Graph classification is an important challenge across various scientific fields, such as drug discovery and materials science, but current methods often struggle with both accuracy and computational cost. Hamed Poursiami, Shay Snyder, and colleagues at George Mason University and Oak Ridge National Laboratory, including Guojing Cong, Thomas Potok, and Maryam Parsa, introduce VS-Graph, a new approach that significantly improves both speed and performance. This innovative framework leverages hyperdimensional computing, a brain-inspired technique, to create systems that rival the accuracy of complex graph neural networks while significantly reducing computational demands. VS-Graph employs a new mechanism to identify key nodes and efficiently aggregate information from across the graph, delivering up to 450x faster training and maintaining high accuracy even with limited computational resources, opening new possibilities for graph analysis on resource-constrained devices.
Hyperdimensional computing for robust graph learning
In this study, we investigate the application of hyperdimensional computing (HDC), also known as vector symbolic architecture (VSA), to graph machine learning. HDC is a biologically-inspired computing technique that uses high-dimensional random vectors to represent data and leverages properties such as distance preservation and robustness to noise for efficient computation. Key features include distributed representation, holographic principles, similarity-based calculations, and fast operations within these high-dimensional spaces. This work introduces GraphHD, an implementation of HDC designed to classify graphs by encoding both the graph structure and node features as high-dimensional vectors. Experiments on standard graph datasets demonstrate that GraphHD achieves competitive or state-of-the-art classification accuracy, with the potential for faster computation and reduced memory usage, especially for large graphs. This study highlights the potential of implementing HDC algorithms on neuromorphic hardware, potentially leading to significant energy savings and enabling real-time graph processing for applications such as anomaly detection and robotics.
Spike diffusion and associative message passing
Scientists have developed VS-Graph, a new framework that combines the efficiency of hyperdimensional computing (HDC) with the expressive power of graph neural networks (GNNs). This innovative method performs graph learning entirely within a high-dimensional vector space, eliminating the need for computationally expensive gradient-based optimization. The researchers designed a spike diffusion mechanism that identifies key nodes based on the topology of the graph, leading to the construction of node representations that accurately reflect their location and importance. The team then implemented an associative message passing scheme that aggregates information from multihop neighbors, allowing for rapid and efficient information exchange. Experiments demonstrate that VS-Graph achieves accuracy competitive with state-of-the-art GNNs on benchmarks such as MUTAG and DD, outperforms previous HDC approaches by 4-5%, and speeds up training by up to 450 times.
VS-Graph delivers fast and accurate graph classification
The research team has developed VS-Graph, a new framework for graph classification that significantly improves computational efficiency and achieves high performance. Experiments demonstrate that VS-Graph significantly reduces both training and inference times while achieving comparable or better accuracy than existing graph neural networks on benchmark datasets. On the NCI1 dataset, VS-Graph completed training in less than 1 second, compared to more than 5 minutes for comparable graph neural networks. VS-Graph is suitable for resource-constrained environments because it maintains high accuracy even when hypervector dimensionality is significantly reduced, unlike comparison approaches that suffer from reduced accuracy.
Vector symbol graph for fast classification
VS-Graph represents a significant advance in graph classification, introducing a vector symbolic framework that bridges the gap between the computational efficiency of hyperdimensional computing and the expressive power of modern graph neural networks. The method leverages new techniques such as spike spreading and associative message passing to build a robust graph-level representation in a single pass and achieves excellent accuracy on benchmarks such as MUTAG and DD. VS-Graph achieves this performance without relying on gradient-based optimization, resulting in faster training times and ultra-low inference latencies. Future research will focus on exploring neuromorphic co-design and specialized hardware accelerators to further enhance the efficiency and scalability of VS-Graph's brain-inspired hyperdimensional computation.
