Research map / Continual learning / Class-incremental
External Knowledge Injection for CLIP-Based Class-Incremental Learning
Da-Wei Zhou, Kai-Wen Li, Jingyi Ning, Han-Jia Ye, Lijun Zhang, De-Chuan Zhan
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 visual knowledge injection (augmentation-based unit) to zero-shot CLIP (ablation) | ENGINE · CIFAR100 B0 Inc10 | accuracy | Improved |
| add textual (GPT-4 description) injection on top of visual injection (ablation) | ENGINE · CIFAR100 B0 Inc10 | accuracy | Improved |
| add post-tuning knowledge injection (top-k pairwise re-ranking) (ablation) | ENGINE · CIFAR100 B0 Inc10 | accuracy | Improved |
| vary prototype noise alpha in {0.15..0.35} and top-k in {3..7} (ablation) | ENGINE · CUB B0 Inc20 | last accuracy | No clear effect |
| replace GPT-4o mini descriptions with Qwen2.5-72B (ablation) | ENGINE · Food B0 Inc10 | accuracy | No clear effect |
| give baselines the same GPT-generated text and augmentation as ENGINE (ablation) | RAPF vs ENGINE · ImageNet-R B0 Inc20 | average accuracy: 81.89 -> 86.22 | Improved |
| give ZS-CLIP the same GPT-generated text (ablation) | ZS-CLIP vs ENGINE · CIFAR B0 Inc10 | average accuracy: 82.26 -> 86.92 | Improved |
| ENGINE without exemplars versus exemplar-based PROOF (20/class) | ENGINE · Aircraft B0 Inc10 | average accuracy: 61.00 -> 69.69 | Improved |
| ENGINE without exemplars versus exemplar-based iCaRL (20/class) | ENGINE · SUN B0 Inc30 | average accuracy: 78.56 -> 85.04 | Improved |
| ENGINE versus strongest baseline | ENGINE · ObjectNet B0 Inc20 | average accuracy: 52.62 -> 59.11 | Improved |
| use OpenAI pretrained CLIP weights instead of LAION-400M (ablation) | ENGINE · UCF B0 Inc10 | accuracy | No clear effect |
| ENGINE vs Finetune CLIP in each stage | Finetune · Aircraft B0 Inc10 | average accuracy: 69.69 -> 3.16 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DualPrompt — DualPrompt: Complementary Prompting for Rehearsal-Free Continual Learning (ECCV 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SAME: Sparse and Anchored Model Editing for Heterogeneous Incremental Learning under Limited Data (CVPR 2026 · similarity 0.87)
- Class-Incremental Learning with CLIP: Adaptive Representation Adjustment and Parameter Fusion (ECCV 2024 · similarity 0.87)
- Hierarchical Semantic Tree Anchoring for CLIP-Based Class-Incremental Learning (NeurIPS 2026 · similarity 0.87)
- Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning (CVPR 2025 · similarity 0.86)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.85)
- PEARL: Input-Agnostic Prompt Enhancement with Negative Feedback Regulation for Class-Incremental Learning (AAAI 2025 · similarity 0.85)
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.
