GANs: The Adversarial Game

A GAN replaces reconstruction loss with a game. The generator maps random noise to fake images; the discriminator — an ordinary classifier — tries to tell fake from real. Both train against each other:

• The discriminator improves → the generator must make better fakes to fool it
• The generator improves → the discriminator must get sharper to keep up

At equilibrium, fakes are indistinguishable from real. No explicit distribution, no reconstruction target — quality emerges purely from competition.