Leveraging conditional generative models in a general explanation framework of classifier decisions

With the increase in use of machine learning classifiers several fields, providing human- understandable explanation their outputs has become an imperative. It is essential to generate trust for day-to-day tasks, especially sensible domains as medical imaging. Although many works have addressed this problem by generating visual maps, they often provide noisy and inaccurate results forcing heuristic regularization unrelated classifier question. In paper, we propose a general perspective overcoming these limitations. We show that can be produced difference between two generated images obtained via specific conditional generative models. Both models are trained using explain database enforce following properties: (i) All first generator classified similarly input image, whereas second generator’s oppositely. (ii) belong distribution real images. (iii) The distances image corresponding minimal so elements only reveals relevant information studied classifier. Using symmetrical cyclic constraints, present different approximations implementations formulation. Experimentally, demonstrate significant improvements with respect state-of-the-art on three public data sets. particular, localization regions influencing consistent human annotations.