Research map / Detection and segmentation / Open-set and few-shot detection
Unlocking Compositional Generalization in Continual Few-Shot 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 COMPOSE, a dual-phase continual few-shot learner that trains slot-based holistic class prototypes on a frozen self-supervised ViT and composes slots via Chamfer matching only at inference, reaching 94.14% noc on CGQA.
Problem
Continual few-shot learners using global embeddings or part-level matching objectives fail to generalize to novel concepts (non-compositional split) and suffer forgetting.
Approach
Strictly decouple representation learning from compositional inference: train slots holistically with prototype cross-entropy plus a cross-correlation decorrelation penalty on a frozen DINOv2 backbone with attention-weighted slot aggregates, then compose slots with bidirectional Chamfer matching at inference without gradient updates.
Key results
- COMPOSE reaches 94.14% noc and 92.92 H_a on CGQA, surpassing CoFiMA's 91.97% noc by 2.17 pp.
- On COBJ, COMPOSE achieves 88.42% noc but trails EASE on H_a (76.83 vs 80.85).
- On FSCIL CIFAR-100, forgetting drops from 27.91% (Comp-FSCIL) to 3.54% with best AA of 84.02%.
Surprising findings
- Novel-concept accuracy (noc) exceeds systematicity accuracy (sys) on CGQA (94.14 vs 90.27), which the authors attribute to DINOv2 pretraining rather than the CGQA curriculum.
- Adding Chamfer matching during training (COMPOSE-CT) lowers noc by 2.51 pp on CGQA, so part-level matching is kept at inference only.
Limitations
- Depends on patch-level semantic geometry; degrades with supervised ViTs, CLIP-style backbones, or domain shift.
- Underperforms EASE on H_a for COBJ natural images, with slot quality identified as the bottleneck.
Experimental setup
Datasets: CGQA COBJ CIFAR-100 ImageNet-R ImageNet-A miniImageNet CUB-200 COCO 2017
Benchmarks: CGQA COBJ CIFAR-100 ImageNet-R ImageNet-A miniImageNet CUB-200
Metrics: noc accuracy H_a sys pro sub non average accuracy forgetting
Baselines: DINOv2-Proto SimpleShot TransSlot MetaProtoNet RanPAC EASE CoFiMA Comp-FSCIL CPrompt ADAM FOSTER DER
Builds on: Slot Attention DINOSAUR Compositional Few-Shot Transfer (CFST) Barlow Twins DINOv2
Closest papers by meaning
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.90)
- Provable Compositional Generalization for Object-Centric Learning (ICLR 2024 · similarity 0.90)
- Composition-Incremental Learning for Compositional Generalization (AAAI 2026 · similarity 0.89)
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.88)
- Revealing the Proximate Long-Tail Distribution in Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.87)
- Revisiting Continuity of Image Tokens for Cross-domain Few-shot Learning (ICML 2025 · 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.