GPU-accelerated Scan Statistics With A Case Study In Radio Astronomy

Outlier detection in data with high variability and with clustering present in multiple scales requires using scan statistics -a process where statistical properties of the data are computed within appropriately sized moving windows that allows detection of outliers (signal) in the presence of bias or background.While this technique has been effectively employed in diverse areas, in this paper we address the computationally expensive case of deriving scan statistics (in particular, median filters) when detecting and cataloging astronomical sources in big-data radio astronomy images such as those produced by the Australian Square Kilometre Array Pathfinder (ASKAP) telescope.We present a GPUoptimized implementation where we introduce a new versatile abstract machine namely Network for Scan Statistics.While high performance computing (HPC) strategies have been adopted in advanced source-finding software packages like Selavy, we demonstrate significant additional improvement (3x) in processing times using our novel methods that employ hardware acceleration, while using fewer resources.This speed-up, compared to the widely used single CPU approach provided in the core libraries of the Common Astronomy Software Applications (CASA), is 229x.We believe that our approach could find applications beyond source-finding and radio astronomy.