储备池计算概述

Survey on Reservoir Computing

针对传统递归神经网络存在训练困难的问题,一种新的递归神经网络的训练方法———储备池计算被提出,这种方法的核心思想是只训练网络部分连接权,其余连接权一经产生就不再改变,网络的训练一般只需要通过求解线性回归问题.广义地说,储备池可以作为一种时序相关的核函数使用,从而完全拓展了其应用领域,使之不再仅仅是递归神经网络训练算法的一种改进.本文在介绍储备池计算基本数学模型的基础上,从储备池计算研究的热点问题——储备池适应性问题的角度,全面地分析了目前储备池计算的研究现状、热点及应用等方面的问题.

A novel training method for recurrent neural networks,which is called reservoir computing,was proposed with the purpose of dealing with difficulties in the training of the traditional recurrent neural networks.The main idea of the reservoir computing is training only parts of the connection weights of the networks,and generating the rest parts randomly.The connection weights generated randomly remain unchanged during the training process.Then training process of the network can be carried out by solving a linear regression problem.The reservoir can be considered as a temporal kernel function which extends the applications of the reservoir computing.In fact,the reservoir computing is not only a modification of the training algorithm to recurrent neural networks.In this paper,we firstly introduce the mathematical model of the reservoir computing and analyze the current related researches and applications in detail in the view of reservoir adaption which has attracted much interest of the researchers recently.