Single-Valued Risk Estimation for Segmentation and Regression Software with Dependent Data

The use of machine learning models, including segmentation and regression algorithms, has grown tremendously in the last decade. When used in safety-critical systems, an adequate risk and reliability analysis is necessary to guarantee safe usage. Hereby, complex models pose a special challenge, as their black-box character hinders several established testing techniques. Consequently, the risk can only be described at a confidence level. In this contribution, we propose a method to obtain a single-valued risk estimation, which upper bounds the failure on demand of segmentation or regression components at high confidence. This enables the usage of classical risk assessment techniques like fault trees or Markov models and their existing tools. With the framework of proper covers, we depart from independent data, providing a more realistic setup than the common viewpoint of independent and identically distributed data. Additionally, by balancing bound and confidence, this method can be viewed as an optimal selection. So, in case the confidence level is not given, one can use our framework to get an automated selection.