Class-incremental learning via prototype similarity replay and similarity-adjusted regularization

The task of incremental learning is to enable machine learning models to continuously learn and adapt to new tasks and data in changing environments while maintaining knowledge of prior tasks. Recently, researchers have proposed a variety of incremental learning methods. Some methods rely on data storage or complex generative models to perform satisfactorily. However, existing incremental learning approaches typically focus on mitigating catastrophic forgetting, with less emphasis on effectively applying old knowledge to facilitate learning new tasks. In this paper, we propose a non-exemplar-based incremental learning approach called Class-Incremental Learning via Prototype Similarity Replay and Similarity-adjusted Regularization (PSSR) to tackle catastrophic forgetting in incremental learning. The essence of PSSR is leveraging prior knowledge of prior tasks to facilitate the acquisition of new tasks. PSSR memorizes a prototype for each old class, representing the class, and learns the new classes based on the similarity between the prototypes and the new class samples during the learning process. The feature space distribution is modified by the old class prototypes to enhance the model’s learning of the new classes. Extensive experiments on three benchmark datasets demonstrate the superior incremental performance of PSSR, with classification accuracy improvements of 2.73%, 3.37%, and 4.21% over state-of-the-art methods. Code available at https://github.com/FutureIAI/PSSR.