Improving generalization of k-means clustering based probabilistic neural network using noise injection

Abstract

In this article, a methodology has been presented to enhance the generalization of the probabilistic neural network (PNN). For the purpose, in this article, I have performed a noise based training over the PNN classifier. Here, while training the PNN, I have injected random Gaussian multiplicative noise in the samples of the data set. This external noise injection mechanism improves the classification accuracy of the data set. Furthermore, to reduce the storage requirement of the network, I have used k-means clustering algorithm, and through this algorithm I have selected a subset of class samples from each class. It reduces the number of stored pattern in the pattern layer. The entire process generates a advanced classifier based on fusion neural network model. To test the classification rightness of the proposed method, eight standard data sets have been used. Proposed model has been compared with conventional PNN classifier. Comparison of result exhibit the ascendancy of the presented method. Wilcoxon signed rank trial also manifests that proposed method improves the performance of the classifier.