Contrastive Explanations

This page describes contrastive explanations. These are necessary and sufficient explanations. A normal explanation from ReX is only sufficient.

From the command line, the basic call is

ReX <image> --script <script.py> --contrastive

Example

This image is a tandem bike

If we run

ReX tandem.jpg --script ../tests/scripts/pytorch_resnet50.py --vv --output tandem_exp.png

we get

Passing the highlighted pixels to the model, against the baseline defined in rex.toml, is enough to get the classification tandem. However, if we remove these highlighted pixels and leave the rest alone, we still get tandem. Why is that? Because these pixels are sufficient to get tandem, but they aren’t necessary. There must be at least one more, independent, explanation for tandem in the image.

We can try to find other sufficient explanations using --multi.

ReX tandem.jpg --script ../tests/scripts/pytorch_resnet50.py --multi --vv --output tandem_exp.png

Each one of these is, by itself, sufficient for tandem. By default, contrastive uses combinations from the set of found sufficient explanations to find a combinations which is also necessary: if we remove these pixels then we no longer have a tandem, if we have only these pixels, we have tandem. This explanations is both sufficient to get tandem and necessary to get tandem.

Notes

contrastive uses multiple explanations under the hood and then tests all combinations of discovered explanations. As a result, it can fail to find a contrastive explanation, especially if there are more sufficient explanations than discovered by --multi. In this case, ReX quits with an error message.

constrastive takes an optional int at the command line, indicating how many spotlights to launch. This defaults to 10, just like with multiple explanations.

The constrastive algorithm is an instance of the set packing problem. As such, in the worst case scenario, it can be quite expensive to compute.