Information Technology Reference
In-Depth Information
SG Gestalt Patterns
This model demonstrates how sequential, piecemeal
inputs can be integrated together into a coherent sum-
mary or gestalt-like distributed representation. This
demonstration validates the idea that multiple constraint
satisfaction processing can take place across inputs oc-
curring over time, and not just across inputs occurring
together at the same time. We will refer to this phe-
nomenon as
temporal integration
.
We think that many important aspects of cognition
require temporal integration over piecemeal inputs. For
example, in forming a visual representation of all but
the most simple scenes, we must integrate over a large
number of visual inputs produced by fixating on differ-
ent aspects of the scene. Clearly, this must be true for
representing temporally extended events (e.g., as cap-
tured by one scene in a movie). In the domain of lan-
guage, almost all processing requires temporal integra-
tion, from the perception of spoken words to sequences
of words (the domain of the present model) to para-
graphs and chapters and entire topics!
An important unresolved question concerns the role
of the context representation in achieving temporal inte-
gration. We know that in the present model, the discrete
updating and maintenance capabilities of the context
representation are essential for its success. In chapter 9,
we saw that these properties of the context representa-
tion are closely associated with those of the prefrontal
cortex. Thus, must we conclude that the frontal cortex
is essential for all forms of temporal integration?
People with frontal lesions or schizophrenia (which
affects frontal processing) are impaired at some tasks
requiring temporal integration (e.g., Shallice, 1988;
Cohen & Servan-Schreiber, 1992). However, frontal
patients are probably not impaired at integrating over
multiple views of a visual scene, or in integrating over
the phonemes of a word in speech perception. In the ab-
sence of more complete data on this issue, we suggest
that temporal integration taking place over fairly short
intervals (e.g., up to a few seconds) can be subserved by
the natural temporal integration properties of a generic
cortical network (e.g., in the posterior cortex), but that
integration over longer periods would require the spe-
cialized frontal cortex active maintenance capabilities.
Some evidence for the temporal integration of the
current input with prior context in the brain comes from
Y
16.00
sc_st_ko.
15.00
te_st_ko.
14.00
bu_st_ko.
13.00
pi_st_ko.
12.00
te_dr_ic.
11.00
bu_dr_ic.
10.00
sc_dr_ic.
9.00
pi_dr_ic.
8.00
pi_at_st.
7.00
te_at_st.
6.00
bu_at_st.
5.00
sc_at_st.
4.00
pi_at_so.
3.00
te_at_so.
2.00
bu_at_so.
1.00
sc_at_so.
0.00
X
0.00
2.00
4.00
6.00
8.00
Figure 10.31:
Cluster plot of the gestalt layer representations
at the end of the probe sentences listed in table 10.16.
shows that the sentences are first clustered together ac-
cording to verb, and then by patient, and then by agent
within that. Furthermore, across the different patients,
there appears to be the same similarity structure for the
agents. Thus, we can see that the gestalt representa-
tion encodes information in a systematic fashion, as we
would expect from the network'sbehavior.
To stop now, quit by selecting
Object/Quit
in the
PDP++Root
window.
,
!
10.7.3
Summary and Discussion
The sentence gestalt model has demonstrated that mul-
tiple constraints from semantics and syntax can inter-
act in shaping a representation of the meaning of a sen-
tence. Because this meaning is captured in a distributed
representation, one would expect it to have all of the ad-
vantages of efficiency and generalization we have come
to expect from such representations, which it indeed ap-
pears to have. This form of sentence representation,
and the processing that produces it, differs consider-
ably from the traditional use of phrase structure trees or
rewrite rules, which are rigid and difficult to construct.
Search WWH ::
Custom Search