Robust Sparse Transfer Learning for Image Classification

Abstract

Transfer learning aims to transfer the knowledge learned from the source to the target data. However, noise-corrupted target data may limit the transfer learning capability. Thus, removing the noise in the data is essential to improve transfer learning performance. This paper proposes robust sparse transfer learning (RSTL) to improve the robustness of transfer learning. The RSTL uses noise-removed target domain data for project learning, where the employed nuclear norm ensures that the clean data matrix and the coefficient matrix are low-rank. The L norm is also adopted to ensure the sparsity of the target domain noise. Further, a reconstructive term is used, which aims to learn a reconstruction coefficient matrix. Extensive experimental evaluations on four datasets verify the promising ability of the proposed method compared with the other methods.