Using additional information to supervise learning in the compressed representation space rather than the raw observation space.