Dynamic experts improve learning new classes without forgetting old
Dynamic LoRA-Experts and Prototype-Ensemble Matching for Class-Incremental Learning
Machine Learning
Summary
Learning new categories without forgetting old ones is a tough problem for AI. The authors created a method called Dynamic LoRA-Experts and Prototype-Ensemble Matching that assigns a special expert for each new group of categories, helping avoid confusion between old and new knowledge. Their method also combines different types of memory representations to better recognize tasks. Tests show this method works well for learning both many new classes at once and tiny new sets with few examples.
What this means in practice
- •For machine learning engineers: Improve continual learning systems to add new classes without retraining on all past data or forgetting old ones.
- •For ai model deployment teams: Deploy adaptable models that efficiently fine-tune per task with low memory and reduced interference for long-lived AI services.
Authors
Hongwei Zhao, Rui Liu, Yansong Liu
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.