Large-Scale Brain Area Functional Organization - Computational Explorations in Cognitive Neuroscience - page 219

things have both subordinate components and are them-
selves a component of a superordinate entity. For ex-
ample, a person's face is composed of components such
as eyes, a nose, and a mouth, and at the same time is
a component of the body. We think that there are spa-
tially invariant representations of the visual aspects of
such objects (e.g., an invariant representation of a nose,
mouth, face, etc.).
The question is therefore how the structural relation-
ship information of the parts as components (e.g., the
relative position of the nose within the face) can be rec-
onciled with the invariant representations of the parts as
objects unto themselves (e.g., Hinton et al., 1986). Are
there separate “relationship” representations that some-
how bind together the invariant nose representation with
the containing face representation? As you can tell,
these kinds of questions presume canonical represen-
tations of the object parts, and the structuralist, sym-
bolic solution is somehow to connect up these canonical
face-part representations within a larger structure for the
entire face, with relationship-binding associations that
convey the relationship information.
In contrast, the multiple, distributed representations
solution to this problem is to allow that there can be
many invariant representations active at some higher
level in the system, instead of just one canonical rep-
resentation. These higher-level representations can all
feed off of lower-level visual feature representations
that encode structural information via the same kind of
limited conjunctive-style representations as were sug-
gested as a solution to the binding problem (we will ex-
plore the role of these kinds of representations in object
recognition in chapter 8). Thus, there can be a shared
representation of the nose in a face and the invariant
nose to the extent that they both build off of the same
lower-level features. The higher-level face representa-
tion can retain the conjunctive information that encodes
the relationship between the nose and the rest of the
face, while the invariant nose representation abstracts
away from this context.
It is also likely that sequential attentional mecha-
nisms can be used to focus on different aspects of hi-
erarchically structured objects. Thus, the state of the
visual system when one is focusing on the face as a
whole will be different than when one is focusing just

on the nose — higher levels of the system can maintain
the invariant representations (and integrate over the se-
quential attentional shifts), while the lower levels are
modulated by attention in a way that emphasizes the
information necessary to process the object in the cur-
rent focus of attention. Thus, when you focus on the
nose, this deemphasizes those lower-level features that
encode the face, while emphasizing those that encode
the nose. This would make the job of abstracting away
from the facial context easier.
As it happens, the example of face recognition is
a particularly interesting one for these issues, because
it appears that the representation of the face is much
more than the sum of the representations of the indi-
vidual parts — faces appear to be encoded more “holis-
tically” (conjunctively) than other objects like houses
(e.g., Farah, 1992). Thus, the representation of a nose
in a face is not likely to have much in common with the
representation of the nose alone.

7.6.5 Recursion and Subroutine-like Processing

Another challenge for dedicated, content-specific rep-
resentations is the problem of recursion or executing
subroutines , where either the same type of processing
(recursion) or a different type of processing (subroutine)
needs to be performed in the middle of a given process-
ing step to obtain a result that is needed before process-
ing can proceed. In a serial computer, the current set
of state variables is simply pushed onto the stack (i.e.,
stored in a temporary memory buffer), and the subrou-
tine (possibly the same one) is called with appropriate
arguments.
This recursion is not easy when data and processing
are not separable, particularly in the case of recursion
where the same type of processing has to be performed
on different data, without forgetting the previous data!
This issue arises in the processing of sentences with
embedded clauses (e.g., “The mouse the cat the dog
bit chased squeaked.”), where one might imagine that
the overall processing of the sentence is composed of
a number of “subroutine calls” to process each clause.
However, as you can probably tell, people cannot easily
parse such sentences, indicating a limitation in exactly
this type of processing.