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


Description:	This feedforward network recognizes printed characters 
===========	of two different fonts. 

The characters are displayed in a matrix of 24x24 input units. Each
output neuron represents one character. There are a number of special
characters and punctuation symbols, all captial letters, all lowercase
letters and all digits for a total of 75 character classes.


Pattern-Files:	font.pat
==============

This file contains 150 input and output patterns for the 75 different
characters (2 characters of each class).


Network-Files:	font.net
==============

The network font.net contains a trained feedforward network for this
task. It consists of 24x24 = 576 input units, two groups of 4x6 hidden
units and 75 output units. The input layer is NOT fully connected with
the hidden layer, because this would yield too many weights and make
the network too large to be used as an example. Instead, only each
unit of the same input row is connected with a hidden unit of the
first group, for a total of 24 input rows (24 hidden units of the
first group). Also, each unit of the same input column is connected
with one hidden unit of the second group, for a total of 24 columns
(24 hidden units of the second group).

All hidden units are fully connected with all output units.


Config-Files:	font.cfg
=============

This example is best displayed with one large 2D network display for
the whole network (without unit numbers or activations) and one 2D
display for the output units with names (classes) and activation
values.


Result-Files: font1.res
=============

This file is an example for Result files. You can analyze them with the tool
analyze in the tools directory.


Hints:
======

This is NOT an example of a 'real world' neural network character
recognition network, but a toy example to be played with. The number
of characters is much too small for the network size.
However, it makes for a visually appealing demonstration of the simulator.

The following table shows some learning functions one may use to train
the network. In addition, it shows the learning-parameters and the
number of cycles we needed to train the network successfully. These
are not necessarily the best parameters. They are given as starting
points for own experiments and should not be cited in comparisons of
learning algorithms or network simulators.

Learning-Function               Learning-Parameters       Cycles

Backpropagation                 2.0                       100
Backpropagation with momentum   0.8  0.6  0.1              30
Quickprop                       0.1  2.0  0.0001           50 
Rprop                           0.6                        50

In this example Backpropagation with momentum seems to be the most
stable learning algorithm giving the least classification error on the
training set.


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