This document addresses the validation of functions, or components, integrating or using machine learning algorithms (AI/ML) through identification of error risks throughout the development cycle of an AI/ML component. These risks are systematically linked to the activities and quality objectives of the development process. Measures for mitigating these risks and criteria for assessing their efficiency at different steps of the development process are therefore proposed.
This document has the following characteristics:
— from a methodological standpoint, it adopts a risk-based approach addressing the inherent risks of machine learning, rather than a process-oriented approach. Two points of particular concern have been addressed: the specification of AI/ML components and the structural separation between learning activities and verification and validation activities in relation to domain experts;
— it describes the specific activities involved in developing an AI/ML component (system-related and planning aspects are not specific and are therefore not covered);
— it sets out the main issues relating to the modulation of the AI/ML assurance level according to the criticality of the applications.
This document limits the relationship between the system level and the AI/ML component to specification and validation matters. An AI/ML component is regarded as a unitary object that is not decomposed into further software components.