基于结点敏感度的单隐含层前馈神经网络结构选择  

Architecture Selection for Single-hidden Layer Feed-forward Neural Networks Based on Sensitivity of Node

作　　者：翟俊海[1,2] 哈明光 邵庆言[1] 王熙照[1,2] 

机构地区：[1]河北大学数学与计算机学院,保定071002 [2]河北省机器学习与计算智能重点实验室,保定0710022

出　　处：《计算机科学》2014年第2期153-156,共4页Computer Science

基　　金：国家自然科学基金项目(61170040);河北省自然科学基金项目(F2013201110;F2013201220);河北大学自然科学基金项目(2011-228043);河北大学教育教学改革研究项目(JX07-Y-27)资助

摘　　要：提出了一种基于结点敏感度的单隐含层前馈神经网络结构选择方法。该方法从一个隐含层结点个数较多的网络开始,首先利用结点敏感度度量隐含层结点的重要性,然后按重要性对隐含层结点由大到小排序,最后逐个剪去不重要的隐含层结点,直到满足预定义的停止条件。该算法的特点是不需要重复训练神经网络,得到的网络结构紧凑,具有较高的泛化能力。在实际数据集和UCI数据集上的实验结果显示,提出的算法是行之有效的。Based on sensitivity of node, an architecture selection for Single-hidden Layer Feed-forward Neural Networks （SLFNNs） was proposed. Beginning from an initial large number of hidden nodes, the proposed algorithm firstly em- ploys the sensitivity to measure the significance of the hidden nodes, and then the hidden nodes are sorted in descending order by their significance, finally all unimportant nodes are pruned. The algorithm will terminate when a predefined stop condition is held. The main feasures of the proposed algorithm include the unnecessity of retraining the SLFNN, the compact architecture and the high generalizition capacity. We experimented the proposed approaches on real world datasets and UCI datasets,and the experimental results show that the proposed method is effective and efficient.

关 键 词：前馈神经网络 结构选择 敏感度 交叉熵 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]