Making and breaking symmetry induces Thatcher Effect in CLIP. Seems a single conv1 coordinate, 0196 (1024) - not a retinal (e.g. Gabor) filter, but used as a positional coordinate - is puppeteering 5 main early attention heads that underly the Thatcher Effect.🙃 #machinelearning t.co/EWpYsxN6l3
- 0 réponses
- 0 partages
- 1 j'aime
- 41 vues
