Zero-Shot Classification by Large-Margin Distance Learning

Abstract

Zero-Shot Learning (ZSL) has increasingly attained a lot of attentions due to the scalability it provides to recognition models for classifying instances from new classes of which no training data have been seen before. This scalability is achieved by providing semantic information about new classes, which could be obtained remarkably easier, with a lower cost, in comparison with gathering a new training set. In other words, ZSL can be seen as a subset of transfer learning. In this paper, a distance function is learnt in order to discriminate the classes with a customized large-margin loss function. We also propose a nonlinear prototype learning approach by defining a theoretical-based, but simple mapping function to generate the class prototypes from associated semantic information. The instances are compared with the generated prototypes for classification considering the learnt distance function. The evaluations on five widely-used ZSL datasets, illustrate the effectiveness and superiority of the proposed method in comparison with the state-of-the-art ZSL approaches, despite of its simplicity.