Improving deep imbalanced classification via large-margin mix-up

Abstract

A large variety of labeled datasets in practice are often severely imbalanced class-wise and it is well-known that modern deep neural networks poorly generalize from such datasets due to overfitting classes of majority training samples. To address the issue, we explore the recent state-of-the-art regularization method, called Mixup, and found that it is also very effective for class imbalanced training. Motivated by this, we propose a novel, yet simple advanced alternative, coined Boundary-Mixup. It also generates synthetic training samples by following the original Mixup principle, but our main idea beyond it is to balance the uncertainty level between classes by generating mixed samples near the decision boundary of the classifier. We demonstrate the effectiveness of Boundary-Mixup for image classification, natural language processing and continual learning tasks, improving the prior baseline methods including the original Mixup.