Abstract: Zero-shot learning for visual recognition, which approaches identifying unseen categories through a shared visual-semantic function learned on the seen categories and is expected to well ...