Transfer Learning via $\ell_1$ Regularization