UCPH Statistics Seminar: Ahcene Boubekki

Speaker: Ahcene Boubekki from Physikalisch-Technische Bundesanstalt (Berlin)

Title: Explaining explanations with feature activations.

Abstract: Recent developments in the field of explainable artificial intelligence (XAI) for image classifiers investigate region-based attribution explanations using ad-hoc superpixel algorithms. In this talk, we will discuss how one can extract a segmentation of the input from the feature activations of the frozen convolutional network to be explained, which reflects the semantic information stored/captured by the network. Building upon this result, we will review how the quantization of the saliency maps using these segmentations eases their interpretation and reveals the inconsistencies of the widely used area-under-the-relevance-curve metric, boosts the weakly supervised object localization performance, opening the door to a (semi?) unsupervised semantic evaluation.