Online Continual Learning with Contrastive Vision Transformer

Online continual learning (online CL) studies the problem of sequential tasks from an online data stream without task boundaries, aiming to adapt new while alleviating catastrophic forgetting on past tasks. This paper proposes a framework Contrastive Vision Transformer (CVT), which designs focal contrastive strategy based transformer architecture, achieve better stability-plasticity trade-off for CL. Specifically, we design external attention mechanism CL that implicitly captures previous tasks’ information. Besides, CVT contains learnable focuses each class, could accumulate knowledge classes alleviate forgetting. Based focuses, loss rebalance between and consolidate previously learned representations. Moreover, dual-classifier structure decoupling current balancing all observed classes. The extensive experimental results show our approach achieves state-of-the-art performance with even fewer parameters benchmarks effectively alleviates