Research map / Label-efficient and robust learning / Recognition and long-tail
Generative Modeling Helps Weak Supervision (and Vice Versa)
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 |
|---|---|---|---|
| Align GAN discrete codes with sample-dependent LF accuracy-weighted pseudolabels | WSGAN · AwA2-A, AwA2-B, DomainNet, MNIST, FashionMNIST, GTSRB, CIFAR10-A, CIFAR10-B | pseudolabel accuracy: AwA2-A: 0.641 -> 0.681; AwA2-B: 0.623 -> 0.699; DomainNet: 0.658 -> 0.661; MNIST: 0.782 -> 0.813; FashionMNIST: 0.737 -> | Helped |
| Align weak supervision with learned discrete latent structure | WSGAN · AwA2-A, AwA2-B, DomainNet, CIFAR10-A, CIFAR10-B | FID: AwA2-A: 41.62 -> 34.71; AwA2-B: 41.62 -> 34.52; DomainNet: 53.98 -> 44.35; CIFAR10-A: 28.93 -> 22.71; CIFAR10-B: 33.50 - | Helped |
| Align weak supervision with learned discrete latent structure | WSGAN · GTSRB | FID: 69.67 -> 73.96 | Hurt |
| Use a sample-dependent accuracy encoder instead of a global LF-weight vector (ablation) | WSGAN-Encoder versus WSGAN-Vector · Label model accuracy across eight main datasets | pseudolabel accuracy: AwA2-A: 0.647 -> 0.681; AwA2-B: 0.645 -> 0.699; DomainNet: 0.661 -> 0.643; MNIST: 0.782 -> 0.813; FashionMNIST: 0.737 -> | Helped |
| Use a sample-dependent accuracy encoder instead of a global LF-weight vector (ablation) | WSGAN-Encoder versus WSGAN-Vector · AwA2-A, AwA2-B, DomainNet, GTSRB, CIFAR10-A, CIFAR10-B | FID: AwA2-A: 36.74 -> 34.71; AwA2-B: 36.79 -> 34.52; DomainNet: 50.16 -> 44.35; GTSRB: 75.27 -> 73.96; CIFAR10-A: 25.70 -> 22 | Mixed |
| Add 1,000 generated images labeled by the learned latent-code pseudolabeler | ResNet-18 downstream classifier · AwA2-A, AwA2-B, DomainNet, MNIST, FashionMNIST, GTSRB, CIFAR10-A, CIFAR10-B | test accuracy increase: AwA2-A: 0.88%; AwA2-B: 2.40%; DomainNet: 2.31%; MNIST: 1.60%; FashionMNIST: 0.29%; GTSRB: 0.40%; CIFAR10-A: 0.04%; CIFAR | Helped |
| Add 1,000 generated images labeled by LFs and WSGAN label model | ResNet-18 downstream classifier · AwA2-A, AwA2-B, DomainNet, MNIST, FashionMNIST, GTSRB | test accuracy increase: AwA2-A: 0.79%; AwA2-B: 3.90%; DomainNet: 1.50%; MNIST: 1.71%; FashionMNIST: 0.34%; GTSRB: 0.02% | Helped |
| Increase the number of generated augmentation samples beyond 1,000 (ablation) | WSGAN-augmented downstream classifier · Image classification datasets | test accuracy | No clear effect |
| Replace the DCGAN base with StyleGAN2-ADA (ablation) | StyleWSGAN · LSUN scene categories | FID: 8.41 -> 7.54 | Helped |
| Replace the DCGAN base with StyleGAN2-ADA (ablation) | StyleWSGAN · CIFAR10-B | FID: 3.85 -> 3.79 | Helped |
| Add a decaying penalty encouraging equal LF weights early in training (ablation) | WSGAN · Latent label estimates | latent label estimate quality | Helped |
| Use batch size 8 or 32 instead of batch size 16 (ablation) | WSGAN · GAN image generation and label modeling | FID and label model accuracy | No clear effect |
Closest papers by meaning
- Fusing Conditional Submodular GAN and Programmatic Weak Supervision (AAAI 2024 · similarity 0.92)
- Learning from weak labelers as constraints (ICLR 2025 · similarity 0.92)
- Fair Generative Modeling via Weak Supervision (ICML 2020 · similarity 0.90)
- From Linear to Nonlinear: Provable Weak-to-Strong Generalization through Feature Learning (NeurIPS 2025 · similarity 0.87)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.86)
- Weak Supervision Performance Evaluation via Partial Identification (NeurIPS 2024 · similarity 0.86)
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.