Erik Wallin, Signalbehandling och medicinsk teknik

Deep learning-based classification systems commonly rely on conditions that are difficult to satisfy for real-world applications. One such requirement is the availability of large-scale, curated, and labeled training data. Another is the absence of unknown classes during training and deployment. Furthermore, many classification systems treat classes as independent, even when they form structured relationships that are important to account for. Overcoming these limitations is central to the practical deployment of these systems.

We address these challenges through five papers that study deep classification under limited supervision, the presence of unknown classes, hierarchical class structures, and combinations thereof. Paper A studies semi-supervised learning, where labeled and unlabeled training data are combined, and proposes a self-supervised component for better utilization of unlabeled data. Papers B and C address unknown classes within semi-supervised learning, enabling learning from realistic, uncurated, unlabeled data. In particular, Paper C proposes a probabilistic method that improves accuracy and uncertainty quantification when detecting unknown samples in this setting. Finally, Papers D and E study hierarchical open-set classification, i.e., assigning unknown classes to appropriate high-level categories of a hierarchy, and propose a method that approximates the predictive distribution over both known classes and higher-level categories. This enables more expressive predictions of unknown samples than binary rejection.

In summary, the included papers propose methods that advance performance on benchmarks for their respective problem settings, while providing empirical analyses that improve understanding of the underlying challenges. Overall, this thesis contributes to more robust and accurate deep classification systems for real-world deployment.