Generative Adversarial Network
Two networks trained against each other: a generator that fakes data and a discriminator that detects fakes.
Definition
A generative adversarial network (GAN) pits two networks against each other. The generator produces synthetic samples from noise; the discriminator tries to tell real data from generated data. Their competition drives the generator toward realistic output.
The two networks must improve in step: if the discriminator becomes too strong too quickly, the generator receives little useful gradient and stalls. Managing this balance, through architecture and loss choices, is what made stable GAN training a research field in itself.
Evaluating generative quality is itself hard, since there is no single correct output; metrics attempt to measure realism and diversity together but each has blind spots. This difficulty, along with training instability, is why the field has partly shifted toward diffusion models with more stable objectives. The adversarial idea nonetheless endures wherever one network is trained to fool another, including in robustness research.
At the ideal equilibrium the generator's samples are indistinguishable from real data and the discriminator can do no better than chance.
Challenges
- Training instability from the adversarial dynamics.
- Mode collapse: the generator produces limited variety.
- Difficulty measuring sample quality objectively.
Why it matters
GANs produced early breakthroughs in realistic image synthesis and data augmentation. Though diffusion models now lead many generative tasks, the adversarial idea remains influential.
Fusion connection
Generative models can augment scarce simulation data with plausible synthetic samples, though for Kronos any generated case is treated as a hypothesis to be checked against physics, never as evidence.