Self-assembled Generative Framework for Generalized Zero-shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generative models are increasingly used for generalized zero-shot learning (GZSL).
• Current generative GZSL models synthesize visual features using semantic information but struggle with class-irrelevant noise in real visual features, hindering discriminability.
• This noise leads to ambiguous synthesized features and reduced model performance.

Purpose of the Study:

• To address the issue of class-irrelevant information in visual features for generative GZSL.
• To propose a novel framework that enhances the discriminability of synthesized visual features.
• To improve the overall performance of GZSL models.

Main Methods:

• Empirical analysis to identify class-irrelevant elements within real visual features.
• Development of a self-assembled generative GZSL framework (SaG) that re-assembles real and synthesized features.
• Introduction of an element-affinity regularizer to guide feature synthesis within the SaG framework.

Main Results:

• The SaG framework effectively identifies and updates class-irrelevant elements in visual features.
• Embedding existing generative GZSL models into SaG significantly boosts their performance.
• One SaG-derived method outperformed 20 state-of-the-art GZSL methods in most experimental cases.

Conclusions:

• The proposed SaG framework offers a robust solution for improving visual feature synthesis in GZSL.
• SaG enhances feature discriminability by mitigating the impact of class-irrelevant information.
• The framework's modular design allows seamless integration with various generative GZSL models, demonstrating broad applicability and effectiveness.