Well, it is good to want something. It gives you a goal. You can't always get what you want.
However, software does what it is designed to do, not what you might hope it would do. Tools like VPA are designed to work with symbolic objects, and only that. Other tools in MATLAB will not employ that limited precision. In fact, since most people have various toolboxes, but NOT all of them, it makes no sense for a toolbox to assume that any specific other toolbox will be installed. (There is at least one of exception to this rule, where certain combinations of toolboxes go together.) So the neural net toolbox would never have been written to assume that VPA exists on any given installation.
At best you can have single and double precision variables in MATLAB for general use, and not every tool even enables the use of single.
It seems that your goal is to limit the accuracy of a result, thus setting a tolerance on how well the fit is accomplished. That you should be able to do in the neural net tools.
