Class-incremental learning is an unsolved problem in the domain of deep learning. The challenging point of class-incremental learning is to maintain the performance of original classes while learning new classes. Without class-incremental learning, deep learning based classification models cannot adaptively evolve to newly introduced environments and this is a significant barrier to deep learning based consumer electronics. In this paper, we address the problem of the softmax layer in the class-incremental learning process and tackle the problem by adaptively regularizing incremental weights. Our method impressively improves the performance of class- incremental learning tasks compared to naive approaches.