Fully Quantized Graph Convolutional Networks for Embedded Applications

Graph Neural Networks (GNNs) demonstrate re- markable proficiency across diverse tasks involving graph data, such as communication networks, physical systems, and chem- ical bonds. However, GNNs require significant computational resources compared to other machine learning approaches. Significant memory is necessary to store graph information. The problem is that GNNs are not well suited for resource constrained devices that have limited memory, computation, and energy. Most current solutions quantize Convolutional Neural Networks (CNNs), such the Dorefa-Net algorithm. There has not been a comprehensive quantization technique proposed for GNNs that fully quantizes all network parameters. In this paper we propose and evaluate using the Dorefa-Net algorithm integrated with the Graph Convolutional Network (GCN). We compare the accuracy performance of several quantization techniques and three commonly used datasets. Our findings reveal that the Dorefa-Net method quantizes network parameters down to 4-bit integers with an acceptable accuracy loss (up to 2%) compared to the base model. However, we find that Dorefa-Net does not perform well for aggressive quantization levels (e.g. 1 and 2 bit), which are necessary for specialized hardware, such as FPGAs. To address this, we propose a modified version denoted Dorefa- Graph. We show that Dorefa-Graph performs better than the other quantization techniques, particularly when aggressively quantized, making it better suited for bespoke hardware.