Eduardo Santos Escriche

The focus of this project is on Geometric Deep Learning, which unifies a broad class of machine learning techniques from the perspectives of symmetries, seeking to provide a general framework for defining inductive biases aimed at specific problems. However, current formulations of this framework present some limitations, e.g., requiring knowledge of the symmetry a priori or being limited to global or exact symmetries that are often not found in real-world problems. As a result, this PhD project aims to explore different avenues for overcoming those limitations, for instance, by defining automatic symmetry discovery methods or by exploring ways of allowing for approximate or local symmetries instead of exact and global ones. In addition, this project will also study the effect of applying such methods to the specific application domain of molecular generation, for which we believe that such alternative symmetry formulations could be particularly beneficial. In summation, this PhD project will explore the role of symmetries in Deep Learning, with the goal of overcoming some limitations of current Geometric Deep Learning formulations and providing possible improvements in the performance or robustness of methods for the task of molecular generation.