Mode collapse in GAN training

WHAT THIS TESTS: Whether you understand why a GAN can produce realistic but non-diverse samples and which interventions actually address the cause.

A GOOD ANSWER COVERS: Mode collapse is the failure where the generator captures only a few modes of the data distribution, producing little variety even though individual samples may look real. It arises from the adversarial dynamics: the generator is rewarded purely for fooling the current discriminator, so it can collapse onto a single output or a small cluster that the discriminator cannot currently reject. The discriminator then adapts, the generator jumps to a new collapsed point, and the two oscillate without ever covering the full distribution. Because each network optimizes against a moving opponent, there is no force explicitly encouraging the generator to spread mass across all modes. Mitigations add such a force or smooth the dynamics: minibatch discrimination lets the discriminator see diversity within a batch and penalize sameness; feature matching changes the generator objective to match intermediate discriminator statistics; unrolled GANs let the generator anticipate the discriminator's update; and the Wasserstein loss with gradient penalty provides smoother, more informative gradients that reduce the incentive to collapse.

COMMON WRONG ANSWERS: Calling it overfitting to the training set; blaming only vanishing gradients; claiming a larger generator alone fixes it; saying it means the discriminator is too weak when it can equally stem from the discriminator being too strong.

LIKELY FOLLOW-UPS: How do you detect mode collapse quantitatively? Why does Wasserstein distance help compared with the original Jensen-Shannon objective? What is the difference between full and partial mode collapse? How does mode collapse interact with FID?

ONE CONCRETE EXAMPLE: Train a GAN on a dataset of ten handwritten digit classes. After a while every random latent vector produces a convincing one. The samples pass the discriminator, the generator loss looks fine, yet the model has collapsed onto a single mode and ignores nine of ten digits. Adding minibatch discrimination so the discriminator penalizes batches that lack variety pushes the generator to cover the remaining digits.