Generative Adversarial Nets
A generator and a discriminator play a minimax game: one produces samples, the other tells real from fake, and both improve. At the equilibrium the generator matches the data distribution, and the paper proves this for the idealised game.
Key ideas
- A learned discriminator is a trainable loss function
- The optimal discriminator turns the game into minimising a divergence
- Training dynamics, not the objective, are the hard part
Why read it now
Adversarial training reframed generation as a game and introduced instabilities that a decade of papers then tried to tame. It is short and worth reading for the proof alone.
Question to keep in mind
Why does the theoretical optimum say nothing about whether alternating gradient steps will reach it?