Control of the behaviour of an image classifier over genetically engineered disorders

摘要

Procedure (100) for determining the proportion of images on which an image classifier (1) with a folding neural network, CNN (1a), assigns a given image (2) to a given class (3a-3c) of a given classification (3) is based, using the steps:Several candidate malfunctions (4a-4c) are provided (110) which have the same dimensionality as an intermediate product (4) of CNN (1a), which has been produced by at least one folding layer (1b) of CNN (1a) from the specified image (2) and still has a spatial correspondence to the given image (2);(a) the candidate's malfunctions (4a-4c) are evaluated with a quality function (5) which measures whether and to what extent the assignment of the given image (2) to the given class (3a-3c) is changed; when the candidate malfunction (4a-4c) produces a transformation (4*) of the intermediate product (4*) and this transformation (4*) is processed (4*) instead of the intermediate product (4) by the image classifier (1);(c) the candidate malfunctions (4a-4c) are optimised with a genetic algorithm which includes at least changes in existing candidate malfunctions (4a-4c) and the fusion of at least two candidate malfunctions (4a-4c) into a new candidate malfunction (4a-4c) (130); where the optimisation objective includes at least a good evaluation (5a) of the candidate malfunctions (4a-4c) by the quality function (5).