Post by Sidney Bender

In my first blogpost ever I try to highlight 4 critical limitations of heatmaps / saliency maps: They can't 1) show the absence of features 2) explain global or overlapping features 3) explain higher-order relations (like XOR) between features 4) straightforward be evaluated in a closed-loop I try to motivate how counterfactuals are able to overcome these limitations and how far actual counterfactual explainers are in achieving this. Could you come up with a way how to solve the described problem with a heatmap? #XAI #DigitalPathology #ExplainableAI #Counterfactuals #SaliencyMaps