Leveraging Balanced Semantic Embedding for Generative Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generative zero-shot learning (ZSL) models synthesize unseen class features but often suffer from seen class feature dominance.
• This bias leads to poor performance in ZSL and unbalanced results in generalized ZSL (GZSL).

Purpose of the Study:

• To develop a novel balanced semantic embedding generative network (BSeGN) for unbiased generalized zero-shot learning (GZSL).
• To address the challenge of generated unseen features being misclassified as seen classes.

Main Methods:

• Introduced a feature-to-semantic embedding module (FEM) with bidirectional contrastive and balance losses for online discrimination of real and fake features.
• Proposed a multilevel feature integration module (mFIM) within a cycle-consistency framework to mitigate domain bias.
• Jointly explored embedding and generative learning for ZSL.

Main Results:

• The proposed BSeGN effectively distinguishes between real seen and fake unseen features.
• Bidirectional contrastive and balance losses ensure balanced interdomain feature prediction.
• Multilevel feature integration mitigates domain bias, enhancing generative learning.
• Extensive evaluations on four benchmarks show BSeGN outperforms state-of-the-art methods.

Conclusions:

• BSeGN offers a novel approach to generative learning in ZSL by integrating balanced semantic embedding.
• The method significantly improves performance in generalized zero-shot learning scenarios.
• This work represents a pioneering effort in jointly applying embedding and generative learning to ZSL.