Encoding preferences for representation learning in computer vision tasks

Jung, Steffen

Vorschau

PDF
Dissertation_Steffen_Jung.pdf - Veröffentlichte Version
Download (34MB)

URN:	urn:nbn:de:bsz:180-madoc-717332
Dokumenttyp:	Dissertation
Erscheinungsjahr:	2025
Ort der Veröffentlichung:	Mannheim
Hochschule:	Universität Mannheim
Gutachter:	Keuper, Margret
Datum der mündl. Prüfung:	2025
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Machine Learning (Keuper 2024-)
Fachgebiet:	004 Informatik
Freie Schlagwörter (Englisch):	machine learning , representation learning , computer vision , regularization
Abstract:	Machine learning is the backbone of modern computer vision systems. The increasing availability of computational power, large-scale datasets, and advances in learning algorithms have enabled models to learn complex patterns directly from data, moving beyond traditional rule-based programming. A key factor in this success is representation learning, which transforms data into compact, meaningful representations for tasks such as image classification, segmentation, and synthesis. However, the No Free Lunch (NFL) theorems state that no universal algorithm can perform well across all tasks without incorporating task-specific knowledge. To build effective models, practitioners must introduce inductive biases that guide learning toward desirable solutions. Regularization is a central tool for this purpose, as it constrains the solution space and encodes preferences into the learning process. This thesis presents novel regularization techniques to guide representation learning in computer vision, focusing on three complementary strategies.