Adapt and Feature Translation for Class-Incremental Learning with Pre-Trained Models

Class-incremental learning (CIL) aims to enable a learning system to continuously learn new information. Although pre-trained models (PTMs) exhibit strong performance in CIL, the lack of data from previously learned classes results in class imbalance and catastrophic forgetting during the updating process. Consequently, integrating new categories while retaining previous knowledge remains a significant challenge. In this paper, we introduce the Adapt and Feature Translation (AFT) method for PTM-based CIL, designed to enhance the stability-plasticity balance. The adapter module is employed to adapt the model for specific downstream tasks, while keeping all backbone parameters fixed during subsequent updates to maintain the stability of the feature extraction. Additionally, actual features of new classes, combined with synthesized features of old classes via translation, are jointly used to retrain the classifier, ensuring a balanced classifier. Experimental results on multiple datasets validate the effectiveness of our method, achieving a new state-of-the-art performance.