They tweak the input to maximize the response of specific neurons somewhere in the middle of the network to figure out what those neurons actually "learned".