A1122
Title: Subsampling in transfer learning
Authors: Jing Wang - University of Connecticut (United States) [presenting]
Abstract: Transfer learning is an emerging field in recent years. Subsampling can be considered as a data selection method for transfer learning to obtain better performances. However, optimal subsampling with potential model misspecification has not been fully investigated, which limits the usage of subsampling algorithms in transfer learning. Subsampling algorithms are developed with potential mean shifts, which connects subsampling under misspecified models with data selection for transfer-learning algorithms. Theoretical analysis implies that the performances of transfer learning estimators are determined by model biases and variances. Therefore, two different subsampling strategies are proposed: one reduces model biases, and the other reduces variances due to subsampling. Two approaches are also proposed to combine the two sampling strategies to further improve the performances of transfer-learning estimators. Non-asymptotic bounds of the proposed estimators are proved. Numerical experiments justify the usage of the proposed transfer learning algorithms with data selection techniques.
