DETERMINING A MASK FOR EXPLAINING A CLASSIFICATION

摘要

The invention relates to a computer-implemented method (400) of determining a mask for explaining a classification of an input instance to a trained classifier. The mask indicates features of the input instance at a layer of the trained classifier for explaining the classification. The mask is updated in multiple outer iterations by updating the mask from a current mask to an updated mask. To this end, an additional mask is determined in an inner optimization. The inner optimization is configured to avoid indicating features already indicated by the current mask.