Towards Supporting AI System Engineering with an Extended Boxology Notation

Abstract

As AI systems grow in complexity, understanding their structure and behaviour becomes increasingly challenging. The boxology notation offers a dataflow-oriented abstraction to simplify AI system representation and help address this challenge. In this work, we explore the potential of extending the boxology notation for AI system engineering and introduce the Boxology Extended Annotation Model (BEAM). BEAM enhances boxology through (i) incorporating auxiliary notations to capture engineering-relevant information and (ii) introducing an additional perspective for AI system risk assessment and mitigation. Furthermore, we developed the BEAM ontology as a machine-readable representation of BEAM to support further functionalities. We evaluated the BEAM approach through (a) an initial feasibility evaluation with students in a classroom setting and (b) applying BEAM on use cases as part of a research project. Positive feedback from both evaluations demonstrates its effectiveness in supporting the AI system engineering process.