Online Continual Learning under Real Concept Drift: A Statistical Perspective

Haihan Xie, Jinhan Xie, Lingzhu Li, Chao Gao, Bei Jiang, Linglong Kong

Keywords: Continual learning, Concept drift, Regularization, Memory buffer, Online algorithm

Abstract: Real-world data often exhibit non-stationarity, prompting growing interest in adaptive learning techniques. Continual learning, which aims to sequentially learn multiple tasks, provides a promising framework to address this challenge. However, learning under real concept drift, where the relationship between inputs and outputs evolves over time, remains relatively underexplored. In this paper, we propose a novel regularization-based method that incorporates a memory buffer to improve robustness against concept drift. Assuming the existence of a common center for the evolving true models, our method jointly constrains current and past task estimates, effectively bridging them to form a stable estimate that incorporates information across tasks. To further adapt to task variability, we develop an online algorithm that dynamically tunes task-specific regularization parameters. We also provide theoretical guarantees by deriving an error bound that characterizes the overall performance of the estimator, explicitly capturing the effects of task-relatedness, memory buffer size, and regularization strength. Extensive experiments demonstrate that our method achieves superior stability–plasticity trade-offs under varying degrees of task similarity.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 8085