Definition
A generative adversarial network (GAN) is a generative model, introduced by Ian Goodfellow and colleagues in 2014, built from two neural networks locked in competition. The model learns to produce new samples that share the distribution of its training data.
Generator vs. Discriminator
The generator transforms random noise vectors, sampled from a latent space, into candidate outputs such as images or audio. The discriminator is trained to tell the generator's fakes apart from genuine training examples. The two are trained adversarially: the generator is not optimized toward a fixed target but to fool the discriminator, while the discriminator sharpens its ability to catch fakes. As training proceeds, both improve until the generator's output becomes hard to distinguish from real data.
Strengths and Pitfalls
GANs are known for producing sharp, high-fidelity samples, which made them dominant for image synthesis. They are also notoriously tricky to train: the adversarial balance can collapse (mode collapse) or fail to converge. Running a GAN on local hardware demands careful tuning, but it keeps both the model and any generated data fully under your control.
GANs share the generative goal with the variational autoencoder, and both draw their inputs from a model's latent space.
In Simple Terms
A generative adversarial network (GAN) is a generative model, introduced by Ian Goodfellow and colleagues in 2014, built from two neural networks locked in competition.…