Would you see a lower rank subspace if the learned weights were just random vectors?