Jamieson, KevinCamilleri, Romain2024-09-092024-09-092024-09-092024Camilleri_washington_0250E_26914.pdfhttps://hdl.handle.net/1773/51878Thesis (Ph.D.)--University of Washington, 2024In an era characterized by the rapid evolution and widespread adoption of machine learning, the imperative to balance innovation with responsible deployment has never been more pressing. This thesis embarks on a multifaceted exploration of efficient and interactive machine learning deployment, with a focus on advancing methodologies for the development of trustworthy and reliable models. Beginning with an in-depth analysis of high-dimensional experimental design and kernel bandits, it investigates the nuances of modeling smooth reward functions in Reproducing Kernel Hilbert Spaces (RKHS) and introduces novel algorithms for regret minimization and pure exploration tasks. This exploration extends to level set estimation in kernel bandits, shedding light on nearly optimal algorithms and their performance characteristics. Building on these foundational principles, the thesis progresses to examine selective sampling for online best-arm identification, enabling to leverage unlabeled and labeled data optimally. Further, it delves into the challenge of non-stationary linear bandits, emphasizing robustness in dynamic environments. Additionally, the thesis explores integrating safety and fairness constraints into the active learning paradigm, empowering the alignment of model development with ethical considerations and real-world constraints. This effort enhances the reliability of machine learning models and contributes to the advancement of responsible AI. In conclusion, this thesis offers a comprehensive exploration of efficient interactive learning deployment, presenting novel insights, methodologies, algorithms, and perspectives to foster responsible innovation in the field.application/pdfen-USnoneActive LearningArtificial IntelligenceBanditsFairnessMachine LearningRobustnessComputer scienceStatisticsComputer science and engineeringThesis