Background

Attempting to create Meta-SAEs on a classification Vit. The hope is that this gives more access to fine-grained features.

Meta-SAE ViT-SAE Training

State of Vision Interp

Work Done

• Wrote some boilerplate Vit and training it currently (it is based off of a google one so may not have 'pure?' results ). I guess purity in this sense is just wondering about the dataset used to train teh other. Though it should not matter on a representation level since dataset has a set of images.But if there are completely different images we will not be able to capture it.
• installed vit-prisma and attempted to use its vit training code
• gave up on MNIST sae since could not figure out how to parse the features --> lack of data (just 9 categories)

Code

Confusions

• How do they eval
• Which layer is good
• Do i Need better saes
• Will I run into the same data issue with these SAEs ?

Next Steps

• Read State of Vision Interp + other background resources
• Train a ViT
  • Smaller Raw One
• Train an SAE on the Vit
• Follow others in evaluating the SAE
  • ViT-Prisma Evals
  • Hugo Fry's ViT Sae

GOAL: Reproduce the Vit Prisma SAE training and then run a Meta-SAE on it.