Convergence rates of training deep neural networks via alternating minimization methods.