Generative Adversarial Nets

TL;DR

Trains two networks against each other, a generator that fabricates samples and a discriminator that tries to spot fakes, until the fakes are convincing.

Why it matters

GANs were the breakthrough that made sharp, realistic image synthesis possible and dominated generative modeling for years. The adversarial-training idea reappears across the field.

Key ideas

• A minimax game: the generator fools the discriminator; the discriminator resists.
• No explicit likelihood is needed, the discriminator provides the training signal.
• At equilibrium the generator's samples are indistinguishable from real data.
• Notoriously tricky to train (instability, mode collapse).

Related concepts