Paper Summary: A boxology of design patterns for hybrid learning and reasoning systems
One of the main research questions in our SemSys group is the design and development of Auditable Semantic AI Systems. In this direction, the combination of Knowledge-based and Machine Learning systems has gained traction in recent times, e.g., to provide explainability for ML models and other application fields. However, while the state of art in this field is rapidly evolving, there is currently little consensus yet on how to classify such systems and evaluate them in a structured way.
Therefore, van Harmelen and ten Teije [1] propose in their paper a set of design patterns, similar to the ones common for software engineering [2] or ontology modeling [3], for systems that combine symbolic knowledge representation techniques with statistical techniques from knowledge representation.
They come up with a graphical boxology notation to categorize possible patterns based on hybrid systems combining reasoning and learning. These patterns include Inductive Logic Programming, Markov Logic networks, ontology learning from text, Logic Tensor Networks, and meta-reasoning systems among others. The patterns start from more simple “basic” building blocks, which are then composed into more complex patterns towards the end.
Using these patterns is an interesting method to categorize such hybrid systems combining Semantic and Machine Learning technologies. One of the strengths is certainly the relatively easy notation consisting of a small subset of symbols and the ease to communicate these different types of systems.
All in all, the paper is an interesting read and suggests one possible way to characterize such hybrid systems. It will be interesting to investigate different approaches to come up with a suitable taxonomy for our application field.
[1] Van Harmelen, F., & Teije, A. T. (2019). A boxology of design patterns for hybrid learning and reasoning systems. arXiv preprint arXiv:1905.12389. [2] Fowler, M. (2002). Patterns of enterprise application architecture. Addison-Wesley Longman Publishing Co., Inc. [3] Gangemi, A., & Presutti, V. (2009). Ontology design patterns. In Handbook on ontologies (pp. 221-243). Springer, Berlin, Heidelberg.