A Deep Neural Network Classifier Based on Belief Theory

Abstract

Classification is a machine learning technique that is used to find the membership of an object to a given set of classes or groups. Neural network (NN) classifier always suffers some issues while examining outliers and data points from nearby classes. In this paper, we present a new hyper-credal neural network classifier based on belief theory. This method is based on credal classification technique introduced in Dempster-Shafer Theory (DST). It allows a data point to belong not only to a specific class but also to a meta-class or an ignorant class based on its mass. The sample which lies in an overlapping region is accurately classified in this method to a meta-class which corresponds to the union of the overlapping classes. Therefore this approach reduces the classification error at the price of precision. But this decrease in precision is acceptable in applications such as medical, defence related applications where a wrong decision would cost more than avoiding some correct decisions. The results and analysis of different databases are given to illustrate the potential of this approach. This idea of hyper-credal classification is extended to convolutional neural network (CNN) classifiers also.