search

pytorch / captum

Model interpretability and understanding for PyTorch
https://captum.ai
BSD 3-Clause "New" or "Revised" License
4.94k stars 499 forks source link

DeepLift RevealCancel Rule Support #657

Open Ycblue opened 3 years ago

Ycblue commented 3 years ago

Hi all,

First of all, thank you very much for the amazing work on Captum and making it available for everyone.

I was wondering about the future plans for support of DeepLifts RevealCancel Rule in general. The algorithm description mentions that the rule is going to be available in later releases, but I haven't been able to find any other info. Is it currently a priority and when can we expect it to be released?

Cheers and all the best,

Yu-Chia

NarineK commented 3 years ago

Thank you for the feature request, @Ycblue! Currently it is not high priority item but we can potentially prioritize it based on the demand.

Ycblue commented 3 years ago

From my understanding, it does improve DeepLift quite a bit. But then again, I just recently got into the literature of DeepLift and DeepShap, so I am not entirely sure if the performance of DeepShap is already better than the full implementation of DeepLift. Also it is very difficult to actually compare these methods when a key feature (imo) of DeepLift is still missing. Please correct me if I am wrong.
I am for sure very interested in this feature and I am sure others are as well!

On another note, the implementation of 'DeepLift Ablation' in this paper (http://proceedings.mlr.press/v119/sixt20a.html) seems to outperform the original formulation. Would it be possible to test this out/implement for Captum as well?

yihao001 commented 3 years ago

Will be great to have the RevealCancel rule implemented!

jmschrei commented 1 year ago

Just checking in on this. It'd be great to have the RevealCancel rule implemented for people in genomics settings.