Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Compositional Prototypical Networks 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 |
|---|---|---|---|
| Learn attribute component prototypes and reuse them to construct class prototypes (ablation) | CPN · CUB and SUN few-shot classification | accuracy: 79.62 ± 0.27 -> 82.04 ± 0.24 (CUB 5-way 1-shot); 71.21 ± 0.29 -> 85.63 ± 0.19 (SUN 5-way 1-shot) | Helped |
| Use randomly initialized component prototypes instead of learned prototypes (ablation) | CPN · CUB and SUN few-shot classification | accuracy: 79.62 ± 0.27 -> 19.43 ± 0.21 (CUB 5-way 1-shot); 71.21 ± 0.29 -> 20.96 ± 0.20 (SUN 5-way 1-shot) | Hurt |
| Adaptively fuse randomly initialized component prototypes with the visual prototype (ablation) | CPN · CUB and SUN few-shot classification | accuracy: 79.62 ± 0.27 -> 79.66 ± 0.27 (CUB 5-way 1-shot); 71.21 ± 0.29 -> 71.21 ± 0.29 (SUN 5-way 1-shot) | No clear effect |
| Adaptively fuse learned component and visual prototypes rather than use either alone (ablation) | CPN · CUB and SUN 5-way 1-shot | accuracy: 79.62 ± 0.27 and 82.04 ± 0.24 -> 84.54 ± 0.22 (CUB); 71.21 ± 0.29 and 85.63 ± 0.19 -> 86.54 ± 0.18 (SUN) | Helped |
| Further optimize learned component prototypes during meta-training (ablation) | CPN · CUB and SUN few-shot classification | accuracy: 84.54 ± 0.22 -> 87.29 ± 0.20 (CUB 5-way 1-shot); 86.54 ± 0.18 -> 88.07 ± 0.17 (SUN 5-way 1-shot) | Helped |
| Replace adaptive fusion with naive concatenation of visual and learned compositional prototypes (ablation) | CPN · CUB and SUN few-shot classification | accuracy: 85.27 ± 0.21 -> 87.29 ± 0.20 (CUB 5-way 1-shot); 86.61 ± 0.18 -> 88.07 ± 0.17 (SUN 5-way 1-shot) | Hurt |
| Use visual plus compositional prototype as the weight-generator input instead of compositional prototype alone (ablation) | CPN weight generator · CUB and SUN 5-way 1-shot | accuracy: 87.29 ± 0.20 vs 87.01 ± 0.20 (CUB); 88.07 ± 0.11 vs 87.26 ± 0.17 (SUN) | Mixed |
| Use visual plus compositional prototype as the weight-generator input instead of compositional prototype alone (ablation) | CPN weight generator · CUB and SUN 5-way 5-shot | accuracy: 92.54 ± 0.14 vs 92.77 ± 0.13 (CUB); 89.21 ± 0.15 vs 89.47 ± 0.15 (SUN) | Mixed |
| Replace Conv4 with ResNet12 | CPN on SUN · SUN 5-way 1-shot and 5-way 5-shot | accuracy: 7.62% and 7.65% performance gains | Helped |
Closest papers by meaning
- Compositional Few-Shot Class-Incremental Learning (ICML 2024 · similarity 0.89)
- Envisioning Class Entity Reasoning by Large Language Models for Few-shot Learning (AAAI 2025 · similarity 0.87)
- A Robust Prototype-Based Network with Interpretable RBF Classifier Foundations (AAAI 2025 · similarity 0.87)
- MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes (CVPR 2024 · similarity 0.86)
- Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.86)
- Revealing the Proximate Long-Tail Distribution in Compositional Zero-Shot Learning (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.