Thomas Jelonek thomas at McRCIM.McGill.EDU
Thu Aug 15 03:40:29 EST 1991

In article <SLEHAR.91Aug10105625 at>, slehar at (Steve Lehar) writes:
> > In your  original posting   you implied   that  backprojections in the
> > visual cortex influenced the functionality of the computational models
> > you described.  However, my  impression  is that  in  the neuroscience
> > community there is very  little consensus regarding  the functionality
> > of feedback in the visual cortex.
> > ...
> > Was there any specific  physiological  findings that formed  the basis
> > for this  model? Note, I am  NOT saying that  feedback  between visual
> > areas is
> This  model  is based  primarily   on   visual  illusions and    other
> psychophysical    findings,    rather than    on   detailed   study of
> neuroanatomy, i.e. the  model is  confirmed by successful reproduction
> and prediction  of visual illusions.    The structure of the model  is
> however consistent with known neurophysiology, and there has even been
> some neurophysiological  evidence for some  kind of feedback,  in that
> illusory contours  (non-existant edges that everyone sees   in certain
> figures) have been linked to  signals in simple  cells.  How could the
> simple cell fire to a  nonexistant edge whose illusory existance could
> only be determined by a more global view of nearby edges?   This seems
> to indicate a feedback interaction between the local  simple cells and
> the more global complex or hypercomplex cells.

I assume you are referring to the work of Peterhans and von der Heydt who
found cells in area V2 (so can I assume that when you refer to feedback that 
you are invoking horizontal connections within V2 as opposed to
backprojections from a `higher' area?) of macaque that responded 
to `illusory contours'. However, I believe they also found that the 
response time for these neurons was so fast that they ruled out any iterative
process. How do reconcile this? I am not saying I agree with them -- a 
process that required 2 or 3 iterations might still be consistent with 
their findings. 

You last sentence is puzzling.  Why do you consider simple cells local and
complex cells global (note that the Hubel and Wiesel notion of hypercomplex
cells does not exist rather one can have simple cells and complex cells that
are either singlely or doublely end-stopped)?  In terms of classical receptive
field they are both local operators. And both of their responses can be modulated by a surround as shown by: Blakemore, Robson, Allman, Van Essen etc. 

> > 
> > I am confused.  Why are you talking about edge detectors when you want
> > to represent lines?  Why can I not talk about receptive fields at coarsely
> > quantized orientations that are selective for lines, i.e. inhibition on
> > both sides? I would represent a curve by its local tangent estimate.
> > Your `+' would be represented by two neurons each selective for lines
> > at different orientations firing simutaneously.  Also, when you are 
> Let me try this again.  If you look at my figure above, each "-", "/",
> "\" and "|"  represents a single receptive  field.   At every point of
> the image the choice is between "-\|/", and there  is  no point on the
> image  where a "X"  or "+" receptive field  would be  more appropriate
> than one of the "-\|/" field.
> What  you seem to be  talking about is a "+"  figure that  is so small
> that at the center crossing point it is covered  by a single receptive
> field.  This figure would be  so small that  it would be unresolvable.
> You are   trying to  resolve  a form  that  is smaller than  a  single
> receptive field!  The only appropriate response to such a figure would
> be...
>     |
>   --o--
>     |
> where   the  "o"  represents   ambiguous   orientation.   This  is  an
> appropriate representation  because  at the  very  center of the cross
> there  is no  locally   detectable   orientation- we only infer   such
> orientations   on  the basis of  the  nearby arms   of the cross, i.e.
> hyperacuity, and the central  point would be  just plain  black.   Now
> according to this model, the nearby arms would influence the ambiguous
> signal  to become a  "|" and a "-", and  if  each  of these signals is
> equally  strong,  then  a "+"   will  result  ("|"  AND    "-"),   but
> psychophysical evidence  shows that  these orientations  compete, such
> that any imbalance between  the orientations  will produce  a stronger
> local signal for one than the other.

I would argue that any representation of a `+' would require an explicite coding of the discontinuity in orientation between the 2 lines.  The 
mechanism to achieve this would be having 2 orientated cells firing 
simutaneously -- there is no ambiguity in orientation we simply require
multiple orientation labels at a point.


More information about the Neur-sci mailing list