03·GAN: adversarial training, generator vs discriminator

Introduction

A Generative Adversarial Network (Goodfellow et al. 2014) is two models trained in a zero-sum game: the generator G(z) maps noise z ~ p(z) to samples x̂; the discriminator D(x) classifies x as "real" vs "fake". They train alternately: D maximises E[log D(x_real)] + E[log(1 − D(G(z)))], G tries to fool D. This lesson breaks down the minimax objective, the global optimum (D* = p_data/(p_data+p_g)), Jensen–Shannon vs KL, Goodfellow’s "non-saturating loss", and the mechanics of alternating training plus how G "learns through gradients flowing back via D".