Page 63 - Foundations of Cognitive Psychology : Core Readings

62 Jay L. McClelland, David E.Rumelhart, and Geoffrey E.Hinton

schemata (Norman & Bobrow, 1976; Rumelhart, 1975).Such knowledge struc-
tures are assumed to be the basis of comprehension.A great deal of progress
has been made within the context of this view.
However, it is important to bear in mind that most everyday situations can-
not be rigidly assigned to just a single script.They generally involve an in-
terplay between a number of different sources of information.Consider, for
example, a child’s birthday party at a restaurant.We know things about birth-
day parties, and we know things about restaurants, but we would not want to
assume that we have explicit knowledge (at least, not in advance of our first
restaurant birthday party) about the conjunction of the two.Yet we can imag-
ine what such a party might be like.The fact that the party was being held in a
restaurant would modify certain aspects of our expectations for birthday par-
ties (we would not expect a game of Pin-the-Tail-on-the-Donkey, for example),
while the fact that the event was a birthday party would inform our expect-
ations for what would be ordered and who would pay the bill.
Representations like scripts, frames, and schemata are useful structures for
encoding knowledge, although we believe they only approximate the underly-
ing structure of knowledge representation that emerges from the class of mod-
els we consider in this chapter.Our main point here is that any theory that tries
to account for human knowledge using script-like knowledge structures will
have to allow them to interact with each other to capture the generative capac-
ity of human understanding in novel situations.Achieving such interactions
has been one of the greatest difficulties associated with implementing models
that really think generatively using script- or frame-like representations.

Parallel Distributed Processing
In theexampleswehave considered, anumberofdifferent pieces of informa-
tion must be kept in mind at once.Each plays a part, constraining others and
being constrained by them.What kinds of mechanisms seem well suited to
these task demands? Intuitively, these tasks seem to require mechanisms in
which each aspect of the information in the situation can act on other aspects,
simultaneously influencing other aspects and being influenced by them.To ar-
ticulate these intuitions, we and others have turned to a class of models we call
Parallel Distributed Processing (PDP) models.These models assume that infor-
mation processing takes place through the interactions of a large number of
simple processing elements called units, each sending excitatory and inhibitory
signals to other units.In some cases, the units stand for possible hypotheses
aboutsuchthingsasthe lettersinaparticular displayorthe syntacticroles of
the words in a particular sentence.In these cases, the activations stand roughly
for the strengths associated with the different possible hypotheses, and the
interconnections among the units stand for the constraints the system knows to
exist between the hypotheses.In other cases, the units stand for possible goals
and actions, such as the goal of typing a particular letter, or the action of mov-
ing the left index finger, and the connections relate goals to subgoals, subgoals
to actions, and actions to muscle movements.In still other cases, units stand
not for particular hypotheses or goals, but for aspects of these things.Thus a