Generative Adversarial Network
A GAN trains a generator and a discriminator against each other: the generator maps noise to fake samples, the discriminator learns real vs fake, and the equilibrium is a generator whose samples match the data distribution.
Definition
Goodfellow et al. (2014) set up a minimax game: min_G max_D V(D, G) = E_x[log D(x)] + E_z[log(1 − D(G(z)))]. At optimum, D cannot distinguish, and G samples from the data distribution. Training is notoriously unstable (mode collapse, vanishing gradients), which later papers (DCGAN, WGAN, StyleGAN) attacked with architecture and better divergences.
Why it matters
GAN is the first widely successful deep generative model that does not require an explicit likelihood. The adversarial idea leaked into domain adaptation, robust learning, and synthetic data. If an “AI” glossary has RAG but no GAN, the catalog is incomplete.
Case
DCGAN made bedroom images that fooled casual viewers in 2015. The same recipe on a tick dataset will mode-collapse to one volatility cluster unless you constrain the critic and audit sample diversity — likelihood-free does not mean assumption-free.
How to read it
Do not trust a single pretty sample. Report coverage (did modes drop out?), a proper distance (FID, MMD), and whether the critic still has gradient. A collapsed GAN looks sharp and is scientifically empty.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- Generative Adversarial NetsGoodfellow et al. · 2014 · NeurIPSThe original GAN minimax formulation.
- Unsupervised Representation Learning with Deep Convolutional Generative Adversarial NetworksRadford, Metz, Chintala · 2016 · ICLRDCGAN — the first widely reproduced image GAN recipe.
- Wasserstein GANArjovsky, Chintala, Bottou · 2017 · ICMLEarth-mover distance to stabilize training.