More Discriminative CNN with Inter Loss for Classification

Recently years, convolutional neural networks (CNN) has been a hot spot in various areas such as object detection, classification. As deep study in CNN, its performance is almost human-competitive. We find that the test accuracy largely depends on the relationship of samples in feature space. Softmax loss is widely used in many deep learning algorithms. However, it cannot directly reflect this kind of relationship. In this paper, we design a new loss function, named inter loss. This inter loss function can maximizes the distance between different classes, analogous to maximizing margin in SVM. By integrating inter loss and softmax loss, larger inter-class distance and smaller intra-class distance can be obtained. In this way, we can significantly improve the accuracy in classification. Impressive results is obtained in SVHN and CIFAR-10 datasets. However, our main goal is to introduce a novel loss function tasks rather than beating the state-of-the-art. In our experiments, other forms of loss functions based on inter and intra class distance is also considered as to demonstrate the effectiveness of inter loss.