Fig. 6. - GALORE explanation architecture (x: Cardinal, xc: Summer Tanager.). Feature activations Fh and Fs are computed for pre-determined layers of the classifier (upper branch) and confidence predictor (lower branch), respectively. Attributions for prediction ha, ambiguous or counter class hb, and confidence score s are computed by attribution functions q(.,.) according to (18), (19), (20). a, b are a class pair of candidate class ambiguities set for deliberative explanations, and y*, yc for counterfactual explanations. These attributions are combined with (5) or (11) to obtain the final map, which is thresholded to produce explanations. Multiple pairs (a,b) are shown for deliberative explanations, where a is Cardinal (Ca), and b Pine Grosbeak (PG), Purple Finch (PF) or Summer Tanager (ST). Counterfactual explanations are obtained by additionally reversing the roles of x and xc and thresholding the discriminant heat maps.
