Bayesian Deep Learning for Pulsar Classification

Abstract

Neural networks are powerful tools for classification that can process vast amounts of data in a very short time. Bayesian neural networks offer the same capacity for classification, as well as being able to return well balanced uncertainties on their predictions. However, these networks are expensive from a computational point of view. A method called dropout, which randomly disconnects neurons during training, can be used as an approximation for bayesian inference, being able to not only perform fast classification tasks but also return the uncertainties associated with these tasks. In this project, I apply the dropout classification method to a classic astronomical task: pulsar classification. This task has been historically done manually, but with the advent of a new generation of extremely powerful telescopes such as the Square Kilometer Array, it will become a necessity to have good automated classifiers. With this issue in mind, I developed BonNet, a small but mighty dropout powered classifier that is able to perform highly accurate and precise classification and is also returning well-balanced uncertainties on its predictions. BonNet was trained and tested on two data subsets of the High Time Resolution Universe (HTRU) survey, HTRU1, and HTRU2. On HTRU1, the final version of BonNet achieved an accuracy of 99%, a precision of 98%, a recall of 99%, and a false positive rate of 0.42%. For HTRU2, BonNet achieved an accuracy of 97%, a precision of 95%, a recall of 80%, and a false positive rate of 0.53%.

Date of Award	16 May 2022
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Albert Zijlstra (Co Supervisor) & Anna Scaife (Main Supervisor)