X · @ylecun
· X / Twitter
RT Grigory Sapunov: 1/ SSL models like DINOv2 are engineering marvels, but they rely on fragile hacks like EMA, stop-gradients, and custom centering t…
RT Grigory Sapunov1/ SSL models like DINOv2 are engineering marvels, but they rely on fragile hacks like EMA, stop-gradients, and custom centering to avoid representation collapse. What if we could replace all these heuristics with pure, mathematically sound optimal transport? 🧵