Slide the damage up to corrupt more squares of a clean pattern. A denoising autoencoder breaks its input on purpose, then learns to repair it.
A denoising autoencoder damages its own input on purpose before feeding it in. This sounds like sabotage. It is the point.
