Contrastive Prototype-Guided Generation for Generalized Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generalized zero-shot learning (GZSL) faces challenges in recognizing unseen classes due to limited training data.
• Current methods often generate synthetic unseen data but suffer from generator bias towards seen classes, impacting generation quality and diversity.

Purpose of the Study:

• To propose a novel method, Prototype-Guided Generation (PGZSL), to improve generalized zero-shot learning by enhancing unseen data generation.
• To address the bias in synthetic data generation and improve both the precision and diversity of generated unseen samples.

Main Methods:

• PGZSL utilizes contrastive prototypical anchors to guide and rectify unseen data generation, ensuring semantic consistency and feature discriminability.
• The Certainty-Driven Mixup technique is introduced to enrich the diversity of generated unseen samples and suppress uncertain boundary samples.

Main Results:

• PGZSL significantly outperforms state-of-the-art (SOTA) methods on five benchmark datasets for both ZSL and GZSL tasks.
• The proposed methods demonstrate improved precision and diversity in generated unseen data compared to existing approaches.

Conclusions:

• Prototype-Guided Generation (PGZSL) offers a more effective approach to generalized zero-shot learning by leveraging unseen class knowledge for data generation.
• The study highlights the importance of guided generation and diversity enhancement for advancing GZSL capabilities.