ISBN 978-3-8439-4804-3

978-3-8439-4804-3, Reihe Informationstechnik

Patrick Schlachter
Machine Learning in an Open World: Intra-Class Splitting and Automatic Measurement Data Plausibility Check

205 Seiten, Dissertation Universität Stuttgart (2021), Softcover, A5

Zusammenfassung / Abstract

Classification problems are present in many real-world application domains, including autonomous driving, medical signal processing or metrology. In recent years, some classification problems were successfully solved by machine learning. However, most state-of-the-art methods fail in real applications when novel or unknown data occur, because they assume a closed world with a fixed set of classes.

In contrast, open set recognition solves classification in an open world by expecting unseen classes during testing. This is highly relevant in practice and can improve safety and robustness. An additional challenge arises when only one training class is available. This is called one-class classification and is deployed in anomaly detection.

This dissertation focuses on machine learning in an open world from a theoretical and application-oriented perspective. In the first part, the generic principle of intra-class splitting is proposed, which enables novel end-to-end deep learning methods for open set recognition and one-class classification. Moreover, new feature learning and active learning methods for one-class classification were developed. Results are state-of-the-art methods for machine learning in an open world that could outperform existing methods on popular image datasets and are easy to use.

The second part deals with the design and implementation of a system for automatic plausibility checks of measurement data, which is an application of one-class classification. Thereby, a methodological foundation was established by combining existing with novel methods, such as ensemble-based online anomaly detection and the detection of relations in multivariate data using hybrid feature learning. Finally, a client and server software was designed and implemented, which is now in productive use.

As a whole, many starting points for further research are offered, including intra-class splitting, hybrid feature learning or machine learning in metrology.