文献类型：期刊文献

中文题名：多任务学习视角下的持续学习稳定性和可塑性均衡策略

英文题名：Multi-task LearningApproach to Stability-Plasticity Balance in Continual Learning

作者：张咪咪[1];冯翔[1];虞慧群[1]

机构：[1]华东理工大学计算机科学与工程系,上海200237

年份：2026

卷号：20

期号：4

起止页码：1091

中文期刊名：计算机科学与探索

外文期刊名：Journal of Frontiers of Computer Science and Technology

收录：;北大核心:【北大核心2023】；

基金：国家自然科学基金重点项目(62136003);国家自然科学基金面上项目(62276097,62372174)。

语种：中文

中文关键词：持续学习;影响函数;多任务学习;双目标优化;帕累托最优

外文关键词：continual learning;influence function;multi-task learning;dual-objective optimization;Pareto optimality

摘要：持续学习(CL)是一种先进的机器学习范式,其目标是模拟和优化人类学习机制,在其整个生命周期中增量地获取新知识、更新旧知识并有效地积累和利用知识。持续学习追求在保留历史任务知识(稳定性,S)的同时,灵活适应新任务(可塑性,P)。为此,模拟并分析了影响函数(IF)计算中的关键环节,分别获得了稳定性(S)和可塑性(P)的样例影响评估。进而,创新性地提出了Pareto SP算法,该算法通过求解双目标优化问题,将这两种样例影响融合起来,旨在找到对SP均达到帕累托(Pareto)最优的融合影响策略。这一策略采取多任务学习(MTL)中多目标优化的思想,将帕累托多任务学习算法推广至持续学习系统,通过并行解决具有不同权衡偏好的子问题,获取一组能够体现竞争任务间不同权衡的代表性帕累托最优解。尤为重要的是,该算法识别并解决了由二阶影响带来的潜在问题,这些影响可能放大重放缓冲区中的随机偏差,影响样本选择过程的有效性。为此,引入了一种新颖的目标选择策略,对核心集选择过程进行正则化,从而增强了核心集的选择效率与效果。实验结果显示,该算法在任务增量与类别增量两类持续学习基准数据集上,均展现出相较于当前最先进方法的显著优势。在有限的计算内存下,Pareto SP算法超越了最先进的方法,在3个不同的持续学习基准中持续实现超过2个百分点的显著改进。
Continual learning(CL)is an advanced paradigm in machine learning designed to simulate and optimize human learning mechanisms.It enables machines to incrementally acquire new knowledge,update prior knowledge,and effectively accumulate and utilize this knowledge throughout their lifecycle.The primary aim of continual learning is to preserve knowledge from previous tasks(stability,S)while maintaining the flexibility to adapt to new tasks(plasticity,P).To this end,this paper simulates and analyzes the critical aspects influencing function computation,obtaining sample influence assessments for both stability(S)and plasticity(P).Subsequently,this paper proposes the innovative Pareto SP algorithm,which integrates these two types of sample influences by addressing a dual-objective optimization problem,with the goal of identifying a fusion influence strategy that achieves Pareto optimality for both stability(S)and plasticity(P).This approach employs the principles of multi-objective optimization within multi-task learning and extends the Pareto multitask learning algorithm to continual learning systems.By simultaneously solving subproblems with distinct trade-off priorities,it generates a representative set of Pareto optimal solutions that capture the balance between competing tasks.Importantly,the algorithm identifies and mitigates the underlying issues arising from second-order effects,which can amplify random biases within the replay buffer and compromise the effectiveness of the sample selection process.To this end,this paper introduces a novel objective selection strategy that regularizes the core set selection process,optimizing the efficacy and efficiency of core set selection.The experimental results reveal that the algorithm exhibits significant advantages over current methods across both class-incremental and task-incremental continual learning benchmark datasets.With limited computational memory,the Pareto SP algorithm surpasses state-of-the-art methods,consistently achieving statistically significant improvements exceeding 2 percentage points across 3 diverse continual learning benchmarks.