Research map / Continual learning / Class-incremental
Adapter Merging with Centroid Prototype Mapping for Scalable Class-Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ACMap, an exemplar-free class-incremental learning method that merges task-specific adapters into one shared adapter via average merging and aligns old prototypes with centroid prototype mapping, giving constant inference time.
Problem
Class-incremental learning with pre-trained models must balance catastrophic forgetting, exemplar-free privacy constraints, and inference scalability, since methods using per-task adapters or prompts slow down as tasks grow.
Approach
Sequentially train an adapter per task starting from the first task's weights, merge them incrementally by weight averaging into a single adapter, and map previous-task prototypes into the current merged subspace by shifting them with the centroid difference of the current task's prototypes computed under the current and original adapters.
Key results
- On ImageNet-R B0 Inc5, ACMap reaches 65.19 average and 56.19 final accuracy vs. EASE's 65.34 / 55.04 (L=10 variant: 65.14 / 56.04).
- Inference time on task 40 of ImageNet-R is 23.5 s for ACMap vs. 916.5 s for EASE, a 39x speedup.
- On CIFAR-100 B0 Inc5, ACMap achieves 92.04 average and 87.81 final accuracy, slightly above EASE's 91.51 / 85.80.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add centroid prototype mapping (ablation) | ACMap with initial weight replacement · CIFAR-100 B0 Inc5 | final accuracy: 86.85 -> 87.73 | Helped |
| add centroid prototype mapping (ablation) | ACMap with initial weight replacement · ImageNet-R B0 Inc5 | final accuracy: 69.88 -> 70.25 | Helped |
| add initial weight replacement (ablation) | ACMap with centroid prototype mapping · CIFAR-100 B0 Inc5 | final accuracy: 86.85 -> 87.73 | Helped |
| add initial weight replacement (ablation) | ACMap with centroid prototype mapping · ImageNet-R B0 Inc5 | final accuracy: 69.80 -> 70.25 | Helped |
| use early stopping at L = 10 instead of no early stopping (ablation) | ACMap · CIFAR-100 B0 Inc5 | final accuracy: 87.80 -> 87.76 | No clear effect |
| use early stopping at L = 10 instead of no early stopping (ablation) | ACMap · ImageNet-R B0 Inc5 | final accuracy: 70.49 -> 70.25 | No clear effect |
| replace task-specific adapters with one merged adapter | ACMap versus EASE · ImageNet-R B0 Inc5 | final accuracy: 70.58 -> 70.49 | Hurt |
| replace task-specific adapters with one merged adapter | ACMap versus EASE · VTAB B0 Inc10 | final accuracy: 93.55 -> 87.56 | Hurt |
| use one merged adapter instead of EASE's task-specific adapters | ACMap versus EASE · ImageNet-R task 40 | inference time: 916.5 -> 23.5 | Helped |
Surprising findings
- On VTAB B0 Inc10, ACMap underperforms EASE (91.21 vs. 93.61 average), attributed to the heterogeneous domains and an overfitting-prone fourth task.
Limitations
- Performance lags EASE on VTAB, where tasks come from very different domains.
- Accuracy declines for later tasks on VTAB due to overfitting on a large fourth task.
- Centroid prototype mapping relies on the assumption that alignment computed for the current task transfers to previous tasks.
Experimental setup
Datasets: CIFAR-100 CUB ImageNet-R ImageNet-A VTAB
Benchmarks: CIFAR-100 CUB ImageNet-R ImageNet-A VTAB
Metrics: average accuracy final accuracy inference time
Baselines: Finetune Finetune Adapter L2P DualPrompt CODA-Prompt SimpleCIL APER EASE
Builds on: EASE APER SimpleCIL model averaging (model soups)
Closest papers by meaning
- Dynamical Adapter Fusion: Constructing A Global Adapter for Pre-Trained Model-based Class-Incremental Learning (NeurIPS 2026 · similarity 0.89)
- MOS: Model Surgery for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.89)
- Dynamic Integration of Task-Specific Adapters for Class Incremental Learning (CVPR 2025 · similarity 0.89)
- CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · similarity 0.88)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.87)
- Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning (CVPR 2024 · 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.