Training Linear Neural Networks: Non-Local Convergence and Complexity Results