Ing. Gustav Šír, Ph.D.

This topic targets the contemporary problem of deep learning with relational data and knowledge representations. Virtually all neural models are limited to representations in the form of fixed-size tensors, which enabled for massive scaling by deploying the underlying matrix transformations to existing data-parallel hardware, particularly the GPUs. However, this form of representation is also considerably limiting, as there are also relational data, omnipresent in the interlinked structures of the Internet and relational databases, that do not succumb themselves easily to the precanned form of (dense) numeric tensors. This generic thesis topic then covers research and development in this direction of extending existing neural models with relational learning capabilities to be applied in a selected relational data domain, to be specified upon discussion w.r.t. the student's interests.

It has been recently proposed by many top AI researchers that incorporating symbolic (logic) learning and reasoning capabilities into neural networks is crucial to achieving more general and safe AI systems. Indeed, we see a rising interest in enriching deep learning models with certain facets of symbolic AI, such as logical reasoning and planning. This generic thesis topic covers the design and development of such methods, combining symbolic AI approaches incorporating relational-logic representations, for both deduction (reasoning, planning) and induction (inductive logic programming), with structured deep learning architectures, such as GNNs and Transformers, while focusing on mechanistic interpretability of their underlying principles. The particular methods of integration will be further detailed upon discussion with the supervisor, who will guide an interested student through the topic.