Defining an Initial Classification Scheme for Non-Deterministic AI Technologies

A growing number of aviation applications using Artificial Intelligence (AI) and Machine Learning (ML) algorithms demonstrate how these methods may increase the accuracy of predictions and identify trends or correlations that are difficult for humans to recognize. This is particularly true when the patterns are buried and must be extracted from a large volume of data samples. While the potential benefits of AI and ML are clear for some applications, the inherent difficulty in explaining the behavior of these algorithms presents challenges for potential certified use in civil aviation. If one cannot explain the AI decision or predict the ML output, one cannot provide guarantees on safety or assurances on system performance.Traditional software and system assurance processes rely on mathematical stability, behavior bounded by physics, and/or deterministic system dynamics that may not be applicable to AI systems. For example, a typical method used in verification is to stress test the system on all possible input values and check/verify if the key system parameters (and output variables) stay within prescribed bounds (i.e., safety limits). Current assurance processes also rely primarily on safety assurance thresholds, which drive validation and verification requirements for approval. While safety is a major factor in determining requirements, other factors addressing the unique technical aspects (e.g., non-deterministic and autonomous) and functional aspects (e.g., purpose and role) of AI technologies should also be considered. Thus, exploring an initial AI classification scheme is a crucial step in maturing research for the future acceptance and approval of different types of AI-driven services. Existing efforts (e.g., European Union Aviation Safety Agency (EASA)) to address requirements for AI technologies have begun to define roadmaps based primarily on level of autonomy. However, our work goes beyond the "level" of human/system interaction by considering a formal structure based on both technical and functional aspects of AI. A multi-facet classification scheme may also facilitate low-risk, high-value, advisory AI use cases by assigning an appropriate level of rigor.The goal of our work is to propose, research, and analyze a classification scheme considering both technical and functional AI characteristics that can be applied to a wide range of AI technologies. Our aim is to document gaps and challenges in defining a robust classification scheme. In our methodology, we employ a three-fold approach. First, based on literature review and subject matter expertise, we identify and define components that form the classification scheme. These classification components include the functionality (i.e., functional criticality) of the application being developed, the role of the AI with respect to a human user (e.g., advisory), and the trustworthiness (e.g., computational complexity) of the AI method utilized. Each of these components will be categorized into factors that can be considered against a variety of AI applications. Second, we develop relationships to describe how the classification scheme may be used to inform validation and verification approaches. We focus on how each of the components may influence the potential level of rigor needed for approval or acceptance of the AI technology. Third, we support our classification scheme by aligning the classifications with realistic aviation use cases.