Real-Time Memory-Bounded A/RED: Discovering Rare Relevant Classes in Streaming Data

We present a memory-bounded, real-time Anomalous/Relevant Event Detection (A/RED) system for indefinitely long data streams. Extending prior work, our algorithm maintains bounded memory using a circular buffer of queried points and accelerates nearest-neighbor search via ball trees, reducing complexity from O(n 2 ) to O(d log(n) + kd) per point. Small clusters are periodically merged to control model size, while a tunable paranoia parameter κ regulates query rate. Evaluated on synthetic (NICE), skewed MNIST, Extended MNIST (EMNIST), and industrial inspection (MVTec) datasets as streams, A/RED achieves near-perfect relevant recall and query precision on well-separated data, and significantly outperforms random querying on realworld distributions — discovering rare relevant classes with reduced user burden. This work enables scalable, human-in-the-loop anomaly detection in continuous, high-volume sensor environments.