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

2026-01-29

AI Summary

This work addresses the limitations of task-specific adapters in class-incremental learning, which suffer from restricted knowledge transfer, high retrieval overhead, and catastrophic forgetting induced by parameter fusion. To overcome these challenges, the authors propose a dynamic adapter fusion mechanism grounded in PAC-Bayes theory. Operating under a frozen pre-trained model, the method dynamically computes fusion coefficients via Taylor expansion and incorporates a global knowledge-aware robust initialization strategy to effectively integrate task-specific, historical global, and initial parameters. This approach balances stability and plasticity, preserving learned knowledge while enabling adaptation to new tasks. Extensive experiments demonstrate that the proposed method significantly outperforms existing approaches across multiple class-incremental learning benchmarks, achieving state-of-the-art performance.

Technology Category

Machine Learning: Transfer, Domain Adaptation, Multi-Task LearningSearch and Optimization: Learning to SearchConstraint Satisfaction and Optimization: Constraint Learning and Acquisition

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingUser Modeling, Personalization and Recommendation: Federated recommendation systems and personalizationSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Class-Incremental Learning (CIL) requires models to continuously acquire new classes without forgetting previously learned ones. A dominant paradigm involves freezing a pre-trained model and training lightweight, task-specific adapters. However, maintaining task-specific parameters hinders knowledge transfer and incurs high retrieval costs, while naive parameter fusion often leads to destructive interference and catastrophic forgetting. To address these challenges, we propose Dynamical Adapter Fusion (DAF) to construct a single robust global adapter. Grounded in the PAC-Bayes theorem, we derive a fusion mechanism that explicitly integrates three components: the optimized task-specific adapter parameters, the previous global adapter parameters, and the initialization parameters. We utilize the Taylor expansion of the loss function to derive the optimal fusion coefficients, dynamically achieving the best balance between stability and plasticity. Furthermore, we propose a Robust Initialization strategy to effectively capture global knowledge patterns. Experiments on multiple CIL benchmarks demonstrate that DAF achieves state-of-the-art (SOTA) performance.

Problem

• Class-Incremental Learning
• Catastrophic Forgetting
• Adapter Fusion
• Pre-Trained Models
• Knowledge Transfer

Innovation

• Dynamical Adapter Fusion
• Class-Incremental Learning
• PAC-Bayes
• Adapter Fusion
• Robust Initialization