Post-Training Neural Network Pruning using Graph Curvature
Researchers have introduced a novel approach to neural network pruning by leveraging graph theory, specifically Ollivier-Ricci curvature (ORC). This method identifies critical data flows and connections within a neural network by analyzing activation patterns and calculating a metric called neural curvature (NC). Experiments on image datasets like MNIST and CIFAR-100 demonstrate that this NC-based pruning can more effectively identify unimportant edges compared to existing techniques. AI
IMPACT Introduces a novel graph-theory-based method for optimizing neural networks, potentially leading to more efficient models.