克服人工神经网络灾难性遗忘的连续学习算法研究

doi:10.12267/j.issn.2096-5931.2026.01.010

信息通信技术与政策 ›› 2026, Vol. 52 ›› Issue (1): 75-83.doi: 10.12267/j.issn.2096-5931.2026.01.010

克服人工神经网络灾难性遗忘的连续学习算法研究

Research on continual learning method for overcoming catastrophic forgetting of artificial neural networks

于达¹^,², 董晓飞¹^,², 曹峰¹^,², 查富生³, 孙立宁³

1.中国信息通信研究院人工智能研究所,北京 100191
2.人工智能关键技术和应用评测工业和信息化部重点实验室,北京 100191
3.哈尔滨工业大学机器人技术与系统全国重点实验室,哈尔滨 150000

收稿日期:2025-11-10 出版日期:2026-01-25 发布日期:2026-01-28
通讯作者: 董晓飞
作者简介:
于达, 中国信息通信研究院人工智能研究所、人工智能关键技术和应用评测工业和信息化部重点实验室工程师,博士,主要从事计算机视觉、机器人技术、智能体、大模型等方面工作
曹峰, 中国信息通信研究院人工智能研究所平台与工程化部主任,高级工程师,主要牵头可信人工智能评测标准体系和能力建设、工程化能力相关评估规范制定与评测等方面工作
查富生, 哈尔滨工业大学机器人技术与系统全国重点实验室教授,博士生导师,主要从事人工智能、机器人视觉、仿生机器人、机器人控制等方面工作
孙立宁, 俄罗斯工程院外籍院士,哈尔滨工业大学机器人技术与系统全国重点实验室教授,博士生导师,国家杰出青年科学基金获得者,主要从事人工智能、微纳机器人、仿生机器人、医疗机器人等方面工作

YU Da¹^,², DONG Xiaofei¹^,², CAO Feng¹^,², ZHA Fusheng³, SUN Lining³

1. Artificial Intelligence Institute, China Academy of Information and Communications Technology, Beijing 100191, China
2. Key Laboratory of Artificial Intelligence Key Technologies and Application Evaluation, Ministry of Industry and Information Technology, Beijing 100191, China
3. State Key Laboratory of Robotics and System, Harbin Institute of Technology, Harbin 150000, China

Received:2025-11-10 Online:2026-01-25 Published:2026-01-28
Contact: DONG Xiaofei

摘要/Abstract

摘要：

传统的人工神经网络训练通常聚焦于封闭、静态的独立同分布数据,并在完成离线训练后执行单一任务。然而,当数据分布随环境不断变化时,模型会忘记在先前任务中学到的知识,即发生“灾难性遗忘”。连续学习作为一个新的学习范式,旨在赋予模型从分布不断变化的数据流中持续学习、累计和巩固知识的能力,使得人工神经网络达到“稳定性-可塑性”平衡,进而克服灾难性遗忘。通过深入分析当前连续学习算法的主要特点,搭建真实机器人实物验证平台,在机器人实物抓取场景下验证连续学习算法的有效性。试验结果表明,将对比相关性保留回放算法应用到机器人实物抓取任务,抓取任务的平均准确率提高26.67%,能更好地帮助机器人执行目标任务。

关键词: 人工智能, 人工神经网络, 连续学习, 灾难性遗忘

Abstract:

Traditional artificial neural network training typically focuses on closed, static, independent and identically distributed data, and performs a single task after completing offline training. However, when the data distribution continuously changes with the environment, the model will forget the knowledge learned from previous tasks, a phenomenon known as “catastrophic forgetting”. As an emerging learning paradigm, continual learning aims to endow models with the ability to continuously learn, accumulate, and consolidate knowledge from data streams with constantly changing distributions. This enables artificial neural networks to achieve a “stability-plasticity” balance, thereby overcoming catastrophic forgetting. Through in-depth analysis of the key characteristics of current continual learning algorithms, a real-world robotic physical verification platform was established. The effectiveness of continual learning algorithms was verified in the scenario of robotic physical object grasping. Experimental results show that when the Contrastive Correlation Preserving Replay (CCPR) algorithm is applied to the robotic physical object grasping task, the average accuracy of the grasping task increases by 26.67%, better assisting the robot in performing the target task.

Key words: artificial intelligence, artificial neural networks, continual learning, catastrophic forgetting

中图分类号:

TP183

于达, 董晓飞, 曹峰, 查富生, 孙立宁. 克服人工神经网络灾难性遗忘的连续学习算法研究[J]. 信息通信技术与政策, 2026, 52(1): 75-83.

YU Da, DONG Xiaofei, CAO Feng, ZHA Fusheng, SUN Lining. Research on continual learning method for overcoming catastrophic forgetting of artificial neural networks[J]. Information and Communications Technology and Policy, 2026, 52(1): 75-83.

导出引用管理器 EndNote|Ris|BibTeX

链接本文:

http://ictp.caict.ac.cn/CN/10.12267/j.issn.2096-5931.2026.01.010

http://ictp.caict.ac.cn/CN/Y2026/V52/I1/75

图/表 7

图1 机械臂实物抓取平台

图2 机械臂连续目标识别抓取任务

图3 任务1:识别并抓取目标“碗”

图4 任务2:识别并抓取目标“苹果”

图5 执行任务1(未采取缓解遗忘的措施)

图6 执行任务1(采取缓解遗忘的措施)

图7 不同抓取位姿生成任务抓取成功率演变

参考文献 15

[1]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[C]// Advances in Neural Information Processing Systems 25 (NIPS 2012). Lake Tahoe: NeurIPS, 2012: 1106-1114.
[2]	MCCLOSKEY M, COHEN N J. Catastrophic interference in connectionist networks: the sequential learning problem[J]. Psychology of Learning and Motivation. 1989, 24: 109-165.
[3]	ALJUNDI R, KELCHTERMANS K, TUYTELAARS T. Task-Free continual learning[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Leuven: IEEE Press, 2019: 11254-11263.
[4]	KIRKPATRICK J, PASCANU R, RABINOWITZ N, et al. Overcoming catastrophic forgetting in neural networks[J]. Proceedings of the National Academy of Sciences, 2017, 114(13): 3521-3526. doi: 10.1073/pnas.1611835114 URL
[5]	ZENKE F, POOLE B, GANGULI S. Continual learning through synaptic intelligence[C]// International Conference on Machine Learning. Sydney: ICML, 2017: 3987-3995.
[6]	LI Z, HOIEM D. Learning without forgetting[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 40(12): 2935-2947. doi: 10.1109/TPAMI.34 URL
[7]	RANNEN A, ALJUNDI R, BLASCHKO M B, et al. Encoder based lifelong learning[C]// Proceedings of the IEEE International Conference on Computer Vision. Venice: ICCV, 2017: 1320-1328.
[8]	RUSU A A, RABINOWITZ N C, DESJARDINS G, et al. Progressive neural networks[J]. arXiv Preprint, arXiv:1606.04671, 2016.
[9]	YOON J, YANG E, LEE J, et al. Lifelong learning with dynamically expandable networks[C]. International Conference on Learning Representations. Vancouver: ICLR, 2018.
[10]	FERNANDO C, BANARSE D, BLUNDELL C, et al. Pathnet: evolution channels gradient descent in super neural networks[J]. arXiv Preprint, arXiv:1701.08734, 2017.
[11]	MALLYA A, LAZEBNIK S. PackNet: adding multiple tasks to a single network by iterative pruning[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Salt Lake City: IEEE Press, 2018: 7765-7773.
[12]	RIEMER M, CASES I, AJEMIAN R, et al. Learning to learn without forgetting by maximizing transfer and minimizing interference[C]// International Conference on Learning Representations. New Orleans: ICLR, 2019.
[13]	YU D, ZHANG M, LI M, et al. Contrastive correlation preserving replay for online continual learning[J]. IEEE Transactions on Circuits and Systems for Video Technology. 2023, 34 (1): 124-139. doi: 10.1109/TCSVT.2023.3285221 URL
[14]	SHIN H, LEE J K, KIM J, et al. Continual learning with deep generative replay[J]. Advances in Neural Information Processing Systems, 2017: 2990-2999.
[15]	LOPEZ-PAZ D, RANZATO M. Gradient episodic memory for continual learning[J]. Advances in Neural Information Processing Systems, 2017: 6467-6476.

克服人工神经网络灾难性遗忘的连续学习算法研究

Research on continual learning method for overcoming catastrophic forgetting of artificial neural networks

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 7

参考文献 15

相关文章 15

编辑推荐

Metrics

本文评价

[1]	陈昕, 张恒升, 付韬, 李果. 工业算网和工业人工智能融合发展现状与趋势研究[J]. 信息通信技术与政策, 2026, 52(1): 31-40.
[2]	张斌. 人工智能时代领导力数字化转型的驱动因素:基于技术接受模型的研究[J]. 信息通信技术与政策, 2026, 52(1): 59-67.
[3]	李想, 孙永华, 李伟健. 人工智能赋能全生命周期智慧风电^*[J]. 信息通信技术与政策, 2026, 52(1): 84-89.
[4]	闫嘉豪. 城市智能算力底座建设现状与发展建议[J]. 信息通信技术与政策, 2025, 51(9): 2-7.
[5]	龙海泉, 王少帅, 李科. “人工智能+”赋能IPv6网络管理场景及技术方案研究[J]. 信息通信技术与政策, 2025, 51(8): 10-14.
[6]	李腾达, 朱紫钰, 韩子奇. “人工智能+”赋能具身智能机器人新形态及关键技术应用[J]. 信息通信技术与政策, 2025, 51(8): 15-25.
[7]	吴盈盈, 宋平, 任璐. 我国“人工智能+”行动的必要性及推进路径[J]. 信息通信技术与政策, 2025, 51(8): 2-9.
[8]	燕江依, 李荪, 樊威, 曹峰. 新一代数据标注产业对“人工智能+”范式创新的作用机理与实践路径研究[J]. 信息通信技术与政策, 2025, 51(8): 26-34.
[9]	胡文娟, 闫昭坤, 孟丹妮, 许珊, 蒋博丞, 许丽姣, 陈广拓, 张义. “人工智能+”推动天然气行业高质量发展路径研究[J]. 信息通信技术与政策, 2025, 51(8): 35-41.
[10]	王锐, 左铠瑞, 蒋亦伟. 人工智能与信息通信业融合发展成效、面临问题与建议[J]. 信息通信技术与政策, 2025, 51(8): 50-54.
[11]	刘志鹏, 张津铭. 从平台经济到人工智能:治理机理比较与治理体系构建[J]. 信息通信技术与政策, 2025, 51(8): 55-63.
[12]	秦思思, 闫东伟, 齐可心, 程阳. 人工智能赋能软件形态演进趋势研究[J]. 信息通信技术与政策, 2025, 51(8): 64-70.
[13]	吴重庆, 杨思哲, 宁庭勇, 胡永波. 上海政务垂类大模型应用探索[J]. 信息通信技术与政策, 2025, 51(8): 78-83.
[14]	徐贵宝, 柏润泚, 易宏杰. 人工智能赋能医康养融合服务生态体系建设研究^*[J]. 信息通信技术与政策, 2025, 51(8): 84-90.
[15]	岳锦疆, 宁庭勇. 人工智能在固态储氢材料中的综合应用[J]. 信息通信技术与政策, 2025, 51(6): 21-26.