Representing Concepts in Networks: The Connectionist Approach • 257HOW ARE CONCEPTS REPRESENTED
IN A CONNECTIONIST NETWORK?
The model in Figure 9.22 was proposed by James McClelland and Timothy Rogers
(2003) to show how different concepts and their properties can be represented in a con-
nectionist network. Although this model is more complex than the one in Figure 9.21, it
has similar components: units, links, and connection weights (although the connection
weights are not shown).
First, let’s compare this model to Quillian and Collins’ hierarchical model in Figure
9.12. The fi rst thing to notice is that both models are dealing with the same concepts.
Specifi c concepts, such as “canary” and “salmon,” shown in blue in Figure 9.12, are rep-
resented on the far left as concept units in Figure 9.22. Also, notice that the properties of
the concepts are indicated in both networks by the following four relation statements:
“is a” (A canary is a bird); “is” (A canary is yellow); “can” (A canary can fl y); and “has”
(A canary has wings). But whereas the hierarchical network indicates these properties
at the network’s nodes, the connectionist network indicates these properties by activity
in the property units on the far right.
Let’s now put our connectionist network to work by presenting a concept, “canary,”
and a relation statement, “can,” indicated by the red units in Figure 9.22. The network’s
goal is to complete the following fi ll-in-the-blank statement: “A canary can _____.”
We can see that this has been achieved by noting that the units for “grow,” “move,”
“fl y,” and “sing” have been activated. Try your hand at connecting the concepts and
relation statements on the left with properties on the right by doing the following
demonstration.DEMONSTRATION Activation of Property Units in a Connectionist Network
Looking at the network in Figure 9.22, focus on the concept units, the relation units, and the
property units. Ignore all of the other units in between. Your task is simply to ask the network
questions by indicating a concept and a relation, and then to note which property units should
be activated when the network answers the question. Determine which property units will be
activated by the following:Concept unit Relation unit Property units- Salmon is a?
- Pine tree has?
- Daisy is?
The demonstration illustrates how a concept’s properties can be represented by
activation of property units in our connectionist network. The connectionist model
proposes that a concept like “canary” is represented not only by activity of the property
units, but also by the pattern of activation of other units within the network. How is
this pattern determined? The answer to this question involves the connection weights,
as we will see when we consider what happens inside the network when we activate a
concept unit and a relation unit.
If we activate “canary” and “can,” these units send activity to the representation
and hidden units. The effect of this activation on the rest of the network depends on the
connection weights between the various units in the network. If all of the connection
weights were 1.0, then many of the units would be activated, including many incorrect
property units. For example, if you start at canary and can and follow the links, signals
would get through to incorrect property units like “tall,” “daisy,” and “green.”
For this model to work, the connection weights have to be adjusted so that when
concept unit “canary” and relation unit “can” are activated, only “grow,” “move,”
“fl y,” and “sing” are activated. This adjustment of weights is achieved by a learningCopyright 2011 Cengage Learning. All Rights Reserved. May not be copied, scanned, or duplicated, in whole or in part.