Zero-Shot Learning with Deep Canonical Correlation Analysis

Zero-shot learning (ZSL) improves the scalability of conventional image classification systems by allowing some testing categories having no training data. One key component is to learn a shared embedding space where both side information of object categories and visual representation of object images can be projected to for nearest neighbor search. Although great progress has been made, existing approaches mainly focus on shallow models, which cannot learn a strong generalized embedding space. To this end, this paper proposes a novel deep embedding model for ZSL, which formulates the embedding space with Deep Canonical Correlation Analysis (DCCA). Specifically, the side information and the visual representation are transformed via two independent deep neural networks, and then they are highly linearly correlated in the final output layer. Extensive experiments on two publicly popular datasets demonstrated the effectiveness and superiority the proposed approach.