Research map / Label-efficient and robust learning / Semi-supervised learning
Debiased Learning From Naturally Imbalanced Pseudo-Labels
Check 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 adaptive debiasing and adaptive marginal loss to FixMatch (ablation) | FixMatch · CIFAR10-LT, γ=100, 30% labels | accuracy: 73.5 -> 80.6 | Helped |
| Add adaptive debiasing to FixMatch (ablation) | FixMatch · CIFAR10-LT, γ=100, 30% labels | accuracy: 73.5 -> 79.6 | Helped |
| Add adaptive marginal loss to debiased FixMatch (ablation) | DebiasPL · CIFAR10-LT, γ=100, 30% labels | accuracy: 79.6 -> 80.6 | Helped |
| Add adaptive debiasing and adaptive marginal loss to FixMatch (ablation) | FixMatch · CIFAR10, 4 labels per class | accuracy: 86.1 -> 94.6 | Helped |
| Add adaptive debiasing to FixMatch (ablation) | FixMatch · CIFAR10, 4 labels per class | accuracy: 86.1 -> 93.3 | Helped |
| Add DebiasPL to existing SSL methods | FixMatch, MixMatch, and UDA · CIFAR10, 4 labels per class | top-1 accuracy: 89.7 ± 4.6 -> 94.6 ± 1.3; 47.5 ± 11.5 -> 61.7 ± 6.1; 29.1 ± 5.9 -> 43.2 ± 5.2 | Helped |
| Use debias factor λ=0.5 rather than λ=0.0 (ablation) | DebiasPL · CIFAR10-LT, γ=100, 30% labels | accuracy: 73.5 -> 80.6 | Helped |
| Vary the debias factor λ across the tested values (ablation) | DebiasPL · CIFAR10-LT, γ=100, 30% labels | accuracy: λ=0.25: 79.5; λ=0.5: 80.6; λ=0.75: 80.5; λ=1.0: 80.5; λ=2.0: 77.7 | Mixed |
| Increase CLIP confidence threshold from τ_clip=0.2 to τ_clip=0.95 (ablation) | CLIP + DebiasPL · ImageNet-1K ZSL | top-1 accuracy: 55.9 -> 67.7 | Helped |
| Use DebiasPL instead of vanilla CLIP zero-shot predictions | CLIP · ImageNet-1K ZSL | top-1 accuracy: 59.6 -> 68.3 | Helped |
| Use FixMatch self-training with CLIP pseudo-labels without debiasing | CLIP · ImageNet-1K ZSL | top-1 accuracy: 59.6 -> 55.7 | Hurt |
| Add DebiasPL to FixMatch | FixMatch w/ EMAN · ImageNet-1K SSL, 0.2% labels | top-1 accuracy: 43.6 -> 47.9 | Helped |
Closest papers by meaning
- CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning (CVPR 2024 · similarity 0.91)
- Debiased Self-Training for Semi-Supervised Learning (NeurIPS 2022 · similarity 0.90)
- Don’t fear the unlabelled: safe semi-supervised learning via debiasing (ICLR 2023 · similarity 0.89)
- LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing (AAAI 2025 · similarity 0.89)
- Rectifying Soft-Label Entangled Bias in Long-Tailed Dataset Distillation (NeurIPS 2025 · similarity 0.87)
- 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.