Research map / Continual learning / Class-incremental

Dynamical Adapter Fusion: Constructing A Global Adapter for Pre-Trained Model-based Class-Incremental Learning

NeurIPS 2026 · Continual learning › Class-incremental · approach: Exemplar-efficient class-incremental learning · 0 citations (OpenAlex)

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 Dynamical Adapter Fusion (DAF), which fuses a robust running-average initialization, the current task adapter, and the previous global adapter into one exemplar-free global adapter for ViT-based class-incremental learning, with fusion coefficients derived from a PAC-Bayes bound and a Taylor/Fisher approximation.

Problem

Pre-trained-model-based class-incremental learning typically keeps per-task adapters that require retrieval at inference and grow in memory, while naive fusion causes interference and forgetting.

Approach

Fuse the optimized task adapter, the previous global adapter, and a recursively averaged historical adapter (robust initialization) using a fusion coefficient beta_t derived by Lagrange multipliers on a PAC-Bayes-motivated objective with a second-order Taylor approximation and Fisher-diagonal Hessian approximation, so only one global adapter is stored and no retrieval is needed.

Key results

• DAF reaches 66.37% final accuracy on ObjectNet (ViT-B/16-IN21K, 10 tasks), 2.75 points above MOS (63.62%).
• On ImageNet-R, DAF achieves 84.01% average accuracy and 79.63% final accuracy, vs. MOS 82.96% and 77.93%.
• Ablation on ImageNet-R: robust initialization raises the DAF global adapter from 78.50% (random init) to 84.01% final accuracy.

Surprising findings

• DAF with random initialization performs worse (71.52% average accuracy) than the last-task adapter baseline (74.03%), showing naive fusion can be destructive without a good initialization.
• TUNA's global-only adapter reaches only 28.64% final accuracy on ImageNet-A, far below its average accuracy, indicating instability for that method.

Limitations

• Requires clear task boundaries during training for the fusion step.
• Optimality claim depends on second-order Taylor and Fisher-diagonal Hessian approximations, which may fail for irregular loss landscapes.
• Performance is bounded by the frozen pre-trained backbone; only ViT backbones are evaluated.

Experimental setup

Datasets: CIFAR-100 ImageNet-R ImageNet-A ObjectNet

Benchmarks: CIFAR-100 ImageNet-R ImageNet-A ObjectNet

Metrics: average accuracy final task accuracy stability plasticity

Baselines: L2P DualPrompt CODA-Prompt SLCA RanPAC SSIAT SimpleCIL APER + Adapter EASE TUNA MOS FOSTER

Builds on: PAC-Bayes bound MOS TUNA APER Fisher information approximation of Hessian

Closest papers by meaning

• Adapter Merging with Centroid Prototype Mapping for Scalable Class-Incremental Learning (CVPR 2025 · 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)
• Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning (CVPR 2025 · similarity 0.88)
• CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · similarity 0.88)
• Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning (CVPR 2025 · similarity 0.87)