Machine learning methods for safety-critical systems with time dependency: With applications to airplane landings and environmental data

Abstract

Artificial intelligence and machine learning are increasingly used in many industries and applications, where more and more important decisions are being made based on machine learning models. However, recent research has shown limitations to many of the commonly used machine learning methods, such as lack of stability and transparency. As machine learning models are increasing their impact on many areas of people's lives, it is critical to ensure safe and responsible use of these methods. This thesis addresses challenges related to the use of machine learning for safety-critical systems. These systems often involve complex physical relationships with a large degree of uncertainty and time dependency. An example of such a system is the Earth’s weather dynamics, where we explore it in the context of safe airplane landings and extreme weather phenomena. The exploration is carried out by considering different ways to create more trustworthy machine learning models. This includes training the models to follow physical laws and mathematical constrictions, so that they behave according to the rules of system. We also explore how the models can provide explanations of their own behavior and information about their uncertainty, so that we know how they work and when we can trust them.