Large-Scale Brain Area Functional Organization - Computational Explorations in Cognitive Neuroscience - page 204

to collect together all the categorical transformations of
visual stimuli into one overall visual form or object pro-
cessing stream. Indeed, it appears that this is just what
the cortex has done (Ungerleider & Mishkin, 1982).
These same types of specialization pressures operate
in many other types of hierarchical structures. For ex-
ample, the functions of a corporation are typically di-
vided into different specialized divisions (i.e., process-
ing streams), because the details of one division (e.g.,
manufacturing) typically have little bearing on the de-
tails of another (e.g., marketing). However, there is a
lot of interdependence within a division. For example,
all the different manufacturing processes must be tightly
coordinated. The same kinds of hierarchical dependen-
cies are also present. For example, a higher-level report
on the status of a given division is produced by summa-
rizing, categorizing, and aggregating (i.e., transform-
ing) the lower-level details at various levels of analysis.

kinesthetic

visual

phone

color

action
oriented

tactile

auditory

phonology

orthography

Figure 7.2: Illustration of how semantic information might
be distributed across a number of more specific processing
pathways, represented here by different sensory modalities.
Linguistic representations (orthography, phonology) are as-
sociated with corresponding distributed activation patterns
across semantics. Figure adapted from Allport, 1985.

Inter-Pathway Interactions

will from their own, making the whole hierarchical no-
tion somewhat inapplicable here (figure 7.1). One can
better conceive of these higher-level association areas
as a heterogenous collection of peers, which intercom-
municate (“associate”) and perform constraint satisfac-
tion processing.

A completely rigid and separate hierarchical structure is
not as effective and flexible as one that has many oppor-
tunities for communication between the different spe-
cialized processing streams at all levels (as many cor-
porations are now discovering). These connections at
lower levels can mutually constrain or inform process-
ing across different pathways to better deal with partial,
noisy, novel, or particularly complex stimuli. For exam-
ple, we will explore the idea in chapter 8 that the visual
form pathway interacts at many levels with the spatial
processing pathway, resulting in the important ability to
focus attention at various spatial scales depending on
where confusions arise in the visual form pathway. By
resolving these confusions at the level at which they oc-
cur (instead of waiting for things to go all the way to
the top of the hierarchy), the system can deal with them
more rapidly and at the appropriate level of detail.

Large-Scale Distributed Representation

The preceding principles converge on the idea that the
knowledge associated with a given item is distributed
widely across a number of different brain areas, some
of which are specialized processing pathways and some
of which are higher-level association areas. Figure 7.2
illustrates a version of this general idea proposed by
Allport (1985), where the representation of semantic
items is distributed across many different specialized
systems that together comprise the “semantic memory
system.” The neuropsychological and neuroimaging ev-
idence generally supports this distributed model (e.g.,
Saffran & Schwartz, 1994).
This notion of distributed representation across large
scale brain areas is similar to the notion of the fine-
grained distributed representations across units that we
explored in previous models.

Higher-Level Association Areas

One further deviation from the general hierarchical
structure comes when one considers the higher levels
of a given processing pathway. These areas will likely
receive as much input from different pathways as they

Both imply that multi-