Research map / Label-efficient and robust learning / Semi-supervised learning
Debiased Self-Training for Semi-Supervised Learning
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 |
|---|---|---|---|
| add debiased self-training (DST) with independent pseudo head and worst-case estimation | FixMatch · 11 tasks supervised pre-trained avg | accuracy: 59.3 -> 71.1 | Helped |
| add DST | FixMatch · 11 tasks unsupervised pre-trained avg | accuracy: 55.6 -> 68.7 | Helped |
| add DST | FixMatch · CIFAR-100, 4 labels/class, supervised pre-trained | accuracy: 53.1 -> 70.4 | Helped |
| replace mutual learning with independent pseudo head (generate vs use pseudo labels separately) (ablation) | DST · CIFAR-100, 4 labels/class, supervised pre-trained | accuracy: 53.4 -> 58.2 | Helped |
| use nonlinear instead of linear pseudo head (ablation) | DST w/o worst-case · CIFAR-100, 4 labels/class, supervised pre-trained | accuracy: 58.2 -> 60.6 | Helped |
| add worst-case estimation of pseudo labeling (ablation) | DST · CIFAR-100, 4 labels/class, supervised pre-trained | accuracy: 60.6 -> 70.4 | Helped |
| add worst-case estimation of pseudo labeling (ablation) | DST · CIFAR-100 unsupervised pre-trained | accuracy: 60.9 -> 68.2 | Helped |
| add DST pseudo head and worst-case estimation | FixMatch · CIFAR-100 worst-20 class pseudo-label accuracy | accuracy: 1.0 -> 34.5 | Helped |
| use nonlinear instead of linear main classifier head (ablation) | FixMatch · CIFAR-100 400/1000 labels | accuracy: 53.1 -> 54.1 (400 labels, supervised); 67.8 -> 67.2 (1000 labels) | No clear effect |
| add DST on top of FlexMatch | FlexMatch · CIFAR-100 and STL-10 | accuracy: +10.7% | Helped |
Closest papers by meaning
- Neighborhood-Regularized Self-Training for Learning with Few Labels (AAAI 2023 · similarity 0.90)
- Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Data (ICLR 2021 · similarity 0.90)
- Don’t fear the unlabelled: safe semi-supervised learning via debiasing (ICLR 2023 · similarity 0.90)
- Debiased Learning From Naturally Imbalanced Pseudo-Labels (CVPR 2022 · similarity 0.90)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.88)
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.87)
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.