Generator: take latent multivar normal as input layer
Output of generator is of same dimension as input iamges
Generative NN: need probability distribution on input layer
Generator typically deconvolutional
Downside of gan. only discriminates between real and fake.
Discriminator: use normal CNN for deep