Research map / Label-efficient and robust learning / Domain adaptation and generalization
Improve Unsupervised Pretraining for Few-Label Transfer
Suichan Li, Dongdong Chen, Yinpeng Chen, Lu Yuan, Lei Zhang, Qi Chu, Bin Liu, Nenghai Yu
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 target-aware pretraining with unlabeled target data | MoCoV2 · 9 datasets mean, 1+9 labels | mean accuracy: 67.49 -> 74.15 | Improved |
| add progressive few-label transfer (ACKMeans eigen-sample annotation) | TUP · 9 datasets mean, 1+9 labels | mean accuracy: 74.15 -> 76.69 | Improved |
| add progressive few-label transfer | TUP · 9 datasets mean, 1+1 labels | mean accuracy: 55.24 -> 59.30 | Improved |
| add sample re-balancing (20% target share) (ablation) | TUP · DTD, 1+9 labels | accuracy: 61.01 -> 66.58 | Improved |
| add sample re-balancing (20% target share) (ablation) | TUP · DTD, 1+1 labels | accuracy: 40.27 -> 44.86 | Improved |
| raise re-balancing ratio from 20% to 50% (ablation) | TUP · DTD, 1+9 labels | accuracy: 66.58 -> 63.86 | Hurt |
| continue unsupervised finetuning on target data instead of joint pretraining (ablation) | MoCoV2 · DTD, 10 labels per class | accuracy: 58.68 -> 58.34 | No clear effect |
| unsupervised finetuning on target data instead of joint pretraining (ablation) | MoCoV2 · STL10, 10 labels per class | accuracy: 90.85 -> 79.34 | Hurt |
| replace unsupervised finetuning with target-aware pretraining (ablation) | MoCoV2 · Pet37, 1 label per class | accuracy: 40.01 -> 57.28 | Improved |
| add target data to pretraining, clustering quality | MoCoV2 · Food101 | cluster accuracy and transfer accuracy: 11.86 -> 39.23 cluster; 28.39 -> 52.25 transfer | Improved |
| use supervised ImageNet-1k pretraining vs ImageNet-100 (better clustering) (ablation) | ResNet-50 pretrained · Pet, 5 labels per class | transfer accuracy: 45.65 -> 77.94 | Improved |
| use target-aware pretraining as initialization for FixMatch | FixMatch · CIFAR10, 10 labels per class | accuracy: 92.15 -> 96.38 | Improved |
Closest papers by meaning
- Unsupervised Selective Labeling for More Effective Semi-Supervised Learning (ECCV 2022 · similarity 0.88)
- Let Go of Your Labels with Unsupervised Transfer (ICML 2024 · similarity 0.86)
- On the Provable Advantage of Unsupervised Pretraining (ICLR 2024 · similarity 0.86)
- Label-Efficient Dataset Pruning via Semi-Supervised Pseudo-Labeling (NeurIPS 2026 · similarity 0.85)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.85)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 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.
