In this paper, we propose to tackle the challenging few-shot learning (FSL)problem by learning global class representations using both base and novelclass training samples. In each training episode, an episodic class meancomputed from a support set is registered with the global representation via aregistration module. This produces a registered global class representation forcomputing the classification loss using a query set. Though following a similarepisodic training pipeline as existing meta learning based approaches, ourmethod differs significantly in that novel class training samples are involvedin the training from the beginning. To compensate for the lack of novel classtraining samples, an effective sample synthesis strategy is developed to avoidoverfitting. Importantly, by joint base-novel class training, our approach canbe easily extended to a more practical yet challenging FSL setting, i.e.,generalized FSL, where the label space of test data is extended to both baseand novel classes. Extensive experiments show that our approach is effectivefor both of the two FSL settings.