Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Understanding Few-Shot Learning: Measuring Task Relatedness and Adaptation Difficulty via Attributes
Minyang Hu, Hong Chang, Zong Guo, Bingpeng Ma, Shiguang Shan, Xilin CHEN
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 |
|---|---|---|---|
| train on 64 selected classes instead of all 100 (remove less-related heterogeneous classes) (ablation) | MatchingNet · miniImageNet->CUB 5-shot | accuracy: 50.0 -> 51.7 | Improved |
| train on 64 selected classes instead of all 100 (remove less-related heterogeneous classes) (ablation) | ProtoNet · miniImageNet->CUB 5-shot | accuracy: 53.7 -> 52.6 | Hurt |
| train on 64 selected classes instead of all 100 (remove less-related heterogeneous classes) (ablation) | RelationNet · miniImageNet->CUB 5-shot | accuracy: 52.6 -> 50.3 | Hurt |
| train on 64 selected classes instead of all 100 (remove less-related heterogeneous classes) (ablation) | MAML · miniImageNet->CUB 5-shot | accuracy: 51.6 -> 49.7 | Hurt |
| train on 64 selected classes instead of all 100 (remove less-related heterogeneous classes) (ablation) | Baseline++ · miniImageNet->CUB 5-shot | accuracy: 52.4 -> 55.5 | Improved |
| add 4 extra labeled samples to hard novel tasks (distance > 0.29, ~5-7% of tasks), test-time intervention | five FSL methods (ProtoNet, RelationNet, MatchingNet, MAML, Baseline++) · miniImageNet->CUB 5-way 1-shot, intervened tasks | accuracy | Improved |
| increase from 1-shot to 5-shot | APNet · CUB 5-way | accuracy: approximately 10% at distance 0.14; around 15% at distance 0.20 | Improved |
| increase average task attribute distance | APNet · CUB and SUN 5-way 1-shot and 5-shot | accuracy | Hurt |
| increase average task attribute distance | MatchingNet, ProtoNet, RelationNet, MAML, Baseline++ · CUB and SUN | accuracy | Hurt |
| increase distance computed with CLIP auto-annotated attributes | five FSL methods · miniImageNet->CUB | accuracy | Hurt |
| remove pattern, shape or texture attributes from distance computation (ablation) | TAD with auto-annotated attributes · miniImageNet->CUB 5-shot | distance-accuracy trend | Mixed |
Closest papers by meaning
- Finding Task-Relevant Features for Few-Shot Learning by Category Traversal (CVPR 2019 · similarity 0.90)
- TADAM: Task dependent adaptive metric for improved few-shot learning (NeurIPS 2018 · similarity 0.88)
- TAFSSL: Task-Adaptive Feature Sub-Space Learning for few-shot classification (ECCV 2020 · similarity 0.87)
- DETA: Denoised Task Adaptation for Few-Shot Learning (ICCV 2023 · similarity 0.86)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.85)
- Towards Few-Shot Adaptation of Foundation Models via Multitask Finetuning (ICLR 2024 · 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.
