权重随机正交化的极速非线性判别分析网络  

Nonlinear Discriminant Analysis Networks with Random and Orthogonalized Input Weights

作　　者：谢群辉 田青 

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京211106 [2]南京信息工程大学计算机与软件学院,江苏南京210044

出　　处：《计算机技术与发展》2018年第1期23-27,共5页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61472186);中国博士后科学基金特别资助项目(20133218110032);南京信息工程大学人才启动基金

摘　　要：极速学习机(Extreme Learning Machine,ELM)以其训练速度快、易实现、泛化性能好等优点受到了广泛关注。然而在数据维度较高的场景,数据中往往蕴含着较多冗余信息,而经典ELM尚未能很好地应对这个问题。此外,经典ELM也未能对标记数据的判别信息有效地加以融合利用。针对传统ELM方法的不足之处,提出一种权重随机正交的判别分析网络(O-ENDA)。在O-ENDA中,一方面对ELM输入层权重施加正交约束,这就降低了输入特征的冗余信息以减低过拟合的风险(尤其在小样本场景下);另一方面将隐层特征与判别分析相融合进行联合学习,实现数据判别信息在ELM中的融合利用。实验结果表明,提出方法在保持数据判别特征的同时能够去除其冗余信息、提高模型的泛化能力并能获得更高的分类精度。Extreme Learning Machines （ELM） has attracted increasing attention due to its fast training speed, simplicity and good gener- alization. However,in applications with high-dimensional features, a large amount of redundant information may exist in the data, which has not been concerned in classical ELM. Moreover,the discriminant information behind data has also not been incorporated in the ELM learning. To overcome the drawbacks of classical ELM, a weight-orthogonal discriminant analysis network （O-ENDA） is put forward. In O-ENDA, the input weights of ELM are restricted to be orthogonal,removing the redundant features to alleviate the risk of over-fit- ring （ especially in the scenario of smaU samples） ; simultaneously the transformed hidden features are combined with discriminant analysis to improve the discriminant ability of O-ENDA. Experiment demonstrates that the proposed approach not only can remove redundant in- put information while preserving the necessary feature information ,but also entirely yields preferable classification accuracy than classical ELM.

关 键 词：极速学习机 线性判别分析 神经网络 降维 分类 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]