Uncertainly quantification for interpretable and reliable machine learning

Abstract

In the booming research field of machine learning and artificial intelligence, uncertainty quantification is often overlooked as an essential quality guarantee. When interacting with and applying artificial intelligence, it is common to evaluate the performance of such learners by metrics such as efficiency, accuracy and other task-specific performance-based metrics without emphasizing the importance of quantifying potential hazards. Uncertainty characterizes the proximity between observations and predictions, providing a measure of how well a model reflects the true underlying data distribution. This thesis places uncertainty quantification at the center of its investigation, and we investigate three interconnecting subareas:. Explanations and visualizations of uncertainties, robustness with regard to uncertainties, and trustworthiness and human interpretability of uncertainties. This interdisciplinary setup is crucial to establishing new connections between the domains. While each of the subjects has been extensively studied in isolation, it is often at the interconnections that new research paradigms emerge: Neural networks as we know them today are a result of a fusion of neuroscience-inspired models of cognition, mathematical formalization, and algorithmic innovation. Without the mathematical groundwork, the advances in computing power and the biological neuron inspiration, this new field of research would not exist today. By interweaving explanation, robustness and trust in the context of uncertainty, in this thesis, we aim to pave the way for engineering practical systems that are both reliable, interpretable, and ultimately trustworthy. In these three areas, we focus on empirical approaches and solutions for important research challenges. The first part of this thesis focuses on visualizing uncertainties of high-dimensional data in an unsupervised setting using the novel proxy for local intrinsic dimensionalities. Furthermore, we show limitations of popular explainable AI methods using a newly constructed open-source dataset that focuses on an ambiguous classification task. We use the proxy of local intrinsic dimensionality as a proxy for the likelihood of adversarial attacks, connecting uncertainties with robustness metrics. In the second part of the thesis, we delve more deeply into the robustness domain by proposing certainty attacks and discussing the independence of adversarial transferability to topological changes in the datasets. We discuss the origin of transferability and possible research directions for future work. The main motivation between the importance of uncertainties stems from the need to calibrate human trust---for the successful application of machine learners, we have to align the trust levels of humans according to their actual performance. This motivates the third part of the thesis, where we discuss trust in AI systems with a special emphasis on uncertainty quantification. Finally, we discuss open challenges regarding uncertainty quantification and outline future work in this particular domain, with special emphasis on explainable AI and robustness.