Radio Galaxy Zoo’s ClaRAN
On the 31 October 2018, Radio Galaxy Zoo published its first end-to-end machine learning system for “Classifying Radio sources Automatically using Neural networks” (ClaRAN). This paper is led by ClaRAN’s developer, Chen Wu, a data scientist at the International Centre for Radio Astronomy Research at the University of Western Australia (ICRAR/UWA), who repurposed the FAST-rCNN algorithm (used by Microsoft and Facebook) to classify radio galaxies. ClaRAN was trained on radio galaxies classified by Radio Galaxy Zoo and so recognises some of the most common radio morphologies that have been classified.
The purpose of ClaRAN is to reduce the number of radio sources that require human visual classification so that future Radio Galaxy Zoo projects will have fewer “boring” sources, thereby increasing the chances of real discoveries by citizen scientists. ClaRAN (and its future cousins) are crucial for future surveys such as the EMU survey which is expected to detect ~70 million radio sources (using the Australian Square Kilometre Array Pathfinder telescope). While Radio Galaxy Zoo has made visual source classifications much more efficient, we will still need to reduce the total survey sample size to a sample for visual inspection that is less than 1% of the 70 million sources.
How does ClaRAN work? ClaRAN inspects both the radio and coordinate-matched infrared overlay in the same fashion as RGZ Zooites, and then determines the radio source component associations in a similar fashion to the RGZ Data Release 1 (DR1) catalogue. As ClaRAN is still in its prototype stage (–analogous to the capabilities of a toddler), it only understands 3 main classes of radio morphologies — sources which have 1-, 2- or 3- separate radio components. ClaRAN was trained to understand these three different radio morphologies through seeing examples of all three classes from the
RGZ DR1 catalogue. The animated gif in the figure below (from the ICRAR press release) describes how ClaRAN “sees” the example radio galaxy.
As we look towards the future, we look forward to teaching ClaRAN some of
the more complex and exotic radio galaxy structures. For that to happen, we need to assemble much larger samples of more complex radio morphology classifications. With your support of Radio Galaxy Zoo, I am sure that we will get there.
Fun fact: did you know that some of the more obscure bugs in the RGZ DR1 catalogue processing was actually found through training ClaRAN? This is because ClaRAN is a good learner and will learn all the small details that we didn’t initially notice. We only discovered these bugs through some of the funny answers that we got out of some of the early testing of ClaRAN.
Thank you very much again to all our Radio Galaxy Zooites for your support. More information on the ICRAR press release for ClaRAN can be found via this link: https://www.icrar.org/claran/