More discriminative convolutional neural network with inter-class constraint for classification

Abstract Recently, convolutional neural network (CNN) has achieved impressive results in object classification tasks. Through various modifications and careful design of the inner structure of CNN, its performance has already become human-competitive, according to recent reports. To some extent, the testing accuracy depends largely on the decision boundaries produced by classifiers, which classify different objects into specific feature spaces. Hence, the relationships among samples in the feature space are critical. However, the softmax loss function that is used in most CNN models, does not directly contain the relationship information. In this paper, we propose a novel loss function, named inter-class constraint loss function, that maximizes the distance between different classes. Together with softmax loss, we can obtain larger inter-class distances and smaller intra-class distances in CNN, thus significantly improving the accuracy in classification. We achieve substantial improvements for the SVHN, CIFAR-10 and CIFAR-100 datasets using our proposed loss function.