Cross-domain few-shot learning: Theories, methods, and the source-free paradigm
Thesis event information
Date and time of the thesis defence
Place of the thesis defence
Auditorium 101A Palotiesali
Topic of the dissertation
Cross-domain few-shot learning: Theories, methods, and the source-free paradigm
Doctoral candidate
Master of Science Huali Xu
Faculty and unit
University of Oulu Graduate School, Faculty of Information Technology and Electrical Engineering, Center for Machine Vision and Signal Analysis (CMVS)
Subject of study
Computer Science and Engineering
Opponent
Professor Moncef Gabbouj, Tampere University
Custos
Emeritus Professor Olli SilvΓ©n, Center for Machine Vision and Signal Analysis (CMVS)
Cross-domain few-shot learning: Theories, methods, and the source-free paradigm
Machine learning degrades under distribution shifts, especially with scarce labels. Cross-domain few-shot learning (CDFSL) aims to generalize to novel classes in unseen domains. This dissertation develops a unified theoretical framework and new recognition methods for CDFSL.
We extend Empirical Risk Minimization to a two-stage optimization that separates intra- and inter-domain challenges, and organize existing methods into four families: π-Extension (instance-level augmentation), Ξ-Adaptation (distribution alignment), π-Constraint (hypothesis-space restriction), and hybrids.
For π-Extension, we propose Inter-Source Stylization Network (ISSNet), which enriches target instances with styles drawn from unlabeled auxiliary domains, improving generalization via style-driven augmentation.
For Ξ-Adaptation, we define source-free CDFSL (SF-CDFSL), where source data are inaccessible, and introduce Information-Maximized Distance-aware Contrastive Learning (IM-DCL) to align a pretrained source model with unlabeled target data using distance-aware contrastive learning and information maximization.
For π-Constraint, we present Step-wise Style Prompt Tuning (StepSPT), which stabilizes learning by constraining the hypothesis space with learnable prompts and step-wise alignment, enabling efficient use of large pretrained models under low-data, source-free conditions.
Overall, the work provides a unified perspective and three complementary methods that enhance robustness and generalization in data-constrained cross-domain scenarios, supporting applications in medical imaging, autonomous systems, and remote sensing.
We extend Empirical Risk Minimization to a two-stage optimization that separates intra- and inter-domain challenges, and organize existing methods into four families: π-Extension (instance-level augmentation), Ξ-Adaptation (distribution alignment), π-Constraint (hypothesis-space restriction), and hybrids.
For π-Extension, we propose Inter-Source Stylization Network (ISSNet), which enriches target instances with styles drawn from unlabeled auxiliary domains, improving generalization via style-driven augmentation.
For Ξ-Adaptation, we define source-free CDFSL (SF-CDFSL), where source data are inaccessible, and introduce Information-Maximized Distance-aware Contrastive Learning (IM-DCL) to align a pretrained source model with unlabeled target data using distance-aware contrastive learning and information maximization.
For π-Constraint, we present Step-wise Style Prompt Tuning (StepSPT), which stabilizes learning by constraining the hypothesis space with learnable prompts and step-wise alignment, enabling efficient use of large pretrained models under low-data, source-free conditions.
Overall, the work provides a unified perspective and three complementary methods that enhance robustness and generalization in data-constrained cross-domain scenarios, supporting applications in medical imaging, autonomous systems, and remote sensing.
Created 15.10.2025 | Updated 20.10.2025