D6Graph Neural Networks

Graph neural networks handle data that is naturally relational: molecules, social networks, knowledge graphs, recommendation systems, anything best described by nodes and the edges between them. Standard CNNs and Transformers assume a grid or a sequence, but a graph has no fixed ordering or neighborhood size, so this track builds architectures that respect that irregular structure directly.

You will represent graphs through adjacency and feature matrices, frame learning as the message, aggregate, and update steps of message passing, and then instantiate that framework as concrete models: GCN's Laplacian smoothing, GraphSAGE's neighborhood sampling, and GAT's attention weights. The track extends to graph Transformers that swap local passing for global attention, and to equivariant and geometric networks whose outputs transform predictably under symmetries, the kind used for 3D molecules.

These methods underpin drug discovery, materials science, traffic and logistics modeling, and large-scale recommendation, and they share the attention machinery you met in sequence modeling.