Large-Scale Brain Area Functional Organization - Computational Explorations in Cognitive Neuroscience - page 217

RC
GS

representations that encode conjunctive information).
The most prominent idea of this type is that the syn-
chronous oscillation of features belonging to the same
object could encode binding information (e.g., Gray,
Engel, Konig, & Singer, 1992; Engel, Konig, Kreiter,
Schillen, & Singer, 1992; Zemel, Williams, & Mozer,
1995; Hummel & Biederman, 1992). However, the
available evidence does not establish that the observed
synchrony of firing is actually used for binding, instead
of being an epiphenomenon. This is important, because
such synchrony (which has been observed) is very likely
to be a natural consequence of simple activation prop-
agation in the spiking neurons of the brain (i.e., neu-
rons that are communicating with each other will tend
to drive each other to spike at roughly the same time).
Another problem with the synchrony-based binding
idea is that it requires entirely new mechanisms for pro-
cessing the bound information. Because feature bind-
ings are transient and dynamic in these systems, any
further processing of the bound representations would
also have to rely on dynamic mechanisms — standard
weight-based detectors and transformations would not
work. For example, if there is some unique conse-
quence or set of associations for red circles that does not
apply to green or blue circles, how can this information
become associated with a representation that only exists
as a relatively fleeting temporal synchronization? This
problem does not arise when a unique pattern of activity
across a set of dedicated representational units is used
(as with the combination of conjunctive representations
scheme described earlier).
Finally, it is essential to realize that in many cases,
people fail to solve the binding problem successfully,
and such failures can provide important clues as to the
underlying representations involved. For example, it is
well known that searching for some combinations of vi-
sual features requires slow, serial-like processing, while
other combinations can be found with fast, parallel-like
speed (Treisman & Gelade, 1980). The simple inter-
pretation of this phenomenon is that people sequentially
restrict their object processing to one object at a time us-
ing spatial attention, so that there is no possibility of a
binding problem between the features of multiple ob-
jects. The details of the visual search process are more
complicated than this simple story, but the basic idea

obj1 obj2 R G B S C T BT

S C 1101100

C S 1101101

S T 1101010

T S 1101011

S C 1011100

C S 1011101

S T 1011011

T S 1011010

C T 1100111

T C 1100110

C T 1010111

T C 1010110

S C 0111101

C S 0111100

S T 0111011

T S 0111010

C T 0110111

T C 0110110

Tab le 7 . 1 : Solution to the binding problem by using rep-
resentations that encode combinations of input features (i.e.,
color and shape), but achieve greater efficiency by represent-
ing multiple such combinations. Obj1 and obj2 show the fea-
tures of the two objects. The first six columns show the re-
sponses of a set of representations that encode the separate
color and shape features: R = Red, G = Green, B = Blue, S
= Square, C = Circle, T = Triangle. Using only these sep-
arate features causes the binding problem: observe that the
two configurations in each pair are equivalent according to the
separate feature representation. The final unit encodes a com-
bination of the three different conjunctions shown at the top of
the column, and this is enough to disambiguate the otherwise
equivalent representations.

units would be required, but only 11 are needed with the
features plus combinations-of-conjunctions scheme de-
scribed here (8 feature units and 3 conjunctive units).
We will explore more complex conjunctive binding
representations that develop naturally from learning in
chapter 8.
Another proposed solution to the binding problem,
mentioned previously, uses processing dynamics to es-
tablish bindings between different feature elements on
the fly (i.e., without using dedicated, content-specific