Enabling generalized zero-shot learning towards unseen domains by intrinsic learning from redundant LLM semantics

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generalized Zero-Shot Learning (GZSL) struggles with domain shift, where unseen classes are misclassified as seen classes.
• Existing GZSL methods are limited to seen domains and do not address cross-domain generalization.
• Information asymmetry arises from redundant class semantics, particularly when using large language models (LLMs).

Purpose of the Study:

• To address the limitations of existing GZSL by introducing Cross-Domain Generalized Zero-Shot Learning (CDGZSL).
• To develop a method that enables GZSL in unseen domains by constructing a common feature space and acquiring shared intrinsic semantics.
• To mitigate information asymmetry caused by LLM-annotated semantics in CDGZSL.

Main Methods:

• Propose Meta Domain Alignment Semantic Refinement (MDASR) for CDGZSL.
• MDASR employs Inter-class similarity alignment to remove non-intrinsic semantics based on inter-class feature relationships.
• MDASR utilizes unseen-class meta generation to preserve intrinsic semantics and maintain connectivity between seen and unseen classes via simulated feature generation.

Main Results:

• MDASR effectively aligns redundant semantic spaces with common feature spaces, reducing information asymmetry in CDGZSL.
• The method demonstrates effectiveness on the Office-Home and Mini-DomainNet datasets.
• Performance was also validated on a self-constructed multi-domain rare animal dataset.

Conclusions:

• MDASR offers a robust solution for Cross-Domain Generalized Zero-Shot Learning.
• The approach successfully tackles domain shift and semantic information asymmetry challenges.
• The shared LLM-based semantics serve as a valuable benchmark for future research in CDGZSL.