Can AI-based Components be Part of Dependable Systems?

Artificial Intelligence and especially Machine Learning have become main topics in the scientific community as well as in industry. These techniques seem to be a solution for complex problems and are suggested even for critical applications such as the medical diagnosis, predictive analytic in finance or autonomous driving. But will it be feasible to employ such techniques in critical systems in the avionics, train, or medical domain taking into account the current regulations in domain-specific standard relating to dependability? Validation, verification and certification in these domains strongly rely on explicit traceability and provability of functional and dependability requirements down to the code level. This seems to be impossible for components derived using e.g. learning approaches. Nonetheless, scientific community and industry do not want to lose the advantages related to AI-based techniques - new ways to ensure the required level of confidence just need to be found. This is a process that requires people with different professional background to work together - and it has started. This paper presents selected aspects relating to the current state of the art with regard to AI and dependable systems and describes ongoing activities and ideas for obstacle detection and routing for autonomous driving at HAW Hamburg.