A General Transductive Regularizer for Zero-Shot Learning

Huaqi Mao (Nanjing University of Science and Technology), Haofeng Zhang (Nanjing University of Science and Technology), Shidong Wang (University of East Anglia), Yang Long (Newcastle University), Longzhi Yang (Northumbria University)

Abstract
Zero Shot Learning (ZSL) has attracted much attention due to its ability to recognize objects of unseen classes, which is realized by transferring knowledge from seen classes through semantic embeddings. Since the seen classes and unseen classes usually have different distributions, conventional inductive ZSL often suffers from the domain shift problem. Transductive ZSL is a type of method for solving such a problem. However, the regularizers of conventional transductive methods are different from each other, and cannot be applied to other methods. In this paper, we propose a General Transductive Regularizer (GTR), which assigns each unlabeled sample to a fixed attribute by defining a Kullback-Leibler Divergence (KLD) objective. To this end, GTR can be easily applied to many compatible linear and deep inductive ZSL models. Extensive experiments on both linear and deep methods are conducted on four popular datasets, and the results show that GTR can significantly improve the performance comparing to its original inductive method, and also outperform some state-of-the-art methods, especially the extension on deep model.

DOI
10.5244/C.33.55
https://dx.doi.org/10.5244/C.33.55

Files
Paper (PDF)

BibTeX