Parallel implementation of graph neural networks
Graph Neural Networks are a powerful machine learning technique for unstructured data. How can we make them scale to supercomputers?
Unlike convolutional neural networks which work on grid structured data such as images, graph neural networks (GNNs) can deal with unstructured connections between data. They have become an indispensable tool for many applications that involve such data, including social network analysis, traffic prediction, NLP, computer vision, and chemoinformatics.
In the recent years many new GNN architectures have been proposed, many of which are powerful but very expensive to compute. In this thesis we will implement modern GNN architectures on powerful parallel computers.
Goal
The goal of this thesis is to implement modern GNN architectures such as PathNN on large parallel systems.
Qualifications
- Experience with C/C++
- Familiarity with parallel programing (OpenMP,MPI) and Pytorch/PyG is helpful
Supervisors
- Johannes Langguth
Collaboration partners
- University of Vienna
- University of Bergen