Research map / Label-efficient and robust learning / Noisy labels
Robustness of conditional GANs to noisy labels
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 |
|---|---|---|---|
| Corrupt generated labels with the known noise channel before presenting generated samples to the discriminator, using a projection discriminator | Conditional GAN · MNIST, noise levels 0.2, 0.3, and 0.5 | generator label accuracy: RCGAN: 0.994; AmbientGAN: 0.940, 0.902, 0.857 | Helped |
| Use a projection discriminator instead of the DCGAN discriminator (ablation) | AmbientGAN-style training · Noisy MNIST | generator label accuracy: RCGAN: 0.994; AmbientGAN: 0.940, 0.902, 0.857 | Helped |
| Concatenate a one-hot label to the first discriminator layer input (ablation) | RCGAN · Noisy MNIST | generator label accuracy | Helped |
| Estimate the noise model jointly with the generator rather than requiring the confusion matrix | RCGAN-U · MNIST and CIFAR-10 | generator quality and label accuracy | Helped |
| Train a conditional GAN directly on noisy labels without noise correction | Biased GAN · Noisy MNIST and CIFAR-10 | generator label accuracy | Hurt |
| Use inverse-confusion-matrix loss correction | Unbiased GAN · Noisy CIFAR-10 | inception score | Hurt |
| Use a linear classifier as the RCGAN-U permutation regularizer (ablation) | RCGAN-U · Noisy CIFAR-10 | noise-model estimation | Hurt |
| Initialize the learned confusion matrix to be diagonally dominant (ablation) | RCGAN-U · Noisy CIFAR-10 | noise-model estimation | Helped |
| Use the proposed cGAN label-recovery procedure with a robustly trained generator rather than relying only on a discriminator-based method | RCGAN label recovery · Noisy MNIST training labels | label recovery accuracy | Helped |
Closest papers by meaning
- Label-Noise Robust Generative Adversarial Networks (CVPR 2019 · similarity 0.94)
- Robust Conditional Generative Adversarial Networks (ICLR 2019 · similarity 0.93)
- Noise Robust Generative Adversarial Networks (CVPR 2020 · similarity 0.91)
- Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions (CVPR 2024 · similarity 0.88)
- Fusing Conditional Submodular GAN and Programmatic Weak Supervision (AAAI 2024 · similarity 0.87)
- Turning Waste into Wealth: Leveraging Low-Quality Samples for Enhancing Continuous Conditional Generative Adversarial Networks (AAAI 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.