Research map / Label-efficient and robust learning / Semi-supervised learning
SemPPL: Predicting Pseudo-Labels for Better Contrastive Representations
Matko Bošnjak, Pierre Harvey Richemond, Nenad Tomasev, Florian Strub, Jacob C Walker, Felix Hill, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, Jovana Mitrovic
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 pseudo-label-based semantic positives to ReLICv2 | ReLICv2 on ImageNet with 1% labels · ImageNet test, 1% labels | top-1 accuracy: 58.1 -> 68.5; +10.4% | Improved |
| Add pseudo-label-based semantic positives to BYOL | BYOL on ImageNet with 1% and 10% labels · ImageNet test, 1% and 10% labels | top-1 accuracy: +3.9%; +3.6% | Improved |
| Use pseudo-labels to select semantic positives for unlabelled data (ablation) | SemPPL on ImageNet · ImageNet test, 1% and 10% labels | top-1 accuracy: 61.9 -> 68.5; 71.1 -> 76.0 | Improved |
| Add semantic positives for labelled data (ablation) | SemPPL on ImageNet · ImageNet test, 1% and 10% labels | top-1 accuracy: 58.1 -> 68.5; 72.4 -> 76.0 | Improved |
| Replace predicted pseudo-labels with ground-truth labels for positive retrieval (ablation) | SemPPL on ImageNet · ImageNet test, 10% labels, 100 epochs | top-1 accuracy: 69.9 -> 71.6; 1.7% | Improved |
| Increase semantic positives from 1 or 2 to 3 (ablation) | SemPPL on ImageNet · ImageNet test, 10% labels | top-1 accuracy: 69.9 -> 69.9 | No clear effect |
| Increase semantic positives to 5 (ablation) | SemPPL on ImageNet · ImageNet test, 10% labels | top-1 accuracy: 69.9 -> 71.0 | Improved |
| Train for 500 rather than 300 epochs (ablation) | SemPPL on ImageNet · ImageNet test, 10% labels | top-1 accuracy: 72.7 -> 72.3 | Hurt |
| Use majority voting across views for pseudo-labels (ablation) | SemPPL on ImageNet · ImageNet test, 10% labels | top-1 accuracy: 69.0 -> 69.9 | Improved |
| Increase k-NN from 1 to larger neighbour counts (ablation) | SemPPL on ImageNet · ImageNet test, 10% labels | top-1 accuracy: 70.0 -> 69.9 (k=2); 69.8 (k=3); 69.8 (k=5); 69.1 (k=10) | Hurt |
| Increase labelled-embedding queue capacity (ablation) | SemPPL on ImageNet · ImageNet test, 10% labels | top-1 accuracy: 67.9 -> 69.9 | Improved |
| Use SemPPL instead of the strongest listed semi-supervised baseline | ResNet-50 1× on ImageNet · ImageNet test, 1% labels | top-1 accuracy: 67.2 -> 68.5; +1.3% | Improved |
Closest papers by meaning
- More Reliable Pseudo-labels, Better Performance: A Generalized Approach to Single Positive Multi-label Learning (ICCV 2025 · similarity 0.89)
- Negatives Make a Positive: An Embarrassingly Simple Approach to Semi-Supervised Few-Shot Learning (TPAMI 2023 · similarity 0.88)
- All Labels Are Not Created Equal: Enhancing Semi-Supervision via Label Grouping and Co-Training (CVPR 2021 · similarity 0.88)
- ST++: Make Self-Training Work Better for Semi-Supervised Semantic Segmentation (CVPR 2022 · similarity 0.87)
- Barely-Supervised Learning: Semi-supervised Learning with Very Few Labeled Images (AAAI 2022 · similarity 0.87)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · 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.
