Jul 23, 2016

link to the paper.

http://arxiv.org/abs/1607.02533

Their method more maximizing the cross entropy yields very peculiar results if you look at the images as the hyper-parameter iterations go up. It seems as if what we are seeing is some convolution with some peculiar form of noise. I wonder what the structure across these adversarial images looks like and whether it could provide some insight into certain patterns in cov net learning?