Neural Network Training Using $\ell_1$-Regularization and Bi-fidelity Data