PI Stories by Emanuele Rodolà

PI Stories is a series of seminars organized by the IECS (former ICT) Doctoral School, aimed at providing the opportunity to the PhD students to learn the success stories of some of the most talented researchers in the world. Details at: iecs.unitn.it/...
Adventures in geometric deep learning
In this talk I will present several recent directions of research in the areas of machine learning and applied AI, the so called "geometric deep learning" subfield; common to these topics are the notions of geometry and structure.
I'll start by showing how data-driven methods can serve as efficient and robust drop-in replacements for classical axiomatic constructions, even when learning is not strictly necessary, for typical tasks of geometry and signal processing; in turn, I'll show how geometry itself can serve as a valid proxy for better machine learning pipelines. I'll then explore some preliminary steps toward the application of geometric deep learning in other fundamental disciplines, such as mathematical physics and structural biology, and conclude with an overview of some work in progress in our lab as well as some wishful thinking on future perspectives