Uncertainty in Machine Learning Applications: A Practice-Driven Classification of Uncertainty

摘要

Software-intensive systems that rely on machine learning (ML) and artificial intelligence (AI) are increasingly becoming part of our daily life, e.g., in recommendation systems or semi-autonomous vehicles. However, the use of ML and AI is accompanied by uncertainties regarding their outcomes. Dealing with such uncertainties is particularly important when the actions of these systems can harm humans or the environment, such as in the case of a medical product or self-driving car. To enable a system to make informed decisions when confronted with the uncertainty of embedded AI/ML models and possible safety-related consequences, these models do not only have to provide a defined functionality but must also describe as precisely as possible the likelihood of their outcome being wrong or outside a given range of accuracy. Thus, this paper proposes a classification of major uncertainty sources that is usable and useful in practice: scope compliance, data quality, and model fit. In particular, we highlight the implications of these classes in the development and testing of ML and AI models by establishing links to specific activities during development and testing and means for quantifying and dealing with these different sources of uncertainty.