Large-Scale Brain Area Functional Organization - Computational Explorations in Cognitive Neuroscience - page 203

Hierarchical Structure

We begin by considering the basic building block for
our model of the cognitive architecture, the neuron-
as-detector presented in chapter 2. Individual neurons
are viewed as having relatively stable representations
that detect some (difficult to define and complex) set of
conditions in their inputs. We saw in chapter 3 that a
layer of such detectors can perform a transformation
of the input patterns that emphasizes some distinctions
between patterns and collapses across or deemphasizes
others. In chapters 4-6 we saw how these transforma-
tions can be shaped by learning so that they both repre-
sent important structural or statistical properties of the
environment and enable tasks to be solved.
Cognition can be viewed as a hierarchical structure
(figure 7.1) of sequences ( layers ) of such transforma-
tions operating on sensory inputs and ultimately pro-
ducing motor outputs ( responses ), or just useful inter-
nal states that provide an interpretation of the environ-
ment, which can be important for subsequent behavior.
As discussed in chapter 4, the sensory input contains
a large quantity of low quality information, so it must
be highly processed before sensible responses or inter-
pretations can be made. For example, the sensory sig-
nals from viewing the same object in two different loca-
tions can have almost nothing directly in common (i.e.,
no overlapping activations), but it nevertheless makes
sense to interpret these signals as representing the same
object. Thus, one wants to transform the inputs to col-
lapse across differences in location, while preserving
distinctions between different objects. As we will see
in chapter 8, a hierarchical sequence of transformations
is necessary to achieve this kind of spatial invariance .
This same process of performing transformations that
emphasize some dimensions or aspects and collapse
across others operates at all levels of processing. For
example, the representations that underlie the meanings
of words (e.g., dog vs. cat, truth vs. fiction) emphasize
those features or properties that define the word, while
collapsing across irrelevant ones. An example of a rel-
evant feature for dogs and cats is physical size, but this
is irrelevant for truth and fiction, whereas the notion of
“reality” is central to truth and fiction, but doesn't affect
cats and dogs that much (witness Garfield and Snoopy).

Pathway 1 Pathway 2

Figure 7.1: Generic hierarchical system, showing two
specialized pathways or streams of processing, with inter-
pathway connectivity.

Specialized Pathways

An important correlate of a hierarchical structure is the
existence of specialized and somewhat distinct process-
ing pathways or streams , which are necessary because
each layer of processing in a hierarchy requires specific
types of transformations to be performed in the pre-
vious layer(s), to then accomplish their particular job.
Thus, subsequent layers build on the transformations
performed in previous ones. In addition, there are typi-
cally many other potential transformations of the input
that are irrelevant for a given transformation. Therefore,
it makes sense to group together all the relevant trans-
formations into one coherent stream.
To illustrate the process of building successive lev-
els of transformations upon each other, let's continue
with the example of visual object recognition. Consider
a layer of processing that transforms visual representa-
tions of digits into the appropriate categorical represen-
tations (e.g., the digit “8”). If this layer is situated in a
specialized pathway after a sequence of transformations
that produce a spatially invariant visual representation,
then it would be relatively simple for one layer of pro-
cessing to transform the spatially invariant pattern of ac-
tivity into a categorical representation of a digit (as we
saw in chapter 3). In contrast, if this transformation was
performed based on the raw sensory input, a huge num-
ber of redundant detectors would be required to process
images in all the different locations. A visual digit cat-
egory transformation does not require olfactory, audi-
tory, or somatosensory information, so it makes sense
to have specialized visual processing pathways that are
distinct from other modalities. In short, it makes sense