Transductive zero-shot learning via knowledge graph and graph convolutional networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) aims to recognize objects from unseen categories by transferring knowledge from seen categories.
• Current ZSL methods struggle with domain shift due to limited seen data and semantic relationships, hindering performance.
• A significant domain shift problem exists in zero-shot learning (ZSL) that limits the performance of deep learning models.

Purpose of the Study:

• To propose a novel transductive zero-shot learning (ZSL) method to address the domain shift problem.
• To leverage Knowledge Graphs (KG) and Graph Convolutional Networks (GCN) for improved ZSL classification.
• To enhance the recognition of unseen object categories by effectively transferring knowledge from seen categories.

Main Methods:

• Construct a knowledge graph where each node represents a category encoded by its semantic embedding.
• Employ a shallow Graph Convolutional Network (GCN) to learn classifiers supervised by seen categories.
• Utilize a Double Filter Module with Hungarian algorithm for clustering unseen samples and assigning pseudo-annotations.
• Incorporate a transductive setting where accurately classified unseen samples update model parameters.

Main Results:

• The proposed transductive ZSL method achieved state-of-the-art performance on benchmark datasets.
• Achieved 47.36% accuracy on AWA2, 30.69% on ImageNet50, and 18.87% on ImageNet100.
• Demonstrated a 4-10% improvement over existing zero-shot learning methods.

Conclusions:

• The proposed KG and GCN-based transductive ZSL method effectively mitigates domain shift.
• The integration of pseudo-annotations and transductive learning significantly boosts classification accuracy for unseen categories.
• This approach offers a promising direction for advancing zero-shot learning capabilities in computer vision.