Physical infrastructures acting “smart” due to embedded IT and interacting with humans, that is, human-cyber-physical systems, require the integration of artificial intelligence (AI), especially of components trained by machine learning (ML). Given the safety-criticality of many physical infrastructures, this provokes a quest for ML validation and verification in context, that is, reflecting the overall application and the system architecture into which the ML components are embedded. In this lecture note, we discuss the diversity of verification patterns thus induced, hint at approaches matching them, and discuss issues of soundness and practical applicability.