Zero-Shot Learning via Attribute Regression and Class Prototype Rectification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) enables classification of unseen classes using knowledge from seen classes.
• Existing ZSL methods often rely on attribute-based knowledge transfer, which can suffer from projection domain shift and hubness problems.
• These issues hinder the performance and accuracy of ZSL models.

Purpose of the Study:

• To propose a novel ZSL approach that addresses the limitations of existing regression-based methods.
• To mitigate the projection domain shift and hubness problems in ZSL.
• To improve the classification accuracy for unseen classes in ZSL tasks.

Main Methods:

• Formulating ZSL as an attribute regression problem with novel enhancements.
• Introducing a class prototype rectification method to bridge seen and unseen classes.
• Employing an alternating learning scheme for joint attribute regression and prototype rectification.
• Proposing a new objective function balancing attribute regression accuracy and class prototype discrimination.

Main Results:

• Successfully mitigated projection domain shift and hubness problems.
• Demonstrated significant effectiveness on public datasets CUB200-2011, SUN Attribute, and aPaY.
• Achieved improved classification performance for unseen classes.

Conclusions:

• The proposed approach offers a robust solution for ZSL by effectively addressing key challenges.
• The novel methods of class prototype rectification and alternating learning enhance ZSL performance.
• Experimental validation confirms the superiority of the approach on benchmark datasets.