Neuro-Symbolic Graph Transformation

Rule-based systems that perform symbolic computations on graph-structured data have been studied for more than five decades and are extremely well understood.

Due to their high level of abstraction, they provide transparency: both the computations themselves and their results can be interpreted by humans.

They also provide algorithmic transparency, due to a rich theory that has been developed over half a century, and resource efficiency However, their discrete nature makes them inflexible and difficult to learn from data.

In contrast, modern neural models of computation are untransparent black boxes whose creation is resource demanding, but they come with powerful machine learning algorithms and can flexibly adapt to unanticipated inputs.

This project´s aim is to integrate the latter into the former to combine their advantages, thus making rule-based methods more general and adaptive without losing transparency and their good algorithmic properties.The foundation of the project will be my ongoing work on rule-based formalisms for the generation and transformation of graphs, which will be taken to a new level by integrating neural methods in two ways: (1) by providing formal hooks which can be instantiated by neural models to restrict or generalize the behavior of the rule-based model and (2) by on-the-fly creation of rules similar to the original rules of a given rule-based system.