Research map / Label-efficient and robust learning / Recognition and long-tail
Prototypical Contrastive Learning of Unsupervised Representations
Junnan Li, Pan Zhou, Caiming Xiong, Steven Hoi
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 |
|---|---|---|---|
| Replace instance-wise contrastive learning with PCL | ResNet-50 representations · ImageNet, 1% labels | top-5 accuracy: 56.9 -> 75.3 | Improved |
| Replace instance-wise contrastive learning with PCL | ResNet-50 representations · ImageNet linear classification | top-1 accuracy: 60.6 -> 61.5 | Improved |
| Replace MoCo representations with PCL representations | ImageNet kNN classification · ImageNet | accuracy: 47.1 -> 54.5 | Improved |
| Replace MoCo representations with PCL representations | ImageNet feature clustering · ImageNet, k-means with k=25000 | AMI: 0.285 -> 0.410 | Improved |
| Replace MoCo pretraining with PCL pretraining | ResNet-50-FPN Faster R-CNN · VOC07 test, detector trained on VOC07 | mAP: 66.4 -> 71.7 | Improved |
| Add the prototypical contrastive term (ablation) | instance-only PCL objective · ImageNet, 1% labels | top-5 accuracy: 56.9 -> 75.3 | Improved |
| Add the prototypical contrastive term (ablation) | instance-only PCL objective · ImageNet linear classification | top-1 accuracy: 60.6 -> 61.5 | Improved |
| Add warm-up using the instance-wise loss (ablation) | proto-only objective · ImageNet, 1% labels | top-5 accuracy: 60.7 -> 72.3 | Improved |
| Add warm-up using the instance-wise loss (ablation) | proto-only objective · ImageNet linear classification | top-1 accuracy: 60.4 -> 60.9 | Improved |
| Add warm-up using the instance-wise loss (ablation) | instance + proto objective · ImageNet, 1% labels | top-5 accuracy: 74.6 -> 75.3 | Improved |
| Add warm-up using the instance-wise loss (ablation) | instance + proto objective · ImageNet linear classification | top-1 accuracy: 61.3 -> 61.5 | Improved |
Closest papers by meaning
- Mutual Contrastive Learning for Visual Representation Learning (AAAI 2022 · similarity 0.88)
- Neighborhood Contrastive Learning for Novel Class Discovery (CVPR 2021 · similarity 0.88)
- Prototypical Contrastive Predictive Coding (ICLR 2022 · similarity 0.88)
- Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.87)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.87)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 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.
