Graham Smith's suggestion

[malsburg@neuroinformatik.ruhr-uni-bochum.de]

If I understand the text correctly, both on the input level and the
output level he has two cells for each of the four feature types, one
for each object.  I presume that after learning, there are no
connections in the system that confuse cells belonging to different
objects; there will be, for instance, no hidden unit to fire in
response to the combination ``red-a - square-b'' (if -a and -b stand
for the two object identities), and correspondingly the output could
not fire erroneously to this false conjunction.  I have actually
discussed this ``solution'' to the binding problem in my article ``Am
I thinking assemblies?'' (Proceedings of the Trieste Meeting on Brain
Theory, October 1984. G.Palm and A.Aertsen, eds. Springer: Berlin
Heidelberg (1986), {\it pp} 161--176).  I then talked about the ``box
solution'' (keeping things not to be confused with each other in
separate boxes with no confusing connections between them).  The main
problem with that ``solution'' is creating those boxes in the first
place (Graham Smith solved this by learning).  Sorting features into
boxes appropriately is another one (this problem is not solved in his
scheme at all, features being sorted already into the -a and -b boxes
in the input patterns).  A third problem is that rigid boxes keep
the system from generalizing appropriately.

The beauty of temporal binding is that a system can deal with a
feature combination even if it occurs for the first time.  For
instance, if a ``red'' unit and a ``square'' unit have been learned to
send activation to some output cell, and a red square occurs for the
first time in life, the two units correlate their signals in time and
summate on the output cell, whereas with a red triangle and a blue
square, the signals will not be synchronized and cannot summate on the
output.