Partition Pooling for Convolutional Graph Network Applications in Particle Physics
This incremental improvement addresses computational bottlenecks for particle physics researchers using graph neural networks, potentially enhancing event reconstruction efficiency.
The paper tackled the performance limitations of convolutional graph networks in particle physics due to high sensor counts by introducing a partition pooling scheme that reduces computational resources, enabling deeper networks and yielding improved performance with less overfitting in simulated neutrino detector reconstructions.
Convolutional graph networks are used in particle physics for effective event reconstructions and classifications. However, their performances can be limited by the considerable amount of sensors used in modern particle detectors if applied to sensor-level data. We present a pooling scheme that uses partitioning to create pooling kernels on graphs, similar to pooling on images. Partition pooling can be used to adopt successful image recognition architectures for graph neural network applications in particle physics. The reduced computational resources allow for deeper networks and more extensive hyperparameter optimizations. To show its applicability, we construct a convolutional graph network with partition pooling that reconstructs simulated interaction vertices for an idealized neutrino detector. The pooling network yields improved performance and is less susceptible to overfitting than a similar network without pooling. The lower resource requirements allow the construction of a deeper network with further improved performance.