Towards classes of architectural dependability assurance for machine-learning-based systems
Max Scheerer, Jonas Klamroth, Ralf Reussner, Bernhard Beckert · 2020
Advances in Machine Learning (ML) have brought previously hard to handle problems within arm's reach. However, this power comes at the cost of unassured reliability and lacking transparency. Overcoming this drawback is very hard due to the probabilistic nature of ML. Current approaches mainly tackle this problem by developing more robust learning procedures. Such algorithmic approaches, however, are limited to certain types of uncertainties and cannot deal with all of them, e.g., hardware failure. This paper discusses how this problem can be addressed at architectural rather than algorithmic level to assess systems dependability properties in early development stages. Moreover, we argue that Self-Adaptive Systems (SAS) are more suited to safeguard ML w.r.t. various uncertainties. As a step towards this we propose classes of dependability in which ML-based systems may be categorized and discuss which and how assurances can be made for each class.