Definition
In GANs, mode collapse occurs when the generator learns to exploit weaknesses in the discriminator by producing a narrow range of plausible samples, ignoring other modes of the data distribution. This results in a lack of diversity in generated content, such as generating only one specific digit in MNIST despite training on all digits, severely limiting the utility of the generative model.
Summary
Mode collapse is a failure mode in Generative Adversarial Networks where the generator produces limited varieties of outputs.
Key Concepts
- GAN Stability
- Distribution Diversity
- Generator Failure
- Discriminator Feedback
Use Cases
- Diagnosing GAN training instability
- Improving image generation diversity
- Analyzing latent space coverage