On the Correctness of Automatic Differentiation for Neural Networks with Machine-Representable Parameters