PulSort: a novel machine learning approach to searching for pulsars

Pulsars are highly magnetised, rapidly rotating neutron stars that are of research interest to astrophysicists. Thus, many large-scale pulsar surveys have been conducted, often leading to huge amounts of unprocessed data. Manually searching for pulsars is labour-intensive and time-consuming. This research describes PulSort, a neural network-based program that aims to solve this problem. PulSort uses ~1600 neurons for image pattern recognition to sort pulsar plots collected from the Pulsar Search Collaboratory database and reduces the number of plots that must be visually inspected by at least one order of magnitude. PulSort was trained on ~11000 manually classified images and mimics humans in identifying the features that are used to classify plots. It differs from earlier machine learning approaches because it uses only two features of candidate plots- the pulses of best profile (PBP) and frequency vs. phase (FP) subplots- whereas earlier solutions have used four or more features. This makes PulSort computationally more efficient. In PulSort, the models for averaged pulse profile and phase/frequency subplots are trained separately and the predictions from each are taken into account to classify the plot appropriately. A simple train-test split was used to partition the dataset and ~1300 images, including a few of millisecond pulsars, were used for testing. The PBP and FP models achieved validation accuracies of 100% and 96.5% respectively during the training stage and correctly classified all the pulsars that were input in the testing stage. The PBP and FP models had false positive rates of 1.62% and 0% respectively. Thus far, PulSort has shown promising results when tested against Pulsar Search Collaboratory data and its performance can be improved as more training data is collected and when the model is tested against data from other pulsar surveys.