Graphs as arrays, message passing, and how they add up to message passing neural networks, explained with lots of pictures and a minimum number of equations.
Graph deep learning sounds exotic, but the core ideas are picture-friendly. This talk follows four parts: how graphs can be represented as arrays, what message passing is (and its linear algebra interpretation), how embedding message passing inside a neural network gives you message passing neural networks, with graph Laplacian and graph attention networks as variations on a theme, and the learning tasks that involve graphs, such as node label prediction and edge prediction. Lots of pictures, a minimum number of equations. If you've used neural networks but never built a graph neural network, you'll leave ready to implement your own GNN layers instead of relying on black box APIs.
Slides are available here.
The GitHub repository can be found here.