摘要
传统径向基函数(RBF)神经网络模型使用完整的隐含层节点进行模型构建时,会因缺乏隐含层节点抽取机制而使得受训模型的泛化性能下降,导致模型更加复杂。为此,提出一种改进的RBF神经网络模型。通过Lasso稀疏约束对隐含层节点和输出层连接权值进行稀疏表示,去除冗余和不相关隐含层节点的同时保留重要的隐含层节点,并使用交叉验证和网格搜索确定收缩参数以优化模型分类性能。实验结果表明,与现有RBF神经网络模型相比,该模型具有更低的计算复杂度和更高的分类精度。
The traditional Radial Basis Function(RBF) neural network model uses all hidden layer nodes to construct the model.In this case,the generalization performance of traditional RBF neural network model is degraded because of the lackness of the effective hidden layer nodes extraction mechanism,which will easily leads to more complicated of model.In order to solve this problem,this paper proposes an improved RBF neural network model.It realizes sparse representation of hidden layer nodes and output layer connection weights by Lasso sparse constraint to remove redundant and uncorrelated hidden layer nodes,and retain important hidden layer nodes.The shrinkage parameter are determined by Cross Validation(CV) and grid search,so as to optimize model classification performance.Experimental results show that RBF neural network model based on Lasso sparse learning can not only reduce the calculation complexity of the model,but also improve the classification accuracy compared with existing RBF neural network model.
作者
崔晨
邓赵红
王士同
CUI Chen;DENG Zhaohong;WANG Shitong(School of Digital Media,Jiangnan University,Wuxi,Jiangsu 214122,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2019年第2期173-177,共5页
Computer Engineering
基金
国家自然科学基金(61772239
61403247)
国家重点研发计划(2016YFB0800803)
江苏省杰出青年基金(BK20140001)
关键词
数据挖掘
Lasso稀疏学习
径向基函数
神经网络
收缩参数
data mining
Lasso sparse learning
Radial Basis Function(RBF)
neural network
shrinkage parameter