There is no understanding of why NN training converges in reasonable time and does not over/underfit. (Unlike for say convex SVMs)