I Think Multimodal Explanations Could Be a Really Interesting New Research Area

I agree that we should work more on explainability and it is really hard. I think the solution has to be kind of multimodal explanations too because ultimately language isn't going to be that great at telling you, oh it's because of these three pixels and this particular edge and this particular combination that the model identified this as a cat. So then even no matter how black box the algorithm is, if you can tie it back to its training data, you will have some explainability there. It's still, it's tricky because if in a litigious society, you might still have to go with the one that kills more people, but you have like a good cover for your organization