A Comparison of Convolutional Neural Networks for RR Lyrae Light Curve Classification

Abstract

Light curves are time series of the brightness of astronomical objects and are fundamental to analyze variable stars. RR Lyrae are a particular type of variable stars that exhibit periodic behavior in their light curves. The Vista Variable in the Via Lactea (VVV) survey aims to understand how our galaxy was formed and finding large quantities of RR Lyrae is key to accomplish this. In this work we evaluate convolutional neural networks for the automatic classification of RR Lyrae using a subset of the light curves of the VVV survey. To address the differences in length between light curves we compare padding, partial-convolution and subsampling based strategies. The experiments show that the best test-set results are achieved using conventional convolutional layers with a global max pooling operator over zero-padded light curves. Future work includes testing with continuous-time convolutions, exploring synergies with feature-based models and evaluating on more classes of periodic variable star.