摘要
本文从子空间变换的角度，研究了前馈网络（FFN）外监督分类器用于分类的机理．在批方式学习最小二乘误差代价函数为零的条件下，证明了线性输出FFN（或线性FFN）外监督分类器的输出节点对应的不同类别权矢量，是相互正交的，而非线性输出FFN外监督分类器对应的类别权矢量位于互反的类别于空间内；证明了网络获得零代价全局最小解的充要条件是R（Y）R（X）．
This paper studies the classification mechanism of outer-supervised feedforward network classifiers (FFNC) from the perspective of the subspace transformation. Under the condition of null least square error cost function via the batch-style learning, it proved that the class weight vectors of linear outputed FFNCs (or linear FFNCs) are mutually orthogonal and that every class weight vector of nonlinear outputed FFNCs must be included in the negative direction of another class subspace. It is shown that sufficient and necessary condition for the networks obtaining null error cost function is proved to be R(Y) R(X).
出处
《计算机学报》
EI
CSCD
北大核心
1998年第7期650-655,共6页
Chinese Journal of Computers
基金
国家自然科学基金!69705001
关键词
前馈网络
分类机理
神经网络
外监督分类器
Feedforward networks, outer-supervised learning, classification, mechanism, bottleneck, orthogonality, global minimum solution
