Learning Competitive and Discriminative Reconstructions for Anomaly Detection