submitted on 15 Feb 2026] Title:Radial-VCReg: More Informative Representation Learning Through Radial Gaussianization Authors:Yilun Kuang, Yash Dagade, Deep Chakraborty, Erik Learned-Miller, Randall Balestriero, Tim G. J. Rudner, Yann LeCun View a PDF of the paper titled Radial-VCReg: More Informative Representation Learning Through Radial Gaussianization, by Yilun Kuang and 6 other authors View PDF HTML (experimental) Abstract:Self-supervised learning aims to learn maximally informative representations, but explicit information maximization is hindered by the curse of dimensionality. Existing methods like VCReg address this by regularizing first and second-order feature statistics, which cannot fully achieve maximum entropy. We propose Radial-VCReg, which augments VCReg with a radial Gaussianization loss that aligns feature norms with the Chi distribution-a defining property of high-dimensional Gaussians. We prove that Radial-VCReg transforms a broader class of distributions towards normality compared to VCReg and show on synthetic and real-world datasets that it consistently improves performance by reducing higher-order dependencies and promoting more diverse and informative representations. Comments: Subjects: Machine Learning (cs.LG) Cite as: arXiv:2602.14272 [cs.LG] (or arXiv:2602.14272v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2602.14272 Focus to learn more arXiv-issued DOI via DataCite (...
