Vision
In this WG we focus on the study of new models to advance in our understanding of vision in the framework of the new techniques as deep learning (DL), geometric deep learning (GDL).
G4.1: Provide a new understanding of the interplay between the Geneo theory (in TDA) and the new machine learning algorithms coming from geometric deep learning with group equivariance.
G4.2: Provide new models for vision via Cartan Geometry, understand its application in DL, GDL.
T4.1: Enhance the Geneo approach to machine learning vision applications, beyond topological data analysis, towards the applications to concrete problems (molecular dynamics, material science).
T4.2: Reframe the GDL approach via symmetric space theories developed in Cartan geometry.
T4.3: Interpret SGD and the metric structure of the model space with Souriau Lie Thermodynamics. Interpret the coadjoint orbits of the symmetry group action as level set of entropy; exploit their symplectic structure to construct further symmetries (group equivariant GDL).