a generative adversarial network is a pair of networks that combine to form a generate system. one of the networks, the generator, maps values from to in order to produce samples from the distribution . a typical scheme samples from a unit Gaussian of moderate dimension and then passes it through a deep network to obtain .
GANs are a kind of implicit model in the sense that samples can be generated but their probabilities are not readily available; the challenge in implicit modeling is to design a loss function that makes it possible to train the model using samples from the distribution, rather than maximizing the likelihood assigned to training examples from the data set.
both the generator and the discriminator are trained simultaneously, with the generator learning to fool the discriminator and the discriminator learning to accurately separate real from fake data. the competition between generator and discriminator can be described in the language of game theory. the idea is that in the equilibrium state of the game, the generator should reproduce the training distribution perfectly, such that the discriminator cannot perform better than random guessing. GANs have worked particularly well for image generation tasks. for example, GANs can create photorealistic, high-resolution images of people who have never existed.