A machine learning method is introduced here to improve the accuracy of brain registration. Generally, different brain regions might need different types or sets of features for registration, which actually can be determined and learned from the brain samples by a machine learning method. In this paper, we focus on investigating the best geometric features required by different brain regions, to match the correspondences and manage the registration procedure hierarchically. Compared to other conventional registration methods where no learning method is employed, our learning-based registration method is able to produce not only more consistent registration on serial images of the same subject, but also more accurate registration on simulated dataset.