摘要
双隐层标准前馈（ＢＰ）网络只要其隐层节点数足够多就能解决任何形式的分类问题．应用标准（ＢＰ）网络识别多模式类分类问题时存在以下缺陷：（１）对不同模式类均使用相同数目的隐层元；（２）增加新模式类后，网络要重新学习；（３）网络识别的机理研究困难．笔者提出了一种局域连接前馈神经网络（ＬＣＮＮ）结构，其隐层神经元与输出神经元之间为局域连接，学习算法与ＢＰ算法类似．ＬＣＮＮ具有以下特点：（１）便于自构网络结构，提高网络的推广能力；（２）便于提取各模式类的不变特性；（３）具有较强的记忆能力，便于实现追加学习．以五种海底沉积层介质类型的分类识别为例，分别利用标准前馈（ＢＰ）网络与ＬＣＮＮ网络进行分类识别，结果表明：ＬＣＮＮ便于自构网络结构，具有追加学习的能力．
The standard forward feeding BP network with two hidden layers can solve any classification problems if it includes enough processing elements.But when used for the classification of several patterns,it has some shortcomings as follows:(1)an identical network structure is used for different patterns;(2) when a new pattern is added to the training set,the network must be trained again;(3) it is difficult to study the mechanism of the network recognition.In the paper,the author puts forward an improved structure of forward feeding network—local domain connection neural network (LDCNN). Its hidden layer neurons are divided into several groups,and each neuron output only connects to one of the groups.Learning algorithm for LDCNN is similar to BP.LDCNN has some advantages as follows:(1)according to the training set,LDCNN can construct the network structure by itself;(2)identical features can be easily obtained from each pattern after LDCNN is properly trained;(3)LDCNN can implement supplementary learning when a new pattern is added to the training set.Five classes of sediments were recognized by using LDCNN and stardard BP network respectively,and it is shown that the learning and testing performance of LDCNN is better than that of standard BP network.
出处
《西安石油学院学报》
1997年第3期49-52,共4页
Journal of Xi'an Petroleum Institute
关键词
神经网络
模式识别
数据处理
前馈网络
Free words]:nerve network,pattern recognition,data processing/[forward feeding neural network,supplementary learning]
