Definition
A Generative Adversarial Network (GAN) is a generative machine-learning architecture, introduced by Ian Goodfellow and colleagues in 2014, in which two neural networks are trained simultaneously in a competitive, zero-sum game. One network learns to manufacture synthetic data; the other learns to tell that synthetic data apart from real examples. The competition drives both toward higher quality until the generated output becomes difficult to distinguish from genuine data.
Generator and discriminator
The generator takes a random noise vector (a sample from a latent space) and attempts to produce an output — typically an image — that looks authentic. The discriminator receives a mix of real training samples and generated fakes and tries to classify each as real or fake. As training proceeds, the generator improves at fooling the discriminator while the discriminator improves at catching fakes, an adversarial feedback loop that ideally converges toward a Nash equilibrium where the generator's output is statistically indistinguishable from real data.
Strengths and limitations
GANs produced some of the earliest photorealistic synthetic images and remain useful for image synthesis, super-resolution, and data augmentation. They are notoriously hard to train, however: failure modes include mode collapse (the generator produces limited variety) and unstable, non-converging training. For many generative tasks they have since been complemented or supplanted by diffusion models and transformer-based generators, though the adversarial training idea remains influential across machine learning.
GANs are one of several generative architectures relevant to anyone running self-hosted models on sovereign hardware. For related concepts, see our entries on the Variational Autoencoder (VAE) and the foundation model.
In Simple Terms
A Generative Adversarial Network (GAN) is a generative machine-learning architecture, introduced by Ian Goodfellow and colleagues in 2014, in which two neural networks are trained…