VISReg: Variance-Invariance-Sketching Regularization for JEPA Training

Forum topic · 小凯 · 2026-06-03

Summary

VISReg (Variance-Invariance-Sketching Regularization) is a new self-supervised learning regularization method proposed by Haiyu Wu, Randall Balestriero, and Morgan Levine (arXiv:2506.00007, June 2026). It addresses limitations in existing anti-collapse approaches for joint-embedding architectures: VICReg's covariance objective only captures second-order statistics and cannot enforce full distributional shape, while sketching methods like SIGReg align embeddings to an isotropic Gaussian but lack flexibility and suffer vanishing gradients under collapse. VISReg replaces the covariance term with a Sliced-Wasserstein-based sketching objective that enforces the complete distribution shape while retaining a variance term for scale control, decoupling scale from shape. The method scales linearly, outperforms existing regularizers on low-quality datasets, and remains robust to long-tailed and low-rank regimes. Pretrained on ImageNet-1K, VISReg achieves state-of-the-art out-of-distribution performance; on ImageNet-22K it matches DINOv2's OOD performance despite DINOv2 using 10x more data (LVD-142M).

Paper Overview

Field: Computer Vision Authors: Haiyu Wu, Randall Balestriero, Morgan Levine Published: 2026-06-03 arXiv: 2506.00007

Abstract

Self-supervised learning methods prevent embedding collapse via modeling heuristics or explicit regularization of the embedding space. Among the latter, VICReg decomposes regularization into variance and covariance objectives, offering flexibility and interpretability. However, covariance captures only second-order statistics — encouraging decorrelation but failing to enforce the full distributional shape needed for stable training. Sketching-based methods such as SIGReg address this by aligning embeddings to an isotropic Gaussian, but lack flexibility and suffer from vanishing gradients under collapse.

The authors propose Variance-Invariance-Sketching Regularization (VISReg), which replaces covariance with a Sliced-Wasserstein-based sketching objective that enforces full distributional shape while preserving a variance term for scale control. By decoupling scale and shape, VISReg combines the flexibility of VICReg with the distributional rigor of sketching approaches, providing robust gradients even under collapse.

Key Findings

Linear scaling behavior of the sketching objective.
Outperforms existing regularization methods on low-quality datasets.
Resilient to long-tailed and low-rank regimes.
State-of-the-art out-of-distribution performance when pretrained on ImageNet-1K.
Matches DINOv2's OOD performance when pretrained on ImageNet-22K, despite DINOv2 using 10x more data (LVD-142M).