Attribute affinity coordination debiasing for generalized zero-shot learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generalized Zero-Shot Learning (GZSL) uses visual-semantic mappings for knowledge transfer from seen to unseen classes.
• Domain transfer in GZSL often introduces bias, hindering generalization to unseen categories.
• Current GZSL methods face challenges in balancing debiasing with performance, limiting transfer effectiveness.

Purpose of the Study:

• To propose a novel framework, Attribute Affinity Coordinated Debiasing (AACD), to address bias in GZSL.
• To improve the generalization ability of GZSL models on unseen classes by correcting incorrect knowledge transfer.

Main Methods:

• The AACD framework models correlations among category-level attributes to capture inter-class affinity structures.
• It employs attribute-informed domain adaptation to enhance visual-semantic interactions.
• Key modules include the Affinity Discriminant Module (ADM) for embedding space guidance and the Affinity Constraint Module (ACM) for intra-class consistency and inter-class separability.

Main Results:

• The AACD framework effectively identifies and corrects biased knowledge transfer.
• Joint integration of modules within the encoder reduces reliance on seen categories and promotes robust domain transfer.
• Experiments on CUB, SUN, and AwA2 benchmarks show substantial improvements using the AACD framework.

Conclusions:

• The proposed AACD framework offers a significant advancement in mitigating bias for Generalized Zero-Shot Learning.
• AACD enhances the model's ability to generalize to unseen categories by leveraging attribute affinity structures.
• The method demonstrates strong performance across multiple standard GZSL benchmarks.