This report is the output of a joint project, entitled ‘Concepts of Design Assurance for Neural Networks’, which examined the challenges posed by using Neural Networks in aviation within the broader context of allowing Machine Learning (ML) and more general Artificial Intelligence (AI) on board aircraft.
The report aims to address challenges related to the use of neural networks in aviation safety-critical systems.
Applicability
The report focuses on machine-learning (ML) systems applied to safety-critical application in aviation embedded systems.
Key achievements
-
Learning assurance
The report introduces a W-shaped development cycle for ‘learning assurance’, building on classical validation and verification (V&V) development assurance steps and visiting each step of a learning process.
-
Generalisation guarantee
The concept of generalisation is crucial for neural networks to ensure expected behaviour during operation.
-
Safety assessment
The report discusses the role of Artificial Intelligence (AI) / Machine Learning (ML) components in safety-critical failure scenarios through a visual-landing use case. It particularly focuses on how to handle incorrect outputs or wrong predictions made by AI/ML components. The report also considers the need for a quantitative safety assessment and a common mode analysis when required by regulations.
-
Advanced concepts
Transfer learning and synthetic data are explored.
Challenges
Addressing data completeness and representativeness, bias/variance trade-off, and robustness in Machine Learning (ML) models is critical for the introduction of Artificial Intelligence (AI) in aviation safety-critical application.
Future work
The report paves the way for practical approaches to certification safety objectives when Machine Learning (ML) / deep learning (DL) is used in aviation applications.
It was followed by a second Innovation Partnership Contract (IPC) report (CoDANN II) investigating complementary aspects and by the EASA Research and Innovation (R&I) MLEAP project.
Quoting this report
EASA and Daedalean, Concepts of Design Assurance for Neural Networks (CoDANN), March 2020.
