自动化学报2023，Vol.49Issue(12)：2467-2480,14.DOI:10.16383/j.aas.c220312
一种基于随机权神经网络的类增量学习与记忆融合方法
A Class Incremental Learning and Memory Fusion Method Using Random Weight Neural Networks
摘要
Abstract
The ability to continual learning(CL)on multiple tasks is crucial for the development of artificial gener-al intelligence.Existing artificial neural networks(ANNs)performing well on a single task are prone to suffer from catastrophic forgetting when sequentially fed with different tasks in an open-ended environment,that is,the connec-tionist models trained on a new task could rapidly forget what was learned previously.To solve the problem,this paper proposes a new metaplasticity-inspired randomized network(MRNet)for the class incremental learning(Class-IL)scenario by relating random weight neural networks(RWNNs)with the relevant working mechanism of biologic-al brain,which enables a single model to learn and remember the unknown task sequence without accessing old task data.First,a general continual learning framework with the closed-form solution is constructed in a feed-forward manner to effectively accommodate new categories emerging in new tasks;Second,a memory-related weight import-ance matrix is formed by referring to the property of synapses,which adaptively adjusts network parameters to avoid forgetting;Finally,effectiveness and efficiency of the proposed method are demonstrated in the class incre-mental learning scenario with 5 evaluation metrics,5 benchmark task sequences,and 10 comparison methods.关键词
连续学习/灾难性遗忘/随机权神经网络/再可塑性启发Key words
Continual learning(CL)/catastrophic forgetting/random weight neural networks(RWNNs)/metaplas-ticity-inspired引用本文复制引用
李德鹏,曾志刚..一种基于随机权神经网络的类增量学习与记忆融合方法[J].自动化学报,2023,49(12):2467-2480,14.基金项目
科技部科技创新2030重大项目(2021ZD0201300),中央高校基本科研业务费专项资金(YCJJ202203012),国家自然科学基金(U1913602,61936004),111计算智能与智能控制项目(B18024)资助Supported by National Key Research and Development Pro-gram of China(2021ZD0201300),Fundamental Research Funds for the Central Universities(YCJJ202203012),National Natural Science Foundation of China(U1913602,61936004),and 111 Project on Computational Intelligence and Intelligent Control(B18024) （2021ZD0201300）
