Research map / Continual learning / Class-incremental
Compositional Few-Shot Class-Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Comp-FSCIL, a cognitive-inspired compositional FSCIL method that represents classes as sets of image-patch primitives, scores classes via linear CKA similarity, and adds a primitive reuse loss that replaces primitives with nearest ones from other classes.
Problem
Few-shot class-incremental learning requires learning novel classes from very few samples without forgetting base classes; the paper asks whether human-like compositional learning from reusable visual primitives can make FSCIL more effective and interpretable.
Approach
Treat backbone patch features as candidate primitives and per-class prototype sets as primitive sets; use linear CKA between the two sets as a composition score for classification, and add a primitive reuse loss that replaces each primitive with its nearest primitive from other classes to encourage cross-class reuse.
Key results
- On miniImageNet with ResNet12, Comp-FSCIL reaches 60.61% last-session accuracy vs. 58.31% for NC-FSCIL, with PD 23.39.
- On CUB200 with ResNet18, last-session accuracy is 61.17% vs. 59.58% for CLOM (PD 19.77 vs 19.99).
- On CUB200 with Swin-T, last-session accuracy is 72.80% vs. 70.50% for CLOM, a gain of about 2.3 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the composition module (ablation) | Comp-FSCIL baseline with ResNet18 · CUB200 last incremental session | Overall accuracy: 57.18 -> 59.25 | Helped |
| add primitive reuse after composition (ablation) | Comp-FSCIL with ResNet18 · CUB200 last incremental session | Overall accuracy: 59.25 -> 61.17 | Helped |
| add the composition module (ablation) | Comp-FSCIL baseline with Swin-T · CUB200 last incremental session | Overall accuracy: 69.18 -> 71.42 | Helped |
| add primitive reuse after composition (ablation) | Comp-FSCIL with Swin-T · CUB200 last incremental session | Overall accuracy: 71.42 -> 72.80 | Helped |
| add the composition module (ablation) | Comp-FSCIL baseline with ResNet12 · CIFAR100 last incremental session | Overall accuracy: 53.98 -> 55.30 | Helped |
| add primitive reuse after composition (ablation) | Comp-FSCIL with ResNet12 · CIFAR100 last incremental session | Overall accuracy: 55.30 -> 59.05 | Helped |
| add the composition module (ablation) | Comp-FSCIL baseline with ResNet12 · miniImageNet last incremental session | Overall accuracy: 57.72 -> 58.84 | Helped |
| add primitive reuse after composition (ablation) | Comp-FSCIL with ResNet12 · miniImageNet last incremental session | Overall accuracy: 58.84 -> 60.61 | Helped |
| replace CKA with power transformed CKA | baseline-trained model evaluated on CIFAR100 · CIFAR100 | Last session accuracy (%): 38.42 -> 39.47 | Helped |
| increase the primitive count from 1 to 16 of the full ratio (ablation) | Comp-FSCIL · CIFAR100 | All Classes accuracy: 33.20 -> 51.61 | Helped |
| increase the primitive count beyond ratio 16 (ablation) | Comp-FSCIL · CIFAR100 | All Classes accuracy: 51.61; 51.49; 51.28; 51.23; 51.60; 51.06; 51.37 | No clear effect |
Surprising findings
- Plain CKA-based comparison without training on it gave 38.42% last-session accuracy on CIFAR100, versus 24.03% for DeepEMD and 18.15% for FRN on a baseline-trained model, with lower time cost.
- Replacing up to about 80% of novel-class primitives with nearest base-class primitives causes little performance drop, suggesting only about 20% of primitives need relearning.
Limitations
- Authors note neglect of many-shot scenarios where primitive updates cannot be ignored.
- Interpretability is shown via qualitative primitive retrieval and CAM visualizations rather than formal evaluation.
Experimental setup
Datasets: miniImageNet CIFAR100 CUB200-2011
Benchmarks: miniImageNet CIFAR100 CUB200-2011
Metrics: last-session accuracy performance drop (PD) base accuracy novel accuracy
Baselines: DeepEMD CLOM SoftNet ALICE SAVC NC-FSCIL WaPR MetaFSCIL GKEAL CPDE RPC
Builds on: CEC NC-FSCIL CLOM CKA
Closest papers by meaning
- Retrieval-Augmented Primitive Representations for Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.88)
- ProCC: Progressive Cross-Primitive Compatibility for Open-World Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.87)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.87)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.87)
- Revealing the Proximate Long-Tail Distribution in Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.87)
- Not Just Object, But State: Compositional Incremental Learning without Forgetting (NeurIPS 2024 · similarity 0.87)
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.