Our new method, Hessian Null Space Continuation (HNC)! Around any trained network there is a large connected set of weights that compute nearly the same function. HNC walks through this set along flat directions of the Hessian. 1/3
How much representational diversity exists around a trained network? A lot, it turns out! Moving a ViT’s weights by 1.3% while preserving its predictions, its representations end up less similar to the original than a random-init ViT’s ‼️ New paper: Hessian Null Space Continuation (HNC) 💫🧵 (1/12)