Conceptual Knowledge Modelling for Human-AI Teaming in Data-Frugal Industrial Environments

When AI interacts with humans in complex environments, such as aerospace manufacturing, safety of operation is of paramount importance. Trustworthiness of AI needs to be ensured through, among other things, explainability of its behaviour and rationale, which is typically a challenge for current deep neural network-based systems. We tackle the knowledge comprehensibility aspect of intrinsic explainability by suggesting a concept-level environment awareness model combining various complementary knowledge sources - statistical learning using dedicated property detectors through publicly available software, and crowd-sourced common-sense knowledge graphs. Our approach also addresses the issue of data-frugal learning, typical for environments with highly specific purpose-built artefacts. We adopt Gärdenfors’s Conceptual Spaces as a cognitively-motivated knowledge representation framework and apply our typicality quantification model in a use case on interpretable classification of manufacturing artefacts.