Research map / Label-efficient and robust learning / Recognition and long-tail
Fusing Conditional Submodular GAN and Programmatic Weak Supervision
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Fuses a noise-aware, classifier-guided conditional GAN with an encoder-based programmatic weak supervision label model, and uses submodular subset selection under an entropy knapsack constraint to choose training data for the classifier.
Problem
Jointly training a label model for programmatic weak supervision and a class-conditional GAN is hampered by non-class-specific InfoGAN latent factors and noisy pseudo labels that destabilize GAN training.
Approach
Replace InfoGAN's latent factors with a noise-aware classifier (symmetric cross-entropy on label-model pseudo labels) that shares weights with the discriminator and guides a conditional GAN and the label model, plus a cost-effective greedy submodular subset selection (generalized graph cut with an entropy knapsack) to pick training samples.
Key results
- On the label model benchmark, Ours (comp) reaches 0.701 posterior accuracy on AWA2 versus 0.670 for WSGAN, a 3.0-point gain.
- Mean FID improves over WSGAN by about 7.08 points across datasets; on GTSRB the subset variant improves FID from 73.96 (WSGAN) to 59.09.
- Ours (sub) reaches FID 14.24 on LSUN versus 14.47 for WSGAN, and 23.13 on AFHQ versus 24.72 for WSGAN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed fused method instead of WSGAN | label model · AWA2 | posterior accuracy: 0.670 -> 0.701 | Helped |
| use the proposed fused method instead of WSGAN | label model · CIFAR10-A | posterior accuracy: 0.872 -> 0.890 | Helped |
| use the proposed fused method instead of WSGAN | conditional GAN · AWA2 | FID: 36.00 -> 25.92 | Helped |
| use the proposed fused method instead of WSGAN | conditional GAN · CIFAR10-A | FID: 22.71 -> 18.00 | Helped |
| add submodular subset selection (ablation) | proposed method · GTSRB | FID: 62.55 -> 59.09 | Helped |
| add submodular subset selection (ablation) | proposed method · DomainNet | FID: 37.42 -> 36.82 | Helped |
| add submodular subset selection (ablation) | proposed method · FashionMNIST | FID: 11.78 -> 11.81 | Hurt |
| add submodular subset selection (ablation) | proposed method · AWA2 | posterior accuracy: 0.701 -> 0.698 | Hurt |
| add submodular subset selection (ablation) | proposed method · DomainNet | posterior accuracy: 0.650 -> 0.653 | Helped |
| add submodular subset selection (ablation) | proposed method · CIFAR10-A | posterior accuracy: 0.890 -> 0.890 | No clear effect |
| vary subset entropy budget ratio η from 0.2 to 0.8 (ablation) | subset-selected method · AWA2 | FID: 29.207 -> 28.956 | Helped |
| vary subset entropy budget ratio η from 0.2 to 0.8 (ablation) | subset-selected method · GTSRB | FID: 61.260; 56.443; 60.255; 59.590 | Mixed |
Surprising findings
- The label model's accuracy changes only marginally between subset selection and using all non-abstained samples, and the subset is not uniformly better for the label model.
- For FID, the best subset budget varies by dataset: smaller subsets work best for some datasets (GTSRB, MNIST, CIFAR10-B) while larger ones work best for others.
Limitations
- Subset selection faces computational challenges for very large datasets, which the authors acknowledge.
- Improvements in label-model accuracy are small and the subset variant is not consistently better than the complete-data variant.
- Experiments are on small image datasets at 32x32 resolution with DCGAN, with only limited high-resolution results.
Experimental setup
Datasets: AWA2 DomainNet CIFAR10 MNIST FashionMNIST GTSRB LSUN AFHQ
Benchmarks: AWA2 CIFAR10 DomainNet FashionMNIST GTSRB MNIST LSUN AFHQ
Metrics: posterior accuracy FID F1 precision recall test accuracy ARI
Baselines: WSGAN Majority Voting MeTaL FlyingSquid Snorkel Hyper-label model Dawid-Skene
Builds on: WSGAN InfoGAN programmatic weak supervision symmetric cross-entropy DCGAN
Closest papers by meaning
- Generative Modeling Helps Weak Supervision (and Vice Versa) (ICLR 2023 · similarity 0.92)
- Learning from weak labelers as constraints (ICLR 2025 · similarity 0.90)
- Label-Noise Robust Generative Adversarial Networks (CVPR 2019 · similarity 0.89)
- Statistical Analysis of an Adversarial Bayesian Weak Supervision Method (NeurIPS 2025 · similarity 0.89)
- Losses over Labels: Weakly Supervised Learning via Direct Loss Construction (AAAI 2023 · similarity 0.89)
- OSSGAN: Open-Set Semi-Supervised Image Generation (CVPR 2022 · similarity 0.88)
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.