Plausibility of output of neural classification networks

摘要

Procedure (100) for the plausibility of the output of an artificial neural network, CNN (1), used as a classifier, with the steps:A majority of images (2) for which the CNN (1) has determined an allocation (3) to one or more classes (3a-3c) of a given classification, as well as the allocation determined by the CNN (1) are provided (110);For each combination of an image (2) and an assignment (3), a localised relevance assessment (2a) of the image (2) is determined by applying a relevance assessment function (4), indicating which parts of the image (2a) have contributed to the assignment (3);A further classifier (5) is trained (130), from an image (2) and a relevance assessment (2a) to determine a reconstruction (3') of the classification (3) to which this relevance assessment refers (2a);(iii) a measure of quality (4a) for the relevance function (4) is determined on the basis of the consistency between the reconstructions (3') and the actual allocations (3).