Few-Shot Learning onder distributieverschuivingen

Few-shot learning (FSL) enables artificial intelligence (AI)-based systems to adapt rapidly to novel test conditions with minimal supervision, in contrast to most contemporary AI methods, which typically encourage retraining (fine-tuning) on large-scale datasets. Many real-world applications are inherently dynamic, with test conditions frequently deviating from the training distribution, leading to distributional shifts. In such contexts, FSL provides a viable approach for adaptation in data-scarce domains, including small-scale industrial applications and healthcare. Despite the theoretical appeal and consistent progress on benchmarks, FSL has yet to achieve widespread practical adoption, largely due to limited understanding of its true generalization capability across diverse real-world conditions and concerns about the reliability and trustworthiness of predictive outcomes, particularly under distributional shifts. This dissertation investigates the generalization mechanics of FSL under non-i.i.d. (non-independent and identically distributed) conditions, focusing on the transition from standard benchmarks to resource-constrained, safety-critical deployments. By leveraging the support set structure to provide a targeted inductive bias, this research explores how FSL can facilitate effective generalization in out-of-distribution (OOD) conditions using minimal supervision. The research formalizes distributional shifts as domain-specific operational variances, designs resource-efficient models for low-latency, on-device inference—thereby addressing privacy concerns by local data preservation—and studies predictive behavior under various shifts by integrating formal uncertainty quantification to mitigate the risks of model overconfidence in dynamic environments. The first study (Chapter 3) examines the capability of FSL to generalize across variations in infrastructure caused by sensor noise, structural changes, and geographical diversity, for an automated railway monitoring system using 3D point clouds, by formalizing application-specific distribution shifts. Experiments validate that FSL maintains high performance under in-distribution shift; however, performance drops under (OOD) shifts, especially when encountering previously unseen infrastructure elements. Nonetheless, it outperforms standard fine-tuning, highlighting its potential for label-efficient adaptation in `specialized' industrial applications such as those characterized by data and resource scarcity. The second study (Chapter 4) designs FALCON, an FSL framework for boundary-precise medical image segmentation in data-scarce clinical scenarios by integrating FSL task-based learning with conventional fine-tuning. The framework leverages priors from natural images and transfers them to the medical domain through boundary-aware adversarial fine-tuning, enabling the use of unlabeled medical data as support examples to enhance salient feature representations. Experimental results demonstrate that this framework achieves improved boundary delineation as measured by Hausdorff distance (HD), compared to model variants using standard cross-entropy, Dice, or Hausdorff loss. Notably, the framework maintains clinical relevance by avoiding artifact-prone stochastic data augmentation while significantly reducing both annotation requirements and computational costs. The third study (Chapter 5) formalizes distributional shifts based on clinician-identified sources of variation in lung cancer detection. It analyzes whether FSL inherently handles uncertainty or requires integration with formal uncertainty estimation methods to achieve reliable predictive outcomes. Findings reveal that while FSL offers robust generalization, its integration into clinical systems warrants caution; specifically, that pairing it with explicit uncertainty measures is essential to provide clinicians with the transparency needed for trustworthy outcomes and clinically meaningful decision-making. Beyond the technical contributions, this research addresses knowledge valorization (Chapter 6) by emphasizing reproducibility, academic dissemination, and broader societal impact, particularly in the global south. A structured pathway for industry knowledge transfer is proposed, underpinned by research-backed financial projections and market analysis, aimed at bridging the gap between academic research and practical, real-world implementation.