Research map / Label-efficient and robust learning / Noisy labels
Label-Noise Robust Generative Adversarial Networks
Check 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 |
|---|---|---|---|
| Incorporate a label-noise transition model into AC-GAN training | rAC-GAN versus AC-GAN · Noisy-label conditional image generation, across CIFAR conditions | conditional-generation quality (Intra FID, GAN-test, GAN-train) | Helped |
| Incorporate a label-noise transition model into cGAN, without a classifier | rcGAN versus cGAN · Noisy-label conditional image generation, across CIFAR conditions | conditional-generation quality (Intra FID, GAN-test, GAN-train) | Helped |
| Use robust GAN variants with transition-model correction | rAC-GAN and rcGAN · CIFAR-10 symmetric noise, overall FID | FID: 32/40 conditions | Helped |
| Use an estimated transition matrix T' instead of a known transition matrix | rAC-CT-GAN and rcSN-GAN · CIFAR-10 noisy-label settings | conditional-generation quality | Helped |
| Use an estimated transition matrix T' at high noise rates | rAC-CT-GAN and rcSN-GAN · CIFAR-100 highly noisy settings | conditional-generation quality | Hurt |
| Add mutual information regularization to the generator (ablation) | rAC-GAN and rcGAN · Symmetric noise with μ = 0.9 | GAN-train: 26.3 -> 70.8; 48.7 -> 60.7 | Helped |
| Increase learning rate from 0.0001 toward 0.0004 (ablation) | rAC-GAN and rcGAN · CIFAR-10 and CIFAR-100, symmetric noise rate 0.9 | quantitative scores | Helped |
| Use an extensively high learning rate of 0.0008 (ablation) | rAC-GAN and rcGAN · CIFAR-10 and CIFAR-100, symmetric noise rate 0.9 | quantitative scores | Hurt |
| Increase batch size (ablation) | rAC-GAN and rcGAN · CIFAR-10 and CIFAR-100, symmetric noise rate 0.9 | quantitative scores | Helped |
| Use a model transition rate within ±0.2 of the real transition rate (ablation) | rAC-CT-GAN and rcSN-GAN · CIFAR-10 and CIFAR-100, symmetric and asymmetric noise | quantitative quality | No clear effect |
| Train with noisy rather than clean Clothing1M labels using transition correction | rAC-CT-GAN versus AC-CT-GAN; rcSN-GAN versus cSN-GAN · Clothing1M noisy | GAN-train: 49.5 -> 51.7; 48.6 -> 49.8 | Helped |
| Use mixed clean and noisy Clothing1M data with transition correction | rAC-CT-GAN and rcSN-GAN · Clothing1M mixed | GAN-train: 52.8 -> 57.0; 51.7 -> 55.0 | Helped |
Closest papers by meaning
- Robustness of conditional GANs to noisy labels (NeurIPS 2018 · similarity 0.94)
- Noise Robust Generative Adversarial Networks (CVPR 2020 · similarity 0.92)
- Robust Conditional Generative Adversarial Networks (ICLR 2019 · similarity 0.91)
- CcGAN: Continuous Conditional Generative Adversarial Networks for Image Generation (ICLR 2021 · similarity 0.90)
- Fusing Conditional Submodular GAN and Programmatic Weak Supervision (AAAI 2024 · similarity 0.89)
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · similarity 0.87)
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.