关于前馈网络分类器隐层单元变换特性的研究

本文研究了线性与非线性前馈网络隐层单元的变换特性,深刻揭示了线性怀非线性网络用于信息表示的机理,证明线性网络的隐输出模式是相关的,而非线性网络的隐输出模式是不相关的,并就非线性网络突破性网络的“瓶颈”行为,给出一个定理,最后,以三元奇偶问题为偶,给出有关实验结果。

On the Study of Transformation Properties of Hidden Units of Feedforward Neural Networks Classifiers

This paper studies the transformation properties of hidden units of linear or nonlinear feedforward networks(FNN) classifiers. The mechanisms of how to organize information in the linear or nonlinear FNNs are disclosed. It is proved that the hidden output patterns in linear FNNs are correlated and the ones in nonlinear FNNs are decorrelated. Moreover, a theorem , which shows that the nonlinear networks break through the "bottleneck " in linear networks and obtain independent hidden outputs, is given. Finally,an example, i. c., parity 3 problem, is used as the simulating data and the related experimental results are presented.