TODO for libneural
------------------

TODO before the next release (+ essential for 1.1, - possible, likely for 2.0)

+ more examples (XOR problem, DSP examples, prediction, other recognition
  stuff);
+ try some different versions of the generalised delta rule algorithm for 
  weight updates - cumulative delta rule looks useful for prediction;
+ Eliminate the neuron and nnlayer classes, replace with faster data
  structures (arrays) in the main nnwork class.
- add optional bias term (initial test indicates this will have a negligible
  positive effect, and may actually slow the training process down);
- add optional momentum term (unless this is done well, it will also slow 
  things down!);
- add alternative transfer functions for hidden and/or output layers - sin,
  tanh, linear;
- allow arbitrary number of hidden layers (each with different numbers
  of elements? Is there any point?); and
+ finish my thesis... ;)
 
TODO in the future
 
- start looking at other network structures (other than BPN); and
- implement hidden-layer node pruning algorithms.