A Generative Model for Zero-Shot Learning via Wasserstein Auto-encoder

Zero-shot learning aims to use the labeled instances to train the model, and then classifies the instances that belong to a class without labeled instances. However, the training instances and test instances are disjoint. Thus, the description of the classes (e.g. text description or class attribute information) is to establish a connection between the training set and the test set to make the model effective. Since real world image annotation requires a lot of manpower and material resources, this setting is very important in the real world. Zero-shot learning can effectively solve the problem of image annotation. Most of the previous methods explored the mapping between visual space and semantic space. Some recent generative methods attempt to generate image features of unseen classes based on auxiliary information, and have achieved good performances. In this paper, we propose to use Wasserstein Auto-encoder (WAE) as a generative model to establish data distribution. Then the generated samples from the generative model are used to classify unseen classes. We test our model on four benchmark datasets including CUB, SUN, AWA2 and aPY, the results of which demonstrate the effectiveness of our model.