Mode collapse in GANs and how to fix it

WHAT THIS TESTS This evaluates depth on adversarial training dynamics, specifically the most notorious GAN failure and whether you know targeted, named fixes rather than vague tuning.

A GOOD ANSWER COVERS Mode collapse occurs when the generator produces only a small number of distinct outputs, or many inputs collapse to nearly identical images, instead of covering the full diversity of the real data distribution. It happens because the generator discovers a few outputs that reliably fool the current discriminator, so it keeps producing them; the discriminator adapts, the generator jumps to another small set, and the two can oscillate without ever covering all modes. Strategies to address it include using the Wasserstein GAN loss, often with a gradient penalty, which provides smoother and more informative gradients and reduces the incentive to collapse. Minibatch discrimination lets the discriminator look at relationships across a batch, so a batch of near-identical samples is easily flagged as fake, penalizing low diversity. Unrolled GANs let the generator optimize against a few lookahead steps of the discriminator, stabilizing the game. Feature matching changes the generator objective to match statistics of intermediate discriminator features rather than directly fooling the output.

COMMON WRONG ANSWERS Confusing mode collapse with vanishing gradients or general non-convergence. Believing more epochs or a larger generator alone cures it. Saying it means overfitting. Citing dropout as the primary remedy.

LIKELY FOLLOW-UPS How does Wasserstein distance differ from Jensen-Shannon divergence here. Why does minibatch discrimination specifically discourage collapse. How would you even detect mode collapse quantitatively.

ONE CONCRETE EXAMPLE Train a GAN on the ten MNIST digit classes. Under mode collapse the generator might output only convincing 1s and 7s, because those reliably fool the discriminator, and never produce 0s, 4s, or 8s, so generated samples lack diversity despite looking individually sharp. Adding minibatch discrimination quickly punishes batches that are all 1s, pushing the generator to spread across all ten digit modes.