Towards zero-shot learning generalization via a cosine distance loss

Abstract With the knowledge learned from some labelled training images, zero-shot learning (ZSL) aims to recognize new visual concepts by leveraging some intermediate information for both seen and unseen classes. Despite the existence of various methods, few work tends to comprehensively evaluate the generalization ability towards some practical data, which is important for their popularization and application. In this paper, we illustrate that it is inadequate and unconvincing to evaluate the ZSL methods using the current fixed data split. 19 existing methods are investigated on two practical datasets using 5-fold cross-validation to evaluate their generalization performances. More specifically, to alleviate the hubness problem in the high-dimensional visual space, we propose a cosine distance-based objective function to learn the transformation from semantic to visual features. Extensive experiments on up to 9 practical subsets demonstrate that the proposed method significantly outperforms the other baseline approaches.