Generative adversarial network
Also: GAN
The generator maps random noise to samples and is rewarded for fooling the discriminator, a binary classifier of real versus fake. This minimax game can yield very sharp samples in a single forward pass, but training is unstable and prone to mode collapse.
min_G max_D E[log D(x)] + E[log(1 − D(G(z)))]
Example
Early on, D might output 0.9 for a real image and 0.2 for a generated one; the generator’s update pushes D(G(z)) toward 1. If it settles on a few outputs that always fool D, that is mode collapse.