Adversarial Training: Theory and Applications

This lecture covers the practical implementation of adversarial training using stochastic subgradient descent, the application of adversarial training for better interpretability in retinopathy classification, an introduction to Generative Adversarial Networks (GANs) for modeling complex distributions, the notion of distance between distributions including the Earth Mover's distance and the Wasserstein distance, and the theory and practice of enforcing 1-Lipschitz of the discriminator in GANs. The lecture also discusses the difficulties of GAN training, historical background on GANs, and the abstract minmax formulation in the context of GAN optimization problems.