Zero-Shot Learning in the Presence of Hierarchically Coarsened Labels

Zero-shot image classification leverages side information including label attributes and semantic class hierarchies to transfer knowledge about fine-grained training classes to fine-grained zero-shot classes. In this paper, we consider the problem of zero-shot learning of fine-grained classes given a mixture of images with fine-grained and coarsened labels. We show how probabilistic hierarchical classification models can be used to simultaneously accommodate fine and coarse-grained labels in the zero-shot learning setting. We show that this approach is robust even to significant levels of coarsening.