aiqbench analysis
This paper advances self-supervised learning by introducing kernel-based regularizers that improve training stability and flexibility for Euclidean representations in Joint-Embedding Predictive Architectures (JEPAs). By deriving closed-form high-dimensional limits for sliced discrepancies, the authors provide a more robust framework for scaling vision learners without relying on standard isotropic Gaussian priors. The work represents a significant methodological contribution to the foundations of machine learning representation theory.
KerJEPA: Kernel Discrepancies for Euclidean Self-Supervised Learning
arXiv:2512.19605v2 Announce Type: replace Abstract: Breakthroughs in self-supervised Joint-Embedding Predictive Architectures (JEPAs) have established that regularizing Euclidean representations toward isotropic Gaussian priors yields provable gains in training stability and downstream generalization. We introduce a new, flexible family of Euclidean joint embedding self-supervised learning algorithms with kernel-based regularizers. One instance of this family corresponds to the SIGReg Epps-Pulley regularizer which approximates a sliced Maximum Mean Discrepancy (MMD) with a Gaussian prior and Gaussian kernel. By expanding the class of viable kernels and priors and computing the closed-form high-dimensional limit of sliced discrepancies, we develop alternative methods with improved training stability and design flexibility and leverage those techniques to train strong vision learners at scale.
Abstract:Breakthroughs in self-supervised Joint-Embedding Predictive Architectures (JEPAs) have established that regularizing Euclidean representations toward isotropic Gaussian priors yields provable gains in training stability and downstream generalization. We introduce a new, flexible family of Euclidean joint embedding self-supervised learning algorithms with kernel-based regularizers. One instance of this family corresponds to the SIGReg Epps-Pulley regularizer which approximates a sliced Maximum Mean Discrepancy (MMD) with a Gaussian prior and Gaussian kernel. By expanding the class of viable kernels and priors and computing the closed-form high-dimensional limit of sliced discrepancies, we develop alternative methods with improved training stability and design flexibility and leverage those techniques to train strong vision learners at scale.
Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
arXiv:2512.19605 [cs.LG]
(or arXiv:2512.19605v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2512.19605 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 206 words
Related coverage
Expanding SPHERE-JEPA: A Family of Statistical Regularizers for the Hypersphere
arXiv · Jun 17, 2026
SPHERE-JEPA: Spherical Prediction with Homogeneous Embeddings
arXiv · Oct 6, 2026
Follow the Geometry, Not the Model: Cold Start Semi-Supervised Learning
arXiv · Sep 16, 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · Sep 17, 2026