In the early stages of visual processing, objects and
scenes are represented by neurons with small visual receptive
fields, which only "see" a small part of the retinal image.
Each neuron can therefore provide information about local
features of a scene, but to describe a scene in terms
of objects requires that these features be combined. Objects
can cover wide areas of visual space and be partially
occluded by other objects, so the problem of binding the
separate representations of object parts into coherent
wholes is not a simple one. The computations involved
in solving the binding problem probably take place in
parallel at multiple levels of visual processing, and
the problem may only be fully "solved" at a high level
in the hierarchy of cortical visual areas.
Some theorists have advanced the view that binding is
a special problem and requires a special solution because
it seems necessary to "tag" each visual neuron to signify
the object to which its activity relates. Each neuron
therefore has to carry two distinct signals, one that
indicates how effective a stimulus is falling on its receptive
field, and a second that tags it as a member of a particular
cell assembly. To make these signals distinct, von der
Malsburg proposed that the "effectiveness" signal would
be carried by a conventional code based simply on neural
firing rates, while the "tag" signal would be created
by synchronizing the spike activity of the neuron with
spikes from other neurons in the same assembly. This theory
has been elaborated by Singer and others, and supported
by a variety of neurophysiological evidence that seems
to show that neurons in the visual pathway, even at rather
low levels of processing, tend to synchronize their firing
under stimulus conditions that might favor perceptual
binding.
First, I consider whether the theory is an a priori
reasonable approach to solving the binding problem, and
conclude that it is at best incomplete. Next, I ask whether
spike synchrony can plausibly be used as an informational
code, and conclude that encoding and decoding information
in this way would be very difficult in the cerebral cortex
because of the rich and massively parallel nature of the
synaptic connections between neurons. I examine the experimental
evidence adduced to support the synchrony hypothesis,
and conclude that the evidence is largely indirect and
has no proven relevance to the issue of binding per
se. I next ask whether the binding problem is truly
of unique difficulty and requires a unique solution, or
is simply one of a number of "hard" problems in perception
that have so far eluded our understanding. I finish by
considering some strategies for solving the binding problem
that do not require the creation of a special neural code.
