Multi-view graph representation with similarity diffusion for general zero-shot learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) enables prediction of unseen classes without training samples, applicable across domains like image recognition and anomaly detection.
• Existing ZSL methods often rely on environment-specific attributes, limiting their generalizability to diverse datasets like ImageNet.
• Current open-knowledge ZSL approaches show limited performance (<10% accuracy) due to insufficient semantics and a notable semantic gap.

Purpose of the Study:

• To propose a novel ZSL method adaptable to general data environments, overcoming limitations of existing attribute-dependent and open-knowledge approaches.
• To enhance semantic representation and bridge the semantic gap between seen and unseen classes for improved zero-shot prediction.
• To achieve state-of-the-art performance in general zero-shot learning tasks.

Main Methods:

• A multi-view graph representation is employed to enrich the semantics of classes.
• An innovative similarity diffusion model is introduced to augment graph representations.
• A feature diffusion method is utilized to bridge the semantic gap and enhance zero-shot prediction capabilities.

Main Results:

• The proposed method achieves new state-of-the-art results on general zero-shot learning tasks across benchmark datasets.
• Experiments demonstrate the effectiveness of the multi-view graph and diffusion models in enhancing semantic understanding.
• Ablation studies confirm the significant contribution of individual modules to the overall performance.

Conclusions:

• The multi-view graph with a similarity diffusion model offers a robust solution for general zero-shot learning.
• The proposed feature diffusion method effectively addresses the semantic gap, improving prediction accuracy for unseen classes.
• This approach advances the applicability of ZSL in real-world, open data environments.