Explaining image classifiers by removing input features using generative models

Chirag Agarwal, Anh Nguyen

Perturbation-based explanation methods often measure the contribution of an input feature to an image classifier's outputs by heuristically removing it via e.g. blurring, adding no…