Research map / Image generation / GANs
On-Line Adaptative Curriculum Learning for 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 adaptive discriminator-mixture weights instead of fixed uniform weights | GAN training on CIFAR-10 (3 discriminators) · CIFAR-10 | mean best FID: 4.41 -> 3.66 | Helped |
| Use adaptive discriminator-mixture weights instead of GMAN | GAN training on CIFAR-10 (3 discriminators) · CIFAR-10 | mean best FID: 4.59 -> 3.66 | Helped |
| Use adaptive discriminator-mixture weights instead of fixed uniform weights | GAN training on CIFAR-10 (5 discriminators) · CIFAR-10 | mean best FID: 3.72 -> 3.34 | Helped |
| Use adaptive discriminator-mixture weights instead of GMAN | GAN training on CIFAR-10 (5 discriminators) · CIFAR-10 | mean best FID: 4.42 -> 3.34 | Helped |
| Use 5 discriminators rather than 3 (ablation) | acGAN · CIFAR-10 | mean best FID: 3.66 -> 3.34 | Helped |
| Use adaptive discriminator-mixture weights rather than a vanilla GAN | GAN training · 25 Gaussians | modes: 17 -> 25 | Helped |
| Use adaptive discriminator-mixture weights rather than uniform weights | GAN training · 25 Gaussians | modes: 20 -> 25 | Helped |
| Use 3 discriminators rather than a vanilla single-discriminator GAN (ablation) | GAN training initialized with a 2-mode generator · 8 Gaussians | recovered modes: Vanilla GAN could only retrieve 2 additional modes; acGAN recovered all (8) modes | Helped |
| Use sample-quality progress as the reward instead of GAN-loss progress (ablation) | acGAN reward function · Experiments across datasets | performance | Helped |
| Keep a moving average of discriminator rewards (ablation) | acGAN · GAN training | mixture stability | Helped |
| Add decaying Gaussian input noise to the discriminators (ablation) | GAN with 3 discriminators · 8 Gaussians | recovered modes: all 8 modes | Helped |
| Use Tanh rather than ReLU in the weak discriminator (ablation) | Soft-acGAN · 8 Gaussians | performance | Hurt |
Closest papers by meaning
- Mixed batches and symmetric discriminators for GAN training (ICML 2018 · similarity 0.92)
- MGAN: Training Generative Adversarial Nets with Multiple Generators (ICLR 2018 · similarity 0.91)
- Collaborative Sampling in Generative Adversarial Networks (AAAI 2020 · similarity 0.91)
- Training Generative Adversarial Networks with Limited Data (NeurIPS 2020 · similarity 0.91)
- Improving GAN Training with Probability Ratio Clipping and Sample Reweighting (NeurIPS 2020 · similarity 0.90)
- MCL-GAN: Generative Adversarial Networks with Multiple Specialized Discriminators (NeurIPS 2022 · similarity 0.90)
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.