Research map / Learning theory and algorithms / Optimal transport
Learning from Bad Data via Generation
Tianyu Guo, Chang Xu, Boxin Shi, Chao Xu, Dacheng Tao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add worst-case generator training to classifier (Ours vs Classifier) | classifier · CIFAR-10 IF=20 | accuracy (%): 81.51 -> 81.94 | Improved |
| add worst-case generator (Ours vs Classifier) | classifier · MNIST IF=100 | accuracy (%): 89.77 -> 92.67 | Improved |
| add worst-case generator (Ours vs Classifier) | classifier · FMNIST IF=100 | accuracy (%): 79.89 -> 83.06 | Improved |
| add DAGAN augmentation | classifier · CIFAR-10 IF=10 | accuracy (%): 85.57 -> 77.82 | Hurt |
| add Triple GAN generator | classifier · CIFAR-10 IF=10 | accuracy (%): 85.57 -> 83.06 | Hurt |
| add Ours to classifier on balanced data | classifier · MNIST IF=1 | accuracy (%): 99.35 -> 99.42 | Improved |
| add Ours to classifier on balanced data | classifier · CIFAR-10 IF=1 | accuracy (%): 93.04 -> 93.01 | No clear effect |
| use ResNet-101 instead of VGG16 as classifier in Ours | Ours · CIFAR-10 G(0.2) | accuracy (%): 85.02 -> 85.87 | Improved |
| replace Mixup with Ours as augmentation | Mixup · CIFAR-10 reduced | accuracy (%): 83.91 -> 84.60 | Improved |
| replace standard GAN with Ours as generator (FID) | GAN · CIFAR-10 IF=10 | FID: 37.96 -> 26.10 | Improved |
| remove BN and WD (classifier only), Ours vs classifier (ablation) | classifier · Reduced CIFAR-10 | accuracy (%): 71.85 -> 75.42 | Improved |
| remove BN from Ours (accuracy with BN+WD to No BN) (ablation) | Ours · Reduced CIFAR-10 | accuracy (%): 84.60 -> 78.01 | Hurt |
Closest papers by meaning
- How Well Generative Adversarial Networks Learn Distributions (JMLR 2021 · similarity 0.87)
- Influence Estimation for Generative Adversarial Networks (ICLR 2021 · similarity 0.87)
- An Error Analysis of Generative Adversarial Networks for Learning Distributions (JMLR 2022 · similarity 0.87)
- Leveraging Contaminated Datasets to Learn Clean-Data Distribution with Purified Generative Adversarial Networks (AAAI 2023 · similarity 0.86)
- Self-Diagnosing GAN: Diagnosing Underrepresented Samples in Generative Adversarial Networks (NeurIPS 2021 · similarity 0.86)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.85)
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.
