DeepGCNs: Making GCNs Go as Deep as CNNs

Area of Science:

• Artificial Intelligence
• Computer Vision
• Graph Neural Networks

Background:

• Convolutional Neural Networks (CNNs) excel in computer vision but struggle with non-Euclidean data.
• Graph Convolutional Networks (GCNs) handle non-Euclidean data but are typically shallow due to training limitations.
• Vanishing gradients restrict the depth of current GCN architectures.

Purpose of the Study:

• To enable the training of very deep Graph Convolutional Networks (GCNs).
• To adapt successful deep learning techniques from CNNs for GCNs.
• To demonstrate the effectiveness of deep GCNs on various non-Euclidean data tasks.

Main Methods:

• Transferred residual/dense connections and dilated convolutions from CNNs to GCNs.
• Developed and trained GCN architectures with up to 112 layers.
• Evaluated deep GCN performance on part/semantic segmentation and node classification tasks.

Main Results:

• Successfully trained very deep GCNs (up to 112 layers).
• Achieved promising performance in part and semantic segmentation on point clouds.
• Demonstrated strong results in node classification for protein function prediction on biological graphs.

Conclusions:

• Deep GCNs, enabled by CNN concepts, significantly enhance performance on non-Euclidean data tasks.
• The proposed methods overcome previous depth limitations in GCNs.
• This research opens new possibilities for GCN applications in diverse fields.