Proefschrift
Open-world Continual Learning via Knowledge Transfer
This thesis investigates Open-world Continual Learning (OWCL), a learning paradigm designed for intelligent systems operating in non-stationary environments with persistent exposure to unknown data.
- Auteur
- Y. Li
- Datum
- 27 januari 2026
- Links
- Thesis in Leiden Repository
OWCL goes beyond a simple integration of continual learning and open-set recognition by explicitly addressing the joint challenges of lifelong adaptation, unknown class discovery, and knowledge preservation under evolving task distributions. The thesis first establishes a mathematical formulation of OWCL and identifies representative real-world scenarios that characterize its openness and non-stationarity. Adopting a knowledge transfer-centric perspective, the thesis proposes three novel models to address complementary aspects of OWCL. Pro-KT introduces a prompt-based mechanism for transferring task-specific and task-agnostic knowledge across tasks, enabling balanced continual adaptation. PEAK extends OWCL to a few-shot setting, jointly addressing limited supervision and ambiguous task boundaries while maintaining robust open-set detection. To unify theory and practice, the HoliTrans framework integrates open-set recognition and continual learning within a shared decision space, providing both strong empirical performance and theoretical guarantees on boundary stability and knowledge retention.Finally, the thesis validates OWCL in a real-world cross-regional fraud detection application, demonstrating its practical relevance and scalability. Overall, this work establishes OWCL as a distinct and essential paradigm for developing adaptive, robust, and lifelong artificial intelligence systems.