脉冲神经网络的时间计算─—世界中心识别原理  

Time Computation of Spiking Networks for World-Centered Representation

作　　者：郭雷[1] 

机构地区：[1]西北工业大学自动控制系,西安710072

出　　处：《电子学报》1998年第11期10-14,36,共6页Acta Electronica Sinica

基　　金：国家杰出青年科学基金

摘　　要：本文表明传统的基于脉冲平均点火车（averagefiringrate）的人工神经网的全并行计算是不妥的，因为它不能进行时间操作我们提出一个脉冲神经网（Spikingnetwork），对神经计算引入空间搜索机制搜索能够将空间坐标转化为时间坐标这使得空间相对坐标也转化为时间相对坐标利用延迟的连接，神经网能够非常容易地完成时间相对坐标的计算设计了一个以世界为中心的模型（World-centeredMedel，WCM），它由空间搜索器和特征转移向量存储器组成WCM是单纯的神经网络。This article argues that full-parallel algorithms of traditional sigmoidal neural networks based on average fire rate which have excluded useful time operation seem improper for the invariance.Thus. we present a spiking network that intreduces a space search mechanism into neural computing. The search can transform space coordinate into time coordinate. Consequently, the operations based relative positions are converted into those based relative time which are very easy for neural realization through delay connections. Accordingly, we develop a world-centered model (WCM) that consists of a space searcher and a feature transfer vector (FTV) memory. WCM is a pure neural network that represents the neural principle of invariant recognition.

关 键 词：神经计算 模式识别 人工神经网络 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.4[自动化与计算机技术—控制科学与工程]