The History of AI

Generative adversarial networks

Era 4 · Deep learning and foundation models (2012–2021) › Generative image models

Lineage
Neurons and networks
Kind
Publication

Who

Université de Montréal

What happened

Ian Goodfellow and seven colleagues at the University of Montreal, Yoshua Bengio among them, propose generative adversarial networks (GANs): two networks trained in competition, one generating fake samples and the other trying to tell them apart from real ones.

Why it mattered

A generator that learns without anyone telling it what a good image should look like, simply by trying to fool an opponent. For several years GANs are the method of choice for generating realistic images.

The full story

The generative models of the day had to compute the probability of each data point, or approximate it with costly methods. Goodfellow and colleagues proposed skipping that calculation with a game between two networks.

The generator takes random noise and produces a sample, such as an image. The discriminator receives images, some real and some from the generator, and estimates which ones are genuine. The generator is trained to make the discriminator wrong, and the discriminator to be right. The authors framed it as a two-player zero-sum game and proved that, in theory, it has a unique solution: the generator reproduces the data distribution exactly and the discriminator can do nothing better than guess at random. Since both pieces are neural networks, everything is trained with backpropagation.

In practice the game was hard to balance. The generator often suffered mode collapse: it learned to produce only a few kinds of image and forgot the rest.

Even so, within a few years GANs went from blurry digits to faces of people who do not exist. The faces generated by Nvidia's StyleGAN in 2019 were compared to deepfakes and raised concerns about fake profiles and image manipulation.

A "noise z" box leads to the "generator," which sends a "fake" arrow to the "discriminator." Another box, "real data," sends a "real" arrow to the same discriminator, which answers "real or fake?" A dashed line runs from that answer back to the generator.

The generator turns noise into fake samples; the discriminator receives fake and real samples and says which is which. The generator learns from the discriminator's successes.

Under the hood

Goodfellow and coauthors set it up as a game between two networks: min_G max_D E[log D(x)] + E[log(1 − D(G(z)))]. The discriminator D tries to output 1 for real data x and 0 for samples G(z) made from noise z; the generator G tries to do the opposite.

The paper proves that if both networks had unlimited capacity, equilibrium would be reached when the generator reproduces the data distribution exactly and the discriminator outputs 1/2 for everything, because it can no longer tell the difference.

Sources

Original source

How to cite

“Generative adversarial networks.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/generative-adversarial-networks

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline