
Revision date:  
Jun 07, 1988



Thesis proposal


Wesley R. Elsberry
Master's candidate, CSE

Committee:

Karan Briggs, CSE  (Graduate Chairman)
Daniel Levine, Mathematics
Lynn Peterson, CSE


Preliminary outline


I.  Introduction

II.  Literature review

III.  Topic proposal

  a.  Topic description

  b.  Topic verification (implementation)

    i.  Application proposal

    ii.  Description

    iii.  Resources needed for accomplishment


************************************

I.  Introduction

  The field of artificial neural network research currently
suffers from several misapprehensions on the part of researchers.
First, communication continues to be sketchy and prone to
misunderstanding, as no clearcut definitions have been attached
to even the most commonly accepted terms and phrases that
comprise ANN jargon.  Researchers will ignore the
interdisciplinary nature of ANN research to promote or denigrate
ANN results in a specialized context.  Often this is done in such
a way that it is not clear that the comments or analysis are only
valid in the specialized context.  Second, the motivations for
research vary wildly, and thus criticisms of models or data often
are intiated on the basis of entirely different goal assumptions.
Finally, much criticism and infighting occurs not because of any
real research related causes, but because of politicking and the
quest for personal power or recognition.  While Kuhn [History of
Scientific Revolutions] may revel in the unfolding byplay, it is
a source of annoyance and an obstacle to good work for others
engaged in this research.  While these misapprehensions may not
be conscious in nature, that does not lessen the negative impact
of the misapprehensions.

  One misapprehension which remains particularly pervasive is the
idea that there exists one 'correct' model for artificial neural
networks.  The biological reality reflects a complex set of
systems which accomplish diverse functions.  No one has suggested
that all biological neural systems operate in the same manner.
Other, more easily apprehensible, biological systems reflect that
variation arises both in structures and mechanisms that perform
functional tasks.  Spiders, insects, fish, birds, and mammals
have all developed methods of flight, yet none are quite the
same.  Other examples can demonstrate that the same mechanism may
be coopted for more than one purpose.  Certainly the expectation
should be that biological neural systems follow this pattern, yet
the prevailing attitude in current ANN research denies this.

  Different models reflect variation in an approach to a single
function, or simply approaches to different functions.
Comparisons which should account for this feature often do not.

  Since various models will have features which make them
preferable for classes of problems, solving problems which can be
divided into subset problems may be best solved through
integration and coordination of differing ANN models.  This
approach is expected to prove more tractable and productive than
attempting to force a solution model to fit a specified problem
complex (or changing the problem specification to fit the model).


II.  Literature review

  Problem solving as McCulloch and Pitts envisioned it
       [from Levine 83]

  As Rosenblatt redefined it
       [from Levine 83 and Rosenblatt ??]

  What Hopfield says about Grossberg [this will be short]
       [from H-T 86]

  What Rumelhart and McClelland say about Hopfield
       [from PDP]

  What Rumelhart and McClelland say about Grossberg
       [from PDP]

  What Grossberg says about everybody else [stated as briefly as
    possible]
       [from Applied Optics article, 87 Cognitive Science
       article]

  Evidences for multi-model integration:

  PDP Ch. 26, p 541:  "A problem with the PDP models presented in
  this book is that they are too specialized, so concerned with
  solving the problem of the moment that they do not ask how the
  whole might fit together.  The various chapters present us with
  different versions of a single, homogeneous structure,
  perfectly well-suited for doing its task, but not sufficient,
  in my opinion, at doing the whole task.  One structure can't do
  the job:  There have to be several parts to the system that do
  different things, sometimes communicating with each other,
  sometimes not."  

  Of course, McClelland here means to have several variants of
  the PDP model performing the functions, and is not per se
  referring to a multi-model approach.  But the admission that a
  single instantiation of a model does not a solution make is
  very important.


III.  Topic proposal

  a.  Topic description
     Use the models of Hopfield, PDP, and Grossberg's ART in an
     integrated manner to solve a problem set that is a complex
     suite of problem classes.  The purpose here is not to
     develop a general tool for such problems, but to demonstrate
     the desirability and applicability of using an integrative
     approach to ANN problem solving.    

  b.  Topic verification (implementation)

    i.  Application proposal

       Possible project 1:  Cryptographic example.  Small problem
       that involves transposition, pattern recognition, and
       feature detection and extraction.  Models used as pre- and
       co- processors for problem-solving.


    ii.  Description
     The data set generated for presentation to the solution
     system may have complex interdependencies which the ANN
     would have to extract.

    iii.  Resources needed for accomplishment

       Computer:
         Available currently:
           Heathkit H-100, MS-DOS, 768K
           Heathkit H-158, MS-DOS (PC comp), 640K
           DEC PDP 11/23, RT-11, 256K

       Languages:  
        Available currently:
         Under MS-DOS:
           Turbo Pascal
           XLISP
           PD-Prolog
           Turbo C
           ECO-C88
           ICON
           MS-FORTRAN
           MASM

         Under RT-11:
           MACRO-11 (assembler)
           DIBOL
         



