MetaOCDN: A Cognition-Inspired Meta Optimized Complementary Dual Networks for Online Continual Concept Drift Adaptation

Zhijie Wu, Husheng Guo, Wenjian Wang

Keywords: open environment, concept drift, streaming data

Abstract: The *Complementary Learning Systems* (CLS) theory points that humans can continuously and efficiently adapt to new tasks through the collaboration between the hippocampus and the neocortex: the former rapidly encodes new knowledge, while the latter extracts structured knowledge by abstract learning. Their synergy enables humans not only to quickly learn new tasks in the short term but also to transfer acquired knowledge across different tasks. Inspired by this theory, we address the challenge of streaming data mining under open environment with concept drift by proposing a cognition-inspired meta optimized complementary dual networks architecture (MetaOCDN), which consists of the Adaptive Fine Tuning Network (AFT-Net) and the Meta Representation Network (MRN-Net). AFT-Net is similar to the hippocampus, selectively fine-tunes key layers based on gradient variations to achieve rapid adaptation to novel concepts; MRN-Net is similar to the neocortex, we design self-supervised duality loss to continuously enhance its deep representation capability, thereby improving generalization to unknown distributions; furthermore, we design MAML-based multi-scale knowledge distillation strategy to facilitate dynamic information flow and knowledge transfer between the two networks. In summary, MetaOCDN provides a brain-inspired collaborative architecture that integrates the rapid responsiveness of AFT-Net with the abstract generalization capacity of MRN-Net, and enhances their interaction through knowledge distillation, thereby achieving a dynamic balance between fast adaptation and stable generalization in non-stationary data streams with concept drift. Extensive experiments demonstrate that MetaOCDN consistently outperforms state-of-the-art baselines across various drift scenarios.

Primary Area: learning on time series and dynamical systems

Submission Number: 23128