Influence scores

[I-Good-Vegetable]

Hi Dr. Bae,

Thanks for releasing the codes!

I'm wondering how to interpret the influence score. I assume it captures the change of loss when down-weighting a train sample (latter one in the picture).

Please let me know if I'm wrong. Many thanks indeed!

Best,
Jason

[pomonam]

Hi Jason, it depends on how you defined the Task (specifically, compute_train_loss and compute_measurement). If you are using the example code (e.g., https://github.com/pomonam/kronfluence/blob/main/examples/uci/analyze.py#L58), you are correct - it should be the latter, as it will directly compute (loss grad)^T H^{-1} (measurement grad) from the code.

[I-Good-Vegetable]

Thank you Dr. Bae! Got it! ❤️