Quantifying Explainability of Saliency Methods in Deep Neural Networks

Regardless, the development of heatmap methods have continued without correspondingly reliable ways to evaluate how one heatmap is better than another. The metrics used to quantify the quality of heatmaps are sometimes indirect, and at other times, qualitative assessment of the quality of heatmaps appear to be possibly given in hind-sight to fit natural reasoning. This often occurs due to the lack of ground-truth heatmaps to verify the correctness of the generated heatmaps. Under such situation, the quality and effectiveness of interpretable heatmaps have nevertheless been demonstrated in several ways.