Information Technology Reference
In-Depth Information
up with ( B in this case, though the correct answer A
was considerably closer than the clearly wrong answer,
C ). Now let's just run through the rest of the quiz.
problems. Indeed, the addition of such task-based learn-
ing might serve as a useful model of the beneficial ef-
fects of production on comprehension. We have proba-
bly all had the experience of feeling as if we understood
something, only to find that this feeling was somewhat
illusory when it came time to describe the idea orally
or in writing. Similarly, the present model has a fairly
fuzzy set of representations based on its purely passive,
receptive experience with the domain. If we included
task-based learning as well, this would be like giving
the model practice (and feedback) articulating various
ideas, which would presumably clarify and sharpen the
representations.
Although the model we explored demonstrates a
number of important principles, it falls short of address-
ing the issue of the topographic organization of differ-
ent kinds of semantic information across different brain
areas. As mentioned in the introduction, this organiza-
tional issue has been a major focus of neuropsycholog-
ical studies, and neural network models such as the one
by Farah and McClelland (1991) have played an impor-
tant, if controversial, role in this field. One could con-
sider extending our model by including recurrent self-
connections within the hidden layer, and exploring the
topographic organization that develops as a result, much
in the same way we explored the topographic organiza-
tion of visual feature detectors in chapter 8.
However, we are somewhat skeptical of such an ap-
proach, because it seems clear that the topographic or-
ganization in the human brain depends on all kinds of
constraints that might be difficult to capture in such a
simple model. Referring back to figure 10.23 and the
distributed semantics idea of Allport (1985), it would
seem that one would have to build models that include
sufficient instantiations of various sensory systems, to-
gether with language processing and functional motor-
based processing. While not impossible, this is a more
formidable challenge that awaits future ambitions.
Turn off the network Display toggle (upper left-
hand corner of the NetView), and Run the NEpoch_2
process through the remainder of the quiz.
You should observe that the network does OK, but
not exceptionally — 60-80 percent performance is typ-
ical. Usually, the network does a pretty good job of
rejecting the obviously unrelated answer C , but it does
not always match our sense of A being better than B .
In question 6, the B phrase was often mentioned in the
context of the question phrase, but as a contrast to it,
not a similarity. Because the network does not have the
syntactic knowledge to pick up on this kind distinction,
it considers them to be closely related because they ap-
pear together. This probably reflects at least some of
what goes on in humans — we have a strong associa-
tion between “black” and “white” even though they are
opposites. However, we can also use syntactic informa-
tion to further refine our semantic representations — a
skill that is lacking in this network. The next section
describes a model that begins to address this skill.
, !
Go to the PDP++Root window. To continue on to
the next simulation, close this project first by selecting
.projects/Remove/Project_0 . Or, if you wish to
stop now, quit by selecting Object/Quit .
10.6.3
Summary and Discussion
Perhaps the most important aspects of this semantics
model are its ability to represent variations in semantic
meaning in terms of rich, overlapping distributed rep-
resentations, and its ability to develop these represen-
tations using Hebbian learning. Because our low-level
visual representations model from chapter 8 relies on
these same features, this helps to reaffirm our belief that
a common set of principles can be used to understand
cortical processing across a wide range of different do-
mains, from visual perception to language in this case.
One limitation of this model is that it does not in-
clude any task-based learning. The inclusion of this
type of learning might help to further refine the dis-
tinctions captured by the semantic representations, and
improve performance on things like the multiple choice
10.7
Sentence-Level Processing
The semantic representations model we just explored
provides one step toward understanding language pro-
cessing at a level higher than individual words. How-
ever, the semantic model was limited in that it com-
Search WWH ::




Custom Search