Deep transductive network for generalized zero shot learning

Abstract

Abstract Zero Shot Learning (ZSL) aims to learn projective functions on labeled seen data and transfer the learned functions to unseen classes by discovering their relationship with semantic embeddings. However, the mapping process often suffers from the domain shift problem caused by only using the labeled seen data. In this paper, we propose a novel explainable Deep Transductive Network (DTN) for the task of Generalized ZSL (GZSL) by training on both labeled seen data and unlabeled unseen data, with subsequent testing on both seen classes and unseen classes. The proposed network exploits a KL Divergence constraint to iteratively refine the probability of classifying unlabeled instances by learning from their high confidence assignments with the assistance of an auxiliary target distribution. Besides, to avoid the meaningless ascription assumption of unseen data on GZSL, we also propose an experimental paradigm by splitting the unseen data into two equivalent parts for training and testing respectively. Extensive experiments and detailed analysis demonstrate that our DTN can efficiently handle the problems and achieve the state-of-the-art performance on four popular datasets.