Concept-Aware Graph Convolutional Network for Compositional Zero-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Compositional zero-shot learning (CZSL) aims to recognize novel concepts by combining known attributes and objects.
• Existing CZSL methods struggle with distribution shifts and domain bias between seen and unseen components.
• Intrinsic variations within primitive concepts (attributes, objects) are often overlooked.

Purpose of the Study:

• To develop a novel method for compositional zero-shot learning that mitigates domain bias and effectively utilizes primitive variations.
• To enhance the model's ability to identify unobservable compositional concepts using prior knowledge.
• To improve performance in both closed-world and open-world CZSL scenarios.

Main Methods:

• Proposed a concept-aware graph convolutional network (GCN) employing cross-attentions for feature extraction from concept-sharing inputs.
• Utilized cosine similarity between visual features and synthetic embeddings to generate feasibility scores for unseen compositions.
• Integrated Earth mover's distance (EMD) to refine concept learning within disentanglers.

Main Results:

• The proposed GCN model demonstrated superior performance on three benchmark datasets: UT-Zappos 50K, C-GQA, and MIT-States.
• Achieved state-of-the-art results in both closed-world and open-world compositional zero-shot learning (OW-CZSL).
• Effectively addressed domain bias and leveraged primitive variations for improved concept recognition.

Conclusions:

• The concept-aware GCN offers a robust solution for CZSL, outperforming existing approaches.
• The method successfully handles the challenges posed by distribution differences and primitive variations.
• The proposed approach advances the field of zero-shot learning, particularly for complex compositional concepts.