Research map / Label-efficient and robust learning / Semi-supervised learning
Solving Semi-Supervised Few-Shot Learning from an Auto-Annotation Perspective
Tian Liu, Anwesha Basu, James Caverlee, Shu Kong
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 confidence and loss temperatures to FixMatch on VLM (ablation) | FixMatch (OpenCLIP) · 5 datasets mean, 4/8/16-shot | accuracy: 39.3 -> 57.7 (4-shot); 57.2 -> 71.2 (16-shot) | Improved |
| run established SSL (FixMatch) on the VLM instead of few-shot finetuning only | FS-FT (OpenCLIP) · 5 datasets mean, 16-shot | accuracy: 69.1 -> 57.2 | Hurt |
| add stage-1 classifier init with loss temperature to FixMatch (ablation) | FixMatch (OpenCLIP) · 5 datasets mean, 4-shot | accuracy: 39.3 -> 56.3 | Improved |
| add stage-2 semi-supervised finetuning with retrieved data (ablation) | SWIFT stage-1 (OpenCLIP) · 5 datasets mean, 4-shot | accuracy: 56.3 -> 68.8 | Improved |
| add stage-3 few-shot finetuning (ablation) | SWIFT stage-2 (OpenCLIP) · 5 datasets mean, 16-shot | accuracy: 77.3 -> 79.7 | Improved |
| apply SWIFT to DINOv2 backbone (ablation) | FixMatch direct (DINOv2) · 5 datasets mean, 4-shot | accuracy: 50.2 -> 78.2 | Improved |
| apply SWIFT instead of DebiasPL directly (ablation) | DebiasPL (OpenCLIP) · 5 datasets mean, 4-shot | accuracy: 39.6 -> 73.1 | Improved |
| add loss temperature to linear probing classifier (ablation) | text-init linear probing · 5 datasets mean, 16-shot | accuracy: 59.9 -> 64.7 | Improved |
| use stage-1 learned classifier for stage-2 instead of text embedding (ablation) | SWIFT stage-2 · 5 datasets mean, 4-shot | accuracy: 66.0 -> 68.8 | Improved |
| SWIFT versus fully supervised finetuning with ground-truth labels | SWIFT · 5 datasets mean, 4/8/16-shot | accuracy: 74.8 -> 71.5 (4-shot); 76.0 -> 79.7 (16-shot) | Improved |
| no loss temperature (T=1.0) on CLIP backbones (ablation) | linear probing semi-Aves · semi-Aves 16-shot | accuracy: CLIP-ViT 53.5 -> 43.9 | Hurt |
| use small loss temperature on ImageNet-pretrained backbones (ablation) | INet-RN50 linear probing · semi-Aves 16-shot | accuracy: 40.7 -> 30.1 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FixMatch — FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.89)
- Revisiting Semi-Supervised Learning in the Era of Foundation Models (NeurIPS 2025 · similarity 0.88)
- Scaling Few-Shot Learning for the Open World (AAAI 2024 · similarity 0.88)
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.88)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.87)
- Erasing the Bias: Fine-Tuning Foundation Models for Semi-Supervised Learning (ICML 2024 · 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.
