NSF Award
2326495
The digital revolution has generated a vast volume of interconnected data, often represented as graphs, which is pertinent to numerous critical real-world applications. This has led to the increasing prevalence of Graph Neural Networks (GNNs), a technique that extends the benefits of Artificial Intelligence (AI) to graph-based applications. GNNs hold promising potential to significantly impact society, from accelerating drug discovery and preventing supply chain disruptions, to averting cascading power grid failures and identifying misinformation on social media. However, the actualization of such potential is currently impeded by computational inefficiencies caused by the colossal size and intricate nature (such as extreme sparsity and irregularity) of graphs, which pose challenges to the practical deployment of GNNs. This project aims to bridge the gap between the computational efficiency required for GNNs and their current performance, primarily due to the uniquely heavy load of communication required in GNN computation. In addition, the project enriches the educational experience of undergraduate and graduate students in the US by enhancing the quality of AI and system-related courses and outreach activities at the University of Rochester and Indiana University. Successful completion of this research project can unlock the immense potential of GNNs to solve problems in fields of medicine, public infrastructure, and economic development, among many other issues critical to the well-functioning of the republic and the prosperity of its economy. <br/><br/>This project aims to develop a revolutionary communication reduction method that organically integrates on-the-fly versatile graph locality enhancement and high-ratio compression through software-hardware co-design. The research is structured around three primary thrusts: (1) The development of an on-the-fly graph locality enhancer via hardware-software co-design, providing significant versatility and additional reductions in communication demands compared to current leading methods. (2) The creation of an efficient lossy compressor that enables high-ratio, error-bounded compression and decompression for graph data, including both graph embedding and topology information. (3) The investigation into methods for effectively combining the graph locality enhancer and graph compressor, allowing them to mutually benefit each other. These strategies together directly address the persistent communication bottlenecks in GNNs and unleash their potential for societal benefits. Moreover, this project aims to resolve the following query: whether a collaborative integration of locality enhancement and data compression, the two most prevalent communication optimization approaches, can provide a ground-breaking solution to general graph problems.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
