Applications and Interpretability of Machine Learning to Identify Gravitational Lenses

Wilde, Joshua William (2023). Applications and Interpretability of Machine Learning to Identify Gravitational Lenses. PhD thesis The Open University.

DOI: https://doi.org/10.21954/ou.ro.00017239

Abstract

In the near future, Euclid and the Vera C. Rubin Observatory will survey over 10⁹ galaxies. Among these galaxies will be thousands of gravitational lenses that will be able to constrain the value of H₀, measure Dark Matter halos, and magnify more distant sources. This is too much data for humans to sift through themselves; machine learning is needed to find these lenses. I have trained CNNs to identify gravitational lenses in simulated Euclid images; one of my CNNs OU-200 performed comparably to a human. I have used interpretability techniques such as occlusion maps, Deep Dream, and guided Grad-CAM to understand what features within the image the CNN has learnt to associate with gravitational lenses. I used OU-200 to search for gravitational lenses in the Hubble WFC3 Infrared Spectroscopic Parallel (WISP) survey. In this survey there was one clear candidate. I then trained a U-net to not only classify images with gravitational lenses but also to highlight the position of the lens within the image.

Viewing alternatives

Metrics

Item Actions

Export

You can export this page using these formats Digital Object Identifier - DOI