Research map / Continual learning / Class-incremental
Language-Inspired Relation Transfer for Few-Shot Class-Incremental Learning
Yifan Zhao, Jia Li, Zeyin Song, Yonghong Tian
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add text-vision prototypical fusion with fixed prompt to visual finetuning (ablation) | LRT base (CLIP ResNet-50) · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 57.16 -> 64.98 | Improved |
| replace fixed handcrafted prompt with learnable context prompt (PT) (ablation) | LRT with M_fuse · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 64.98 -> 67.01 | Improved |
| add imagined contrastive loss (ablation) | LRT with PT+FT+M_fuse · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 67.01 -> 67.88 | Improved |
| add graph relation transformation module M_graph (ablation) | LRT with PT+FT+M_fuse · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 67.88 -> 75.94 | Improved |
| use text prompt tuning only with fixed vision encoder instead of visual finetuning (ablation) | LRT base · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 57.16 -> 52.87 | Hurt |
| fuse prototypes by simple averaging (Proto Add.) instead of learnable tau (ablation) | M_graph+M_fuse · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 75.94 -> 51.25 | Hurt |
| fix tau = 1 (Static) instead of learnable (ablation) | M_graph+M_fuse · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 75.94 -> 51.10 | Hurt |
| add M_fuse to M_graph only (ablation) | LRT · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 72.61 -> 75.94 | Improved |
| increase incremental shots from 1 to 5 (ablation) | LRT · miniImageNet FSCIL (5-way 5-shot) | incremental-class accuracy (%): 17.12 -> 39.32 | Improved |
| increase incremental shots from 1 to 5 (ablation) | LRT · miniImageNet FSCIL (5-way 5-shot) | base-class accuracy (%): 86.40 -> 82.68 | Mixed |
| scale loss weights lambda_im or lambda_m by 0.1x or 10x (ablation) | LRT · miniImageNet FSCIL (5-way 5-shot) | average accuracy (%): 75.94 -> 71.89 | Hurt |
| increase prompt length from 2 to 4 (ablation) | LRT · miniImageNet FSCIL (5-way 5-shot) | base/incremental accuracy (%): 82.68 vs 84.28 base, 35.42 -> 39.32 incremental | Mixed |
Closest papers by meaning
- VT-FSL: Bridging Vision and Text with LLMs for Few-Shot Learning (NeurIPS 2025 · similarity 0.87)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.86)
- Remodeling Semantic Relationships in Vision-Language Fine-Tuning (AAAI 2026 · similarity 0.86)
- Envisioning Class Entity Reasoning by Large Language Models for Few-shot Learning (AAAI 2025 · similarity 0.86)
- A Vision Check-up for Language Models (CVPR 2024 · similarity 0.86)
- RelationAdapter: Learning and Transferring Visual Relation with Diffusion Transformers (NeurIPS 2025 · 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.
