| 摘 要: 液体状态机(Liquid State Machine, LSM)具有实时计算和仿生的特点,在处理时间序列数据上具有巨大潜力。为了研究如何提高神经网络模型训练性能,降低计算复杂度,文章首先梳理和回顾了近几年相关研究文献,其次提出硬件实现和软件模型两个优化思路,并总结了不同优化方法的优势与不足,硬件和软件上的优化可以提高神经网络模型学习性能和训练速度,但依然存在可控性差、算法最优解未知等问题,最后针对以上问题对未来的研究方向进行了展望,可为时间序列数据处理和模式识别领域提供优化思路。 |
| 关键词: 脉冲神经网络;储备池层;液体状态机;遗传算法 |
|
中图分类号: TP183
文献标识码: A
|
| 基金项目: 国防科技工业局国防基础科研计划(JCKYS2020DC202);河北省自然科学基金(F2022208002);河北省高等学校科学技术研究重点项目(ZD2021048) |
|
| Research Advances in Liquid State Machine |
|
ZHANG Yongqiang1,2, NI Shanshan2, SONG Meilin2, MAN Menghua1
|
(1.Shijiazhuang Campus of Army Engineering University of PLA, Shijiazhuang 050003, China;
2.School of Information Science and Engineering, Hebei University of Science and Technology, Shijiazhuang 050018, China)
zyq@hebust.edu.cn; nss_work@163.com; songmeilin1997@qq.com; manmenghua@126.com
|
| Abstract: Liquid State Machine (LSM), characterized by real-time computation and biomimetics, has great potential in processing time series data. In order to study how to improve the training performance of neural network models and reduce computational complexity, relevant research literature in recent years is firstly reviewed in this paper. Then, two optimization ideas, hardware implementation and software model, are proposed, and the advantages and disadvantages of different optimization methods are summarized. Hardware and software optimization can improve the learning performance and training speed of neural network models, but there are still problems such as poor controllability and unknown algorithm optimal solutions. Finally, the future research direction is prospected for the above problems, which can provide optimization ideas for the field of time series data processing and pattern recognition. |
| Keywords: spiking neural network; reserve pool layer; Liquid State Machine; genetic algorithm |
