Artificial Neural Networks are Blackbox routines. Are you allowed to use them?

One of the most often heard complaints against the usage of Artificial Neural Networks is the circumstance of being a blackbox routine and not being physically motivated. This means that one cannot understand the dependences of the input and output parameters only by formal observation of the model.

In the following only some of the most relevant arguments against that complaint are listed here.

Neural Networks are not only a regression function but rather a method for knowledge discovery

People reducing Artificial Neural Networks to the resulting implementation model miss the main part of the story. First of all Neural Networks are - like all the other Machine Learning procedures as well - a method for identification and quantification of the parameters effecting the systems behaviour. That way they are - contrary to the stated complaint - a methods to retrieve knowledge about behaviour and dependences of a given complex system.

The resulting functions are by-products, which can be but do not need to be used for further implementation. If it turns out that the given problem is not that complex as initially thought, it is not really necessary to use an Artificial Neural Network for implementation purpose and nothing can be said against going for a compact analytical or rule based model. But using Nerual Networks (or other general Machine Learning models) is a Top-Down approach with reduced risc in running into extensive trial and error scenarios. 

A trained Neural Network is simple and deterministic

Once a Artificial Neural Network is trained and not adapted anymore, it is a clear and simple deterministic function,with which all arbitrary tests and validations can be done, which are necessary to proof systems functionality and robustness.

"Analytical" and "rule based" approaches are not automatically whitebox

Vice versa very often the mistake is made, that analytical or rule based models are treated per se as whitebox models. Even one often can formally and theoratically investigate them into the depth of each branch. But practically they are often blackbox as well. A ruleset or an analytical equation with dozens or even hundrets of parameters (i.e. like it is often the case for classification or regression trees) may formally be a whitebox, but it is not automatically guaranteed that its behaviour and dependencies can be understood completely.

Futher it has to be remarked that a conglomerate of physically motivated components does not necessarily compose to an overall physical model. For complex systems the reductionists approach often tends to fail.

"Analytical" is not sufficient for having control

The circumstance of having an analytical model - and formally a whitebox model - does not allow to forget the homework.

"Analytical" means that sensitivity analysis can be done by building any kind of derivatives. But one really has to do this and draw the right conclusions, because highly nonlinear equations are not automatically understood hereby. Like chaos theory is showing, a system cannot be controlled only due to analytically describing the system with neat equations.

Ironically validation is often forgotten in the existence of analytical solutions, because it is thought to have everything under control with the formal whiteboxes.

The story is not only about Neural Networks, Machine Learning as a whole is the message

And last but not least the whole story is about Machine Learning at all and not only about Neural Networks. If one is training rulesets for example, the resulting implementation models are formally whitebox (just in case that the form is so much important).

Process Embeddment

Like for almost all flexible and powerful tools, the danger of abused and misleaded application is existent. For prevention the application is embedded into a proper process scheme, which is oriented on the right and conform formulation of the functional requirments including robustness as well as the validation and safeguarding of these. Validation and the prevention against specialized and overtuned, fragile solutions is immanent part of the development process.