BP神经网络的改进及其用于手写数字识别的研究

模式识别是神经网络最有前景的应用领域之一,本文主要讨论如何提高多层神经网络 BP(Back-propagation)算法的学习速度以及该算法用于手写数字识别的研究.文中提出了局部连接的网络结构,并对基于特征输入和基于点阵输入两种神经网络分类器的特点进行了比较,针对神经网络的识别机制、识别能力和自适应学习,进行了深入讨论.本文还给出容错能力的概念,用以描述神经网络对非学习样本的分类机制.所有研究工作是在作者研制的 SSNN 神经网络仿真软件上进行的.

IMPROVED BP NEURAL NET AND ITS APPLICATION TO HANDWRITTEN NUMERAL RECOGNITION

Pattern recognition is one of the most promising fileds of neural network application.This paper is concerned with several improvements to neural network based on back-propagation (BP) algorithm for handwritten numeral recognition. Neural Network classifiers with feature input and pixel input are studied,and a comparison between the two classifier is made.The recognition mechanics, recognition ability and adaptive learning of the proposed neural network are experimentally studied.The concept of error tolerance is emphasized to express the recongnition mechanism of the neural network.This research is carried out. with the help of the SSNN-software simulator of neural network developed by the authors of this paper,too.