brain12.doc

本例是一个神经网络仿真应用程序

	1     1111    1111
	1        1       1
	1     1111     111
	1     1          1
	1     1111    1111

      These numerals can be represented using 0's and 1's as:

      0010    1111    1111
      0010    0001    0001
      0010    1111    0111
      0010    1000    0001
      0010    1111    1111

      Note: For illustrative purposes we are using whole numbers, '0' and
      '1'.  Keep in mind that in the input file these numbers must be
      presented as decimal values ('0.0' and '1.0').

      As each of the numerals consist of 20 0's and 1's, a network is
      constructed that consisted of 20 inputs.  To represent a 1 the
      following is presented to the network:

	  00100010001000100010

      This is taken from the above representation on the numeral one.

      Similarly for a 2:

	  11110001111110001111

      and for a 3:

	  11110001011100011111.

      For the output layer we use 3 units.  If the network has recognized
      the numeral one, the first output unit's result should be one (or
      close to one) and the second and third output units should be zero
      or close to zero.  If the network has recognized the numeral two, the
      second output node should be one or close to one (i.e.  'on') and the
      first and third units should be zero or close to zero (i.e.  'off').
      Similarly, if the numeral three is recognized the first and second
      output units should be 'off' and the third output unit 'on'.



      The Brain v1.2 - User's Manual                               Page 15
      ____________________________________________________________________

      From experimentation we've found that the minimum number of hidden
      units needed to correctly learn the above numerals was three.  The
      input file (available in the distribution package as 'char123.net')
      needed to learn the three numerals (1, 2, and 3) is:

      1                 First input node is # 1. -|
      20                Last input node is # 20. -|-->(i.e. 20 input units).
      21                First hidden unit.   -|
      23                Last hidden unit.    -|--> (i.e. 3 hidden units).
      24                First input unit. -|
      26                Last input unit.  -|--> (i.e. 3 output units).
      3                 Number of training examples.
      train             Indicates training the network (rather than testing).
      0.0 0.0 1.0 0.0   -|
      0.0 0.0 1.0 0.0    |
      0.0 0.0 1.0 0.0    |--> The input pattern for the numeral one.
      0.0 0.0 1.0 0.0    |
      0.0 0.0 1.0 0.0   -|

      1.0 1.0 1.0 1.0   -|
      0.0 0.0 0.0 1.0    |
      1.0 1.0 1.0 1.0    |--> The input pattern for the numeral two.
      1.0 0.0 0.0 0.0    |
      1.0 1.0 1.0 1.0   -|

      1.0 1.0 1.0 1.0   -|
      0.0 0.0 0.0 1.0    |
      0.0 1.0 1.0 1.0    |--> The input pattern for the numeral three.
      0.0 0.0 0.0 1.0    |
      1.0 1.0 1.0 1.0   -|

      1.0 0.0 0.0       The output needed when the numeral one is presented.
      0.0 1.0 0.0       The output needed when the numeral two is presented.
      0.0 0.0 1.0       The output needed when the numeral three is presented.

      Note: the comments on the right should not be included in the file.
      Each input pattern is presented on 5 lines to aid in visualization
      of each numeral; presentation of each pattern on a single line would
      make no difference.

      To run through the tutorial, enter 'start-me' at the DOS prompt.
      Enter 'char123.net' when the input file is requested.  Press 'n'
      when asked if you want to load a set of saved weights.  Learning of
      the training set will then start (i.e.  it is now learning how to
      distinguish between the numerals 1, 2, and 3).

      The network should take less than one thousand training cycles (i.e.
      less than one thousand presentations of the 3 input patterns) to
      learn the 3 patterns.  If it is taking longer it has most likely
      become stuck (as described in the section 'Training the Network') and
      you should stop the network (by pressing 'control-c') and start the
      training again by restarting the program.


      The Brain v1.2 - User's Manual                               Page 16
      ____________________________________________________________________



      After training has been completed you are given the option of saving
      the weights.  The output of the network is then displayed.  The
      output will look something similar to the following:

	  Training example 1
	  0.920013 0.012343 0.211126
	  Training example 2
	  0.000129 0.834562 0.123983
	  Training example 3
	  0.300403 0.203044 0.970030

      From the output we can see that for the first example the first
      output unit is close to being fully 'on' (i.e.  close to 1) while
      the other two output units are close to being fully 'off' (i.e.
      close to 0).  This indicates that the network has recognized that
      the first training example is, in fact, a numeral one.

      With the 2nd training example, the first and third output units are
      'off' while the 2nd output unit is 'on', indicating we have
      identified the numeral two.

      A similar output is noted for the third training example: the first
      and second output units are 'off' with the third being 'on', showing
      the network correctly identifying the numeral three.

      Once the network is trained, the weights can be saved and then used
      to test the network on unseen data (see "Testing the Network").  For
      this example we can test the network using different representations
      of the numerals 1, 2, and 3 (i.e. alter their shapes by moving,
      shrinking, or expanding them).

      Most likely you will notice that the performance on the test data is
      poor.  To get around this problem try expanding the training set by
      adding more examples of the numerals 1, 2, and 3, and by retraining and
      retesting the network.


      The Brain v1.2 - User's Manual                               Page 17
      ____________________________________________________________________


			General Texts on Neural Networks
			--------------------------------

      Hertz, J., Krogh, A., and Palmer, R. (1991).  Introduction to the
      Theory of Neural Computation.  Addison-Wesley: Redwood City,
      California.  ISBN 0-201-50395-6 (hardbound) and 0-201-51560-1
      (paperbound)

      Hecht-Nielsen, R. (1990).  Neurocomputing.  Addison Wesley.

      Aleksander, I. and Morton, H. (1990).  An Introduction to Neural
      Computing.  Chapman and Hall.  (ISBN 0-412-37780-2).

      Beale, R. and Jackson, T. (1990).  Neural Computing, an
      Introduction.  Adam Hilger, IOP Publishing Ltd : Bristol.  (ISBN
      0-85274-262-2).

      Dayhoff, J. E. (1990).  Neural Network Architectures: An
      Introduction.  Van Nostrand Reinhold: New York.

      McClelland, J. L. and Rumelhart, D. E. (1988).  Explorations in
      Parallel Distributed Processing: Computational Models of Cognition
      and Perception (software manual).  The MIT Press.

      McCord Nelson, M. and Illingworth, W.T.  (1990).  A Practical Guide
      to Neural Nets.  Addison-Wesley Publishing Company, Inc.  (ISBN
      0-201-52376-0).

      Orchard, G.A.  & Phillips, W.A.  (1991).  Neural Computation: A
      Beginner's Guide.  Lawrence Earlbaum Associates: London.

      Wasserman, P. D. (1989).  Neural Computing: Theory & Practice.  Van
      Nostrand Reinhold: New York.  (ISBN 0-442-20743-3)


      The Brain v1.2 - User's Manual                               Page 18
      ____________________________________________________________________


			       License Agreement
			       -----------------

      You may use the distribution version of The Brain for evaluation
      purposes only.  Once you start using this program regularly for
      educational, commercial, or private use you must register this
      product (see the file 'register.doc').  You may share this
      distribution version with anyone you choose so long as it is
      unaltered, and so long as you follow the distribution policy
      outlined in the file 'vendor.doc'.

      You are not permitted to share or otherwise distribute, in whole or
      in part, the registered version.  By registering this product you
      acknowledge that this product represents a trade secret and agree to
      protect it.  Misuse of a registered version is subject to collection
      of 100 times the registration fee and all legal fees and costs.
      Licenses are not transferable and may not be modified.

      DP COMPUTING DISCLAIMS ALL WARRANTIES RELATING TO THIS PRODUCT,
      WHETHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY IMPLIED
      WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE,
      AND ALL SUCH WARRANTIES ARE EXPRESSLY AND SPECIFICALLY DISCLAIMED.
      NEITHER DP COMPUTING NOR ANYONE ELSE WHO HAS BEEN INVOLVED IN THE
      CREATION, PRODUCTION, DELIVERY, OR DISTRIBUTION OF THIS PRODUCT
      SHALL BE LIABLE FOR ANY INDIRECT, CONSEQUENTIAL, OR INCIDENTAL
      DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE SUCH SOFTWARE
      EVEN IF DP COMPUTING HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
      DAMAGES OR CLAIMS.  IN NO EVENT SHALL DP COMPUTING'S LIABILITY FOR
      ANY DAMAGES EVER EXCEED THE PRICE PAID FOR THE LICENSE TO USE THE
      SOFTWARE REGARDLESS OF THE FORM OF THE CLAIM.  THE PERSON USING THE
      SOFTWARE BEARS ALL RISKS AS TO THE QUALITY AND PERFORMANCE OF THE
      SOFTWARE.

      It is the users responsibility to determine whether the product will
      work reliably on their equipment and for their specific needs.
      That is the purpose of this evaluation version.  DP Computing does
      not imply in any manner that this software is suitable for any given
      application or purpose.

      If any bugs are found please let us know and return a copy of the
      product with the bugs in it and we will do our best to fix it.

      If you don't agree with these conditions, delete this product from
      your disks.


      The Brain v1.2 - User's Manual                               Page 19
      ____________________________________________________________________


				 Support Policy
				 --------------

      DP Computing is fully committed to providing the best possible
      support for our clients.  If you have any problems at all please
      feel free to contact DP Computing or one of our support sites
      (listed in the file 'support.doc').

      We will provide support for registered users for up to 3 months
      following the registration of the product.  This support is for
      correcting bugs in the software and manuals and does not include
      advice on how to solve various problems using the neural network.

      Questions and advice on "The Brain" and neural networks in
      general may be answered depending upon time constraints.  DP
      Computing is also available for contract work on the use of neural
      networks to solve specific problems.  Please call for details.

			      Distribution Policy
			      -------------------

      All DP Computing products are protected under Australian and
      International (c) copyright laws.  As a shareware distributor you
      have permission to distribute any of DP Computing's shareware
      products as long as:

	      - it is kept in its present electronic form.

	      - it is clearly identified as shareware.

	      - all copyright notices remain intact.

	      - no file in this distribution package is modified or
		deleted.

	      - we don't request you stop.

      You may archive the programs, unarchive them, use your own
      installation routines, include them with other programs on a disk,
      etc., so long as you follow the above rules.

      If you are a shareware distributor we would appreciate a copy of
      your catalog and the disk on which the program is placed.  In
      return, we will send you all updates to the program.  BBS owners are
      also urged to drop us a line so that we can keep you up to date with
      future releases.  Shareware distributors and bulletin board systems
      may be named in our documentation as distribution sites if so
      desired.



      The Brain v1.2 - User's Manual                               Page 20
      ____________________________________________________________________

      If you have any questions, complaints, or concerns please contact me:

      David Perkovic
      DP Computing
      PO Box 712
      Noarlunga Center SA 5162
      Australia

      Ph: [phone] (International)     
	     08 326 4364 (Within Australia)

      Internet: [e-mail]
		[e-mail]


      International:

      If you are a publisher interested in supporting or translating our
      products please contact us for further information.  Anyone
      interested in providing registration and/or distribution services
      outside of Australia please contact us at the above address.


      Site licenses available.




			       About This Manual
			       -----------------

      This manual was written by David Perkovic (author of The Brain) and
      laid out / edited by Eugene Mallay.

      Eugene Mallay is a freelance writer and editor specializing in
      handbooks and manuals.  He can be contacted at:

	  Internet:      [e-mail]                  Voice: [phone]
	  Surface Mail:  945 Midland Ave, Suite 1003     Fax:   [phone]
			 Scarborough, Ontario, Canada
			 M1K 4G5

      International clients are welcomed.