On the Stability of the Jacobian Matrix in Deep Neural Networks