Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Coarsely-Labeled Data for Better Few-Shot Transfer
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 |
|---|---|---|---|
| add parent-consistent filtering of pseudo-labels (ablation) | self-training student · iNat2019-CL pseudo-label grouping | AMI: 0.4258 -> 0.7260 | Improved |
| add parent-consistent filtering of pseudo-labels (ablation) | self-training student · tieredImageNet-CL pseudo-label grouping | AMI: 0.4620 -> 0.7352 | Improved |
| add parent-consistent filtering of pseudo-labels (ablation) | self-training student · CIFAR-100-CL pseudo-label grouping | AMI: 0.3695 -> 0.8293 | Improved |
| add parent-consistent filtering (PAS vs Self-training) (ablation) | PAS · iNat2019-CL all-way 5-shot | top-1 per-class accuracy: 42.17 -> 43.27 | Improved |
| add parent-consistent filtering (PAS vs Self-training) (ablation) | PAS · CIFAR-100-CL all-way 5-shot | top-1 per-class accuracy: 42.43 -> 48.42 | Improved |
| use PAS instead of multitask training for coarse labels | representation learning · iNat2019-CL all-way 5-shot | top-1 per-class accuracy: 41.42 -> 43.27 | Improved |
| use coarsely-labeled data with PAS instead of baseline | representation learning · all-way 5-shot, three datasets average | top-1 per-class accuracy: 8.31 points gain over Baseline | Improved |
| train only on coarse supercategory labels (Repr-Coarse) instead of fine labels | representation learning · tieredImageNet-CL all-way 5-shot | top-1 per-class accuracy: 53.36 -> 37.19 | Hurt |
| reduce base classes from 398 to 144 (Baseline) (ablation) | Baseline · iNat2019-CL all-way 5-shot | top-1 per-class accuracy: 39.22 -> 28.37 | Hurt |
| reduce base classes from 398 to 144 (PAS) (ablation) | PAS · iNat2019-CL all-way 5-shot | top-1 per-class accuracy: 43.27 -> 37.31 | Hurt |
| add PAS with unfiltered pseudo-labels on unseen supercategories | representation learning · iNat2019-CL (unseen supercategories) all-way 5-shot | top-1 per-class accuracy: 35.72 -> 40.73 | Improved |
| use coarser labels (kingdom or order) in PAS filtering instead of genus (ablation) | PAS · iNat2019-CL | top-1 per-class accuracy: AMI 0.73 -> 0.64 -> 0.45 | Hurt |
Closest papers by meaning
- Boosting Few-Shot Visual Learning With Self-Supervision (ICCV 2019 · similarity 0.89)
- Few-Shot Learning via Learning the Representation, Provably (ICLR 2021 · similarity 0.88)
- Self-Supervision Can Be a Good Few-Shot Learner (ECCV 2022 · similarity 0.88)
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.86)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · similarity 0.86)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 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.
