Grounding deep neural network predictions of human categorization behavior in understandable functional features: The case of face identity.