Hierarchical Feature Generating Network for Zero-Shot Learning by Knowledge Graph

Zero-Shot Learning (ZSL) has received much attention and has achieved great success. Most of existing ZSL methods transfer knowledge learned from seen classes to unseen classes by utilizing shared side information, such as, annotated attribute vectors, word embeddings, etc. Recently, the most popular method in ZSL is utilizing generative models to do semantic augmentation for unseen classes. However, these models generate features in one step which will lead to the domain shift problem and don’t leverage rich shared information between seen and unseen classes in the knowledge graph. Thus we construct a hierarchical generative model that synthesizes features for unseen classes layer by layer instead of one-step like previous ZSL work. Experimental results on various datasets show that our method can significantly improve the performance compared with the state-of-the-art ZSL models.