Latency-Aware Online Continual Learning for Non-Stationary Data Streams

Online continual learning (CL) is beneficial for learning incrementally from continuous data streams without forgetting previously learned knowledge. However, current online CL approaches have overlooked the time cost of online data collection and model adaptation, resulting in a high-latency service response, especially in high-velocity non-stationary data streams. In this work, we aim to realize latency-aware online CL for non-stationary data streams, and propose a two-stage time-scale optimization for online data collection and model adaptation. In the first stage with uncertain data arrivals, we propose an optimal stopping algorithm with a logarithmic regret bound to make an irrevocable decision on when to stop data collection. To minimize the training time of model adaptation for stability-plasticity trade-off in the second stage, we introduce a bidirectional data selection algorithm with a logarithmic approximation, to greedily determine which samples to select from both newly collected data and the previous ones. Extensive evaluations demonstrate that our proposed approach consistently outperforms the-state-of-art solutions, improving the accuracy by 16.8% on average and reducing the latency by up to 6.2 times.