A strategy for developing a pair of diverse deep learning classifiers for use in a 2-classifier system

Abstract

A multi-classifier system requires the individual classifiers to be both accurate and diverse. There is always a trade-off between accuracy and diversity. One of the suitable ways to enforce diversity between the classifiers is to use different neural network architectures, hyperparameters and training dataset. However, in general, these individual classifiers are trained by mapping the label into a binary vector such that all the incorrect classes are penalized equally. In this paper, we propose a methodology of generating the label vector into two different forms by exploiting the similarity between the classes. One of these forms of label vector encode values such that the misclassification into a class of lower similarity is penalized more than the misclassification into a class of higher similarity, whereas the second form of label vector encode values to do the opposite. These two forms of label vectors are then used for training two individual classifiers. The results show that the pair of classifiers developed with this proposed approach, when used in a 2-classifier system, yields better performance in terms of both diversity and accuracy as compared to the pair of classifiers trained with the standard approach of penalizing all the incorrect classes equally via binary label vector.