Representing Concepts in Networks: The Connectionist Approach • 255concepts for the network in (b) than in (a) because the links are shorter in (b). But the
lengths of the links can be determined by a number of factors, including a person’s past
experience with fi re engines or other types of vehicles. Unfortunately, there are no defi -
nite rules for determining these lengths—or, for that matter, for determining things like
how long activation remains after it spreads, or how much total activation is needed to
trigger a node. This means that by appropriately adjusting factors such as the length of
the links and how long activation lasts, the model can “explain” many different results.
But if a theory can explain almost any result by adjusting various properties of the
model, what has it really explained? This question is what led P. N. Johnson-Laird and
coworkers (1984) to criticize semantic network theories and to conclude that these
theories are “too powerful to be refuted by empirical evidence.” This is a way of saying
that it is diffi cult to falsify the theories. (See Anderson & Bower, 1973, and Glass &
Holyoak, 1975, for additional semantic network theories.)
One of the characteristics of science is that models are constantly being revised
to meet the challenges posed by new results or by criticisms such as Johnson-Laird’s
assessment of semantic network theory. But sometimes, instead of simply revising an
existing theory, researchers propose a whole new approach. This is what happened in
the 1980s, when a new approach called connectionism started moving to the forefront.Representing Concepts in Networks: The Connectionist Approach
Criticism of semantic networks, combined with new advances in understanding how
information is represented in the brain, led to the emergence of a new approach to
explaining how knowledge is represented in the mind. In two volumes, both titled
Parallel Distributed Processing: Explorations in the Microstructure of Cognition
(McClelland & Rumelhart, 1986; Rumelhart & McClelland, 1986), James McClelland
and David Rumelhart proposed a new approach called connec-
tionism. This approach has gained favor among many research-
ers because (1) it is based on how information is represented in
the brain; and (2) it can explain a number of fi ndings, including
how concepts are learned and how damage to the brain affects
people’s knowledge about concepts.WHAT IS A CONNECTIONIST MODEL?
Connectionism is an approach to creating computer models for
representing concepts and their properties based on character-
istics of the brain. These models are also called parallel distrib-
uted processing (PDP) models because, as we will see shortly,
they propose that concepts are represented by activity that is
distributed across a network.
An example of a simple connectionist network is shown in
● Figure 9.21. The circles are units. These units are inspired by
the neurons found in the brain. The lines are connections that
transfer information between units, and are roughly equivalent
to axons in the brain. Like neurons, some units can be activated
by stimuli from the environment, and some can be activated by
signals received from other units. Units activated by stimuli from
the environment (or stimuli presented by the experimenter) are
input units. In the simple network illustrated here, input units
send signals to hidden units, which send signals to output units.
An additional feature of a connectionist network is con-
nection weights. A connection weight determines how signals
sent from one unit either increase or decrease the activity of● FIGURE 9.21 A parallel distributed processing (PDP)
network showing input units, hidden units, and output units.
Incoming stimuli activate the input units, and signals travel
through the network, activating the hidden and output units.
Activity of the output units is indicated by their colors, with
blue and green representing low activity and red, higher
activity. The patterns of activity that occur in the output units
are determined both by the initial activity of the input units
and by the connection weights that determine how strongly
the hidden and output units will be activated by incoming
activity. Connection weights are not shown in this fi gure.Output unitsInput unitsHidden unitsCopyright 2011 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part.