=============================================================================
	README file for the example files letters.xxx
=============================================================================


Description:	This network is a toy letter recognition network.
============	

This network is one of our favourite networks to display the SNNS user
interface, because all windows conveniently fit onto the screen. It is
NOT an example of a 'real world' letter recognition network. The
network input is are 5x7 binary input matrix. The network has 10
hidden units in one hidden layer which are fully connected to the
input and to the output units. The 26 output units each represent one
captial letter and show an output of 1 if the input pattern is of the
proper class, else 0.


Pattern-Files:	letters.pat
==============	

The pattern-file letters.pat contains 26 training patterns (one
exemplar of each capital letter). The patterns here have binary values
of 0 and 1 but SNNS treats all inputs and outputs as real valued.

Because each pattern is given only once and there are no noisy
patterns this pattern file cannot be used for generalization.


Network-Files:	letters.net
==============	letters3D.net

Both networks contain trained network files with the same topology.
	35 input neurons
	10 hidden neurons
	26 output neurons
They differ only in their assignment of neurons to SNNS display layers
and the use of a 2D or 3D display in the configuration file. The
first network letters.net uses one 2D display only, letters3D.net
several 3D displays and a 3D display.


Config-Files:	letters.cfg
=============	letters3D.cfg
 
The configuration file letters.cfg uses one 2D display only,
letters3D.cfg several 3D displays and a 3D display.


Topology: 35 Input-Neurons
          10 Hidden-Neurons
          26 Output-Neurons


Hints:
======

The following table shows some learning functions one can use to train
the network. In addition, it shows the learning parameters and the
number of cycles needed to train the network successfully.

These parameters have not been obtained with extensive studies of
statistical significance. They are given as hints to start your own
training sessions, but should not be cited as optimal or used in
comparisons of learning procedures or network simulators.


Learning-Function               Learning-Parameters       Cycles

Std.-Backpropagation            2.0                      150
Backpropagation with Momentum   0.8  0.6   0.1           100
Quickprop                       0.2  1.75  0.0001         50  
Rprop                           0.2                       50


=============================================================================
	End of README file
=============================================================================