Announcing the availability of a hyperplane animator

[pratt@cs.rutgers.edu]

                     -----------------------------------
				 Announcing 
			   the availability of an
		  X-based neural network hyperplane animator
                     -----------------------------------

			  Lori Pratt and Paul Hoeper
			    Computer Science Dept
			      Rutgers University


    Understanding neural network behavior is an important goal of many
    research efforts.  Although several projects have sought to translate
    neural network weights into symbolic representations, an alternative
    approach is to understand trained networks graphically.  Many
    researchers have used a display of hyperplanes defined by the weights
    in a single layer of a back-propagation neural network.  In contrast to
    some network visualization schemes, this approach shows both the
    training data and the network parameters that attempt to fit those
    data.  At NIPS 1990, Paul Munro presented a video which demonstrated
    the dynamics of hyperplanes as a network changes during learning.  This
    video was based on a program implemented for SGI workstations.

    At NIPS 1991, we presented an X-based hyperplane animator, similar
    in appearance to Paul Munro's, but with extensions to allow for
    interaction during training.  The user may speed up, slow down, or
    freeze animation, and set various other parameters.  Also, since it
    runs under X, this program should be more generally usable.

    This program is now being made available to the public domain.  The
    remainder of this message contains more details of the hyperplane
    animator and ftp information.

------------------------------------------------------------------------------

1. What is the Hyperplane Animator?

The Hyperplane Animator is a program that allows easy graphical display
of Back-Propagation training data and weights in a Back-Propagation neural 
network.

Back-Propagation neural networks consist of processing nodes interconnected 
by adjustable, or ``weighted'' connections.  Neural network learning consists 
of adjusting weights in response to a set of training data.  The weights 
w1,w2,...wn on the connections into any one node can be viewed as the 
coefficients in the equation of an (n-1)-dimensional plane.  Each non-input 
node in the neural net is thus associated with its own plane.  These hyperplanes
are graphically portrayed by the hyperplane animator.  On the same graph it
also shows the training data. 


2. Why use it?

As learning progresses and the weights in a neural net alter, hyperplane
positions move.  At the end of the training they are in positions that
roughly divide training data into partitions, each of which contains only
one class of data.  Observations of hyperplane movement can yield valuable 
insights into neural network learning.  

3. How to install the Animator.

Although we've successfully compiled and run the hyperplane animator on
several platforms, it is still not a stable program.    It also only
implements some of the functionality that we eventually hope to include.
In particular, it only animates hyperplanes representing input-to-hidden 
weights.  It does, however, allow the user to change some aspects of
hyperplane display (color, line width, aspects of point labels, speed of
movement, etc.), and allows the user to freeze hyperplane movement for
examination at any point during training.

How to install the hyperplane animator:

  1. copy the file animator.tar.Z to your machine via ftp as follows:

     ftp cs.rutgers.edu (128.6.25.2)
     Name: anonymous
     Password: (your ID)
     ftp> cd pub/hyperplane.animator
     ftp> binary
     ftp> get animator.tar.Z
     ftp> quit

  2. Uncompress animator.tar.Z

  3. Extract files from animator.tar with:
     tar -xvf animator.tar

  4. Read the README file there.  It includes instructions for running
     a number of demonstration networks that are included with this
     distribution.

DISCLAIMER:
  This software is distributed as shareware, and comes with no warantees 
whatsoever for the software itself or systems that include it.  The authors 
deny responsibility for errors, misstatements, or omissions that may or 
may not lead to injuries or loss of property.  This code may not be sold 
for profit, but may be distributed and copied free of charge as long as 
the credits window, copyright statement in the ha.c program, and this notice 
remain intact.

-------------------------------------------------------------------------------