Research map / Continual learning / Class-incremental
Random Amalgamation of Adapters for Flatter Loss Landscapes: Towards Class-Incremental Learning with Better Stability
Yao Deng, Xiang Xiang, Jiaqi Gui
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 randomized adapter amalgamation (RAA) (ablation) | adapter CIL baseline · ImageNetA | last accuracy: 62.43 -> 64.01 | Improved |
| add randomized adapter amalgamation (RAA) (ablation) | adapter CIL baseline · ImageNetR | last accuracy: 79.38 -> 80.22 | Improved |
| replace SAM with SAM+ (ablation) | adapter CIL baseline · ImageNetA | last accuracy: 63.53 -> 63.86 | Improved |
| add vanilla SAM (ablation) | adapter CIL baseline · ImageNetA | last accuracy: 62.43 -> 63.53 | Improved |
| add SAM+ on top of RAA (ablation) | RAA · ImageNetR | last accuracy: 80.22 -> 80.31 | Improved |
| add SAM+ plug-in to existing CIL methods | Ease, RanPAC, SLCA, SSIAT and others · mean of 4 datasets | last accuracy: Ease 75.88 -> 76.35 | Improved |
| increase adapters from 2 to 4 (ablation) | RAA · ImageNetR | last accuracy: 80.31 -> 78.05 | Hurt |
| increase adapters from 2 to 3 (ablation) | RAA · CUB200 | last accuracy: 89.94 -> 90.18 | No clear effect |
| RAA vs strongest baselines | RAA · five CIL benchmarks | average of last accuracy: 78.24 -> 79.73 | Improved |
| ViT-B/16-1K instead of 21K backbone (ablation) | adapter CIL · ImageNetA | last accuracy: 62.43 -> 64.88 (ImageNetA); 91.48 -> 90.72 (CIFAR100) | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RanPAC — RanPAC: Random Projections and Pre-trained Models for Continual Learning (NeurIPS 2023) · better on 7 of 7 shared results
- SLCA — SLCA: Slow Learner with Classifier Alignment for Continual Learning on a Pre-trained Model (ICCV 2023) · better on 5 of 5 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
- Adapter Merging with Centroid Prototype Mapping for Scalable Class-Incremental Learning (CVPR 2025 · similarity 0.88)
- Dynamical Adapter Fusion: Constructing A Global Adapter for Pre-Trained Model-based Class-Incremental Learning (NeurIPS 2026 · similarity 0.87)
- Flatten Long-Range Loss Landscapes for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.87)
- MOS: Model Surgery for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.86)
- Make Continual Learning Stronger via C-Flat (NeurIPS 2024 · similarity 0.85)
- Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes (CVPR 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.
