Abstract
We propose a novel residual network called competitive residual network (CoRN) for image classification. The proposed network is composed of residual units having two identical blocks each containing convolutional filters, batch normalization, and a maxout unit. The maxout unit enables the competition among the convolutional filters and reduces the dimensionality of the convolutional layer. The proposed network outperforms the original residual network by a significant margin and test errors on benchmark datasets (CIFAR-10/100 and SVHN) are comparable to the state-of-the-art. Using the ensemble network, we achieve a test error of 3.85% on CIFAR-10, 18.17% on CIFAR-100 and 1.59% on SVHN. (C) 2020 The Korean Institute of Communications and Information Sciences (KICS). Publishing services by Elsevier B.V.