Research map / Continual learning / Class-incremental
Feature Decomposition-Recomposition in Large Vision-Language Model for Few-Shot Class-Incremental Learning
Zongyao Xue, Meina Kan, Shiguang Shan, Xilin Chen
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 holistic prototype calibration with similar base prototypes (ablation) | Prototypical Network on CLIP · CUB200-1-shot | NAcc: 19.86 -> 31.56 | Improved |
| replace holistic calibration with fine-grained keyword-segment recomposition (ablation) | FDR · CUB200-1-shot | NAcc: 31.56 -> 38.40 | Improved |
| ground keyword segments with CLIP attention and value matrices instead of random segments (ablation) | FDR · CUB200-1-shot | NAcc: 32.08 -> 38.40 | Improved |
| use attention-based keyword grounding instead of naive embedding product (ablation) | FDR · CUB200-1-shot | NAcc: 36.25 -> 38.40 | Improved |
| increase segment size q to 200 or 512 (ablation) | FDR · CUB200-1-shot | NAcc | Hurt |
| increase number of base segments r beyond 50 (ablation) | FDR · CUB200-1-shot | NAcc | No clear effect |
| set calibration ratio alpha to 0.5 versus lower or higher (ablation) | FDR · CUB200-1-shot | NAcc | Improved |
| apply FDR calibration instead of TEEN on frozen CLIP | CUB200 5-shot · CUB200-5-shot | NAcc: 58.83 -> 62.79 | Improved |
| apply FDR in trained setting versus EASE on miniImageNet | FDR · miniImageNet (5-shot) | AvgAcc across sessions: 89.2 vs 93.9 for EASE | Hurt |
Closest papers by meaning
- Few-Shot Class-Incremental Learning via Training-Free Prototype Calibration (NeurIPS 2023 · similarity 0.91)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.90)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.89)
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.89)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- VT-FSL: Bridging Vision and Text with LLMs for Few-Shot Learning (NeurIPS 2025 · similarity 0.88)
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.
