Low-shot Learning via Covariance-Preserving Adversarial Augmentation Networks