Research map / Continual learning / Class-incremental
Revisiting Class-Incremental Learning with Pre-Trained Models: Generalizability and Adaptivity are All You Need
Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De‐Chuan Zhan, Ziwei Liu
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 |
|---|---|---|---|
| adapt the pre-trained model with Adapter in the first stage and concatenate with frozen PTM features | SimpleCIL · ImageNet-R B0 Inc5 | last-stage accuracy: 54.55 -> 64.33 | Improved |
| adapt with SSF and concatenate with PTM features | SimpleCIL · ObjectNet B0 Inc10 | last-stage accuracy: 53.59 -> 56.64 | Improved |
| adapt with Finetune and concatenate with PTM features | SimpleCIL · ObjectNet B0 Inc10 | last-stage accuracy: 53.59 -> 48.34 | Hurt |
| adapt with Finetune and concatenate with PTM features | SimpleCIL · OmniBenchmark B0 Inc30 | last-stage accuracy: 73.15 -> 65.03 | Hurt |
| replace sequential finetuning with frozen PTM prototype classifier | Finetune · CIFAR100 B0 Inc5 | last-stage accuracy: 20.17 -> 81.26 | Improved |
| use frozen PTM prototype classifier instead of DualPrompt | DualPrompt · CUB200 B0 Inc10 | last-stage accuracy: 66.54 -> 86.73 | Improved |
| add APER SSF to classical CIL methods setting without exemplars | SimpleCIL · ImageNet-A B0 Inc20 | last-stage accuracy: 49.44 -> 52.67 | Improved |
| adapt in all stages instead of only the first (ablation) | APER w/ Adapter · CIFAR100 B0 Inc10 | accuracy | Hurt |
| add adapted model features to PTM features (APER) instead of SimpleCIL-PTM or SimpleCIL-Adapted alone (ablation) | SimpleCIL · ImageNet-A Base100 Inc5 | accuracy | Improved |
| project APER features to 30 dims with PCA (ablation) | APER w/ SSF · ObjectNet B100 Inc5 | accuracy | No clear effect |
| use APER with domain-incremental SSF | SimpleCIL · Office-Home | last-stage accuracy: 75.85 -> 83.14 | Improved |
| VPT-Shallow adaptation | SimpleCIL · Office-Home | last-stage accuracy: 75.85 -> 63.26 | 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 7 of 7 shared results
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
- Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning (CVPR 2025 · similarity 0.90)
- Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning (NeurIPS 2025 · similarity 0.89)
- Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning (CVPR 2024 · similarity 0.89)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.89)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.89)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.89)
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.
