Haar Wavelet Feature Compression for Quantized Graph Convolutional Networks.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Graph convolutional networks (GCNs) are essential for analyzing unstructured data but face computational challenges with large graphs.
• Standard convolutional neural networks (CNNs) also encounter high computational costs, often addressed by quantization.
• Aggressive quantization in GCNs can significantly impair network performance.

Purpose of the Study:

• To reduce the computational cost of GCNs for large-scale graph data.
• To explore alternative compression techniques beyond aggressive feature map quantization.
• To maintain or improve GCN performance while enhancing computational efficiency.

Main Methods:

• Proposed a novel approach combining Haar wavelet compression with light quantization for GCNs.
• Applied Haar wavelet transforms to compress feature maps, reducing computational load.
• Evaluated the method on diverse graph-based tasks including node and point cloud classification, and segmentation.

Main Results:

• The proposed Haar wavelet compression and light quantization method significantly outperformed aggressive feature quantization.
• This hybrid approach demonstrated superior performance across various GCN applications.
• Achieved substantial reductions in computational cost without compromising accuracy.

Conclusions:

• Haar wavelet compression combined with light quantization offers an effective strategy for optimizing GCNs.
• This method provides a viable solution for deploying GCNs in resource-constrained environments.
• The approach successfully addresses the performance degradation issues associated with aggressive quantization in GCNs.