Research map / Label-efficient and robust learning / Semi-supervised learning
Information-Theoretic Characterization of the Generalization Error for Iterative Semi-Supervised Learning
Haiyun He, Hanshu Yan, Vincent Y. F. Tan
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 iterative pseudo-labelled unlabelled data when classes are easy to distinguish | Iterative self-training with pseudo-labelling · CIFAR-10 horse-ship / automobile-truck | gen-error: converges to around 0.25 after 5 iterations | Improved |
| add iterative pseudo-labelling on 10-class digits | ResNet-6 self-training · MNIST | gen-error: converges to around 0.25 after 20 iterations | Improved |
| add pseudo-labelled data for hard-to-distinguish classes | ResNet-10 self-training · CIFAR-10 cat-dog | gen-error | Hurt |
| add l2 regularization with weight decay 0.0005 (ablation) | ResNet-10 self-training · CIFAR-10 cat-dog | test accuracy: test accuracy improved by 0.6% | Improved |
| increase weight decay (ablation) | ResNet-10 self-training · CIFAR-10 cat-dog | average gen-error over last 10 iterations | Improved |
| use all 9,500 unlabelled images in every iteration instead of sampling 2500 (ablation) | ResNet-10 self-training · CIFAR-10 horse-ship | gen-error | Hurt |
| use unlabelled data when class variance is small (sigma^2=0.6^2) | bGMM iterative SSL · Synthetic bGMM d=2, n=10, m=1000 | gen-error: decreases significantly from t=0 to t=1 then converges to a non-zero constant | Improved |
| use unlabelled data when class variance is large (sigma^2=3^2) | bGMM iterative SSL · Synthetic bGMM d=2, n=10, m=1000 | gen-error | Hurt |
| add l2 regularization lambda = 0, 0.1, 0.5, 2 (ablation) | bGMM iterative SSL · Synthetic bGMM sigma=3, d=2 | gen-error | Improved |
| reuse labelled data in every iteration (ablation) | bGMM iterative SSL · Synthetic bGMM sigma=0.6, n=10, m=1000 | gen-error | No clear effect |
| reduce unlabelled data from m=1000 to m=100 (ablation) | bGMM iterative SSL · Synthetic bGMM with labelled data reuse | gen-error: m=100 higher than m=1000 | Hurt |
| add pseudo-labelled data when labelled set is large (n above 30) | bGMM iterative SSL · Synthetic bGMM m=1000, sigma=0.6, d=2 | gen-error: gen_0 smaller than gen_1 for n larger than 30 | Hurt |
Closest papers by meaning
- Improved Generalization in Semi-Supervised Learning: A Survey of Theoretical Results (TPAMI 2022 · similarity 0.89)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.88)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.86)
- Unlocking the Potential of Unlabeled Data in Semi-Supervised Domain Generalization (CVPR 2025 · similarity 0.86)
- Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning (NeurIPS 2024 · similarity 0.86)
- Semi-Supervised Regression with Heteroscedastic Pseudo-Labels (NeurIPS 2025 · 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.
