Faker and the critic.
A GAN pits two neural networks against each other to generate new, realistic data from scratch.
Imagine a forger trying to paint a fake Van Gogh, and a detective trying to spot the fake. The forger gets better by learning from the detective's mistakes, until the detective can't tell the difference. That's a GAN: two AIs playing a game where one creates and the other critiques.
GANs power deepfakes, realistic video game graphics, and medical image enhancement. They let machines create convincing new content without being explicitly programmed for every detail.
Many think the generator and discriminator are trained separately, but they actually train together in a zero-sum game. Another misconception: GANs always produce perfect results, but they often suffer from mode collapse or training instability.
A generative adversarial network consists of a generator network that produces synthetic data and a discriminator network that distinguishes real from fake. They are trained simultaneously via adversarial loss, where the generator aims to fool the discriminator and the discriminator aims to correctly classify. This minimax game drives the generator to produce data indistinguishable from the true distribution.