A Continual Learning Method for Reducing Class Interference Based on Replay

Although deep neural networks perform well on many individual tasks, they suffer from catastrophic forgetting when learning new tasks continually. Recently, various continual learning methods have been proposed, and some approaches based on replaying memory data achieve promising performance. Maximally Interfered Retrieval is a strong replay-based baseline, however, exists class interference due to class imbalance and lacks the ability to generalize to the real class-incremental scenario. In this paper, aiming at these problems, we design Class cumulative Classifier to replace the shared output layer of the original network, which makes it closer to practical applications, and Class balanced Buffer to address the class imbalance of stored samples. In addition, we propose Retrospect strategy to further improve the accuracy. Experimental results on benchmark datasets show that our method outperforms several strong baselines and is more suitable for complex datasets with more classes.