• About

Questions

When NN adapts by transfer learning, what jnterpretablr neurons are removed and what aren't?

Identify component label by analogous relations

• Transformers:

GPT-2-Small Circuits:

• Are there consistent patterns in inputs that allow us to identify which subject, among given subjects, the model will output? Motivated by tests from test_prompt_most_recent_S.ipynb

  • HYPOTHESIS: It may be that the “default” would output earliest.

• How does adding a duplicate name change from outputting “earliest subject” to “latest subject”?

  • HYPOTHESIS: ???

• MLP:

• CNN:

Superposition

• If a neuron activates for 2 unrelated objs, is there actually something they have in common? Perhaps down the line in a circuit? Or a “shared role” (inhibition, eg) is used for them?
  • Expm by finding commonalities in their circuits

Hypotheses

[ Evidence level: ? how to quanify this in comparable ways, if possible ? ]