Manually wiring neural networks to detect objects would be extremely difficult because: (1) deciding what connects to what and which weights to use is complex, and (2) intermediate feature layers must be good not just for the target task but for many tasks, making hand-wiring prohibitively difficult