Towards White Box Deep Learning

Maciej Satkiewicz

Deep neural networks learn fragile "shortcut" features, rendering them difficult to interpret (black box) and vulnerable to adversarial attacks. This paper proposes semantic featur…