Logical Neural Networks and Formal Ontologies: integration perspectives  

10 June 2025 | Aula Consiglio - Online | 11:00 | Matteo Codiglione  (FBK-3DOM) 

Abstract
Logical Neural Networks (LNNs) – as proposed by Riegel et al. in 2020 – have been widely discussed as architectures combining the essential properties of both neural networks (adaptability and learning) and symbolic logic (transparency and inference), in particular by enabling weighted and resilient reasoning which features the flexibility required for practical deployment in real-world scenarios. Moreover, LNNs may be employed to perform Inductive Logic Programming (ILP), thereby opening the path to applications that go beyond standard “passive” reasoning. Formal Ontologies, for their part, are well-established as data frameworks that facilitate precise and domain-specific semantic modeling and that support reasoning processes based on explicit rule languages (e.g., SWRL). While LNNs achieve interpretability at the expense of certain rigidities in architectural design—and thus would benefit from a robust semantic foundation capable of guiding such structuring—Formal Ontologies attain semantic and inferential precision through strict, classical logic rules and would greatly benefit from more adaptable reasoning mechanisms and from automated rule induction and discovery. In this seminar, we will only briefly introduce the underlying technologies and rather concentrate on the conceptual ways in which they may compensate for each other’s limitations and on the features which an integrated system would display. 

Bio
Matteo Codiglione is a first-year Ph.D. student (IECS, University of Trento) at the 3DOM unit in FBK, where he was previously employed during the final year of his Master’s studies. He holds a Master’s degree in Philosophy - with a background in ontology, philosophy of mind and philosophy of language - and is currently involved in the development of 3Dont, an ontology-based framework for the semantic interpretation and management of 3D data.