Two networks playing a forgery game taught computers to create

orig. “Generative Adversarial Nets” · Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio

Generative Models Intermediate 4 min read Written, reviewed by Marginalia Editorial

In the margin

Heads up

Long before modern image generators, this paper had two networks compete, and the result was machines that could make new pictures from scratch.

What's going on

The setup is a game between two networks. One, the generator, tries to make fake images. The other, the discriminator, tries to tell fakes from real photos. As each gets better, the generator is pushed to make more convincing images, until its output starts to look real. This was one of the first methods that could generate believable new images instead of just sorting existing ones.

Why it matters

This kicked off the wave of AI that creates rather than classifies. The adversarial idea showed up in art tools, photo editing, and data generation for years. Newer methods like diffusion have taken over for images, but the question it asked, how do you teach a machine to make something new, is still central.

Source

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio, University of Montreal

View on arXiv ↗ Read the PDF ↗

Still fuzzy on a sentence?

Paste it and we'll explain it even more simply.

Test your understanding

Member notes Sign in ↗

ME Marginalia Editorial TEAM

The editorial team picked this paper because it is worth your time, and an AI drafted this note. We check each one and rewrite it when the draft misses the mark.