geometric deep learning

Geometric deep learning is a field of machine learning in which the machine learns from complex data represented through graphs or multi-dimensional points, possibly in non-euclidean spaces.

Graphs are a way of representing systems of related objects, therefore graph data processing has applications in many fields, like biology, transportation, natural language processing and social networks. Computing and learning on graphs is an emerging research topic also in computer vision. Point clouds, for example of a human body, are implicitly big graphs where vertices are the scanned points and edges are the (semantic) relations that link such points to form the surface. An increasingly active field of research regards the family of machine learning techniques called Graph Neural Networks (GNN) that can be used also to interpret point clouds. Among them, transformer architecture, a particular instance of GNNs, are now competitive with Convolutional Neural Network models. Geometric machine learning is strongly related to graph machine learning and graph representation learning.

The development of models for geometric deep learning is a research area of interest in TeV with applications in scene understanding and in human behavior analysis. Results are in shape analysis and articulated object pose recognition.