Pre-requisites:
A solid background in differential geometry or geometric computing will be advantageous but is not required.
Students who haven't followed any related courses (Differential Geometry I, Scientific Visualization, ...) can follow the seminar but should be willing to invest more time.

Description:
Geometric deep learning is a broad and emerging research paradigm concerned with the derivation and study of neural network architectures that respect the invariances and symmetries in data.
Indeed, many real-world tasks come with essential pre-defined regularities arising from the underlying low-dimensionality and structure of the physical world.
Capturing these regularities via unified geometric principles has been shown to provide sizable empirical improvements.
Examples of such geometric architectures include graph neural networks as well as models conditioned on data that reside on curved manifolds where vector space operations are not naturally admissible.

The goal of this seminar will be to obtain in-depth knowledge about the core methodology in geometric deep learning as well as an overview of state-of-the-art methods.
Students will acquire practical skills in reading, presenting, explaining, and discussing scientific papers.
The seminar may be used as a preparation for an MSc thesis topic.

Literatur

Michael M. Bronstein, Joan Bruna, Taco Cohen, Petar Velickovic (2021) Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges. arXiv:2104.13478

Zusätzliche Informationen