Is it safe to say that in ML, use of NNs is more about writing code that designs NNs, evaluates the results, and modifies the designs to optimize some desired meta-values, like accuracy, efficiency, etc?