Robust zero-shot learning with source attributes noise

Because of the challenge of collecting labelled training data, zero-shot learning (ZSL) which transfers semantic knowledge represented by category attributes from seen classes to recognize unseen classes has received a lot of attention recently. Existing methods assume that the source attributes are completely correct in zero-shot learning. However, the source attributes in practice may contain noise due to assignment by mistake. In this paper, we develop a novel robust zero-shot learning method (RZSL) for ZSL. It seeks the lowest rank representation among all the candidates and extracts the relational knowledge from data structure in attribute space. RZSL not only settles the semantic shift problem efficiently but also deals with the noisy source attributes problem well. We conduct experiments on two standard image datasets and our results demonstrate the efficacy of the proposed method.