Generative Models: GANs

This lecture covers Generative Adversarial Networks (GANs), which model probability distributions over random variables. GANs consist of a generator and a discriminator playing a two-player game. The generator aims to produce realistic samples, while the discriminator tries to distinguish between real and fake samples. The lecture explains the optimization process of GANs as a minmax game and introduces the concepts of Nash equilibrium and Differential Nash Equilibrium. It also discusses the challenges of using Jensen-Shannon divergence in GANs and presents the Wasserstein distance as an alternative. Additionally, it explores Conditional GANs (CGANs) for generating samples conditioned on additional information like class labels. The lecture concludes with a discussion on Diffusion Models, an alternative to GANs, which progressively add noise to input data to generate samples.