Turning a variational bound into a trainable generator-versus-critic game

This note continues a deep dive into Prof. Prathosh AP's lectures on the mathematical foundations of generative AI. It recalls that the previous week ended with a sample-based lower bound on the f-divergence, which still required maximizing over an unrestricted family of functions. The author explains that a neural network can serve as the scoring function, turning the problem into a two-player game. The analogy offered is a fraud-detection team trying to separate real transactions from a fraudster's fakes. The generator is trained to make its fakes indistinguishable from real data, while the critic only ever sees samples.