🤖 AI Summary
This study addresses the challenges of cross-task cumulative interference and inference alignment deviation in parameter-efficient fine-tuning for class-incremental learning. To this end, we propose DLEPEM, a data-free replay framework that introduces a dynamic LoRA expert allocation mechanism to construct task-specific low-rank adaptation modules, thereby mitigating catastrophic forgetting. Furthermore, a prototype ensemble matching strategy is designed to integrate prototypes from the frozen pre-trained model with those from adaptive LoRA experts, enabling reliable task-level discrimination. Extensive experiments demonstrate that DLEPEM achieves superior performance on both standard and few-shot class-incremental learning benchmarks.
📝 Abstract
Class-Incremental Learning (CIL) aims to continuously learn new classes without forgetting previously acquired knowledge. Parameter-efficient fine-tuning with pre-trained models reduces parameter overhead but can suffer from cumulative interference and suboptimal alignment between inference samples and specialized modules. We propose Dynamic LoRA-Experts and Prototype-Ensemble Matching (DLEPEM), a two-stage rehearsal-free framework. DLEPEM allocates a task-specific LoRA-Expert for each incremental task to reduce cross-task interference, then combines frozen pre-trained-model prototypes with task-adaptive LoRA-Expert prototypes for reliable task-level discrimination. Experiments on standard CIL and Few-Shot CIL benchmarks demonstrate strong performance under the evaluated protocols.