Triple Verification Network for Generalised Zero-shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Conventional Zero-shot Learning (ZSL) methods exhibit performance degradation in Generalised Zero-shot Learning (GZSL), where models must recognize both seen and unseen classes.
• Class-level Over-fitting (CO) is identified as a significant factor contributing to this performance degradation in GZSL.

Purpose of the Study:

• To investigate the impact of Class-level Over-fitting (CO) on Generalised Zero-shot Learning (GZSL) performance.
• To propose a novel approach for ZSL that mitigates CO and enhances performance in both ZSL and GZSL settings.

Main Methods:

• The study frames ZSL as a Triple Verification problem.
• A unified optimization of regression and compatibility functions is proposed, integrating two primary ZSL approaches.
• Complementary losses are utilized for mutual regularization to address the CO problem.
• A deep extension paradigm is applied to linear models.

Main Results:

• The proposed method effectively mitigates Class-level Over-fitting (CO) in GZSL.
• The unified optimization approach demonstrates significant performance improvements over state-of-the-art methods.
• The deep extension paradigm applied to linear models yields superior results.

Conclusions:

• The unified optimization of regression and compatibility functions, coupled with complementary losses, effectively tackles Class-level Over-fitting (CO).
• The proposed deep extension paradigm significantly advances the state-of-the-art in both ZSL and GZSL scenarios.
• This research offers a robust solution for improving zero-shot learning capabilities, particularly in generalised settings.