多维特征空间中模式分类神经网络的一种设计方法

提出了一种在多维空间Ｒｎ中模式分类神经网络设计的新方法。可确定网络的拓扑结构，包括隐层、隐元个数及连接权系数。特别是由于二次阈值神经元的特例——超球面阈值神经元的使用，对特征空间Ｒｎ的分割比使用超平面阈值神经元时需利用Ｖｏｒｏｎｏｉ图进行凸区域分割简便的多。因而设计成的神经网络结构清楚直观。

A Method for Designing Pattern Classification Neural Network in Rn Space

A new method for designing pattern classification neural network in R n space is proposed. The proposed method can determine the topological structure of neural network including the hidden layers, hidden units and link weight coefficients. Especially, dividing the characteristic space R n by using the ultra sphere threshold neural unit, being a particular case of quadratic neural unit, is much easier than by using Voronoi Fig to divide the characteristic space in convex region when the ultra plane threshold neural unit is used. As a result, the designed neural network structure is clear and intuitive.