Applying saliency-map analysis in searches for pulsars and fast radio bursts

¹ National Astronomical Observatories, Chinese Academy of Sciences, A20 Datun Road, Chaoyang District, Beijing 100101, PR China
² University of Chinese Academy of Sciences, Beijing 100049, PR China
e-mail: zhangchao215@mails.ucas.ac.cn
³ CSIRO Data61, Sydney, NSW 2015, Australia
⁴ CSIRO Astronomy and Space Science, Australia Telescope National Facility, Box 76 Epping, NSW 1710, Australia
⁵ CSIRO Scientific Computing, Sydney, NSW 2015, Australia
⁶ CAS Key Laboratory of FAST, National Astronomical Observatories, Chinese Academy of Sciences, Beijing 100012, PR China
⁷ NAOC-UKZN Computational Astrophysics Centre (NUCAC), University of KwaZulu-Natal, Durban 4000, South Africa
⁸ Purple Mountain Observatory, Chinese Academy of Sciences, Nanjing 210008, PR China

Received: 2 December 2019
Accepted: 22 May 2020

Abstract

Context. We investigate the use of saliency-map analysis to aid in searches for transient signals, such as fast radio bursts and individual pulses from radio pulsars.

Aims. Our aim is to demonstrate that saliency maps provide the means to understand predictions from machine learning algorithms and can be implemented in pipelines used to search for transient events.

Methods. We implemented a new deep learning methodology to predict whether any segment of the data contains a transient event. The algorithm was trained using real and simulated data sets. We demonstrate that the algorithm is able to identify such events. The output results are visually analysed via the use of saliency maps.

Results. We find that saliency maps can produce an enhanced image of any transient feature without the need for de-dispersion or removal of radio frequency interference. The maps can be used to understand which features in the image were used in making the machine learning decision and to visualise the transient event. Even though the algorithm reported here was developed to demonstrate saliency-map analysis, we have detected a single burst event, in archival data, with dispersion measure of 41 cm⁻³ pc that is not associated with any currently known pulsar.