摘要
持续学习作为一种在非平稳数据流中不断学习新任务并能保持旧任务性能的特殊机器学习范例,是视觉计算、自主机器人等领域的研究热点,但现阶段灾难性遗忘问题仍然是持续学习的一个巨大挑战。围绕持续学习灾难性遗忘问题展开综述研究,分析了灾难性遗忘问题缓解机理,并从模型参数、训练数据和网络架构三个层面探讨了灾难性遗忘问题求解策略,包括正则化策略、重放策略、动态架构策略和联合策略;根据现有文献凝练了灾难性遗忘方法的评估指标,并对比了不同灾难性遗忘问题的求解策略性能。最后对持续学习相关研究指出了未来的研究方向,以期为研究持续学习灾难性遗忘问题提供借鉴和参考。
Continual learning,as a special machine learning paradigm that continuously learns new tasks in non-stationary data streams and can maintain the performance of old tasks,is a hot research topic in fields such as visual computing and autonomous robotics,but at this stage,the catastrophic forgetting problem is still a great challenge for continuous learning.This paper conducted a review study on the catastrophic forgetting problem of continual learning,analyzed the mechanism of catastrophic forgetting problem mitigation and explored the catastrophic forgetting problem solving strategies at three levels,included regularization strategy,replay strategy,dynamic architecture strategy and joint strategy,in terms of model parameters,training data and network architecture.According to the existing literature,this paper condensed the evaluation index of the catastrophic forgetting method and compared the performance of solving strategies for different catastrophic forgetting problems.Finally,it pointed out the future research direction of continual learning,to provide references for the study of continuous learning catastrophic forgetting problems.
作者
袁坤
张秀华
溥江
杨静
李斌
李少波
Yuan Kun;Zhang Xiuhua;Pu Jiang;Yang Jing;Li Bin;Li Shaobo(School of Mechanical Engineering,Guizhou University,Guiyang 550025,China;State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025,China;School of Physics&Mechantronic Engineering,Guizhou Minzu University,Guiyang 550025,China)
出处
《计算机应用研究》
CSCD
北大核心
2023年第5期1292-1302,共11页
Application Research of Computers
基金
国家重点研发计划资助项目(2018AAA0101800)
国家自然科学基金资助项目(62166005)
教育部重点实验室开放项目(黔教合KY字[2020]245
黔教合KY字[2020]248)
贵州省高层次留学人才项目(高层次人才择优资助项目202109号)
贵州省自然科学基金资助项目(黔科合基础-ZK[2022]一般130)。
关键词
持续学习
灾难性遗忘
正则化策略
重放策略
动态架构策略
continual learning
catastrophic forgetting
regularized strategy
replay strategy
dynamic architecture strategy