Replacing Fragile Heuristics in DINOv2 with Pure Optimal Transport

RexDouglass · x · 2026-07-30

The thread discusses the technical limitations of current Self-Supervised Learning (SSL) models like DINOv2. The author notes that while these models are engineering marvels, they rely on fragile hacks such as EMA, stop-gradients, and custom centering to prevent representation collapse.

It raises the question of whether these empirical heuristics could be replaced entirely by pure, mathematically sound optimal transport theory to fundamentally improve training stability.

Original post →

More from Research

Research channel →