A generator that turns noise into samples and a discriminator that judges them real or fake, trained against each other in a minimax game. The tension is the point and also the difficulty: unstable training and mode collapse, where the generator serves one convincing output.
Card 1 of 259. Answer: A generator that turns noise into samples and a discriminator that judges them real or fake, trained against each other in a minimax game. The tension is the point and also the difficulty: unstable training and mode collapse, where the generator serves one convincing output.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
A generator and a discriminator. The generator G turns random noise z into synthetic samples; the discriminator D takes a sample and outputs the probability that it is real. Ian Goodfellow and colleagues introduced the setup in the 2014 NIPS paper "Generative Adversarial Nets", likening the pair to counterfeiters and police.
The two play a minimax game on one value function: min over G, max over D of E[log D(x)] + E[log(1 − D(G(z)))]. Training alternates a step that makes D better at telling real from fake with a step that changes G to fool it. With unlimited capacity, the optimum has the generator's distribution equal to the data's and D outputting 1/2 everywhere.
In practice that optimum is hard to reach. Each network's target moves as the other learns, so losses oscillate instead of falling, and the generator can collapse onto a few outputs that fool the current D. Since the losses say little about quality, practitioners inspect samples and track metrics such as Fréchet Inception Distance (FID).
