Shapiro, LindaShic, FrederickLi, Beibin2022-04-192022-04-192022-04-192022Li_washington_0250E_23952.pdfhttp://hdl.handle.net/1773/48476Thesis (Ph.D.)--University of Washington, 2022Many machine learning (ML) models are trained on specific datasets for specific tasks. While traditional transfer learning can adapt to new datasets when labeled data are adequate, adapting to small datasets is still a challenging task. Researchers have applied multi-task learning, meta-learning, weakly-supervised learning, self-supervision, generative adversarial training, and active learning for various data adaptation applications. However, a unified data adaptation framework has yet to be developed. This study proposes a unified framework that can adapt to small datasets in a dynamic environment. Our framework, with a versatile encoder and various decoders, can simultaneously learn from source datasets and estimate confidence for novel data samples. We apply the framework to real-world medical imaging, affective computing, eye-tracking analysis, and database management applications.application/pdfen-USCC BYadaptationcomputer visiondata shiftdeep learningmachine learningneural networkComputer scienceComputer engineeringComputer science and engineeringThesis