What is GAN mode collapse, its causes, and two mitigations?

WHAT THIS TESTS: This question probes whether you understand GAN training as a dynamic system rather than a static minimax formula. Interviewers want to see that you know mode collapse is not just a surface symptom but a structural consequence of adversarial imbalance and divergence choice. They also check if you can name concrete GAN-specific fixes rather than generic deep learning tricks.

A GOOD ANSWER COVERS: First, define mode collapse precisely: the generator learns to produce a narrow subset of the true data distribution, ignoring other modes. In the original GAN paper this was called the Helvetica scenario. Second, explain the underlying causes. One cause is discriminator-generator imbalance: if the discriminator learns too slowly, the generator finds a small set of outputs that consistently fool it and stops exploring. Another cause is the Jensen-Shannon divergence used in standard GAN loss, which is too lenient when the generator outputs identical or near-identical samples. A third cause is oscillatory cycling: the generator and discriminator engage in a rock-paper-scissors dynamic where the generator collapses to one mode, then another, without ever covering the full distribution. Third, describe two distinct mitigation techniques. Option one is Wasserstein GAN, which replaces JS divergence with Earth Mover's distance to provide stable gradients and a more meaningful loss surface. Option two is mini-batch discrimination, where the discriminator evaluates diversity across an entire batch rather than single samples, directly penalizing homogeneity. Other valid distinct techniques include unrolled GANs, which optimize the generator against future discriminator states to discourage short-term exploitation, and the two time-scale update rule, which uses different learning rates for generator and discriminator to stabilize convergence.

COMMON WRONG ANSWERS: Do not describe mode collapse as simple memorization or overfitting. Overfitting means peaked distributions around training points, while mode collapse is a collapse to a few modes that may not even be training points. Do not blame it solely on insufficient data unless you explain why the adversarial dynamic specifically exploits that scarcity. Do not offer generic fixes like dropout or data augmentation without explaining how they interact with the adversarial game.

LIKELY FOLLOW-UPS: An interviewer might ask why Wasserstein GAN helps specifically with mode collapse versus just training stability. They might ask you to compare mini-batch discrimination to batch normalization, or to sketch how unrolled GANs change the computational graph. They might also ask how mode collapse manifests in modern diffusion models or LLM fine-tuning.

ONE CONCRETE EXAMPLE: Imagine a GAN trained on MNIST. A collapsed generator might only output digit 7 in slightly different styles, because it found a narrow sweet spot that fools the current discriminator. Even though the dataset contains ten digits, the generator has collapsed to one mode. Using mini-batch discrimination, the discriminator would see that the entire batch contains only sevens and reject it, forcing the generator to explore other digits.