Course contentsShow
← Machine Learning and Deep Learning
Lesson 1477 of 3,53833. Generative Adversarial NetworksFree lesson

Mode Collapse Problem

When the generator produces limited variety by collapsing to a few modes of the data distribution.

Mode Collapse Problem

What you'll learn: You'll understand what mode collapse is in GANs and why it limits the diversity of generated samples.

What Is Mode Collapse?

Mode collapse occurs when a GAN's generator learns to produce only a small subset of the possible data variations, rather than capturing the full diversity of the training distribution. Instead of generating many different kinds of realistic samples, the generator "collapses" to producing just a few successful outputs over and over.

Think of it like a student who discovers that the teacher loves one particular essay style. Instead of learning to write about many topics, the student just keeps submitting slight variations of that one winning essay. They get good grades, but they're not actually learning the full curriculum.

Why Does This Happen?

Remember that the generator and discriminator are playing a minimax game. The generator's goal is to fool the discriminator—not necessarily to capture all modes (varieties) of the real data. If the generator finds a few samples that consistently trick the discriminator, it may exploit those winners and ignore other parts of the data distribution.

For example, if you're training a GAN on handwritten digits (0-9), mode collapse might cause it to generate only "3s" and "7s" because those particular outputs reliably fool the discriminator at that moment in training.

The Problem This Creates

Mode collapse severely limits a GAN's usefulness. You want your generator to produce diverse, varied outputs that span the entire real data distribution—not just memorize a handful of safe examples. This lack of variety means the model hasn't truly learned the underlying patterns.

Key Takeaway: Mode collapse is when the generator produces limited variety by repeatedly generating a few successful samples instead of capturing the full diversity of the training data distribution—a fundamental challenge in GAN training.