Dfa GNN
DFA-GNNs represent a novel approach to training Graph Neural Networks (GNNs) that bypasses the limitations of backpropagation, offering improved efficiency, scalability, and biological plausibility. Current research focuses on enhancing GNN performance through architectural innovations like incorporating global confidence degrees, hierarchical priors, and multi-granularity representations, as well as optimizing training methods for large-scale datasets and resource-constrained environments. These advancements are impacting diverse fields, including financial fraud detection, medical diagnosis, and material science, by enabling faster, more accurate, and privacy-preserving analysis of complex, graph-structured data.
Papers
LSM-GNN: Large-scale Storage-based Multi-GPU GNN Training by Optimizing Data Transfer Scheme
Jeongmin Brian Park, Kun Wu, Vikram Sharma Mailthody, Zaid Quresh, Scott Mahlke, Wen-mei Hwu
Improving Prediction of Need for Mechanical Ventilation using Cross-Attention
Anwesh Mohanty, Supreeth P. Shashikumar, Jonathan Y. Lam, Shamim Nemati