Generative Models

Generative Adversarial Networks

Goodfellow et al., 2014 — proposed training two networks against each other, a generator and a discriminator, as a way to learn to generate realistic data. Dominated image generation for most of the following decade.

"Generative Adversarial Networks" (Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, and colleagues, 2014) introduced the GAN — reportedly conceived during a bar conversation and implemented the same night, and the architecture that made realistic image generation possible for the first time.

What problem it solved

Before this paper, generative models — ones that produce new data rather than classify existing data — mostly relied on explicitly modeling a probability distribution over the data and sampling from it, an approach that scaled poorly to complex, high-dimensional data like natural images. There was no training setup that let a network implicitly learn to produce realistic samples without first solving the much harder problem of writing down an explicit density function for what "realistic" meant.

The key idea

Turn generation into a two-player game instead of a direct optimization problem. Train two networks simultaneously, with opposing goals:

  • The generator takes random noise as input and tries to produce outputs realistic enough to fool the discriminator.
  • The discriminator is an ordinary classifier, trained to tell the generator's fake outputs apart from real training examples.

Each network's training signal comes from the other: the generator improves by getting better at fooling an increasingly capable discriminator, and the discriminator improves by getting better at catching an increasingly capable generator. Neither network needs an explicit model of what "realistic" means — that entire problem is outsourced to the adversarial game itself, which the paper formalizes as a minimax objective with a Nash equilibrium as its target rather than a simple minimum.

Why it mattered

GANs became the dominant approach to high-fidelity image generation for roughly the next decade, producing results — photorealistic faces, style transfer, image-to-image translation — that earlier generative approaches couldn't match. The adversarial training setup itself proved fragile (mode collapse, unstable convergence, no single loss to monitor), which is a large part of why diffusion models eventually displaced GANs for state-of-the-art image generation — but the core idea, that a learned discriminator can substitute for a hand-designed loss function, outlived the specific architecture and reappears across generative modeling more broadly.

Authors: Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio (Université de Montréal)

Read the paper: arXiv:1406.2661

See also: Ian Goodfellow · Yoshua Bengio

Learn more: GAN · Generative Models

On this page