Advancing machine learning in astrophysics

Thesis

This thesis explores four projects applying supervised deep learning to help answer astrophysical questions.

I first consider faint tidal features. Tidal features are a long-lasting signature of galaxy mergers, making them useful for measuring merger rates. However, current automated methods struggle to detect faint tidal features in complex galaxies. I use convolutional neural networks to identify galaxies with tidal features in the CFHTLS- Wide Survey, improving on previous methods applied to the same dataset. I show that my networks can identify which pixels are associated with tidal features, potentially enabling researchers to not only identify but also characterise tidal features.

I then turn to Galaxy Zoo, a citizen science project measuring galaxy morphology. Galaxy Zoo is being gradually outpaced by the increasing scale of new surveys. Au- tomated classifiers can be trained using volunteer responses; however, such classifiers are often unable to consider uncertainty in either volunteer responses or predictions, leading to wasted volunteer effort and overconfident classifications. I introduce a probabilistic approach that allows classifiers to flexibly express uncertainty. I use this probabilistic approach to build a machine learning system that ‘asks’ volunteers to label the galaxies it could best learn from. I relaunch Galaxy Zoo with images from the Dark Energy Camera Legacy Survey and run my system live, collecting 1.8 million volunteer responses. My final models are around 99% accurate on every question for galaxies with confident volunteer answers and are otherwise correctly uncertain.

Next, I help the Canadian Hydrogen Intensity Mapping Experiment (CHIME) detect fast radio bursts. CHIME only attempts to detect FRB above a signal-to- noise threshold of σ = 8.5, in part for lack of expert time to review candidates. I created and launched a citizen science project to classify the 7.8 ≤ σ < 8.5 signal-to- noise candidates detected by CHIME each week. Candidates found by this project may be the most distant fast radio bursts ever detected, which I hope will serve as useful cosmological probes of the intergalactic medium.

Finally, I show that neural network emulation can efficiently recover posteriors of galaxy parameters from photometry. Galaxy SED simulators are too slow to use MCMC inference on large samples. I train a neural network to emulate an SED simulator, providing both faster likelihood evaluations and known gradients. These gradients can then be used for efficient Hamiltonian Monte Carlo inference.

Together, these projects show how deep learning can help astronomers make ef- fective use of limited and uncertain data.

Authors: Walmsley, M

• DOI: 10.5287/ora-koy98q49j
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: astronomy; machine learning; galaxies; computer vision; deep learning
• Subjects: astronomy; machine learning; galaxies; computer vision; deep learning