Review and Critical Analysis of Ontologies for Artificial Intelligence Systems

With the rising popularity of artificial intelligence-based solutions, it is becoming important not only to deploy machine learning models/pipelines with a good accuracy, but also to be able to control and manage their documentation and information related to monitoring, performance tracking, etc. Moreover, crucial aspects of data that is to be used by said applications, needs to be captured and represented in an organized and standardized way. These include, among others, provenance, access restrictions, usage limitations, format. Being able to (pseudo-)formally represent (and communicate) properties of data and applications should enhance efficient MLOps and facilitate governance of artificial intelligence-based systems. This is of particular importance for multi-stakeholder ecosystems, such as Internet of Things deployments, in which collaboration is required. The question thus arises, how such meta-level description of artificial intelligence system components and data that they consume, process and produce, can be realized. In this context, we investigate state-of-the-art of related ontologies, as well as several non-ontological, but important, approaches to represent knowledge about artificial intelligence models/components, related processes and data, to provide transparency into model’s development and system performance. We analyze for which elements of such overall description there are ontologies for potential reuse and where there are gaps that need to be filled to capture the ontological representation of all pertinent aspects of artificial intelligence-based system.