Cross-class generative network for zero-shot learning

Abstract

Zero-shot learning (ZSL) aims to recognize unseen classes when no training samples are provided for these classes. A traditional approach to solving ZSL is to generate samples of unseen classes, transforming it into a supervised task. However, the quality of these pseudo-samples is crucial for model performance. In this paper, we propose a novel network, called cross-class generative network, which includes two novel end-to-end models, to generate high-quality samples for unseen classes. Unlike previous work, our proposed models directly generate samples of unseen classes via samples of seen classes. As a result, generated samples are distributed more similarly to real samples. In addition, we propose an intra-class entropy to measure the discrepancy degree for selecting suitable source–target pairs. To the best of our knowledge, this intra-class entropy is proposed in ZSL for the first time. Our models include two versions, non-adversarial and adversarial ones, to support and explore different scenarios. We conduct extensive experiments on five benchmark datasets. A comprehensive comparison with state-of-the-art methods shows the superiority of our proposed models.