Chunhui Ding

Aojun Lu, Tao Feng, Hangjie Yuan et al.

Continual Learning (CL) seeks to enable neural networks to incrementally acquire new knowledge (plasticity) while retaining existing knowledge (stability). Although pre-trained models (PTMs) have provided a strong foundation for CL, existing approaches face a fundamental challenge in balancing these two competing objectives. Current methods typically address stability by freezing the PTM backbone, which severely limits the model's plasticity, particularly when incoming data distribution diverges largely from the pre-training data. Alternatively, sequentially fine-tuning the entire PTM can adapt to new knowledge but often leads to catastrophic forgetting, highlighting the critical stability-plasticity trade-off in PTM-based CL. To address this limitation, we propose Adapting PTMs before the core CL} process (ACL), a novel framework that introduces a plug-and-play adaptation phase prior to learning each new task. During this phase, ACL refines the PTM backbone by aligning embeddings with their original class prototypes while distancing them from irrelevant classes. This mechanism theoretically and empirically demonstrates desirable balance between stability and plasticity, significantly improving CL performance across benchmarks and integrated methods. Code is available at https://github.com/byyx666/ACL_code.

Aojun Lu, Junchao Ke, Chunhui Ding et al.

Deep neural networks, despite their remarkable success, remain fundamentally limited in their ability to perform Continual Learning (CL). While most current methods aim to enhance the capabilities of a single model, Inspired by the collective learning mechanisms of human populations, we introduce Evolving Continual Learning (ECL), a framework that maintains and evolves a diverse population of neural network models. ECL continually searches for an optimal architecture for each introduced incremental task. This tailored model is trained on the corresponding task and archived as a specialized expert, contributing to a growing collection of skills. This approach inherently resolves the core CL challenges: stability is achieved through the isolation of expert models, while plasticity is greatly enhanced by evolving unique, task-specific architectures. Experimental results demonstrate that ECL significantly outperforms state-of-the-art individual-level CL methods. By shifting the focus from individual adaptation to collective evolution, ECL presents a novel path toward AI systems capable of CL.