Generative Adversarial Networks

Why this one

Read this after Auto-Encoding Variational Bayes, and before Wasserstein GAN. The lasting idea is not that GANs make pretty images, or that every generator should be trained this way. It is that you can replace an explicit likelihood with a learned test: if a discriminator can tell your samples from real data, the generator has a usable gradient. People often remember the drama, mode collapse, unstable training, cherry-picked samples, and miss the clean move underneath. The paper turns modeling into a game, which is why it still shows up anywhere we train systems against critics, rewards, evaluators, or learned losses. It is worth your evening because it gives you a sharp mental model for a whole family of modern tricks: when the metric you want is hard to write down, learn an opponent that exposes the gap.