Abstract
Transfer learning for nonparametric regression is considered. We first studythe non-asymptotic minimax risk for this problem and develop a novel estimatorcalled the confidence thresholding estimator, which is shown to achieve theminimax optimal risk up to a logarithmic factor. Our results demonstrate twounique phenomena in transfer learning: auto-smoothing and super-acceleration,which differentiate it from nonparametric regression in a traditional setting.We then propose a data-driven algorithm that adaptively achieves the minimaxrisk up to a logarithmic factor across a wide range of parameter spaces.Simulation studies are conducted to evaluate the numerical performance of theadaptive transfer learning algorithm, and a real-world example is provided todemonstrate the benefits of the proposed method.