VISReg Targets More Stable, Generalizable SSL

VISReg, short for Variance-Invariance-Sketching Regularization, is a new regularization-based self-supervised learning method aimed at improving generalization while mitigating representation collapse. The posts say it seeks to break the usual stability-performance trade-off by reducing reliance on training tricks such as EMA and teacher-student style setups.

2026-07-14 ~ 2026-07-15 · 2 related posts

• VISReg: Improving Self-Supervised Representation Collapse — 机器之心 · 2026-07-14
• VISReg Enhances Self-Supervised Generalization — ylecun · 2026-07-15