[bearzoo]

They are doing nothing but starting with random noise, and then learning a representation of an image that will maximize the probability in the output layer (by suggesting to the network that this noise should have actually been recognized as a banana or what have you) and back propagating changes into the input layer. Essentially, this has been happening since 2003 in the natural language processing world where we learn 'distributed representations' of words by starting with random representations of words, and learning them by context by back propagating changes into the input layer. Very cool though.