Research map / Learning theory and algorithms / Submodular and game algorithms
On Characterizing GAN Convergence Through Proximal Duality Gap
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use proximal duality gap instead of duality gap to monitor training | WGAN · MNIST, CIFAR-10, CELEB-A | Pearson correlation with IS: MNIST: -0.752 -> -0.892; CIFAR-10: -0.368 -> -0.854; CELEB-A: -0.213 -> -0.524 | Helped |
| Use proximal duality gap instead of duality gap to monitor training | WGAN · MNIST, CIFAR-10, CELEB-A | Pearson correlation with FID: MNIST: 0.845 -> 0.957; CIFAR-10: 0.213 -> 0.738; CELEB-A: 0.263 -> 0.699 | Helped |
| Increase λ beyond the small-value range (ablation) | Proximal DG for converged WGAN · CELEB-A, CIFAR-10, MNIST | Proximal duality gap: DGλ increases sharply past 10¹ for CELEB-A and CIFAR-10, and past 10⁴ for MNIST | Hurt |
| Use λ in the range [10⁻²,1] rather than larger values (ablation) | Proximal DG for converged WGAN · MNIST, CIFAR-10, CELEB-A | Proximal duality gap: λ in [10⁻²,1] | No clear effect |
| Increase the number of discriminator updates per generator update (ablation) | WGAN · MNIST | Proximal duality gap: N in the range -10 to 10; DGλ increases with N | Hurt |
| Increase the number of discriminator updates per generator update (ablation) | WGAN · MNIST | Generated sample quality: N in the range -10 to 10 | Hurt |
Closest papers by meaning
- Do GANs always have Nash equilibria? (ICML 2020 · similarity 0.93)
- Finding Mixed Nash Equilibria of Generative Adversarial Networks (ICML 2019 · similarity 0.92)
- Many Paths to Equilibrium: GANs Do Not Need to Decrease a Divergence At Every Step (ICLR 2018 · similarity 0.91)
- A Convex Duality Framework for GANs (NeurIPS 2018 · similarity 0.90)
- A Domain Agnostic Measure for Monitoring and Evaluating GANs (NeurIPS 2019 · similarity 0.90)
- DO-GAN: A Double Oracle Framework for Generative Adversarial Networks (CVPR 2022 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.