@phdthesis{Julka2025, author = {Julka, Sahib}, title = {Towards Data Efficiency and Controllable Representations for Deep Learning in Resource-Constrained Domains}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:739-opus4-16030}, school = {Universit{\"a}t Passau}, pages = {20 ungez{\"a}hlte Seiten, 131 Seiten, 17 ungez{\"a}hlte Seiten}, year = {2025}, abstract = {The deployment of artificial intelligence (AI) in specialised domains such as planetary science and healthcare, as well as in low-resource NLP settings, faces two fundamental challenges: label scarcity and data scarcity. Label scarcity stems from the high cost of expert annotation, the scarcity of domain experts, and the infeasibility of crowdsourcing, particularly in complex tasks requiring specialised knowledge. In parallel, data scarcity stems from the inherent difficulty of acquiring sufficient raw data, whether due to limited observational opportunities, environmental and technical barriers, or stringent privacy constraints. Together, these limitations impede the broader adoption of AI in these fields. Many existing approaches to label efficiency, such as active learning, rely on problem-specific heuristics and often, as a design choice, employ naive uncertainty estimations—typically at the instance level. However, such methods can lead to redundant or suboptimal sample selection by ignoring structural data properties and failing to account for representational diversity. In practice, they often perform no better than random sampling. For data synthesis, generative models face their own set of challenges. Despite their promise for synthetic data generation, these models frequently lack mechanisms to disentangle generative factors at the representation level, limiting their controllability. Additionally, standardised evaluation metrics to assess the quality of disentanglement remain underdeveloped, limiting their practical utility. These limitations highlight the need for advancements in data-efficient machine learning and controllable generative modelling, focusing on domain-specific validity and rigorous evaluation. This thesis contributes to addressing these challenges by proposing tailored solutions in two key directions. First, for data-efficient learning, a deep active learning (DAL) framework is introduced to enhance label efficiency by prioritising the most informative samples for annotation. Unlike traditional per-sample approaches, this framework aggregates uncertainty across larger data segments—such as orbital intervals in planetary science—allowing it to capture contextual variations. This method reduces labelled data requirements by up to 90\% in the case of boundary crossing detection at Mercury's magnetosphere. To further improve sampling diversity, a GAN-based concept drift detection method is integrated into the DAL framework, leveraging uncertainty and diversity together to offer a sampling method that outperforms random sampling. Additionally, foundation models such as the Segment Anything Model (SAM) are employed for zero-shot annotation to generate high-quality pseudo-labels, which are subsequently used to train a domain-specific model via knowledge distillation. This approach significantly enhances data efficiency, reducing the need for annotated samples several times over in the tested scenario of image segmentation for geological mapping. Furthermore, large language models (LLMs) are explored as active annotators for linguistic tasks in low-resource languages, achieving near-baseline performance while reducing annotation costs by up to 40x. Second, the thesis investigates methods to induce controllability in generative models, enabling the production of high-fidelity, controllable synthetic data. Conditional generative adversarial networks (CGANs) and disentangled representation learning techniques (DRL) are explored, particularly in the context of pedestrian trajectory prediction in the mobility domain, where controlled synthesis of diverse motion patterns is critical. Additionally, the work examines existing metrics for evaluating disentanglement and identifies critical limitations in them. A novel metric, the Exclusivity Disentanglement Index (EDI), is proposed as an improved standardised measure. Based on the principle of exclusivity in factor-code relationships, this metric offers advantages over existing alternatives in terms of efficiency and robustness. By advancing data-efficient learning and controllable generation strategies, this thesis aims to bridge the gap between AI's vast potential and its practical adoption in resource-constrained environments. These contributions pave the way for transformative applications in planetary science, healthcare, and beyond, where label and data scarcity have long been barriers to progress.}, language = {en} }