Knowledge aggregation networks for class incremental learning

Most existing class incremental learning methods rely on storing old exemplars to avoid catastrophic forgetting. However, these methods inevitably face the gradient conflict problem, the inherent conflict between new streaming knowledge and existing knowledge in the gradient direction. To alleviate gradient conflict, this paper reuses the previous knowledge and expands the branch to accommodate new concepts instead of fine-tuning the original models. Specifically, this paper designs a novel dual-branch network called Knowledge Aggregation Networks. The previously trained model is frozen as a branch to retain existing knowledge, and a consistent trainable network is constructed as the other branch to learn new concepts. An adaptive feature fusion module is adopted to dynamically balance the two branches’ information during training. Moreover, a model compression stage maintains the dual-branch structure. Extensive experiments on CIFAR-100, ImageNet-Sub, and ImageNet show that our method significantly outperforms the other methods and effectively balances stability and plasticity.