Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-Supervised Learning of Visual Features by Non-Parametrically Predicting View Assignments With Support Samples
Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, Michael Rabbat
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 ME-max regularization (ablation) | PAWS · ImageNet 1% labels, ResNet-50, 100 ep | top-1 accuracy: 52.9 -> 63.8 | Improved |
| add ME-max regularization (ablation) | PAWS · ImageNet 10% labels, ResNet-50, 100 ep | top-1 accuracy: 73.6 -> 73.9 | Improved |
| add a prediction head during pre-training (ablation) | PAWS · ImageNet 10% labels, ResNet-50, 100 ep | top-1 accuracy: 74.2 -> 73.9 | No clear effect |
| add a prediction head during pre-training (ablation) | PAWS · ImageNet 10% labels, ResNet-50, 200 ep | top-1 accuracy: 75.2 -> 75.0 | No clear effect |
| enlarge the support set from 960x4 to 1000x16 images (ablation) | PAWS · ImageNet 10% labels, ResNet-50, 100 ep | top-1 accuracy: 72.0 -> 74.5 | Improved |
| sample 960 classes x 4 images instead of 448 classes x 8 (ablation) | PAWS · ImageNet 10% labels, ResNet-50, 100 ep | top-1 accuracy: 70.1 -> 72.0 | Improved |
| train 300 instead of 200 epochs (ablation) | PAWS · ImageNet 10% labels, ResNet-50 | top-1 accuracy: 75.0 -> 75.5 | No clear effect |
| train 300 instead of 200 epochs (ablation) | PAWS · ImageNet 10% labels, ResNet-50 (2x) | top-1 accuracy: 77.8 -> 77.7 | No clear effect |
| fine-tune a classifier instead of nearest-neighbour classification | PAWS · ImageNet 1% labels, ResNet-50, 200 ep | top-1 accuracy: 63.2 -> 66.1 | Improved |
| widen ResNet-50 from 1x to 4x | PAWS · ImageNet 10% labels, 200 ep | top-1 accuracy: 75.0 -> 79.0 | Improved |
| use PAWS 100 epochs instead of SwAV 800 epochs | PAWS vs SwAV · ImageNet 1% labels, ResNet-50 | top-1 accuracy: +9.9% | Improved |
| small-batch PAWS (batch 256, 8 GPUs) versus large batch with similar support (ablation) | PAWS · ImageNet 10% labels, ResNet-50, 100 ep | top-1 accuracy: 70.1 -> 70.2 | No clear effect |
Closest papers by meaning
- All Labels Are Not Created Equal: Enhancing Semi-Supervision via Label Grouping and Co-Training (CVPR 2021 · similarity 0.88)
- Curriculum Labeling: Revisiting Pseudo-Labeling for Semi-Supervised Learning (AAAI 2021 · similarity 0.87)
- RoPAWS: Robust Semi-supervised Representation Learning from Uncurated Data (ICLR 2023 · similarity 0.87)
- Multiview Pseudo-Labeling for Semi-Supervised Learning From Video (ICCV 2021 · similarity 0.87)
- Revisiting Semi-Supervised Learning in the Era of Foundation Models (NeurIPS 2025 · similarity 0.85)
- Semi-Supervised Multi-View Multi-Label Learning with View-Specific Transformer and Enhanced Pseudo-Label (AAAI 2025 · similarity 0.85)
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.
