Research map / Label-efficient and robust learning / Semi-supervised learning
Not All Unlabeled Data are Equal: Learning to Weight Data in Semi-supervised Learning
Zhongzheng Ren, Raymond Yeh, Alexander Schwing
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 |
|---|---|---|---|
| Learn per-example weights instead of using a shared UDA weight | UDA · CIFAR-10 and SVHN labeled-data splits | test error rate | Improved |
| Learn per-example weights instead of using a shared FixMatch weight | FixMatch (CTA) · CIFAR-10 and SVHN reported splits | test error rate | Improved |
| Learn per-example weights for UDA | UDA with BERT · IMDb, 20 labeled examples, sequence length 128 | test error rate: 8.98$\pm$0.26 -> 8.51$\pm$0.14 | Improved |
| Use per-example weights rather than tune a single shared weight (ablation) | Proposed method with UDA · CIFAR-10 labeled-data splits | test error rate | Improved |
| Use masked Adam updates rather than vanilla Adam (ablation) | Per-example weight optimizer · CIFAR-10, UDA | validation error rate | Improved |
| Use masked Adam updates rather than SGD (ablation) | Per-example weight optimizer · CIFAR-10, UDA | validation error rate | Improved |
| Increase validation-set size (ablation) | Proposed method with UDA · CIFAR-10, 250 and 4000 labeled examples | test error rate | Improved |
| Set inner steps N to 30 rather than 100 (ablation) | Proposed method with UDA · CIFAR-10, 4000 labeled examples | validation error rate: 3.42 -> 5.13 | Hurt |
| Set inner steps N to 300 rather than 100 (ablation) | Proposed method with UDA · CIFAR-10, 4000 labeled examples | validation error rate: 3.42 -> 4.59 | Hurt |
| Set outer-loop step size eta to 0.1 rather than 0.01 (ablation) | Proposed method with UDA · CIFAR-10, 4000 labeled examples | validation error rate: 3.42 -> 6.16 | Hurt |
| Set outer-loop step size eta to 0.001 rather than 0.01 (ablation) | Proposed method with UDA · CIFAR-10, 4000 labeled examples | validation error rate: 3.42 -> 4.10 | Hurt |
| Use the exact inverse Hessian of only the last layer for influence estimation (ablation) | Proposed method with UDA · CIFAR-10 | test error rate: 4.51% and 4.85% -> 4.43% | Improved |
Closest papers by meaning
- Improved Generalization in Semi-Supervised Learning: A Survey of Theoretical Results (TPAMI 2022 · similarity 0.89)
- Learning to Learn in a Semi-Supervised Fashion (ECCV 2020 · similarity 0.89)
- SimPLE: Similar Pseudo Label Exploitation for Semi-Supervised Classification (CVPR 2021 · similarity 0.89)
- InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning (ICML 2024 · similarity 0.89)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.88)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · 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.
