Reviews: Learning Transferrable Representations for Unsupervised Domain Adaptation

Regarding the proposed model, below are the major concerns: 1. In Abstract and Introduction, the authors highlighted several times that transfer learning or domain adaptation aims to align the mismatch between the training and testing data distributions, such that good generalization can be obtained across domains or tasks. In the problem setup, the authors further explicitly state that \hat{x}_i and x_i follow different distributions p_s and p_t, respectively. However, different from some existing methods, like [19] and [Pan etal., Domain adaptation via transfer component analysis, IEEE TNN, 2011], the proposed model indeed does not explicitly minimize the distance or align the mismatch between the training and testing distributions. There is no guarantee that based on the new representation, the mismatch issue between distributions can be addressed.