计算机科学

计算机科学 ›› 2023, Vol. 50 ›› Issue (11A): 230300012-7.doi: 10.11896/jsjkx.230300012

• 人工智能 • 上一篇    下一篇

基于原型回放和动态更新的类增量学习方法

张禹1, 曹熙卿2,3, 钮赛赛2,3, 许鑫磊1, 张倩1, 王喆1   

  1. 1 华东理工大学信息科学与工程学院 上海 200237
    2 上海航天控制技术研究所 上海 201109
    3 中国航天科技集团公司红外探测技术研发中心 上海 201109
  • 发布日期:2023-11-09
  • 通讯作者: 王喆(wangzhe@ecust.edu.cn)
  • 作者简介:(1061941314@qq.com)
  • 基金资助:
    上海市科技计划项目(21511100800,20511100600);国家自然科学基金项目(62076094);国防科技领域基金项目(2021-JCJQ-JJ-0041);中国航天科技集团公司第八研究院产学研合作基金资助项目(SAST2021-007)

Incremental Class Learning Approach Based on Prototype Replay and Dynamic Update

ZHANG Yu1, CAO Xiqing2,3, NIU Saisai2,3, XU Xinlei1, ZHANG Qian1, WANG Zhe1   

  1. 1 School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China
    2 Shanghai Aerospace Control Technology Institute,Shanghai 201109,China
    3 Research and Development Center of Infrared Detection Technology,China Aerospace Science and Technology Corporation,Shanghai 201109,China
  • Published:2023-11-09
  • About author:ZHANG Yu,born in 1996,postgraduate.His main research interests include incremental learning and deep learning.
    WANG Zhe,born in 1981,Ph.D,associate professor,is a member of China Computer Federation.His main research interests include pattern recognition and image processing.
  • Supported by:
    Shanghai Science and Technology Program(21511100800,20511100600),National Natural Science Foundation of China(62076094),Chinese Defense Program of Science and Technology(2021-JCJQ-JJ-0041) and China Aerospace Science and Technology Corporation Industry-University-Research Cooperation Foundation of the Eighth Research Institute(SAST2021-007).

PDF (PC)

581

摘要： 灾难性遗忘问题在增量学习场景中普遍存在,而对旧知识的遗忘会严重影响模型在整个任务序列上的平均性能。因此,针对在增量学习过程中原型偏移引起的旧知识遗忘问题,提出了一种基于原型回放和动态更新的类增量学习方法。该方法在原型更新阶段保留新类的原型后,进一步采用动态更新策略对旧类的原型进行实时更新。具体地,在学习新任务后,该策略基于当前可访问数据的已知偏移,来实现在旧类原型中存在的未知偏移的近似估计,并最终完成对旧类原型的更新,从而缓解原始的旧类原型与当前的特征映射间的不匹配。在CIFAR-100和Tiny-ImageNet数据集上的实验结果表明,所提出的基于原型回放和动态更新的类增量学习方法能够有效地减少对旧知识的灾难性遗忘,提高模型在类增量学习场景中的分类性能。

关键词: 类增量学习, 原型更新, 知识蒸馏, 原型回放, 灾难性遗忘

Abstract: The problem of catastrophic forgetting is prevalent in incremental learning scenarios,and forgetting of old knowledge can severely affect the average performance of the model over the entire task sequence.Therefore,a class incremental learning approach based on prototype replay and dynamic update is proposed to address the problem of old knowledge forgetting caused by prototype offset in the incremental learning process.This method further updates the prototypes of the old classes in real time using a dynamic update strategy after retaining the prototypes of the new classes in the prototype update phase.Specifically,after learning the new task,the strategy achieves an approximate estimation of the unknown bias present in the old-class prototypes based on the known bias of the currently accessible data,and finally completes the update of the old-class prototypes,thus being able to alleviate the mismatch between the original old-class prototypes and the current feature mapping.Experimental results on CIFAR-100 and Tiny-ImageNet datasets show that the proposed class incremental learning approach based on prototype replay and dynamic update is effective in reducing catastrophic forgetting of old knowledge,thus improving the classification performance of the model in class incremental learning scenarios.

Key words: Class incremental learning, Prototype update, Knowledge distillation, Prototype replay, Catastrophic forgetting

中图分类号: 

  • TP391
[1]AKHTAR N,MIAN A.Threat of adversarial attacks on deep learning in computer vision:A survey[J].IEEE Access,2018,6:14410-14430.
[2]HE T,ZHANG Z,ZHANG H,et al.Bag of tricks for image classification with convolutional neural networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2019:558-567.
[3]BAEK S H,HEIDE F.Polka lines:Learning structured illumination and reconstruction for active stereo[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2021:5757-5767.
[4]FAYEK H M,LECH M,CAVEDON L.Evaluating deep learning architectures for speech emotion recognition[J].Neural Networks,2017,92:60-68.
[5]LIPPI M,MONTEMURRO M A,DEGLI ESPOSTI M,et al.Natural language statistical features of LSTM-generated texts[J].IEEE Transactions on Neural Networks and Learning Systems,2019,30(11):3326-3337.
[6]DABBAGH N,CASTANEDA L.The PLE as a framework for developing agency in lifelong learning[J].Educational Technology Research and Development,2020,68(6):3041-3055.
[7]ZHANG T,WANG X,LIANG B,et al.Catastrophic interference in reinforcement learning:A solution based on context division and knowledge distillation[J].IEEE Transactions on Neural Networks and Learning Systems,2022:1-15.
[8]MASANA M,LIU X,TWARDOWSKI B,et al.Class-incremental learning:survey and performance evaluation on image classification[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2022,45(5):5513-5533.
[9]CHEN H,WANG Y,HU Q.Multi-Granularity Regularized Re-Balancing for Class Incremental Learning[J].IEEE Transactions on Knowledge and Data Engineering,2023,35(7):7263-7277.
[10]LI K,WAN J,YU S.CKDF:Cascaded knowledge distillation framework for robust incremental learning[J].IEEE Transactions on Image Processing,2022,31:3825-3837.
[11]LIN H,FENG S,LI X,et al.Anchor Assisted Experience Replay for Online Class-Incremental Learning[J].IEEE Transactions on Circuits and Systems for Video Technology,2022,33(5):2217-2232.
[12]LIN G,CHU H,LAI H.Towards better plasticity-stabilitytrade-off in incremental learning:a simple linear connector[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2022:89-98.
[13]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.
[14]ALJUNDI R,BABILONI F,ELHOSEINY M,et al.Memoryaware synapses:Learning what (not) to forget[C]//Proceedings of the European Conference on Computer Vision.2018:139-154.
[15]JOSEPH K J,KHAN S,KHAN F S,et al.Energy-based Latent Aligner for Incremental Learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2022:7452-7461.
[16]YOON J,YANG E,LEE J,et al.Lifelong Learning with Dy-namically Expandable Networks[C]//International Conference on Learning Representations.2018.
[17]YAN S,XIE J,HE X.Der:Dynamically expandable representation for class incremental learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2021:3014-3023.
[18]MALLYA A,DAVIS D,LAZEBNIK S.Piggyback:Adapting a single network to multiple tasks by learning to mask weights[C]//Proceedings of the European Conference on Computer Vision.2018:67-82.
[19]REBUFFI S A,KOLESNIKOV A,SPERL G,et al.icarl:Incremental classifier and representation learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2017:2001-2010.
[20]SHIN H,LEE J K,KIM J,et al.Continual learning with deep generative replay[C]//Advances in Neural Information Processing Systems.2017:1-10.
[21]BELOUADAH E,POPESCU A.Il2m:Class incremental learning with dual memory[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.2019:583-592.
[22]ZHU F,ZHANG X Y,WANG C,et al.Prototype augmentation and self-supervision for incremental learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2021:5871-5880.
[23]YU L,TWARDOWSKI B,LIU X,et al.Semantic drift compensation for class-incremental learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2020:6982-6991.
[24]LEE H,HWANG S J,SHIN J.Self-supervised label augmentation via input transformations[C]//International Conference on Machine Learning.2020:5714-5724.
[25]YANG K,YAU J H,LI F F,et al.A study of face obfuscation in imagenet[C]//International Conference on Machine Learning.2022:25313-25330.
[26]ULLAH A,ELAHI H,SUN Z,et al.Comparative analysis ofAlexNet,ResNet18 and SqueezeNet with diverse modification and arduous implementation[J].Arabian Journal for Science and Engineering,2022,47:2397-2417.
[27]CASTRO F M,MARÍN-JIMÉNEZ M J,GUIL N,et al.End-to-end incremental learning[C]//Proceedings of the European Conference on Computer Vision.2018:233-248.
[28]HOU S,PAN X,LOY C C,et al.Learning a unified classifier in-crementally via rebalancing[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2019:831-839.
[29]CHATZIMPARMPAS A,MARTINS R M,KERREN A.t-visne:Interactive assessment and interpretation of t-sne projections[J].IEEE Transactions on Visualization and Computer Graphics,2020,26(8):2696-2714.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
渝ICP备16013121号-4
地址：重庆市渝北区洪湖西路18号    邮编：401121  
电话：023-63500828（编辑部） 023-67039612（会议/专题） 023-67039623（财务/发票）
E-mail：jsjkx12@163.com
本系统由北京玛格泰克科技发展有限公司设计开发