Abstract: Advances in machine learning expose the limits of purely predictive models and motivate a shift toward decentralized, scalable, causal approaches. Such approaches often require learning on complex, expanding networks that are hard to analyze and design. Topological Deep Learning (TDL) provides tools that combine algebraic topology, non-Euclidean geometry, and category theory to handle this complexity. In TDL, networks are built from cells that generalize graph nodes and can represent agents, neural units, or sensors arranged in hierarchical structures with rich interaction patterns. Such representations help reveal structure that traditional models overlook and support more reliable reasoning. The seminar will introduce TDL principles, show how to infer latent cell complexes capturing higher-order relations, and present cellular sheaves that encode relative causal knowledge across cells.

Host: Prof. Cesare Alippi