The Path to Safe Machine Learning for Automotive Applications

Recent rapid advancement in machine learning (ML) technologies have unlocked the potential for realizing advanced vehicle functions that were previously not feasible using traditional approaches to software development. One prominent example is the area of automated driving. However, there is much discussion regarding whether ML-based vehicle functions can be engineered to be acceptably safe, with concerns related to the inherent difficulty and ambiguity of the tasks to which the technology is applied. This leads to challenges in defining adequately safe responses for all possible situations and an acceptable level of residual risk, which is then compounded by the reliance on training data. The Path to Safe Machine Learning for Automotive Applications discusses the challenges involved in the application of ML to safety-critical vehicle functions and provides a set of recommendations within the context of current and upcoming safety standards. In summary, the potential of ML will only be unlocked for safety-related functions if the inevitable uncertainties associated with both the specification and performance of the trained models can be sufficiently well understood and controlled within the application-specific context.