Figure 104

A sample of the materials used by Medin et al. (1990). The sample stimulus T is attributionally similar to A because they both have a shaded circle. B has not got this attributional similarity, but B does share a matching relation—same-colour elements—with T.

Extending similarity models to include relations

The contrast model has stood the test of time well. However, recently, a number of studies have shown that it might have to be adjusted in a number of respects. Tversky's model assumes that concepts can be adequately characterised by attribute lists (as do many of the theories reviewed earlier). Yet, in Chapter 9 we saw that relational concepts might also be important. Traditionally, relations (like flies, on-top-of, connected-to) would be treated as attributes in a concept definition. Running against this treatment, some recent research has examined similarity judgements of stimuli where attributes and relations have been separated out (see Davenport & Keane, 1999; Goldstone, Medin, & Gentner, 1991; Markman & Gentner, 1993a, b; Markman & Wisniewski, 1997; Medin, Goldstone, & Gentner, 1990, 1993).

For instance, Medin et al. (1990) gave subjects the stimuli shown in Figure 10.4. In these experiments, subjects had to choose whether the A or B stimulus was more similar to T. In each case, one of the stimulus options always shared a unique attribute with the T stimulus (e.g., in Figure 10.4, A and T share the unique attribute of having a shaded circle) and the other shared a unique relation with T (e.g., in Figure 10.4, B and T share the unique relation same-colour elements). They found that the stimulus with the relational similarity tended to be chosen as the more similar of the two, indicating that people were sensitive to relations in their judgements and that they also weighted relational-matches as being more important. Further research along these lines has found support for what Goldstone et al. (1991) call the MAX hypothesis; that attributional and relational similarities are pooled separately and shared similarities affect judged similarity more if the pool that they are in is relatively large.

This research has been used to show that similarity is more like analogy (see Chapter 15), that people map the relational structure of one concept onto that of another, using so-called structural alignment (see Markman & Gentner, 1993a, b; Gentner & Markman, 1997). It is fair to say that this view has now gained considerable currency in the treatment of similarity (but see Davenport & Keane, 1999).

A connectionist model of concept learning

Most of the computational models of concepts we have mentioned have been semantic networks in the Collins and Quillian style. More recently, several connectionist models have been used to model concepts.

Indeed, there is a close relationship between semantic networks and some types of connectionist nets (see the section on localist representations in Chapter 9). Connectionist nets make good concept learners because they can learn from specific instances and implement similarity mechanisms automatically. Their strengths lie in extracting the commonalities between a set of examples. In feedforward networks that use back-propagation of errors (see Chapter 1), the network learns the central tendency of a set of target examples and encodes this as a pattern of activation in the network. Other forms of networks, called interactive activation nets (IAC), can find the commonalities between a set of concepts by the way activation is passed between the nodes of the network. McClelland (1981) provides a neat demonstration of how such networks can manifest many of the properties of human conceptual systems.

McClelland's IAC net was given an attribute description of the individuals in the Jets and Sharks gangs from West Side Story (see Table 10.2). In the network, each attribute is represented as a node; so there are nodes for gang names (Jets, Sharks), for education (junior high, college, high school) and jobs (pusher, burglar, bookie). Attributes that are related are grouped into "pools" (see Figure 10.5); so, the pusher, burglar, and bookie nodes are grouped together in a pool because they are all occupations, and all the names of the people are in a "name pool". There is a special "person pool" for each individual in the list. This pool contains nodes, which do not encode any specific attribute, that stand for a particular individual. The links between nodes within a pool are all inhibitory or negative (these links are not explicitly shown in Figure 10.5). This means that the Jets node in the "gang pool" will pass negative activation to the Sharks node and vice versa. So, if one of these nodes has a high activation then it will force the activation of the other node down.

In this network, an individual is encoded by establishing excitatory links between the individual node and the nodes for the attributes of that individual. So, to encode Art, excitatory links are established between the _Art person node and the Art name-node, the Jets node, the 40s age-node, the junior-high education-node, the single marital-status node, and the pusher job-node (see Figure 10.5). This means that if activation is high at one of these nodes it will pass positive excitation to all of its connected nodes. Each of these nodes will then attempt to force down the activation of nodes within its respective pool, via the inhibitory links within the pool.

After the nodes and links have been built in this network it goes through a number of cycles where activation is passed between all the nodes in the network. In this way the activation in one node will interact with the activation in all the other nodes. Typically, after a number of cycles the network will settle into a state where the activation levels in the nodes are relatively unchanging. This network model can be used to demonstrate a number of properties of human conceptual systems. For example, if you want to determine the attributes of a given individual you

Stop Anxiety Attacks

Stop Anxiety Attacks

Here's How You Could End Anxiety and Panic Attacks For Good Prevent Anxiety in Your Golden Years Without Harmful Prescription Drugs. If You Give Me 15 minutes, I Will Show You a Breakthrough That Will Change The Way You Think About Anxiety and Panic Attacks Forever! If you are still suffering because your doctor can't help you, here's some great news...!

Get My Free Ebook


Post a comment