Speaker: Fabio Galasso (Roma La Sapienza Univ.)

Representation learning is an important part of modern computer vision. Literature assumes a default Euclidean space, thus a manifold based on regular grids. Only most recently, hyperbolic spaces have enabled techniques to reach and surpass the state-of-the-art, supporting learning with hierarchical structures and uncertainty, also a by-product of hyperbolic representation learning. I will introduce our most recent work that leverages Hyperbolic Neural Networks for anomaly detection, self-supervised learning of actions, active learning of semantic segmentation, and reinforcement learning of robot navigation in social environments.