Evaluation

[Wienhannover]

Hi, thanks for the amazing work, could you please point out how "Insertion and Deletion" scores are calculated in the code?

[arpita-chowdhury-osu]

Currently, the codebase does not implement the Insertion and Deletion score computation, but I can add that functionality. Since this work focuses on an interpretable method rather than a traditional explainable method, the evaluation procedure differs from the standard approach described in prior explainability literature, as outlined in the paper.

For each correctly classified image, we progressively prune attention heads until we are left with a minimal set of heads such that masking any one of them would change the model’s prediction. These remaining heads are considered the required heads for the decision. We then aggregate the attention maps from these heads and normalize the resulting map. Based on the normalized scores, pixels are ranked from high to low importance. Using this ranking, we perform the insertion and deletion procedures to measure how the model’s prediction changes as informative pixels are added or removed, which allows us to compute the corresponding scores.