Generative Adversarial Networks: Adversarial Objectives and Training Dynamics

Variational autoencoders optimise a single reconstruction-plus-regularisation objective. Generative adversarial networks (GANs) take a different route: a generator $G$ learns to synthesise data that fools a discriminator $D$, while $D$ learns to separate real from fake. The result is a minimax game whose training dynamics — alternating updates, moving targets, mode collapse — differ sharply from ordinary supervised learning. For a quant desk exploring synthetic scenario generation or data augmentation, understanding those dynamics is the difference between interpreting a pretty sample plot and knowing whether the generator has actually covered the data distribution.