Error-correcting output codes for multi-class classification based on Hadamard matrices and a CNN model

Abstract

The error-correcting output codes(ECOC) is the ensemble method for the multi-class classification problem. In some applications of ECOC, Hadamard matrices are used because they have good properties to apply ECOC. However, due to the difficulties of the construction of Hadamard matrices of some specific orders, only Hadamard matrices of order a power of 2 constructed by Kronecker product were usually used. In this paper, we consider Hadamard matrices of various orders to the Hadamard ECOC to determine which orders of Hadamard matrices give a good performance compared to the number of classes. We apply the Hadamard ECOC to the image datasets including CIFAR-10, CIFAR-100, subsets of CIFAR-100, and EMNIST-Letters. We also compare the result of the Hybrid Hadamard ECOC with the Hadamard ECOC. For our experiment, a convolutional neural network(CNN) is chosen as a base learner, which is largely used in the image classification problems.