一种 PNN 网络的 EM 训练算法

AN EFFICIENT EM TRAINING ALGORITHM FOR PROBABILITY NEURAL NETWORKS

提出一种概率神经网络（ＰＮＮ）的ＥＭ（ＥｘｐｅｃｔａｔｉｏｎＭａｘｉｍｉｚａｔｉｏｎ）训练算法．ＰＮＮ网为一四层前馈网，它构成一个贝叶斯分类器，实现多类分类的贝叶斯判别，它把输入的样本模式，经网络变换为输出的分类判决．其网络节点对应于贝叶斯后验概率公式的各个变量．此ＰＮＮ网络用高斯核的Ｐａｒｚｅｎ窗函数作为核密度函数，网络参数训练由ＥＭ算法实现，其学习方式为类间的监督学习和类内的非监督学习．实验表明了此网络及其学习算法在分类应用中的有效性．

An Expectation-Maximization(EM) training algorithm to estimate the parameters of a special Probability Neural Network(PNN) structure which forms a multicatolog Bayes classifier is proposed in this paper. The structure of PNN is a four-layer Feedforward Neural Networks(FNN), where the Parzen gaussian probability density function is regarded as a internal node. In this way, the EM algorithm is extended to deal with supervised learning one muticatolog of the neural network Gaussian classifier. The computational efficiency and the numerical stability of the training algorithm benefit from the well_established EM framework. The effectiveness of the proposed network architecture and its EM training algorithm are assessed by an experiment.