Complete Embedding visualizations

jbloomAus / DecisionTransformerInterpretability

Interpreting how transformers simulate agents performing RL tasks

MIT License

61 stars 15 forks source link

Open jbloomAus opened 1 year ago

jbloomAus commented 1 year ago

[x] 1. Dot product between each output action.
[x] 2. Dot product between each input action.
[ ] 3. Dot product between each time embedding.
[x] 4. Dot product between each input state. 980*980. Create filtered version of this.
[ ] 5. Eigenvalues of the state embedding? Label them?

Use Cosine Similarity not dot product.