GAN Optimal Discriminator JS Identity
An ideal discriminator's objective is exactly a shifted Jensen-Shannon divergence, with perfect separation as a boundary case.
Free theory articleAn ideal discriminator's objective is exactly a shifted Jensen-Shannon divergence, with perfect separation as a boundary case.
Concept & interpretation
The Generative Adversarial Network (GAN) objective connects adversarial classification to statistical divergence. This module explores the exact relationship between the optimal discriminator's objective value and the Jensen-Shannon (JS) divergence.
Interactive Experiment
You can interactively explore how changes in the generated distribution affect the discriminator's objective and its connection to the Jensen-Shannon divergence using theinteractive model.
- Select the Distributions control to explore the "Shifted", "Disjoint" or "Same" boundary cases.
- Adjust the Overlap mixing control to continuously interpolate between the base distribution and a perfectly overlapping distribution.
- Observe the GAN divergence identity residual, confirming it remains exactly zero across all configurations.