基于异方差的PNN训练算法的研究

给出了一种基于高斯核函数有不同协方差的混合模型的异方差 PNN的实现。传统的 ML 估计方法和 EM算法在训练异方差 PNN存在数学上的困难甚至导致算法失败。从协方差的根源出发提出了协方差限制的简单 EM改进算法 ,又基于杰克刀技术提出了鲁棒性的 EM改进算法。此外又从信息论出发提出了基于最小相对熵的训练新算法。闭集文本自由说话人辨认试验证明了提出的模型及其算法的正确性。

Research on Heteroscedastic PNN Training Algorithms

The heteroscedastic PNN that is a mixture of Gaussian basis function having different variances is described.When training heteroscedastic PNN,the traditional maximum likelihood estimation approach and expectation maximization algorithms often encounter numerical difficulties and make algorithms failure.So there approaches are proposed to train heteroscedastic PNN,including the simple improved EM algorithm based on variance limiting,the robust improved EM algorithm based on Jack-knife methods,and the novel training algorithm based on minimum relative entropy of information theory.The experiments results of the closed-set text-independent speaker identification system indicate that the proposed models and algorithms improve identification accuracy.