SPINN: a straightforward machine learning solution to the pulsar candidate selection problem

Abstract

We describe SPINN (Straightforward Pulsar Identification using Neural Networks), a high-performance machine learning solution developed to process increasingly large data outputs from pulsar surveys. SPINN has been cross-validated on candidates from the southern High Time Resolution Universe (HTRU) survey and shown to identify every known pulsar found in the survey data while maintaining a false positive rate of 0.64%. Furthermore, it ranks 99% of pulsars among the top 0.11% of candidates, and 95% among the top 0.01%. In conjunction with the peasoup pipeline (Barr et al. in preparation), it has already discovered four new pulsars in a re-processing of the intermediate Galactic latitude area of HTRU, three of which have spin periods shorter than 5 milliseconds. SPINN's ability to reduce the amount of candidates to visually inspect by up to four orders of magnitude makes it a very promising tool for future large-scale pulsar surveys. In an effort to provide a common testing ground for pulsar candidate selection tools and stimulate interest in their development, we also make publicly available the set of candidates on which SPINN was cross-validated.

Detection of dispersed radio pulses: a machine learning approach to candidate identification and classification

“…Furthermore, while recall values were very high, the F-measures obtained for MPNs were consistently lower than all learners except SMO, with a large variance. Note that ANNs are one of the most common machine learning techniques applied to the problem of radio pulsar detection in periodicity searches, and were used in each paper discussed in Section 3.2 that performed machine learning Bates et al 2012;Zhu et al 2014;Morello et al 2014).…”

“…As our focus is on machine learning, we only provide reviews of papers that use machine learning techniques Bates et al 2012;Zhu et al 2014;Morello et al 2014). The fact that these papers were all published in the last five years indicates that intelligent algorithms are becoming the new standard for pulsar classification.…”

“…Another recent work by Morello et al (2014) presented the classification results from a pulsar ranking system called Straightforward Pulsar Identification using Neural Networks (SPINN). SPINN uses a customized ANN trained on 1,196 observations of pulsars from the HTRU all-sky pulsar survey combined with 90,000 randomly selected negative observations.…”

“…Artificial neural networks (ANNs) have been used in several related papers working with periodicity searches Bates et al 2012;Zhu et al 2014;Morello et al 2014), as mentioned in Section 3.2. The ANN we used is the Java implementation of a Multilayer Perceptron Network (MPN), which classifies instances using the supervised learning method of backpropagation and a sigmoid activation function in all neural nodes (Bishop 1995).…”

“…Automated approaches are the only feasible way to deal with big data, and offer many potential advantages to streamline the discovery process, e.g., by triggering the rapid follow up of candidates at multiple wavelengths to constrain their origins. Machine learning 1 algorithms have been applied to automatically detect pulsars in periodicity searches Bates et al 2012;Zhu et al 2014;Morello et al 2014). However, machine learning approaches with single-pulse candidates are lacking.…”