256 • CHAPTER 9 Knowledge
● FIGURE 9.22 A connectionist
network proposed by Rogers and
McClelland (2003, 2004). Activation
of a concept unit and a relation
unit creates activity throughout
the network that culminates in
activation of property units. (Source:
T. T. Rogers & J. L. McClelland, Semantic
Cognition: A Parallel Distributed Processing
Approach, p. 56, Figure 2.2. Copyright ©
2004 Massachusetts Institute of Technology.
Reprinted by permission of The MIT Press.)living thing
plant
animal
tree
flower
bird
flower
pine
oak
rose
daisy
robin
canary
sunfish
salmonpine
oak
rose
daisy
robin
canary
sunfish
salmonISA
is
can
haspretty
tall
living
green
red
yellowgrow
move
swim
fly
singbark
petals
wings
feathers
scales
gills
roots
skinRelationRepresentation HiddenConceptPropertyis aiscanhasthe next unit. These weights correspond to what happens at a synapse that transmits
signals from one neuron to another (Figure 2.4, page 27). In Chapter 7 we saw that
some synapses can transmit signals more strongly than others and therefore cause a
high fi ring rate in the next neuron (Figure 7.16, page 190). Other synapses can cause
a decrease in the fi ring rate of the next neuron. Connection weights in a connectionist
network operate in the same way. High connection weights result in a strong tendency
to excite the next unit, lower weights cause less excitation, and negative weights can
decrease excitation or inhibit activation of the receiving unit. Activation of units in a
network therefore depends on two things: (1) the signal that originates in the input
units and (2) the connection weights throughout the network.
The effect of connection weights is illustrated by the differences in activation of the
output units in the network in Figure 9.21, indicated by the colors, with highly activated
units indicated by red, and less activated units by blue and green. Although the connection
weights are not shown, differences between different connections are what is causing the
differences in activity of the units. It is these differences, and the pattern of activity they
create, that are responsible for a basic principle of connectionism: A stimulus presented
to the input units is represented by the pattern of activity that is distributed across the
other units. If this sounds familiar, it is because it is similar to the distributed representa-
tions in the brain we have described for neural coding (Chapter 2, page 36), attention
(Chapter 4, page 107), and memory (Chapter 5, page 140; Chapter 7, page 191). Now
that we have used the simple network in Figure 9.21 to introduce the basic principles of
connectionist networks, we will consider how some specifi c concepts are represented in
the more complex connectionist network shown in ● Figure 9.22.Copyright 2011 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part.