Transductive Zero-shot Learning with New Semantic Embeddings

Abstract

Zero-shot learning (ZSL) is an important machine learning paradigm that trains models to handle samples from classes that are unseen during training. Most ZSL methods transfer knowledge about the seen to unseen classes using some semantic representations, such as annotated attribute vectors or word embeddings. However, such information is either difficult to acquire or lack of linkage to visual information. We thus propose a novel transductive ZSL paradigm that makes use of feature maps to distill less expensive but more informative semantic embeddings of seen and unseen classes. Experimental results on different datasets show that this method can significantly improve the performance compared to the latest ZSL models.