Pulsar-candidate Selection Using a Generative Adversarial Network and ResNeXt

Abstract

Pulsar research has been a hot topic in the area of astronomy since they were first discovered. Pulsar discovery is fundamental for pulsar research. While pulsars are now visible across the electromagnetic spectrum, pulsar searches with modern radio telescopes are most promising. As the performance of astronomical instruments improves, the number of pulsar candidates detected by modern radio telescopes grows at an exponential rate. The application of artificial intelligence to the field of pulsar-candidate identification can automatically and efficiently address the identification problem with enormous amounts of data. However, there are still significant challenges in enhancing the accuracy of deep-learning-based pulsar-candidate identification. These problems result primarily from the fact that real pulsar data is scarce: the number of candidates that can be successfully identified as real pulsars (positive samples) is much smaller than those candidates that turn out to not be pulsars but instead radio-frequency interference or noise (negative samples). This makes it difficult to train a machine-learning model that can accurately select those candidates that are real pulsars. Therefore a novel pulsar-candidate identification framework is proposed that combines a deep convolutional generative adversarial neural network (DCGAN) and a deep aggregation residual network (ResNeXt). To overcome sample imbalance, the DCGAN is utilized to generate images that approximate real pulsars, while observed and generated candidates are employed together to train the pulsar-candidate identification model ResNeXt. Experiments on the HTRU Medlat data set back up the framework’s performance. The precision, recall, and F1-score of the framework are 100%.