Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
MRSP: Learn Multi-Representations of Single Primitive for Compositional Zero-Shot Learning
Dongyao Jiang, Hui Chen, Haodong Jing, Yongqiang Ma, Nanning Zheng
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 |
|---|---|---|---|
| learn multiple representations per primitive (state, object, composition nodes) instead of one (ablation) | MRSP with cross-attention · MIT-States closed-world | AUC: 12.6 -> 23.8 | Improved |
| use cross-attention decoder instead of 3-layer MLP (ablation) | MRSP with single representation · MIT-States closed-world | AUC: 8.2 -> 12.6 | Improved |
| add pre-norm in GCN (ablation) | MRSP · MIT-States closed-world | AUC: 20.0 -> 23.8 | Improved |
| widen GCN first layer from 4096 to 8192 (ablation) | MRSP · MIT-States closed-world | AUC: 20.9 -> 23.8 | Improved |
| deepen GCN beyond 2 layers (ablation) | MRSP · MIT-States closed-world | AUC: 23.8 -> 22.5 (d8192 x 2) | Hurt |
| use 3 MLP layers instead of 2 in cross-attention (ablation) | MRSP · MIT-States closed-world | AUC: 23.8 -> 22.5 | Hurt |
| use 8 or 32 attention heads instead of 16 (ablation) | MRSP · MIT-States closed-world | AUC: 23.8 -> 22.4 (8 heads), 22.0 (32 heads) | Hurt |
| combine primitive scores additively instead of multiplicatively at inference (ablation) | MRSP · MIT-States closed-world | AUC: 19.7 -> 23.8 | Improved |
| use stronger backbone ViT-L/14 instead of ResNet18 (ablation) | MRSP · MIT-States closed-world | AUC: 9.9 -> 23.8 | Improved |
| MRSP versus Troika on best unseen accuracy | MRSP · MIT-States closed-world | best unseen accuracy: 53.0 -> 52.7 | No clear effect |
Closest papers by meaning
- ProCC: Progressive Cross-Primitive Compatibility for Open-World Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.93)
- Retrieval-Augmented Primitive Representations for Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.92)
- Beyond Seen Primitive Concepts and Attribute-Object Compositional Learning (CVPR 2024 · similarity 0.91)
- A Dynamic Learning Method towards Realistic Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.89)
- Beyond Image Classification: A Video Benchmark and Dual-Branch Hybrid Discrimination Framework for Compositional Zero-Shot Learning (CVPR 2025 · similarity 0.89)
- Compositional Few-Shot Class-Incremental Learning (ICML 2024 · similarity 0.89)
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.
