Selection of Pattern Neurons for a Probabilistic Neural Network by Means of Clustering and Nearest Neighbor Techniques

The architecture of the probabilistic neural network (PNN) depends on the input data cardinality. This is the effect of the arrangement of neurons in the network's pattern layer. Thus, in order to cope with large data classification tasks, PNN's structure must be minimized. In this work, new algorithm for the PNN's architecture simplification is proposed. The approach is realized in two phases. First, a k-means data clustering is performed and initial PNN's pattern neurons are appropriately selected using obtained centers. Second, based on the nearest neighbor to the determined centers, final data records are chosen and used to activate the pattern neurons. The algorithm is applied to the classification tasks of four repository data sets. PNN is trained by means of a conjugate gradient procedure with a Gaussian kernel function utilized for neurons' activation. The performance of original and reduced network is compared using a 10-fold cross validation method. The outcomes are also collated with state-of-the-art results.