Generative Adversarial Nets
ReadTrains a generator against a discriminator as a two-player game instead of maximizing likelihood. Read it to understand adversarial training and why it is unstable.
My notes
Read it straight after the VAE paper, which was the right order: one model reasons about likelihood explicitly, the other refuses to and just asks "can you fool a critic?".
The Game
min_G max_D E_x[ log D(x) ] + E_z[ log(1 − D(G(z))) ]
Notes
- At the global optimum G reproduces the data distribution and D outputs ½ everywhere. Getting there is the hard part.
- Early in training log(1 − D(G(z))) saturates, so they recommend G maximize log D(G(z)) instead. Small practical note, huge effect.
- Mode collapse — G finds one output that fools D and repeats it — is mentioned almost in passing. It became the story of the next five years.
Relevance today: diffusion replaced GANs for images, but the "learned critic" idea lives on in reward models and LLM-as-judge setups.