Uncertainty Quantification in Advanced Machine Learning Approaches

Abstract

Artificial intelligence (AI) systems perform critical tasks in various safety-critical (e.g., medical devices, mission-control systems, and nuclear power plants). Uncertainty in the system may be caused by various reasons. Uncertainty quantification (UQ) approaches are essential for minimising the influence of uncertainties on optimisation and decision-making processes. Estimating the uncertainty is a challenging issue. Various machine-learning approaches are used for uncertainty quantification. This chapter comprehensively views uncertainty quantification approaches in machine learning (ML) techniques. Various factors cause uncertainty, and their possible solutions are presented. The uncertainty analysing approaches of the different machine learning methods, such as regression, classification, and segmentation, are discussed. The uncertainty optimisation process is broadly categorised into backward and forward approaches. The subsequent sections further classify and explain these backward and forward uncertainty approaches.