Abstract: To address the difficulty of balancing stability and plasticity in existing continual learning methods, a multi-model collaborative continual learning framework based on a filtering mechanism was proposed. The framework was composed of a main model, an auxiliary model, and a student model. The main model was designed to focus on predicting high-confidence outputs, the auxiliary model was used to memorize difficult samples that the main model failed to learn effectively, and the student model was employed to inherit and integrate the recognition ability of a cascaded teacher model constructed from the main and auxiliary models. The three models were collaboratively optimized during each task iteration. Experimental results showed that, in task-incremental scenarios, the proposed three-model framework achieved higher average accuracy, BWT (Backward Transfer), and FWT (Forward Transfer) than typical methods such as EWC (Elastic Weight Consolidation), GEM (Gradient Episodic Memory), iCaRL (incremental Classifier and Representation Learning), DER (Dark Experience Replay), and CBP (Continual BackPropagation), demonstrating stronger memory stability and adaptive flexibility. In class-incremental scenarios, the framework also exhibited good applicability on both the 10-class Fashion-MNIST and the 100-class CIFAR-100 datasets, where its average accuracy was significantly higher than that of the compared methods.

Key words: continuous learning, catastrophic forgetting, knowledge distillation, multi-model, stability, plasticity

摘要： 针对现有的持续学习方法在稳定性与可塑性之间往往难以兼顾的问题，提出一种基于筛选机制的多模型协同持续学习框架，由主模型、辅助模型与学生模型组成。其中，主模型着眼于预测输出高可信度结果，辅助模型记忆主模型难以学习的困难样本，而学生模型继承与整合主模型和辅助模型级联构成的教师模型的识别能力，三者在每一轮任务过程中协同优化。实验结果表明，在任务增量场景下，该三模型框架在Fashion-MNIST和CIFAR-10数据集上的平均准确率、BWT（BackWard Transfer）、FWT（ForWard Transfer）等指标均优于典型EWC（Elastic Weight Consolidation）、GEM（Gradient Episodic Memory）、iCaRL（incremental Classifier and Representation Learning）、DER（Dark Experience Replay）、CBP（Continual BackPropagation）等方法，体现出更强的稳定记忆与灵活适应能力；在类增量场景下，无论是只有10类的Fashion-MNIST还是100类的CIFAR-100数据集，三模型框架均有好的适用性，它的平均准确率亦显著高于对比方法。

关键词: 持续学习, 灾难性遗忘, 知识蒸馏, 多模型, 稳定性, 可塑性