Steve Lehar slehar at park.bu.edu
Wed Aug 7 23:56:39 EST 1991

> Do you  have    any references on  a  functional   significance  for
> backprojections   in    the    visual    cortex?   I    believe   to
> neurophysiologists there  is still a  great deal  of   mystery as to
> their functional significance.  I am not familiar with the Grossberg
> model,  but   do  you  have any  references  from   the neuroscience
> literature?

There are two models relating to this issue, one is by Gail Carpenter,
who together with Grossberg,  worked out the Adaptive Resonance Theory
(ART) [1].  This is an abstract computational model to demonstrate the
principle of resonant feedback in neural systems.  You see,  there are
plenty of neural models that illustrate how a pattern of activation in
a neural layer can be transformed  to a different representation in an
adjacent layer by synaptic connections with the appropriate pattern of
synaptic weights.  Many such schemes suggest transformations  from low
level raw data to higher layer more  abstract representations, where a
single cell in the  higher  layer might represent a particular pattern
of activation in the lower layer.

One thing  that has  plagued  this  kind of model is  that  real world
sensory patterns can vary a lot, and are  often contaminated  by noise
and ambiguities, resulting  in unreliable  performance.  What the  ART
model shows  is  how resonance  between  the   higher and lower  level
representations will enhance those features that are in common between
them, and suppress  those that are  different.  For instance,  say you
had a single cell in the higher level  that responded to a  pattern in
the lower layer like the one on the left.  If it were given  a pattern

	 XXX			 XXX
	X   X			X   XX
	X   X			    X
	X   X			X   X

like the one on the right, it would fire weakly.  If no OTHER top-down
cell fired any better  to this pattern, then  the top-down connections
from this  cell would suppress  the  extraneous parts of  the  pattern
while boosting the  missing parts,  eventually restoring the  original
pattern, which in turn would  fully stimulate the  higher  level cell.
So when a  pattern  is  finally  recognized,  it  is as a result  of a
resonant   matching  between the noisy and    imperfect input and   an
idealized  model  at the higher  level.   This  kind of mechanism also
illustrates top-down priming, how you can be made  more sensitive to a
particular  low level pattern by  priming at  the high level, and how,
with enough priming  you can  be  hyptnotized into hallucinating  that
pattern  even if it  is not there.   The  "adaptive" part of the model
shows how new patterns can be learned by  letting the synapses grow by
Hebbian-like learning during the resonant matching phase.

This is a very high-level or abstract model, but the same principle of
bottom-up and top-down matching is illustrated by the Boundary Contour
System /  Feature Contour   System  model of  Grossberg  and  Mingolla
[2],[3],[4],  which models the  visual system specifically.  Similarly
to ART, low level  edge  detectors detect local   edges, while  higher
level  detectors detect the  joint  firing of  low  level detectors if
their pattern of firing suggests a  continuous larger edge.   If there
were a gap in the original edge, the high  level recognition would try
to  close the gap by connecting  the line across  it.  There is strong
psychophysical evidence for this kind of  operation in natural vision,
and this model has been able to reproduce  and predict a wide range of
visual phenomena.

> Why is it inconsistent to have 2 lines of different orientations at a 
> point?  The visual scene is filled with such examples.

I am talking about  edge detectors  at  the very limit  of resolution-
i.e. the smallest edges you can make out.  Now although the  middle of
this "+" symbol has a vertical and a horizontal line, in order for you
to  see this shape,  your visual  system must  be able  to discern the
boundaries of the  entire   form,  i.e. your oriented detectors   must
perceive a pattern something like this...

	     |  |
	     /  \
	----     ----
	____     ____
	    \    /
	     |  |

at the center  of the "+", i.e.  each  line is made  up of a  pair  of
edges, one light-to-dark, and the other dark-to-light.  At this scale,
there are no points in the image without a specific orientation.

I have a text file  describing Grossberg's vision model  , the BCS, in
an informal and intuitive  way, I would   be happy to send a   copy by
email on request.

[1] Carpenter,  Gail & Grossberg,   Stephen.    A  MASSIVELY  PARALLEL
Computer  Vision, Graphics,  and  Image Processing (1987),  37, 54-115
Academic Press, Inc.

[2]  Grossberg, Stephen   &    Mingolla,  Ennio.  NEURAL  DYNAMICS  OF
Perception & Psychophysics (1985), 38 (2), 141-171.

[3] Grossberg, Stephen & Mingolla, Ennio. NEURAL  DYNAMICS  OF SURFACE
Computer Vision, Graphics and Image Processing (1987) 37, 116-165.

[4] Grossberg, Stephen & Todorovic, Dejan.  NEURAL DYNAMICS OF 1-D AND
2-D BRIGHTNESS  PERCEPTION Perception  and  Psychophysics  (1988)  43,
(O)((O))(((               slehar at park.bu.edu               )))((O))(O)
(O)((O))(((    Steve Lehar Boston University Boston MA     )))((O))(O)
(O)((O))(((    (617) 424-7035 (H)   (617) 353-6741 (W)     )))((O))(O)

More information about the Neur-sci mailing list

Send comments to us at biosci-help [At] net.bio.net