Generative Adversarial Network

A generative model where a generator fakes data and a discriminator tries to catch the fakes, trained against each other.

Generative adversarial networks (GANs) were introduced by Ian Goodfellow and colleagues in 2014 and are a class of generative model — a model that learns to produce new data rather than just classify it. A GAN has two networks: a generator that turns random noise into samples, and a discriminator that judges whether a sample is real or generator-made. The two train in turns, with the generator trying to fool the discriminator and the discriminator trying to catch it; ideally the generator ends up learning the true data distribution. GANs produce an image in a single forward pass, so they're fast, but training is unstable and prone to mode collapse (generating only a few kinds of samples), and in recent years diffusion models have mostly replaced them for image and video generation. The ‘discriminator as a judge’ idea is still common in robotics: both Generative Adversarial Imitation Learning (GAIL) and Adversarial Motion Priors (AMP) use a discriminator to judge how closely a robot's motion resembles demonstration data, then feed that judgment to reinforcement learning as a reward.

ExampleTraining a simulated humanoid with AMP: the discriminator compares motion clips produced by the policy to motion-capture data, giving higher reward the more human-like they look, which teaches the character natural walking and running gaits.

Also called

GAN

Related

Generative Model · Generative Adversarial Imitation Learning · Adversarial Motion Priors · Diffusion Model · Variational Autoencoder · Entropy Collapse / Mode Collapse

Sources

Generative Adversarial Networks (Goodfellow et al., arXiv 1406.2661)
AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control (arXiv 2104.02180)