Neural networks Ensemble for multi-dimensional classification

Abstract

Artificial neural network (ANN) is devised from human neural system in the field of machine learning. Several techniques including ensemble have been proposed to improve ANNs performance. In this thesis, we show three types of ensembles for classification task which approaches in the style of divide and conquer: pairwise, hierarchy, helper. The pairwise ensemble is using binary classifiers to solve one classification problem. The hierarchical ensemble is composed of networks which are independently trained to solve each small part of the problem. And the other ensemble method uses a traditional network that is assisted by a helper, trained to solve easy classification that is modified from original classification. The experiments are conducted with the MNIST database and the results show the problem and possibility of the style of divide and conquer in neural network.