Abstract
Collecting training images for all visual categories isnot only expensive but also impractical. Zero-shot learning (ZSL), especially using attributes, offers a pragmatic solu-tion to this problem. However, at test time most attribute-based methods require a full description of attribute asso-ciations for each unseen class. Providing these associa-tions is time consuming and often requires domain specific knowledge. In this work, we aim to carry out attribute-based zero-shot classification in an unsupervised man-ner. We propose an approach to learn relations that cou-ples class embeddings with their corresponding attributes. Given only the name of an unseen class, the learned rela-tionship model is used to automatically predict the class-attribute associations. Furthermore, our model facilitatestransferring attributes across data sets without additional effort. Integrating knowledge from multiple sources resultsin a significant additional improvement in performance. We evaluate on two public data sets: Animals with Attributes and aPascal/aYahoo. Our approach outperforms state-ofthe-art methods in both predicting class-attribute associations and unsupervised ZSL by a large margin.