The study presents a new framework for Class-incremental learning (CIL), specifically for Long-Tailed Class-Incremental Learning (LT-CIL). The proposed Task-aware Expandable (TaE) framework dynamically allocates and updates task-specific trainable parameters, learning diverse representations from each incremental task while resisting forgetting. This is achieved through the majority of frozen model parameters. The paper also introduces a Centroid-Enhanced (CEd) method to guide the update of these task-aware parameters, designed to minimize intra-class feature distances while maximizing inter-class feature distances. The study concludes that the TaE framework achieves state-of-the-art performance in experiments conducted on CIFAR-100 and ImageNet100.
Publication date: 9 Feb 2024
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2402.05797
