compilation/installation tools:
===============================
* provide Makefile instead of compile.sh script which is only provisory.
* compile.sh: libnnetfileio.a is not linked correctly when compiling test_file
              why?

documentation:
=============
* WRITE IT! ;o)
* provide UML diagrams.

nnet.hpp/nnet.cpp:
==================
* since we use lots of classes defined in the STL without really needing
  most of their functionalities, it might be better to introduce service
  classes especially written for our needs. Attention: this will have an
  impact not only on this library, but on the whole package!
* check the information hiding of the implemented classes: check which
  access and functionality methods do have to be public. Some might be
  declared as beeing private by declaring some other class as a friend
  (e.g. Synapse::GetWeightedVal() can be private if Neuron is
   a friend of Synapse).
  Attention: Don't forget that the classes Neuron, Synapse and NNet
  are written for generic use - independent of the netwok structure!
* exentually set the parameters for the activation function in order to have
  the standard sigmoid function 1/1+exp(-x), currently we have 2/1+exp(-x) - 1.
* allow only synaptical weights between 0 and 1?
* Back Propagation is a supervised learning process. Implement at least one
  unsupervised learning algorithm (Hebbian Learning?). Eventually
  reinforcement learning as well (nearly unsupervised, network is only "told"
  whether the output is correct or not).
* derive the neuron class in order to implement simple McCulloch and Pitts
  neurons (step function as activation function)?

nnetfileio.hpp/nnetfileio.cpp:
==============================
* functions SaveNNet/LoadNNet: change output file format to XML.

further development:
====================
* write a parser in order to interactively manage creation, training,
  iteration, saving and loading of neural networks.
* divide learning and iteration processes into threads, such that they
  may be executed "simultaneously" (ideally on a cluster or a multiprocessor
  computer).
  This would allow to implement the refractory period (temporary "sleep" mode
  after the neuron fired).
* implemet dynamic neural networks.

some theoretical thoughts about neural networks:
================================================
* what about complex signals? how should weights look like? how activation
  functions? how would back propagation be affected? how hebbian learning?
  should we use the polar notation of complex values?
