A Joint Label Space for Generalized Zero-Shot Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-Shot Learning (ZSL) faces challenges due to the discrete nature of one-hot labels, hindering the preservation of inter-class relationships.
• Conventional ZSL methods often fail to jointly optimize visual, semantic, and label spaces, leading to performance degradation.

Purpose of the Study:

• To propose a novel approach, Attributing Label Space (ALS), that directly reconciles visual and semantic spaces within the label space for improved ZSL.
• To address the performance gap caused by the disconnect between attribute and label spaces in existing ZSL algorithms.

Main Methods:

• ALS maps images and attributes into a common space using one-hot labels of seen classes as prototypes during training.
• The method optimizes mappings independently, resulting in low computational complexity and reduced influence from attribute correlations.
• During testing, the discrete label space constraint is removed, enabling composition of unseen class labels for Generalized ZSL (GZSL).

Main Results:

• ALS demonstrates improved performance across five benchmark datasets compared to state-of-the-art methods.
• The approach proves highly discriminative for Generalized ZSL (GZSL) tasks.
• Experiments show that mapping features directly into labels, rather than attributes, enhances visual embedding training.

Conclusions:

• Attributing Label Space (ALS) offers a more effective pathway for Zero-Shot Learning by directly leveraging the label space.
• The proposed method provides a more reasonable and challenging framework for real-world GZSL applications.
• ALS significantly advances the state-of-the-art in ZSL by unifying visual and semantic information within a discriminative label space.