Enhancing the classification accuracy by intra-concentration and the distance between the class boundaries instead of the class centers

Abstract

Abstract The softmax loss function is a commonly used loss function in the field of classification, which aims to increase the angle between two classes in feature space. However, it has some limitations such as class overlap and treating all misclassifications equally, and issue with imbalanced classes. Recently, the I2CS (Intra concentration and inter-separability) loss function has been proposed with a different approach from the softmax loss function, which is compressing data at the center and increasing class distance through the class center, which makes it able to overcome some of the limitations such as class-imbalanced problems, outliers and discover samples of unseen classes. Nevertheless, it still suffers from class overlap problem. Therefore, we have designed a new loss function with a novel approach to not only overcome the limitations of the softmax loss function but also address the class overlap issue of I2CS, and be effective in dealing with class imbalances. Furthermore, our purpose loss function has been thoroughly tested on a variety of standard benchmark datasets such as MNIST, CIFAR, and LFW as well as on imbalanced MNIST class, showcasing enhanced performance when contrasted with the softmax loss function and other widely-used loss functions.