Announcement
A 1.3% weight shift reportedly preserves a ViT's predictions while altering its representations
A post says the altered ViT's representations were less similar to the original's than a randomly initialized ViT's.
TLDR
A post introducing the paper “Hessian Null Space Continuation” says changing a trained vision transformer's (ViT's) weights by 1.3% preserved its predictions. The altered model's representations became less similar to the original's than those of a randomly initialized ViT, the post says.
Combined views
22.3K
2 Sources, first seen ago
