Work Done

• Established rudimentary interp via EMNIST classification labels
• Fixed up some of the experiment code to sagve the indices

• got preliminary findings that there is a optimal depth for meta-saes

• Found that the average amount of activations do increase the deeper you go which implies some level of fine-grained-ness

• The amount of activaitons of the max hit their peak at 2 depths in.

Confusions

• Am I actually doing what I think I am doing?
  • what I am trying to do: Take a trained MNIST model, feed it MNIST data, train an SAE on those. Track the values with EMNIST data to see what patterns there are outside of numbers ( in this case a balanced dataset of letters and numbers)
  • the max activatoins are seeing if tehre are any neurons taht fire exclusively on one label OR if there are some basic trends like letters and numbers

Next Steps

• Plot some of the data using the techniques from GPT
• Create a preliminary writeup using the charts and any other ones to show what i have so far
  • REPLICATE, REPLICATE, REPLICATE with other datasets (like fruit or other EMNISTs) --> basically take all the classification datasets
• Also need to look at the analysis for the meta-sae paper

Cool Stuff

• Fruit Classification Dataset: Has 141 Classes
• SAE Vision Explainability
• Transcoders as a way to interpret MLP layers