Ranking Synthetic Features for Generative Zero-Shot Learning

Abstract

Zero-Shot Learning (ZSL) is an emerging learning paradigm that addresses the problem of recognizing unseen classes during training. Several studies have shown ZSL can be improved using synthetic samples of unseen classes, usually generated with a GAN and conditioned on some high- level descriptions of the desired class. This paper proposes a new generative adversarial network architecture to improve synthetic feature generation by applying a ranking step at training time. We combined two classifiers' results at the zeroshot classification step to ensure improved classification accuracy. Then we evaluated the proposed architecture using the widely used dataset AWA. Our results show an improvement of classification accuracy of 2.3% in ZSL setting and 0.15% in GZSL setting compared to the state-of-the-art.