Large-Scale Brain Area Functional Organization - Computational Explorations in Cognitive Neuroscience - page 205

ple units/areas participate in representing a given thing,
and that each unit/area represents multiple things (cf.
section 3.3.2). This similarity across the fine-grained
properties within an area and the large-scale properties
across areas again suggests that the brain has a fractal
quality, where the small-scale structure is replicated at
the larger scale.
This widely distributed view of knowledge con-
flicts with both popular intuitions and the computer
metaphor, which tend to favor the idea that there is a
single canonical representation of a given knowledge
item, with all of the features and attributes stored in
one convenient location. This bias toward assuming a
canonical representation leads to some of the problems
discussed in the last section of this chapter, where we
show how a more distributed model avoids these prob-
lems.

ture of the transformations also implies a certain amount
of stability of over time. This stability enables a rich set
of content-specific associations to be built up over time
in the connectivity among different representations.
Compare this situation with that of a standard se-
rial computer, where programs (processing) and data
(knowledge) are explicitly separated, and processing
typically operates generically on whatever data it is
passed (e.g., as arguments to a function). The advan-
tage of such a system is that it is relatively concise and
flexible , so that a given function need only be written
once, and deployed in a wide range of different situa-
tions. The ability to do arbitrary variable binding (e.g.,
by passing arguments to a function) is an important con-
tributor to this flexibility. (We will return to this later.)
Although there are obvious advantages to flexibility and
generality, one disadvantage is that it becomes difficult
to treat different stimuli in accordance with their spe-
cific properties and consequences — one has to resort
to messy sequences of if-then constructs and elaborate
representational structures that make clear exactly what
is different about one stimulus compared to others.
Thus, there is a basic tradeoff between specificity
and knowledge-dependency on one hand, and gener-
ality and flexibility on the other. It appears that the
brain has opted to optimize the former at the expense of
the latter. This is interesting, given the consensus view
that it is precisely the inability to deal with this type
of content-specific real world knowledge that led to
the failure of traditional symbolic (computer metaphor
based) models of human cognition (e.g., Lenat, 1995).
Thus, it seems that getting all the details right (e.g.,
knowing the practical differences between tigers and
trees) is much more important for surviving in the world
than having the kind of flexibility provided by arbitrary
variable binding. One compelling demonstration of this
point in the domain of language comprehension is made
by these two sentences:

Dedicated, Content-Specific Processing and
Representations

A basic property of neural computation is that knowl-
edge and processing are intimately interrelated. As the
preceding discussion makes clear, processing amounts
to performing transformations on activity patterns,
where these transformations are shaped by accumulated
experience (knowledge) through learning. An impor-
tant consequence of this interrelationship is that pro-
cessing is relatively dedicated and content specific for
particular types of activation patterns. In other words,
unlike a computer, the brain does not have a general
purpose CPU.
An advantage of the dedicated, content-specific na-
ture of neural processing is that the selection of trans-
formations to apply in a given situation is determined
directly by the specific stimulus (activation pattern) be-
ing processed. This makes it easy for the system to treat
different stimuli according to their specific properties
and consequences. As we saw, this content-specificity
becomes even more important as layers are integrated
into elaborate hierarchies of processing stages, because
subsequent stages can then come to depend on particu-
lar types of transformations from their input layers, and
can in turn reliably provide specific transformations for
subsequent layers. The dedicated, content-specific na-

Time flies like an arrow.
Fruit flies like a banana.

Clearly, specific, real-world knowledge is necessary to
produce the two very different interpretations of the
words flies and like in these sentences. However, we
will see that the sacrifice of flexibility in favor of speci-