稀疏连接的异步池计算网络  

Sparsely connected asynchronous reservoir computing network

作　　者：薄迎春[1] 张欣[1] 刘宝[1] 王平[1] Yingchun BO;Xin ZHANG;Bao LIU;Ping WANG(College of Control Science and Engineering,China University of Petroleum(Huadong),Qingdao 266580,China)

机构地区：[1]中国石油大学(华东)控制科学与工程学院,青岛266580

出　　处：《中国科学：信息科学》2021年第5期764-778,共15页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:21606256)资助项目。

摘　　要：针对池计算网络的构建问题,提出了一种稀疏连接的异步神经元池构造方法,该方法将多个子神经元池顺序连接,并在子神经元池之间设置滞后环节,以实现各子神经元池对输入信息的异步处理,进而构成串行的记忆.为实现信息高效传输,子神经元池之间采用稀疏的连接方式.实验表明,所提方法能够有效地提高神经元池的记忆容量,易于解决长时依赖问题.此外,该结构能够使神经元池产生丰富的动力学行为,对初始参数也有较好的鲁棒性.In order to solve the reservoir computing network construction problem, a sparsely connected asynchronous neuron reservoir construction method is proposed. The method connects several sub-reservoirs sequentially and sets lag links among sub-reservoirs in order to handle input signals asynchronously in sub-reservoirs,and further constitutes serial memory. In order to achieve efficient information transmission, sparse connections are used among sub-reservoirs. Experimental results show that the proposed method can effectively improve the memory capacity of reservoir and it is easy to deal with long-term dependence problems. In addition, the proposed structure makes the reservoir produce more abundant dynamic behavior and has better robustness to the initial parameters.

关 键 词：人工神经网络 池计算 记忆 鲁棒性 动力学 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]