The key of zero-shot learning is to use the visual-semantic embedding to transfer the knowledge from seen classes to unseen classes. In this paper, we propose to build the visual-semantic embedding by maximizing the mutual information between visual