End-to-End Supervised Zero-Shot Learning with Meta-Learning Strategy

Abstract

Zero-shot learning (ZSL) is a challenging but practical task in the computer vision field. ZSL tries to recognize new unknown categories by provided with training data from other known categories. Recently, the ZSL problem can be solved in a supervised learning way by using deep generative models to synthesize data as the training data for unknown categories. In this work, we design an end-to-end supervised ZSL method in which the data generation network and the object classification network are trained jointly. To enhance the generalization performance of the proposed supervised ZSL method, meta-learning strategy is introduced to mitigate the domain shift problem between the synthesized data and the real data of unknown categories. Experimental results on ZSL standard datasets demonstrate the significant superiority of the end-to-end strategy and the meta-learning strategy for the proposed model in ZSL tasks.