Geometric Deep Learning

Resources

Main Idea

Certain Deep Learning problems or applications admit symmetries. Geometric Deep Learning is the field of encoding these symmetries into architectures. Fundamentally, the symmetries are described by Groups in Abstract Algebra. For instance, Convolutional Neural Networks are invariant to some translation group. Similarly, Graph Neural Networks are invariant to the numbering of their nodes or edges. When a graph is represented by an Adjacency Matrix, a function of this adjacency matrix should be invariant to any renumbering of nodes (which is described by some Group).