Zero-Shot Learning on Semantic Class Prototype Graph.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-Shot Learning (ZSL) relies on semantic embedding spaces for visual recognition.
• Current ZSL methods often use Euclidean or cosine distances, which are suboptimal in high-dimensional, sparse spaces.
• Suboptimal distances lead to issues like hubness and domain shift, hindering ZSL performance.

Purpose of the Study:

• To develop an optimal distance metric for semantic embedding spaces in ZSL.
• To address the limitations of existing distance metrics, specifically hubness and domain shift.
• To enhance the accuracy and robustness of visual recognition in ZSL.

Main Methods:

• Proposing a novel manifold distance calculated on a semantic class prototype graph.
• Incorporating the intrinsic semantic structure (semantic manifold) of class prototypes.
• Introducing a new regularization term into a ranking loss-based embedding model, using unseen class prototypes to mitigate bias.

Main Results:

• The proposed manifold distance effectively captures the semantic structure of class prototypes.
• The regularization term successfully reduces domain shift by preventing bias towards seen prototypes.
• Extensive experiments demonstrated significant performance improvements over state-of-the-art methods across four benchmarks.

Conclusions:

• The novel manifold distance and regularization approach significantly advance Zero-Shot Learning.
• This method offers a more robust and accurate solution for visual recognition tasks with unseen classes.
• The findings suggest a new direction for optimizing distance metrics in high-dimensional semantic spaces.