Dr. Kiri Wagstaff

Jet Propulsion Laboratory, National Aeronautics and Space Administration, USA

Machine Learning for Discovery in Radio and Optical Astronomy Investigations

Abstract: Machine learning provides the ability to quickly sift through large data sets to identify observations of known scientific interest and highlight unexpected observations that could lead to new discoveries. I will describe two astronomy projects that leverage machine learning to achieve these goals. The V-FASTR system was designed to enable real-time detection and classification of transient radio pulses in data from the VLBA. V-FASTR also seeks to enable potential discoveries of new radio sources, such as pulsars, star forming regions, and others we haven’t yet categorized. Machine learning is used to filter out spurious detections and reduce the human effort needed to review the detections that arrive each day. Annotations made by reviewers are used to re-train the machine learning classifier each night, enabling continual improvement in classifier performance and in human time saved. Second, we have developed a processing pipeline for galaxy observations from the Dark Energy Survey to identify outliers and generate explanations to help determine whether they (1) indicate an upstream data collection or processing issue or (2) are of scientific interest. Both systems aim to accelerate the process of scientific discovery by directing human attention to where it is most needed.