Research map / Continual learning / Class-incremental
Representation-Steered Incremental Adapter-Tuning for Class-Incremental Learning with Pre-Trained Models
Jiarui Zhao, Libo Huang, Xiangqi Li, Zhulin An, Chuanguang Yang, Yu Wang, Boyu Diao, Yongjun Xu
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 orthogonal loss to the shared-adapter baseline (ablation) | SSIAT-style baseline · ImageNet-R B0I20 | average accuracy: 83.63 -> 84.15 (avg), 79.38 -> 78.32 (final) | Mixed |
| add L2 alignment loss with residual autoencoder projector (ablation) | SSIAT-style baseline · ImageNet-A B0I20 | final accuracy: 62.43 -> 65.04 | Improved |
| combine orthogonal and alignment losses (IRS) (ablation) | SSIAT-style baseline · ImageNet-R B0I20 | final accuracy: 79.38 -> 80.52 | Improved |
| keep residual identity path in the projector (ablation) | IRS projector · ImageNet-A B0I20 | final accuracy: 51.48 -> 65.31 | Improved |
| add base representation-steering loss L_RS (no warm-up) (ablation) | RSIAT with IRS · ImageNet-R B0I20 | average accuracy: 84.86 -> 85.36 | Improved |
| add warm-up schedule for the RS coefficient (ablation) | RSIAT · ImageNet-R B0I20 | average accuracy: 85.36 -> 86.92 | Improved |
| add warm-up schedule for the RS coefficient (ablation) | RSIAT · ImageNet-A B0I20 | final accuracy: 65.50 -> 66.23 | Improved |
| increase adapter bottleneck from 64 to 256 (dp=384) (ablation) | RSIAT · ImageNet-R B0I20 | average accuracy: 86.92 vs 87.04 | No clear effect |
| use RSIAT (exemplar-free) instead of exemplar-based traditional CIL methods | ViT-B/16-IN21K · CIFAR100 B0I10 | final accuracy: 89.26 -> 92.20 | Improved |
| replace SSIAT with RSIAT | ViT-B/16-IN21K · ImageNet-A B0I20 | average accuracy: 70.83 -> 74.89 | Improved |
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.
- RanPAC — RanPAC: Random Projections and Pre-trained Models for Continual Learning (NeurIPS 2023) · 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
- Adapter Merging with Centroid Prototype Mapping for Scalable Class-Incremental Learning (CVPR 2025 · similarity 0.88)
- Dynamic Integration of Task-Specific Adapters for Class Incremental Learning (CVPR 2025 · similarity 0.87)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.87)
- Dynamical Adapter Fusion: Constructing A Global Adapter for Pre-Trained Model-based Class-Incremental Learning (NeurIPS 2026 · similarity 0.87)
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.87)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 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.
