This is a real interesting talk by Ian Goodfellow. He gives an overview applications of a GAN and some underlying theory. He makes a reference to a game theory approach. This is because in a GAN model you have two players essentially. A discriminator and a generator that have their own policies. A discriminator attempts to find counterfeits that are produced by the generator, while the generator attempts to fool the discriminator. Through game theory, we are able to find out because of the strategies that each player takes, the generator will always win in the end.
Ian brings up an interesting point that the discriminator should be a little stronger than the generator. He states it’s because the discriminator dictates what the generator learns. Furthermore, he mentions that the model has issues when it comes to counting things in pictures.
Possible uses of GAN:
What Ian meant by having multiple correct answers is that your lists of possible outcomes allowed from your model is more than one. For example, if you are trying to predict cats, dogs, and human. We might not be able to distinguish that a given input is either one, but we are able to say that it is possible that the input is one of the things in the list. Therefore, the input is accepted.
There are tools like word2vec that are able to turn words into a vector in an embedded space. You are able to do the same thing with GANs but with pictures. For example, given a picture of a man with glasses, subtract a picture of a man, and a picture of a woman with glasses is returned.
http://on-demand.gputechconf.com/gtc/2017/video/s7502-ian-goodfellow-generative-adversarial-networks.mp4