How to trick deep learning algorithms into doing new things