Evolving attention Exploring the use of genetic algorithms and attention for gravitational wave data science

Norman, Michael 2023. Evolving attention Exploring the use of genetic algorithms and attention for gravitational wave data science. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (49MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (469kB)

Abstract

This thesis investigates the application of machine learning, particularly artificial neural networks, to gravitational-wave data analysis. It explores the optimization of these networks using genetic algorithms for hyperparameter optimization. Additionally, it tests the performance of attention-based networks in tasks such as compact binary coalescence detection and parameter estimation of overlapping pairs of compact binary coalescences. The thesis provides a contextual background on gravitational wave science and details the principles of neural networks, emphasizing their effectiveness as a data analysis method. It introduces custom software for rapid dataset generation and reviews previous work in the field, demonstrating the ineffectiveness of unspecialized networks. The genetic algorithm Dragonn is presented as a general optimization method for neural networks in this context. Experiments show the marginal improvement of attention-based networks over convolutional networks for gravitational-wave analysis. Finally, the thesis investigates cross-attention models to estimate parameters from overlapping signals, showing promising results for future parameter estimation techniques.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Physics and Astronomy
Subjects:	Q Science > QC Physics
Uncontrolled Keywords:	Gravitational-wave, data analysis, machine learning, gravitational-wave science, artificial neural networks, genetic algorithms, neural networks, hyperparameter optimization, low-latency burst, burst search pipeline, MLy, attention-based networks, binary coalescence, parameter estimation, cross-attention models, Dragonn, Skywarp, CrossWave, dataset generation, convolutional networks, overlapping signals, machine learning algorithms, machine learning integration, parameter estimation techniques.
Funders:	EPSRC
Date of First Compliant Deposit:	9 July 2024
Last Modified:	09 Jul 2024 15:53
URI:	https://orca.cardiff.ac.uk/id/eprint/170459

Actions (repository staff only)

Edit Item