Geometric Deep Learning

Modern deep learning has had tremendous success in applying complex neural networks to problems from a

wide range of disciplines, such as computer vision and protein folding. Geometric deep learning deals with

incorporating symmetries into deep learning architectures. A symmetry of features is a transformation that is

guaranteed not to change the labels. Symmetries are ubiquitous in many machine learning tasks. For example,

in computer vision the object category is unchanged by shifts, so shifts are symmetries in the problem of visual

object classification. In computational chemistry, the task of predicting properties of molecules independently

of their orientation in space requires rotational invariance. This course gives and understanding of the theoretical

basis underlying geometric deep learning. Furthermore, the course includes implementation of geometric

components and as well as applying geometric deep learning on real-world data.

Content:

• Geometry
• Geometric priors
• Learning on graphs and sets
• Group-equivariant learning
• Learning on manifolds

Learning objectives:

Upon completion of the course the student be able to

• understand the basic principles of geometric deep learning
• implement geometric deep learning algorithms
• compare various approaches in geometric deep learning
• read and critically assess geometric deep learning papers
• apply and evaluate geometric deep learning methods on real data sets