Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Adaptive Poincaré Point to Set Distance for Few-Shot Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use APP2S instead of the strongest listed baseline | ResNet-12 few-shot classifier · mini-ImageNet 5-way 5-shot | accuracy: 82.71±0.31 -> 83.42±0.15 | Helped |
| Use APP2S instead of the strongest listed baseline | ResNet-12 few-shot classifier · tiered-ImageNet 5-way 5-shot | accuracy: 86.03±0.58 -> 86.23±0.15 | Helped |
| Replace point-to-point prototype distance with point-to-set distance (ablation) | Hyperbolic ProtoNet · tiered-ImageNet 5-way 5-shot | accuracy: 79.11±0.22 -> 83.14±0.17 | Helped |
| Add adaptive relation weights fφ (ablation) | Hyperbolic P2S without fφ · tiered-ImageNet 5-way 5-shot | accuracy: 83.14±0.17 -> 84.88±0.17 | Helped |
| Add signature generator fω (ablation) | Hyperbolic P2S with fφ · tiered-ImageNet 5-way 5-shot | accuracy: 84.88±0.17 -> 86.23±0.15 | Helped |
| Use hyperbolic rather than Euclidean geometry (ablation) | Adaptive P2S · tiered-ImageNet 5-way 5-shot | accuracy: 81.96±0.18 -> 86.23±0.15 | Helped |
| Add learned set-to-set distance fζ using local feature maps (ablation) | APP2S without fζ · tiered-ImageNet 5-way 5-shot | accuracy: 84.12±0.13 -> 86.23±0.15 | Helped |
| Use local feature maps instead of global feature vectors (ablation) | APP2S · mini-ImageNet 5-way 5-shot | accuracy: 80.13±0.11 -> 83.42±0.15 | Helped |
| Increase Poincaré-ball curvature from 0.01 to 0.7 (ablation) | APP2S · tiered-ImageNet 5-way 5-shot | accuracy: 84.19 -> 86.23 | Helped |
| Apply image flipping augmentation (ablation) | Hyperbolic ProtoNet · mini-ImageNet 5-way 1-shot | accuracy: *60.65 -> *59.97 | Hurt |
| Apply image flipping augmentation (ablation) | Hyperbolic ProtoNet · tiered-ImageNet 5-way 1-shot | accuracy: *67.38 -> *67.84 | No clear effect |
| Use APP2S rather than Hyperbolic ProtoNet | ResNet-12 classifier under outlier contamination · tiered-ImageNet 5-way 5-shot with outliers | accuracy | Helped |
Closest papers by meaning
- TADAM: Task dependent adaptive metric for improved few-shot learning (NeurIPS 2018 · similarity 0.89)
- Improved Few-Shot Visual Classification (CVPR 2020 · similarity 0.88)
- Hyperbolic Vision Transformers: Combining Improvements in Metric Learning (CVPR 2022 · similarity 0.88)
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.86)
- Unsupervised Feature Learning with Emergent Data-Driven Prototypicality (CVPR 2024 · similarity 0.85)
- Robust Hyperbolic Learning with Curvature-Aware Optimization (NeurIPS 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.