Adversarial strategy for transductive zero-shot learning

Abstract

Zero-shot learning (ZSL) realizes unseen object recognition by transferring knowledge from seen classes to unseen classes under common semantic space assumption, such as attribute space and semantic word vector space. Previous works used seen image feature and semantic representation but ignored unseen test image features to learn a projection from visual space to semantic space, which may lead to the domain shift problem, i.e., due to disjoint seen and unseen classes, the projection learnt from auxiliary dataset is biased when applied directly to the target dataset. In this paper, adversarial strategy is proposed with an instantiation to deal with domain shift problem. It is described as a two-player game in which player 1 is projector while player 2 is classifier. Projector expects to learn a projection from visual space to semantic space with good semantic preservation property while classifier expects to achieve high accuracy. The adversarial meaning comes from the design of parallel structure between loss function on training samples and that on test samples, semantic compatibility of these two loss functions and loss function in classifier, and distribution alignment. A theoretical analysis is provided and experiments are performed on benchmark datasets to ascertain the effectiveness of adversarial strategy.