Global Explanations for Image Classifiers (Student Abstract)

We hypothesize that deep network classifications of complex scenes can be explained using sets of relevant objects. We employ beam search and singular value decomposition to generate local and global explanations that summarize the deep model's interpretation of a class.