Research map / Image generation / GANs
A Domain Agnostic Measure for Monitoring and Evaluating GANs
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 the duality gap to monitor training rather than relying on generator and discriminator losses | GAN training · Mixture of Gaussians | convergence and instability detection: stable models: 0.04, 0.14, 0.03; unstable models: 13, 1.22, 12.09 | Helped |
| Use the minimax loss to assess generator sample quality | GAN generator · RING, SPIRAL, GRID | correlation with covered modes: -0.97 (RING), -0.93 (SPIRAL), -0.95 (GRID) | Helped |
| Use the minimax loss instead of Inception Score to detect mode invention and intra-mode collapse | GAN generator evaluation · CIFAR10 mode sensitivity | mode-change sensitivity | Helped |
| Use the minimax loss to detect visual sample distortions | GAN generator evaluation · CIFAR10 image distortions | visual quality sensitivity | No clear effect |
| Add regularization to the GAN training objective | GAN on a 2D submanifold Gaussian mixture embedded in 3D · 2D submanifold Gaussian mixture | duality gap progression and collapse: regularized version stable beyond 200K iterations; unregularized version collapses after 50K iterations | Helped |
| Evaluate with the duality gap and minimax metric instead of image-specific scores | GAN models · CIFAR10 model ranking | model ranking agreement: NS + SN rank 1; NS + SN + GP rank 2; NS rank 3; WGAN + GP rank 4 | No clear effect |
| Use duality gap and minimax scores on cosmology data instead of relying on FID | WGAN-GP · Cosmology | Pearson correlation with domain-specific scores: Dual gap: 0.53, 0.38, 0.71; FID: 0.40, 0.21, 0.34 | Helped |
| Use a mixture of previous and future training snapshots rather than only past snapshots to approximate the duality gap (ablation) | Snapshot-based duality-gap approximation · GAN training | approximation quality | Helped |
| Use only past training snapshots to approximate the duality gap (ablation) | Snapshot-based duality-gap approximation · GAN training | approximation quality | Hurt |
Closest papers by meaning
- Geometry Score: A Method For Comparing Generative Adversarial Networks (ICML 2018 · similarity 0.92)
- Quantitatively Evaluating GANs With Divergences Proposed for Training (ICLR 2018 · similarity 0.91)
- On Characterizing GAN Convergence Through Proximal Duality Gap (ICML 2021 · similarity 0.90)
- Are GANs Created Equal? A Large-Scale Study (NeurIPS 2018 · similarity 0.90)
- The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS 2024 · similarity 0.89)
- SMaRt: Improving GANs with Score Matching Regularity (ICML 2024 · similarity 0.88)
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.