Dynamic Task Subspace Ensemble for Class-Incremental Learning

Deep-learning models are expected to continually learn new concepts without forgetting old ones in real-world applications with shifting data distributions. However, the notorious catastrophic forgetting often occurs. Recently, methods based on task subspace modeling have been developed to address this issue by gradually adding new subspaces to learn new concepts. In this paper, we reveal that such task-subspace-modeling methods may suffer from the inter-task confusion issue, leading to degraded performance in challenging class-incremental learning settings. Concerning addressing the forgetting issue of deep learning models, we propose a two-stage framework called Dynamic tAsk Subspace Ensemble (DASE), the first stage of which involves the dynamic expansion of the extractor network for memory efficiency, while the second stage delivers dynamic learning and aggregation of diverse features. To further enhance the discriminative capacity of the aggregated features for both historical and new classes, we also introduce new feature-enhancement techniques. Experimental results demonstrate that our method achieves state-of-the-art CIL performance on natural image datasets (CIFAR-100 and ImageNet) and Synthetic Aperture Radar image datasets (MSTAR and OpenSARShip).