Common variable star classifiers are built only with the goal of producing the correct class labels, leaving much of the multi-task capability of deep neural networks unexplored. We present a periodic light curve classifier that combines a recurrent neural network autoencoder for unsupervised feature extraction and a dual-purpose estimation network for supervised classification and novelty detection. The estimation network optimizes a Gaussian mixture model in the reduced-dimension feature space, where each Gaussian component corresponds to a variable class. An estimation network with a basic structure of a single hidden layer attains a cross-validation classification accuracy of ~99%, on par with the conventional workhorses, random forest classifiers. With the addition of photometric features, the network is capable of detecting previously unseen types of variability with precision 0.90, recall 0.96, and an F1 score of 0.93. The simultaneous training of the autoencoder and estimation network is found to be mutually beneficial, resulting in faster autoencoder convergence, and superior classification and novelty detection performance. The estimation network also delivers adequate results even when optimized with pre-trained autoencoder features, suggesting that it can readily extend existing classifiers to provide added novelty detection capabilities.
Download