sourced story
10 June 2014Primary source · 2 sourcesWell documented

Goodfellow invents Generative Adversarial Networks

Two neural networks compete, a counterfeiter and a detective, and both get better

On the timeline · around 10 June 2014 · The Statistical and Machine Learning TurnThe Statistical and Machine Learning TurnTransformers and the Generative AI WaveGoodfellow invents Generative Adversarial Networks200820102012201420162018

Quick facts

Lead author
Ian Goodfellow
Institution
University of Montreal
Co-authors
Includes Yoshua Bengio

What happened

Ian Goodfellow, then a PhD student at the University of Montreal, conceived Generative Adversarial Networks after a late-night conversation with friends at a Montreal bar, Les 3 Brasseurs, about whether a computer could generate photos on its own; he went home and coded a working version that night. The paper, co-authored with Jean Pouget-Abadie, Mehdi Mirza, and others including Yoshua Bengio, described a framework pitting two neural networks against each other: a generator that produces fake data, and a discriminator that tries to tell the fakes from real training examples. Goodfellow compared it to a counterfeiter trying to produce undetectable fake currency against a police force trying to catch the fakes; as training proceeds, both networks improve until the generator's output is statistically indistinguishable from real data. The framework required no explicit description of the data's structure, only the adversarial contest between the two networks.

Why it matters

GANs gave AI its first practical general method for generating convincing new images, and their adversarial training idea influenced deepfake technology, data augmentation, and later generative image tools before diffusion models largely superseded them.

How we know

The original paper describes the generator-discriminator framework and minimax game directly; MIT Technology Review's profile of Goodfellow independently corroborates the bar conversation and same-night coding session as Goodfellow's own account of the invention.

Sources

See something wrong? . Corrections with a source get fixed fastest.

Part of a timelineArtificial Intelligence30 events · From a wartime theory of neurons to machines that write, paint, and fold proteinsView all →