=============================================================================
	README file for the example files	art1_letters.xxx
=============================================================================


Description:	ART1 letters network
============

The ART1 letters network shows the self-organized classification of
input patterns by an ART1 network. The input patterns are identical
with the patterns of the 'letters' pattern file, except that they
consist only of input patterns. The input is a 5x7 binary input
matrix, each input representing a different captial letter of the
alphabet. Each input pattern exists only once, there is no noise in
the input.
The ART1 network as implemented in SNNS differs from the standard ART1
network in that it tries to implement the functionality of the reset
box not algorithmically, but in the form of additional reset neurons:
  The leftmost 5x7 layer is the input layer,
  the next 5x7 layer is the comparison layer (F1 layer)
  the next 5x7 layer ist the recognition layer (F2 layer)
  the remaining two layers (delay layer and reset layer) and the delay
  units d1, ... d3 on top are needed for proper synchronization of the
  reset component.
See the SNNS user manual for a more detailed description of the ART1
implementation in SNNS.


Pattern-Files:	art1_letters.pat
==============

The pattern-file letters.pat contains 26 binary input patterns with values
of 0 and 1 representing capital letters in a 5x7 input matrix.


Network-Files:	art1_letters.net
==============

This network file contains a trained ART1 network for the letter
classification task described above. The standard configuration file
for this network is letters.cfg (one 2D display only).

You may generate your own ART1 network with the BIGNET tool from the
Info-Panel of SNNS. This automatically generates all units and the
necessary connections.

Because the unit types and link structure are highly specialized in
ART1 we strongly urge you only to use this tool to generate ART1
networks in SNNS.


Config-Files:	art1_letters.cfg	(one 2D display only)
=============	art1_letters3D.cfg	(one 2D display, one 3D display)

The drawing of the 3D display is relatively slow for this network, so
you probably want to work only with the 2D display once you know how
the units are connected.
The 3D display is a nice example for a moderately complicated 3D
network layout of a non-homogeneous network.


Result-Files:	(none)
=============


Hints:
======

Read the chapter about ART1 in the SNNS manual very carefully!

Note that ART1 needs a special network initialization function 
	(REMOTE panel: OPTIONS select init function: ART1_Weights).
Note that there exist two different ART1 update functions: 
	(REMOTE panel: OPTIONS select update function: ART1_Synchronous
						or     ART1_Stable)
Note that ART1 needs a special learning function:
	(REMOTE panel: OPTIONS select learning function: ART1)
These should already be set when loading the example ART1 network.

Use a high vigilance parameter $\rho$ (e.g. $\rho$ = 0.9 or 0.95),
otherwise all examples will be grouped into only a few classes. The
small value in the figure 'Setting the ART1 learning parameter $\rho$
in the SNNS manual is misleading.

Note that several input patterns are proper subsets of other
patterns. It is interesting to watch how the 'smaller' pattern erodes
the bitmap of the larger pattern until the former is no longer similar
to the smaller pattern and is assigned a different neuron.

The assignment of input patterns to recognition layer neurons appears
counterintuitive at a first glance but can be explained by the above
erosion effect.

There exists additional documentation in form of the diploma
thesis of Kai-Uwe Herrmann (in German), available via anon. ftp
from our public ftp server ftp.informatik.uni-stuttgart.de under
/pub/SNNS/NN-papers-german/herrmann_kaiuwe_DA.ps.Z


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