The world is being quietly rearranged by people who write very long documents.


The title they went with On the Complexity of Optimal Graph Rewiring for Oversmoothing and Oversquashing in Graph Neural Networks Noisy translates that to

Graph neural networks hit a computational wall


Researchers proved that the two main ways to fix broken deep neural networks on graph data are mathematically impossible to solve perfectly—they're NP-hard problems, meaning no known efficient algorithm exists. This explains why researchers use approximations and heuristics in practice instead of searching for perfect solutions, which would take forever.
This is a theoretical ceiling: it establishes that certain fundamental improvements to how neural networks process graph-structured data (like social networks or molecular structures) cannot be solved exactly at scale, which justifies engineering shortcuts and tells practitioners where to stop looking for better exact algorithms.

If you insist
Read the original →