互连权值灰度阶有限的神经网络的蒙特卡洛学习算法  

MONTE CARLO LEARNING ALGORITHM FOR NEURAL NETWORKS WITH LIMITED GREY LEVELS OF INTERCONNECTION WEIGHTS

作　　者：沈琴婉[1] 黄五群[2] 陈天崙 张延炘[3] 

机构地区：[1]南开大学计算中心,天津300071 [2]南开大学物理系,天津300071 [3]南开大学现代光学研究所,天津300071

出　　处：《模式识别与人工智能》1994年第2期119-123,共5页Pattern Recognition and Artificial Intelligence

基　　金：国家攀登计划认知科学(神经网络)重大关键项目;国家自然科学基金

摘　　要：本文提出一种能在互连权值的灰度阶有限条件下实现神经网络学习的蒙特卡洛学习算法。用该方法研究了随机样本模式的学习样本正确识别率,发现对给定学习样本数,此识别率不仅与神经元数有关,而且与权值的灰度阶有关,将此方法用于具有旋转不变性模式识别能力的级联模型,对四种不同飞行器的144个学习样本,在灰度阶为6的情况下得到了满意的结果。A Monte Carlo algorithm to learn under the restriction of limited grey levels of interconnection weights in a neural network is presented. The correct recognition ratio(CRR) for random training samples has been investigated with the algorithm. It is found that the CRR for a given number of training sampes depends on not only the number of neurons, but also the grey levels of the weights. By using the algorithm to the cascaded model of pattern recognition with rotation invariance, it is shown that a satisfactory result of recognition can be obtained with 6 grey levels for 144 training samples of four different kinds of aircraft.

关 键 词：灰度阶 神经网络 蒙特卡洛浊 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]