Grigory Sapunov(@che_shr_cat):1/ 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? 🧵

Grigory Sapunov

@che_shr_cat

PhD in AI | GDE in AI/ML | CTO Intento | Author "Deep Learning with JAX" 📝 ML insights: 🤖 Daily AI paper reviews:

加入 June 2009

1.6K 正在关注 4.1K 粉丝

Grigory Sapunov@che_shr_cat

2026.07.29 20:18

1/ 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? 🧵

显示更多