GAN architecture: generator and discriminator roles

WHAT THIS TESTS: Whether you can articulate the two-network adversarial structure and the opposing objectives that drive a GAN to learn a data distribution.

A GOOD ANSWER COVERS: A GAN has two networks trained simultaneously. The generator takes a random latent vector, usually sampled from a Gaussian or uniform prior, and transforms it into a sample in data space such as an image. It never observes real examples directly; it only learns through the discriminator's feedback. The discriminator is a binary classifier that takes a sample and outputs the probability that it came from the real dataset rather than the generator. Training is a minimax game: the discriminator updates to maximize its accuracy at separating real from fake, while the generator updates to maximize the probability that the discriminator labels its fakes as real. At the theoretical optimum the generator's distribution matches the real data distribution and the discriminator outputs one half everywhere because it cannot tell them apart.

COMMON WRONG ANSWERS: Saying the two networks cooperate or share a single loss; claiming the generator is trained with a pixel reconstruction or autoencoder loss; forgetting the latent noise input; describing the discriminator as a regressor of image quality rather than a real-versus-fake classifier.

LIKELY FOLLOW-UPS: Why is the non-saturating generator loss often used instead of the original minimax loss? What is mode collapse and how does it arise? How do you know when to stop training given there is no single converging loss? Why is GAN training notoriously unstable?

ONE CONCRETE EXAMPLE: Train a GAN on handwritten digits. The generator turns a 100-dimensional noise vector into a 28x28 image; early on it outputs noise and the discriminator easily flags it. As gradients flow back from the discriminator, the generator sharpens strokes until its digits pass as real. If it learns to produce only convincing eights and ignores other digits, that is mode collapse, signaling the adversarial balance has broken.