Informed Machine Learning: Integrating Prior Knowledge into Data-Driven Learning Systems

Machine Learning is an important method in Artificial Intelligence (AI). It has shown great success in building models for tasks like prediction or image recognition by learning from patterns in large amounts of data. However, it can have its limits when dealing with insufficient training data. A potential solution is the additional integration of prior knowledge, such as physical laws, logic rules, or knowledge graphs. This leads to the notion of Informed Machine Learning (Informed ML). However, the field is so application-driven that general analyses are rare. The goal of this PhD thesis is the unification of Informed ML through general, systematic frameworks. In particular, the following research questions are answered: 1) What is the fundamental concept of Informed ML, and how can existing approaches be structurally classified, 2) is it possible to integrate prior knowledge in a universal way, and 3) how can the benefits of Informed ML be quantified, and what are the requirements for the injected knowledge? First, a concept for Informed ML is proposed, which defines it as learning from a hybrid information source that consists of data and prior knowledge. A taxonomy that serves as a structured classification framework for existing or potential approaches is presented. It considers the knowledge source, its representation type, and the integration stage into the ML pipeline. The concept of Informed ML is further extended to the combination of ML and simulation towards Hybrid AI. Then, two new methods for a universal knowledge integration are developed. The first method, Informed Pre-Training, allows to initialize neural networks with prototypes from prior knowledge. Experiments show that it improves generalization, especially for small data, and increases robustness. An analysis of the individual neural network layers shows that the improvements come from transferring the deeper layers, which confirms the transfer of semantic knowledge (Informed Transfer Learning). The second method, Geo-Informed Validation, checks models for their conformity with knowledge from street maps. It is developed in the application context of autonomous driving, where it can help to prevent potential predictions errors, e.g., in semantic segmentations of traffic scenes. Finally, a catalogue of relevant metrics for quantifying the benefits of knowledge injection is defined. Among others, it includes in-distribution accuracy, out-of-distribution robustness, as well as knowledge conformity, and a new metric that combines performance improvement and data reduction is introduced. Furthermore, a theoretical framework that represents prior knowledge in a function space and relates it to data representations is presented. It reveals that the distances between knowledge and data influence potential model improvements, which is confirmed in a systematic experimental study. All in all, these frameworks support the unification of Informed ML, which makes it more accessible and usable - and helps to achieve trustworthy AI.