Two roles
A generator produces samples. A discriminator assigns scores intended to distinguish generated samples from examples in the data. Their training objectives interact.
Feedback shapes learning
A generator can receive a learning signal through a discriminator. Follow a single generated scalar through a fixed toy discriminator below: change either the sample or the discriminator threshold and see its score change.
INTERACTIVE EXPLORATION
Follow a generated sample to the discriminator
Choose a generated scalar and the threshold of a fixed toy discriminator. Its score is sample minus threshold; sigmoid turns that score into a number between zero and one.
A toy comparison
Real GAN training alternates updates to two networks. This illustration performs neither update; one score cannot establish sample quality or coverage of the data distribution.
Check your understanding
Does one high score prove a GAN has learned the whole data distribution?
No. A single score or sample is insufficient to establish coverage or quality.