Generalization accuracy of probabilistic neural networks compared with backpropagation networks

摘要

The authors demonstrate that probabilistic neural networks (PNN) and backpropagation networks (BPN) generalize comparably for a wide variety of low- and high-dimensional artificial databases. A training time advantage is most important when exploring new databases and preprocessing techniques to determine classification accuracies for potential applications. Since it is demonstrated that classification accuracy can be determined equally well using either PNN and BPN, it is advantageous to use PNN for this stage of a classification project. It is during this phase that most of the effort goes into training a network and a relatively small amount of effort goes into testing and evaluating the accuracy of the network. The PNN is based on the Bayes strategy for decision making and Parzen window estimation, and is asymptotically Bayes-optimal within a given feature space.