SEARCHING FOR PULSARS USING IMAGE PATTERN RECOGNITION

Abstract

In this paper, we present a novel artificial intelligence (AI) program that\nidentifies pulsars from recent surveys using image pattern recognition with\ndeep neural nets---the PICS (Pulsar Image-based Classification System) AI. The\nAI mimics human experts and distinguishes pulsars from noise and interferences\nby looking for patterns from candidate. The information from each pulsar\ncandidate is synthesized in four diagnostic plots, which consist of up to\nthousands pixel of image data. The AI takes these data from each candidate as\nits input and uses thousands of such candidates to train its ~9000 neurons.\nDifferent from other pulsar selection programs which use pre-designed patterns,\nthe PICS AI teaches itself the salient features of different pulsars from a set\nof human-labeled candidates through machine learning. The deep neural networks\nin this AI system grant it superior ability in recognizing various types of\npulsars as well as their harmonic signals. The trained AI's performance has\nbeen validated with a large set of candidates different from the training set.\nIn this completely independent test, PICS ranked 264 out of 277 pulsar-related\ncandidates, including all 56 previously known pulsars, to the top 961 (1%) of\n90008 test candidates, missing only 13 harmonics. The first non-pulsar\ncandidate appears at rank 187, following 45 pulsars and 141 harmonics. In other\nwords, 100% of the pulsars were ranked in the top 1% of all candidates, while\n80% were ranked higher than any noise or interference. The performance of this\nsystem can be improved over time as more training data are accumulated. This AI\nsystem has been integrated into the PALFA survey pipeline and has discovered\nsix new pulsars to date.\n