CGUN-2A: Deep Graph Convolutional Network via Contrastive Learning for Large-Scale Zero-Shot Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hierarchical classification of natural creatures can be represented as knowledge graphs (KGs).
• Graph Convolutional Networks (GCNs) show promise for zero-shot learning by enabling knowledge transfer.
• Deep GCNs suffer from Laplacian over-smoothing, degrading performance in zero-shot image classification.

Purpose of the Study:

• To mitigate the Laplacian over-smoothing problem in deep GCNs for zero-shot image classification.
• To improve node feature discriminability and reduce invalid node aggregation within deep graph networks.
• To propose a novel Contrastive Graph U-Net (CGUN-2A) framework incorporating attention-based pooling.

Main Methods:

• Developed a top-k graph pooling method utilizing a self-attention mechanism for controlled node aggregation.
• Introduced a dual structural symmetric knowledge graph to enhance latent space node representations.
• Integrated these methods into a contrastive learning framework, creating the CGUN-2A model with Attention-based graph pooling (Att-gPool) layers.

Main Results:

• The proposed CGUN-2A model effectively alleviates the Laplacian over-smoothing problem.
• Experiments on a large-scale zero-shot image classification dataset demonstrated significant performance boosts.
• Achieved a 17.5% relative improvement in Hit@1 accuracy compared to baseline models on ImageNet21K.

Conclusions:

• The novel approach of specific node aggregation and homogeneous graph comparison positively impacts deep graph networks.
• The CGUN-2A model offers a robust solution for enhancing zero-shot image classification performance.
• The findings highlight the potential of attention mechanisms and dual structural graphs in addressing GCN limitations.