Next-generation science instruments such as the SKA, LSST, and terrestrial sensor networks will dramatically increase the volume of collected data. This enables detection of very rare transient anomalies, but also creates new challenges since comprehensive storage is impossible and analysis must occur in real time.
Dr. Thompson will discuss machine learning approaches for online anomaly detection in data streams. Pattern recognition triages the incoming data for comprehensive analysis of candidate events, retaining robustness against changing noise conditions and interferences. Examples from radio astronomy (the Very Long Baseline Array Fast Transients Experiment) demonstrate the practical benefits of an adaptive approach.