Classification of radio galaxies is looking at their shape, their morphology. Studying the shapes of galaxies in one thing we use to learn more about the formation and evolution of those great megacities of hundreds of billions stars, black holes and other cosmic beasts.
Radio galaxies are galaxies that have very interesting features that are visible in radio wavelengths. Those features often spread far beyond the visible galaxy, and carry information on what is inside the galaxy, such as central supermassive black holes, how the galaxy evolves over time, and how it interacts with the surrounding vast expanses of intergalactic space.
Classifying galaxies’ shapes was a task best left to humans until in recent years progress in machine learning was made possible by cheaper computing and memory. Since, many machine learning algorithms have been developed, each best at one task or another.
A type of machine learning algorithms performing extremely well at image analysis is called “Convolutional Neural Networks”, CNNs for short. In this paper, the team compared the performance of different CNNs for the classification of radio galaxies based on observations taken by radio telescopes. Two main things were tested. 1. whether the algorithms were overfitting, i.e. optimistically seeing features in the images of galaxies that aren’t there, and 2. the practical aspects of applying CNNs for the purpose of radio galaxy classification.
They find that different implementations of CNNs performed differently as radio galaxy classifiers. This work will help them, and other researchers in radio astronomy use the most effective machine learning algorithms in the context of radio galaxy classification.
The paper has been accepted for publication in the Monthly Notices of the Royal Astronomical Society.
Reference:
Becker B., Vaccari M., Prescott M. and Grobler, T. L.
CNN Architecture Comparison for Radio Galaxy Classification
Preprint