Semantic Consistent Embedding for Domain Adaptive Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptation struggles with differing labels between source and target domains.
• Open-set domain adaptation can detect but not classify new categories in the target domain.

Purpose of the Study:

• To introduce Domain Adaptive Zero-Shot Learning (DAZSL) for recognizing all target domain categories using only source domain supervision.
• To address the challenge of simultaneous knowledge transfer across categories and domain styles.

Main Methods:

• Proposed a novel end-to-end learning mechanism: Three-way Semantic Consistent Embedding (TSCE).
• TSCE embeds source, target domains, and semantic space into a shared space using domain-irrelevant categorical prototypes.
• Employed mutual information maximization for target domain feature alignment and a ranking-based mechanism to prevent catastrophic forgetting.

Main Results:

• TSCE effectively aligns domain differences and facilitates knowledge transfer to unseen classes.
• The ranking-based mechanism maintains semantic and shared space consistency without target domain supervision.
• Experimental validation on I2AwA and I2WebV datasets demonstrated significant effectiveness.

Conclusions:

• DAZSL, powered by TSCE, offers a robust solution for domain adaptation with unseen classes.
• The proposed method successfully transfers knowledge across domains and categories, outperforming existing approaches.
• The approach is validated by strong experimental results on benchmark datasets.