awslabs / awslabs/aws-cv-task2vec
why are task2vec embeddings using diagonal of FIM better than using raw activations?
Open
- Dominant language
- Python
- Stars
- 126
- Forks
- 31
- PR merge metrics
- No merged PRs in 30d
Description
I was curious if there were experiments in favour justifying diagonal of the FIM vs activations.
Contributor guide
Assessment
This issue has not been assessed yet.