Recent Advances in Class-Incremental Learning

Dejie Yang, Minghang Zheng, Weishuai Wang, Sizhe Li, Yang Liu

Abstract: A large number of deep learning models have been applied in a wide range of fields nowadays. However, most existing models can only generalize to the categories in the training set and are unable to learn new categories incrementally. In practical applications, new categories or tasks will constantly emerge, which requires models to continuously learn new category knowledge like humans while maintaining existing category knowledge. Such the learning process, i.e., class-incremental learning (CIL), abstracts more attention from the research community. CIL faces several challenges, such as imbalanced data distribution, limited model memory capacity, and the catastrophic forgetting of category representation. Therefore, we provide an up-to-date and detailed overview of CIL methods in this survey, including data-based, model-based, and representation-based approaches. We also discuss the impact of pre-trained models on CIL and compare the latest methods on widely-used benchmarks. Finally, we summarize the challenges and future directions of CIL.