自编码器结合持续学习:现状、挑战与展望  

Autoencoders Combined with Continual Learning:Development,Challenges,and Prospects

作　　者：吴美君 杨新[1,3] 潘超凡 李天瑞[4] 寇纲[2] WU Mei-Jun;YANG Xin;PAN Chao-Fan;LI Tian-Rui;KOU Gang(Innovation Laboratory of Cognitive Computing and Crowd Intelligence,Southwestern University of Finance and Economics,Chengdu 611130;Department of Business Administration,Southwestern University of Finance and Economics,Chengdu 611130;Department of Computing and Artificial Intelligence,Southwestern University of Finance and Economics,Chengdu 611130;Department of Computing and Artificial Intelligence,Southwest Jiaotong University,Chengdu 610097)

机构地区：[1]西南财经大学认知计算与群智协同创新实验室,成都611130 [2]西南财经大学工商管理学院,成都611130 [3]西南财经大学计算机与人工智能学院,成都611130 [4]西南交通大学计算机与人工智能学院,成都610097

出　　处：《计算机学报》2025年第2期317-357,共41页Chinese Journal of Computers

基　　金：国家自然科学基金青年科学基金项目(62406259);国家自然科学基金面上项目(62476228);国家自然科学基金国际(地区)合作研究与交流项目(71910107002);湖湘高层次人才聚集工程项目(2024RC4008);四川省科技厅中央引导地方科技发展项目(2024ZYD0180)资助。

摘　　要：近年来,许多研究利用自编码器进行增量式学习,以在面对新的数据分布、类别或任务时平衡模型的稳定性与可塑性。这些研究从多个角度推动了持续学习的发展。同时,持续学习的范式通过优化策略促进了自编码器架构的改进,实现了自编码器与持续学习之间的相互促进。目前,自编码器与持续学习的结合在多个领域都影响深远。本文对近五年来的相关研究进行了综述,概述了自编码器的类型与特点,持续学习的常见增量场景与主要挑战,并对二者在不同领域的应用情况进行了详细介绍。最后,本综述对当前研究的优点、局限性以及未来应用的前景进行了总结,旨在为推动持续学习与自编码器的结合与发展提供有价值的参考。In recent years,many studies have employed autoencoders for continual learning,aiming to balance the stability and adaptability of models when faced with new data distributions,categories,or tasks.These studies have propelled the development of continual learning from various perspectives.At the same time,the paradigm of continual learning has facilitated improvements in autoencoder architectures through optimization strategies,thereby achieving mutual enhancement between autoencoders and continual learning.The integration of autoencoders and continual learning has shown a significant impact across various research.This paper reviews the relevant research over the past five years,summarizing the types and characteristics of autoencoders,the incremental scenarios,and the main challenges of continual learning.Additionally,it offers a detailed overview of their applications in different industries.Finally,this review summarizes the advantages,limitations,and prospects,aiming to provide valuable insights for advancing research on developing continual learning and autoencoders.

关 键 词：持续学习 自编码器 灾难性遗忘 知识传输 模型优化 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]