时间分割的池计算网络及其动力学
Temporally segregated reservoir computing and its dynamics
摘要点击 26  全文点击 34  投稿时间:2018-05-26  修订日期:2018-10-10
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2018.80390
  2019,36(8):1315-1321
中文关键词  人工神经网络  池计算  深度学习  短时记忆  时间序列预测
英文关键词  artificial neural networks  reservoir computing  deep learning  short-term memory  time series prediction
基金项目  
学科分类代码  
作者单位E-mail
薄迎春 中国石油大学信息与控制工程学院 boyingchun@sina.com 
张欣 中国石油大学信息与控制工程学院  
刘宝 中国石油大学信息与控制工程学院  
中文摘要
      为解决针对给定任务构建合适的神经元池问题, 提出了一种时间分割的神经元池设计方法, 该方法将多个 子神经元池顺序连接, 每两个相邻的子神经元池之间嵌入一个滞后环节以构成时间分割的神经元池,每个子神经元 池只需处理一段时间的信息, 从而达到复杂记忆任务分解的目的. 输出层可对各子神经元池的状态进行整合以获 取不同时段的输入特征. 对多阶层振荡器的实验表明, 在宏观参数相同的情况下, 时间分割的池计算网络比常规池 计算网络具有更强的记忆能力, 能够产生更加多样化的动力学行为.
英文摘要
      To solve the problem of building a suitable reservoir for a given task, a temporally segregated neuron reservoir design method is put forward, where the single reservoir in a standard reservoir computing is divided into a series of subreservoirs that are connected one by one in sequence. Time delay modules are inserted between every two adjacent subreservoirs to construct a temporally segregated reservoir, and each sub-reservoir processes only the information in a certain corresponding time span. By this method, a complex memory task can be decomposed into a series of simple memory tasks. The output layer can acquire the input characteristics in different time spans by integrating the different internal states in all sub-reservoirs. The numerical experiments on modeling the multiple superimposed oscillator demonstrate that the temporally segregated reservoir is capable of possessing stronger memory capability and generating more diverse dynamical behavior than the standard reservoir computing with the same hyper parameters.