Paying Attention to Radio Astronomy Data

  • Micah Bowles

Student thesis: Master of Science by Research


The development of next generation instruments in radio astronomy poses a massive data processing challenge. The quantity of data which will be produced demands that fast, reliable and trusted automation processes are deployed. In this work I design, develop and present attention gated CNN as a possible solution by evaluation on a FRI/FRII radio galaxy classification task. The presented models have added transparency, as the scaling attention maps used by the model for classification can be extracted by the user to investigate what regions of the input the model is using to make its outgoing classification.Additionally, the trained attention gated CNN models are shown to perform on par with the current state of the art models on the available data set, but with only half the number of network parameters. I show that the attention maps align with the regions a human expert would use to classify the sources. Various implementations and their respective attention maps are investigated. Attention maps are presented throughout various stages of the training process to help enhance the users understanding of the model. Model attention is evaluated across the entire test set and is shown to behave as expected on various subsets of testing data. Finally, I extract and analyse eight sources which the model consistently miss-classifies to demonstrate that deep learning can be used both to evaluate data for population studies, and to highlight abnormal and potentially interesting sources for future studies.
Date of Award31 Aug 2021
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorPaddy Leahy (Supervisor) & Anna Scaife (Main Supervisor)


  • Convolutional Neural Network
  • Artificial Intelligence
  • Explainable
  • Machine Learning
  • Attention
  • Classification
  • Galaxies
  • Radio
  • Computer Vision

Cite this