A virtual node provides a graph neural network with a shared location for collecting and distributing information across a graph. Based on the observation that in the image domain, multiple such `registers’, which are working similar to virtual nodes, improve learning (ICLR 2024), this topic evaluates the corresponding question for graph learning. While multiple VNs have appeared in the literature, this project specifically investigates the impact of multiple VNs, potentially with learned connections as in the NeurIPS paper on larger and newer datasets, as well as its impact on transformer-based architectures.

Starting literature:

• Neural Atoms (ICLR 2024)
• Probabilistic Graph Rewiring via Virtual Nodes (NeurIPS 2024)
• Vision Transformers Need Registers (ICLR 2024)