一种概率映射网络的EM训练算法

• 作 者: 熊汉春 贺前华 李海洲
• 机 构: 华南理工大学电子与通信工程系,华南理工大学电子与通信工程系,华南理工大学电子与通信工程系 广州 510641,广州 510641,广州 510641
• 出 处: 电子科学学刊
• 发表日期: 1999年02期
• 关键词: 训练算法 EM 监督学习 贝叶斯准则 贝叶斯策略 概率映射网络 高斯核 网络参数 多类分类 神经网络分类器

本文提出一种概率映射网络(PMN)的EM(Expectation Maximization)训练算法。PMN为一个四层前馈网。它构成一个贝叶斯分类器,实现多类分类的贝叶斯判别,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PMN用高斯核函数作为密度函数,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。最后的实验表明此网络及其学习算法在分类应用中的有效性。

This paper presents an EM (Expectation Maximization) training algorithm of Probabilistic Mapping Network (PMN). PMN is a four-layer feedforward network. It forms a Bayesian classifier, which implements Bayesian classification of multiple classes. The input sample model is transformed into output classification decision by network. The nodes in the network correspond to the variables of Bayesian posterior probability formula. The PMN uses Gaussian kernel function as the density function, and the network parameter training is implemented by the EM algorithm. The learning mode is supervised learning between classes and unsupervised learning within the class. The final experiment shows the effectiveness of this network and its learning algorithm in classification application.