Hierarchical Feature Learning

Deep networks learn simple features early and combine them into complex ones later.

How do deep networks typically learn features across their layers?