摘要
本文讨论了不分层的通用前馈网络（GFFN），并提出了一种作为模式分类器的排序学习前向掩蔽（SLAM）模型及其算法实验结果表明，这种网络作为模式分类器用时，学习时间远小于各种改进的BP网络而且所需使用的神经元数量也有显著的减少本文还介绍了这种SLAM模型在应用双阈值神经元DTN时进一步减少神经元数量的实验结果及其网络结构和学习算法，以及这种模型的模式分类器所具有的不断扩展与改善的能力论文还介绍了SLAM模型模式分类器在CASSANDRA-I小型神经计算机上实现的实验结果：在256维输入空间1024个随机样本的分类情况，学习时间约3小时20分，判别时间为0．007秒．
On the basis of discussing dineral Feed-FOrward Network(GFFN), a Sequential Learng Ahead Masking Medel and its relevant algorithm for pattern classification are proud. By adapting this model to the pattern clssifier, the computer simulation results show that not only the convergence sped and performance of the network are much better than the existing modified BP algorithms, but also the required network scale is reduced greatly. Moreover, Double-Threshold Neuron(DTN) has been applied to SLAM network for pattern classification and the SLAM pattern classifier shows potential of continuously growing and improving. At last, the SLAM pattern classifier has been implemented on the CASSANDRA-I micro-neurocomputer and the results are provided.
出处
《电子学报》
EI
CAS
CSCD
北大核心
1998年第8期1-6,共6页
Acta Electronica Sinica
基金
新加波国家半导体公司支持课题
关键词
通用前馈网络
SLAM模型
模式识别
神经网络
General Feed-Forward Network(GFFN), SLAM model, Pattern recognition, Neuralnetwork, Pattern classifier
