An Overview of the Research Landscape in the Field of Safe Machine Learning

The applicability of ML components in safety-critical systems will significantly depend on whether it will be possible to provide a comprehensive proof of their safety. Three research questions (RQ) are answered in order to provide a starting point for future activities towards the risk assessment of safety-critical systems containing ML components. First, special emphasis was placed on the design of a literature search strategy in order to enable quantitative insights into a representative set of publications (RQ1). Taxonomy analysis, bibliographic data visualization and compiled findings of reviews were therefore combined. A categorization of identified methods towards improving, ensuring and assessing safe machine learning was developed (RQ2). Then, a comparison was made with those methods for safety-critical software that are recommended by the functional safety standard IEC 61508 (RQ3). The comparison and quantification of research activity revealed imbalances within separate research areas. The conclusion is drawn that for the safety assessment of ML systems a comprehensive toolbox of combined methods needs to be converged from procedures within existing safety standards and the broad spectrum of methods proposed by current scientific publications.