Abstract

Continual learning from non-stationary data streams remains a fundamental challenge for autonomous agents. We revisit continual learning by proposing a formal framework that addresses the dual challenge of context alternation and within-context drift. We illustrate this on context-aware anomaly detection, where detecting anomalies alone is insufficient: the system must determine whether detected anomalies are hazardous within specific, evolving contexts. To address this, we introduce a benchmark with alternating contexts and smooth distribution drifts, and propose a solution leveraging topology-preserving vector quantization (GNG-T) to simultaneously track context-specific representations and register so-called contexts. Contexts encode both a structural characterization of the current distribution and context-dependent classifiers. Experimental results demonstrate that explicit context recognition significantly improves recovery after context switches and enables effective anomaly hazard classification, outperforming approaches that continuously train a single classifier. The code for the experiments is available online.