Applying Deep Learning to Fast Radio Burst Classification

Abstract

Upcoming Fast Radio Burst (FRB) surveys will search $\sim$10\,$^3$ beams on sky with very high duty cycle, generating large numbers of single-pulse candidates. The abundance of false positives presents an intractable problem if candidates are to be inspected by eye, making it a good application for artificial intelligence (AI). We apply deep learning to single pulse classification and develop a hierarchical framework for ranking events by their probability of being true astrophysical transients. We construct a tree-like deep neural network (DNN) that takes multiple or individual data products as input (e.g. dynamic spectra and multi-beam detection information) and trains on them simultaneously. We have built training and test sets using false-positive triggers from real telescopes, along with simulated FRBs, and single pulses from pulsars. Training of the DNN was independently done for two radio telescopes: the CHIME Pathfinder, and Apertif on Westerbork. High accuracy and recall can be achieved with a labelled training set of a few thousand events. Even with high triggering rates, classification can be done very quickly on Graphical Processing Units (GPUs). That speed is essential for selective voltage dumps or issuing real-time VOEvents. Next, we investigate whether dedispersion back-ends could be completely replaced by a real-time DNN classifier. It is shown that a single forward propagation through a moderate convolutional network could be faster than brute-force dedispersion; but the low signal-to-noise per pixel makes such a classifier sub-optimal for this problem. Real-time automated classification may prove useful for bright, unexpected signals, both now and in the era of radio astronomy when data volumes and the searchable parameter spaces further outgrow our ability to manually inspect the data, such as for SKA and ngVLA.