Research map / Label-efficient and robust learning / Domain adaptation and generalization
Self-training For Few-shot Transfer Across Extreme Task Differences
Cheng Perng Phoo, Bharath Hariharan
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 |
|---|---|---|---|
| Self-train the source representation on unlabeled target-domain images, retaining teacher pseudo-label groupings and adding SimCLR | ResNet-10 transfer from miniImageNet · BSCD-FSL 5-way 1-shot, averaged across target datasets | accuracy: +2.9 points on average | Improved |
| Add SimCLR self-supervision to STARTUP (ablation) | ResNet-10 STARTUP · ISIC 5-way 5-shot | accuracy: 46.48 ± 0.61 -> 47.22 ± 0.61 | Improved |
| Add SimCLR self-supervision to STARTUP (ablation) | ResNet-10 STARTUP · EuroSAT 5-way 1-shot | accuracy: 62.90 ± 0.83 -> 63.88 ± 0.84 | Improved |
| Use unlabeled images from the target domain instead of EuroSAT unlabeled images (ablation) | ImageNet-trained ResNet-18 STARTUP · CropDisease 5-way 1-shot | accuracy: 70.40 ± 0.86 -> 85.10 ± 0.74 | Improved |
| Remove the base-dataset training loss (ablation) | ResNet-10 STARTUP · ISIC 5-way 5-shot | accuracy: 47.22 ± 0.61 -> 45.73 ± 0.61 | Hurt |
| Replace SimCLR with rotation self-supervision (ablation) | ResNet-10 STARTUP · CropDisease 5-way 1-shot | accuracy: 75.93 ± 0.80 -> 70.53 ± 0.84 | Hurt |
| Increase the amount of unlabeled target-domain data (ablation) | STARTUP · ISIC few-shot transfer | accuracy | Improved |
| Change the student initialization strategy among scratch, teacher initialization, and teacher embedding with a random classifier (ablation) | STARTUP without SimCLR · BSCD-FSL datasets | accuracy | Mixed |
| Initialize the student from the teacher and omit SimCLR (ablation) | STARTUP · tieredImageNet-less 5-way 1-shot | accuracy: 57.29 ± 0.83 -> 60.39 ± 0.86 | Improved |
| Train the student from scratch and omit SimCLR (ablation) | STARTUP · tieredImageNet-more 5-way 1-shot | accuracy: 57.68 ± 0.89 -> 63.29 ± 0.90 | Improved |
| Apply STARTUP rather than direct transfer | ResNet-10 transfer · miniImageNet 5-way 1-shot | accuracy: 54.18 ± 0.79 -> 54.20 ± 0.81 | No clear effect |
| Train STARTUP using unlabeled data from another domain (ablation) | ImageNet-trained ResNet-18 STARTUP · CropDisease 5-way 1-shot | accuracy: 70.40 ± 0.86, underperforming naive Transfer | Hurt |
Closest papers by meaning
- Understanding Cross-Domain Few-Shot Learning Based on Domain Similarity and Few-Shot Difficulty (NeurIPS 2022 · similarity 0.91)
- Boosting Few-Shot Visual Learning With Self-Supervision (ICCV 2019 · similarity 0.90)
- Task-Level Self-Supervision for Cross-Domain Few-Shot Learning (AAAI 2022 · similarity 0.90)
- Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference (CVPR 2022 · similarity 0.90)
- Revisiting Continuity of Image Tokens for Cross-domain Few-shot Learning (ICML 2025 · similarity 0.88)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.88)
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.
