Architecture
GAN (Generative Adversarial Network)
A generative model architecture consisting of a generator and discriminator that are trained adversarially — the generator creates samples while the discriminator distinguishes real from generated.
Explained at five levels
Level 1
Two AIs that play a game — one makes fake pictures and the other tries to catch the fakes, and they both get better at their jobs.
Level 2
An AI system where two networks compete: one creates fake content (like images) and the other judges if it's real or fake. This competition makes both better.
Level 3
A generative model architecture consisting of a generator and discriminator that are trained adversarially — the generator creates samples while the discriminator distinguishes real from generated.
Level 4
A min-max game between a generator G and discriminator D, where G learns to map noise to realistic samples and D learns to distinguish real from synthetic data, converging at a Nash equilibrium.
Level 5
A two-player minimax optimization: min_G max_D E[log D(x)] + E[log(1-D(G(z)))] — subject to mode collapse, training instability, and evaluation challenges, with variants (WGAN, StyleGAN, BigGAN) addressing these via architectural and objective modifications.
Definitions are educational summaries. Terminology can vary by source and context.
Sources
- NIST Trustworthy and Responsible AI Resource Center: Glossary — National Institute of Standards and Technology. Accessed 2026-07-20.