gan-mode-collapse-vs-vanishing-gradient

IN premise — entries/2026/06/21/wiki-Generative_adversarial_network-chunk-3.md

Created 2026-06-21T09:55:49+00:00

Mode collapse (generator produces limited modes) and vanishing gradient (generator cannot learn) are opposite GAN failure modes caused by the discriminator being too weak vs. too strong, respectively

Dependents

These beliefs depend on this one:

• OUT gan-critical-domain-deployment-responsibly-viable — GAN applications in critical domains (particle physics simulation, synthetic medical imaging) would be responsibly deployable — their practical benefits (accelerating expensive simulations, overcoming privacy barriers) would justify deployment.
• IN gan-game-theoretic-foundations-fragile-beyond-original — GAN game-theoretic foundations are fragile beyond the original formulation — equilibrium equivalence (minimax, maximin, Nash) holds only for the original game and not its variants, Nash equilibria are not guaranteed to exist in general (Farnia & Ozdaglar 2020), and the two dominant failure modes (mode collapse and vanishing gradients) represent opposed destabilizing forces that the equilibrium theory does not resolve.
• OUT gan-reliable-for-safety-critical-deployment — GANs can be reliably deployed in safety-critical domains (medical imaging, physics simulation) given their asymptotic consistency guarantees and demonstrated cross-domain applications.
• OUT gan-stable-convergence-guaranteed — GAN training reliably converges to the unique equilibrium where the generator matches the data distribution, given TTUR and asymptotic consistency guarantees.
• OUT gan-theory-practice-alignment — GAN theory and practice are aligned — the complete analytical characterization (optimal discriminator, JSD minimization, unique equilibrium) accurately predicts GAN behavior in real-world deployment across physics simulation, medical imaging, and creative applications.
• OUT generative-models-provide-unlimited-training-data — Generative models (GANs, VAEs, diffusion) would provide effectively unlimited synthetic training data — generating privacy-preserving medical images, augmenting scarce datasets, and enabling training without data collection barriers — fundamentally resolving the labeled-data bottleneck that constrains supervised learning.