IRIS-HEP Fellow: Sofia Graziano
Fellowship dates: May – Aug, 2021
Home Institution: University of Pennsylvania
Project: Developing Symmetric Graph Neural Networks for Charged Particle TrackingCharged particle tracking is essential in characterizing particles produced in colliders; traditional tracking algorithms scale up poorly, and new methods need to be developed. One approach is to use geometric deep learning to classify connections between tracker hits as true or false, and then link them together to form final track candidates. This can be done using graph neural networks (GNNs) by first con- structing a graph of tracker events and then processing the graph with an intelligent network (IN) or similar architecture. Graphs are a natural representation of particle data because hits can be represented as nodes and track segments can be represented as edges. This project proposes constructing the GNN, and implementing the function that the GNN learns on using the equivariant approach. I will investigate the rotational, CPT, and other symmetries that the dataset should have and construct and train the GNN to be equivariant to these symmetries to help constrain the network size and improve the accuracy of the machine learning algorithm.
More information: My project proposal
Savannah Thais (Princeton University)
Daniel Murnane (LBL)
- 1 Nov 2021 - "Rotationally Equivariant Graph Neural Networks", Sofia Graziano, IRIS-HEP Topical Meetings Recording: Rotationally Equivariant Graph Neural Networks