Adding "Gifsplanation via Latent Shift: A Simple Autoencoder Approach to Counterfactual Generation"

[ieee8023]

🚀 Feature

(I am interested in coding this, just need some guidance for a pull request)

Having an implementation of the "Latent Shift" method integrated into captum. Paper: https://openreview.net/forum?id=rnunjvgxAMt

You can look at some example gifs and 2d attribution maps here: https://mlmed.org/gifsplanation/
You can try a demo here too! https://colab.research.google.com/github/mlmed/gifsplanation/blob/main/demo.ipynb

Motivation

I am the author of this work and I would like to enable more people to use it. I have the current source code here: https://github.com/mlmed/gifsplanation

The approach uses a trained encoder/decoder pair along with a classifier to generate counterfactual explanations (and 2d attribution maps). I imagine this would integrate well into your library because it works with any blackbox encoder/decoder and classifier.

Pitch

1. Are you interested to accept a pull request for this?
2. How should it fit into the library? (can you provide some method stubs that I can implement)

The challenge I see is how best to add the autoencoder into things because it needs to be an extra argument. If you don't have an idea for the best interface for this I can brainstorm one.

Here is an overview of how the method works:
+++

And here is how 2d attribution maps are constructed: