Generative adversarial networks

The story is that Ian Goodfellow had the idea in a Montréal bar, argued about it, went home, and had it working by the morning. The paper posted on 10 June 2014 sets up two networks. A generator turns random noise into an image; a discriminator tries to tell generated images from real ones. Each is trained to beat the other, and at the equilibrium the generator's output is indistinguishable from the data. There is no explicit probability model, no Boltzmann machine to sample from, just a game.

The first results were blurry digits and faces. Within four years, with Alec Radford's convolutional design and NVIDIA's progressive growing, GANs produced photographs of people who did not exist at a resolution no one could fault, and the term deepfake entered the language. They also generated music, molecules and video, and their training was notoriously unstable.

Yann LeCun called adversarial training the most interesting idea in machine learning in ten years. It was displaced in image generation by diffusion models around 2021, which were easier to train, but GANs fixed the idea that generation was the next frontier, and they fixed the public expectation that machines could make convincing pictures.