Deep Semisupervised Zero-Shot Learning with Maximum Mean Discrepancy.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) is crucial for visual recognition with limited labeled data.
• Existing ZSL methods struggle with inadequate semantic representation and differing data distributions between seen and unseen categories.
• Label embeddings often fail to fully capture the nuances of category semantics.

Purpose of the Study:

• To propose a novel deep semisupervised method for zero-shot learning.
• To address the limitations of traditional label embeddings in semantic representation.
• To overcome the distribution differences between seen and unseen instances in ZSL.

Main Methods:

• A deep semisupervised approach is developed, considering inter-modality heterogeneity and intra-modality correlation.
• Category semantics are better captured by replacing original labels with textual descriptions.
• The maximum mean discrepancy is minimized to align seen and unseen instance distributions.

Main Results:

• The proposed method demonstrates significant improvements over existing approaches on benchmark datasets.
• Effective handling of semantic heterogeneity and distribution gaps is achieved.
• Enhanced zero-shot classification performance is validated on CU200-Birds and Oxford Flowers-102 datasets.

Conclusions:

• The novel deep semisupervised method offers a more robust solution for zero-shot learning.
• Utilizing textual descriptions and minimizing distribution discrepancies are key to improved performance.
• The approach advances the state-of-the-art in zero-shot visual recognition.