Ch. 9

Question 1

Learning Objectives

Answer 1

9-1
Explain why it is difficult to decide if a particular object belongs to a particular category, such as “chair,” by looking up its definition.
9-2
Describe how prototypes and exemplars influence our knowledge and use of categories.
9-3
Explain how the properties of various objects are “filed away” in the mind.
9-4
Evaluate how “networks” relate to conceptual information.
9-5
Describe how information about different categories is stored in the brain—including an ability to compare the various approaches and theories.
9-6
Explain the hub-and-spoke model.

Question 2

Knowledge

Answer 2

Explicit awareness of obtained and mentally available information about the world (exterior) and ourselves (interior) that was obtained through experience.

When we say we “know” something, that word refers to knowledge.

Question 3

Conceptual knowledge

Answer 3

Knowledge that enables people to recognize objects and events and to make inferences about their properties.

Conceptual knowledge involves answering questions such as the following:

When we encounter a new item or event in the world, how do we come to know what kind of thing it is?

How do we tell which items in our environment are horses, bicycles, trees, lakes, newspapers?

How do we tell dolphins from sharks, or planets from stars?

What makes a lemon a lemon?

What are the various kinds of “things” in the world?

This knowledge exists in the form of concepts

Question 4

Concepts

Answer 4

A mental representation of a class or individual. Also, the meaning of objects, events, and abstract ideas. An example of a concept would be the way a person mentally represents “cat” or “house”

Question 5

category

Answer 5

Groups of objects that belong together because they belong to the same class of objects, such as “houses,” “furniture,” or “schools.”

One way we organize concepts is in terms of categories.

. A category includes all possible examples of a particular concept.

Question 6

categorization

Answer 6

The process by which objects are placed in categories.

By placing an item in a category, it helps us to better understand that item and other items in that category.

Categories have therefore been called “pointers to knowledge”

Once you know that something is in a category, whether “dog,” “car,” “gas station,” or “sea dragon,” you can focus your energy on specifying what is special about this specific object.

Categorization helps us understand what is happening in the environment, and it plays an essential role in enabling us to take action.

Being able to place things in categories can also help us understand behaviors that we might otherwise find baffling.

These various uses of categories testify to their importance in everyday life. Without categories, we would have a very difficult time dealing with the world.

Categorization becomes more difficult if you encounter something unfamiliar.

This process becomes even more difficult if a person suffers a brain injury that makes it difficult or impossible to identify different objects or to know the purpose or function of these objects. Once we understand that there are situations in which categorization becomes difficult, we can then acknowledge that recognizing and understanding these difficulties is the first step to uncovering the mechanisms of categorization.

Question 7

3 major sections of the difficulties of categorization and the mechanisms involved in day-to-day categorization.

Answer 7

Each of the three sections tells a story that involves a different approach to categorization.

First, we consider a behavioral approach that originated with a series of experiments in the 1970s, which have helped us understand how we place objects in different categories and which have shown that “not all objects are created equal.”

Next, we consider the network approach to categorization that began in the 1960s, inspired by the emerging field of computer science, which created computer models of how categories are represented in the mind.

Finally, we take a physiological approach, which looks at the relationship between categories and the brain.

We will learn that each approach provides its own perspective on categorization—and that all three together provide a more complete explanation of categorization than any one approach can on its own.

Question 8

consider the following questions about the basic properties of categories:

Answer 8

How are different objects, events, or ideas assigned to a particular category?

How can categories be defined?

Why do we say that “not all things in categories are created equal”?

Question 9

definitional approach to categorization

Answer 9

The idea that we can decide whether something is a member of a category by determining whether the object meets the definition of the category.

Definitions work well for some things, such as geometric objects. Thus, defining a square as “a plane figure having four equal sides, with all internal angles the same” works.

However, for most natural objects (such as birds, trees, and plants) and many human-made objects (like chairs), definitions do not work well at all.

The problem is that not all the members of everyday categories share similar features. So, although the dictionary definition of a cup as “a small bowl-shaped container from which someone can drink something, typically having a handle” may sound reasonable, there are many objects we call “cups” that do not meet that definition.

Question 10

family resemblance

Answer 10

In considering the process of categorization, the idea that things in a particular category resemble each other in a number of ways. This approach can be contrasted with the definitional approach, which states that an object belongs to a category only when it meets a definite set of criteria.

Allows for some variation with a category.

In a series of experiments beginning in the 1970s, Eleanor Rosch and colleagues used the idea of family resemblance as a jumping off point for experiments that investigated the basic nature of categories. One of the early ideas to emerge from these experiments is the idea of prototypes.

Question 11

prototype approach to categorization

Answer 11

The idea that we decide whether something is a member of a category by determining whether it is similar to a standard representation of the category, called a prototype.

What is a typical member of a category? Eleanor Rosch (1973) proposed that the “typical” prototype is based on an average of members of a category that are commonly experienced.

For example, the prototype for the category “birds” might be based on some of the birds you usually see, such as sparrows, robins, and bluejays but does not necessarily look exactly like any one of them. Thus, the prototype is not an actual member of the category but is an “average” representation of the category.

The prototype approach to categorization represents a great advance over the definitional approach because it is reinforced by a wealth of experimental evidence that all items within a category are not the same.

Question 12

prototype

Answer 12

A standard used in categorization that is formed by averaging the category members a person has encountered in the past.

Question 13

typicality

Answer 13

The degree to which an item is representative or characteristic of a particular category.

Of course, not all birds are like robins, bluejays, or sparrows. Owls, buzzards, and penguins are also birds. Rosch describes these variations within categories as representing differences in typicality . High typicality means that a category member closely resembles the category prototype (it is like a “typical” member of the category). Low typicality means that the category member does not closely resemble a typical member of the category.

Rosch (1975a) quantified this idea by presenting participants with a category title, such as “bird” or “furniture,” and a list of about 50 members of the category. The participants’ task was to rate the extent to which each member represented the category title on a 7-point scale, with a rating of 1 meaning that the member is a very good example of what the category is, and a rating of 7 meaning that the member fits poorly within the category or is not a member at all.

The 1.18 rating for sparrow reflects the fact that most people consider a sparrow to be a good example of a bird (Figure 9.5a). The 4.53 rating for penguin and 6.15 rating for bat reflect the fact that penguins and bats are not considered good examples of birds. Similarly, chair and sofa (rating = 1.04) are considered very good examples of furniture, but mirror (4.39) and landline home telephone (6.68) are poor examples.

It is worth noting that a landline home telephone would have been common in 1975. As this example may not be relatable to your experiences, you could imagine the concept of “trash can” instead. A trash can is also technically furniture. There is typically one in the kitchen and the bathroom. However, like a mirror, a trash can is unlikely to meet the definition of “furniture” as well as a chair or sofa.

The idea that a sparrow is a better example of “bird” than a penguin or a bat makes sense. However, Rosch went beyond this rather obvious result by doing a series of experiments that demonstrated differences between good and bad examples of a category.

Question 14

How well do good and poor examples of a category compare to other items within the category?

Answer 14

The following demonstration is based on an experiment by Rosch and Carolyn Mervis (1975).

If you responded like Rosch and Mervis’s participants, you assigned many of the same characteristics to chair and sofa. For example, chairs and sofas share the characteristics of having legs, having backs, you sit on them, they can have cushions, and so on. When an item’s characteristics have a large amount of overlap with the characteristics of many other items in a category, this means that the family resemblance of these items is high. But when we consider items like mirror, trash can, and telephone, we find that there is far less overlap, even though mirror and telephone were both classified by Rosch and Mervis as “furniture” (Figure 9.5b). Little overlap with other members of a category means the family resemblance is low.

Rosch and Mervis concluded from their results that there is a strong relationship between family resemblance and prototypicality. Thus, good examples of the category “furniture,” such as chair and sofa, share many attributes with other members of this category; poor examples, like mirror and telephone, do not. In addition to the connection between prototypicality and family resemblance, researchers have determined a number of other connections between prototypicality and behavior.

Demonstration Family Resemblance

Rosch and Mervis’s (1975) instructions were as follows: For each of the following common objects, list as many characteristics and attributes as you can that you feel are common to these objects.

Method Sentence Verification Technique

The procedure for the sentence verification technique is simple. Participants are presented with statements and are asked to answer “yes” if they think the statement is true and “no” if they think it is not.

Question 15

sentence verification technique

Answer 15

A technique in which the participant is asked to indicate whether a particular sentence is true or false. For example, sentences like “An apple is a fruit” have been used in studies on categorization.

When Smith and colleagues (1974) used this technique, they found that participants responded faster for objects that are high in prototypicality (like apple for the category “fruit”) than they did for objects that are low in prototypicality (like pomegranate; Figure 9.6). This ability to judge highly prototypical objects more rapidly is called the typicality effect .

Question 16

typicality effect

Answer 16

The ability to judge the truth or falsity of sentences involving high-prototypical members of a category more rapidly than sentences involving low-prototypical members of a category.

Question 17

Prototypical Objects Are Named First

Answer 17

When participants are asked to list as many objects in a category as possible, they tend to list the most prototypical members of the category first (Mervis et al., 1976). Thus, for “birds,” sparrow would be named before penguin.

Question 18

Prototypical Objects Are Affected More by Priming

Answer 18

priming occurs when presentation of one stimulus facilitates the response to another stimulus that usually follows closely in time.

Rosch (1975b) demonstrated that prototypical members of a category are more affected by a priming stimulus than are non-prototypical members. The procedure for Rosch’s experiment is shown in Figure 9.7. Participants first heard the prime, which was the name of a color, such as “green.” Two seconds later they saw a pair of colors side by side and indicated, by pressing a key as quickly as possible, whether the two colors were the same or different.

The side-by-side colors that participants saw after hearing the prime were paired in three different ways:

(1)
colors were the same and were good examples of the category (primary reds, blues, greens, etc.; Figure 9.7a);

(2)
colors were the same but were poor examples of the category (less rich versions of the good colors, such as light blue, light green, etc.; Figure 9.7b); and

(3)
colors were different, with the two colors coming from different categories (for example, pairing red with blue).

The most important result occurred for the two “same” groups. In this condition, priming resulted in faster “same” judgments for the prototypical (good) colors (reaction time, RT = 610 ms) than for the non-prototypical (poor) colors (RT = 780 ms). Thus, when participants heard the word green, they judged two patches of primary green as being the same more rapidly than two patches of light green.

Rosch explains this result as follows: When participants hear the word green, they imagine a “good” (highly prototypical) green (Figure 9.8a). The principle behind priming is that the prime will facilitate the participants’ response to a stimulus if it contains some of the information needed to respond to the stimulus. This apparently occurs when the good greens are presented in the test (Figure 9.8b), but not when the poor greens are presented (Figure 9.8c). Thus, the results of the priming experiments support the idea that participants create images of prototypes in response to color name

Question 19

exemplar approach to categorization

Answer 19

The approach to categorization in which members of a category are judged against exemplars—examples of members of the category that the person has encountered in the past.

The exemplar approach can explain many of Rosch’s results, which were used to support the prototype approach. For example, the exemplar approach explains the typicality effect (in which reaction times on the sentence verification task are faster for better examples of a category than for poorer examples) by proposing that objects that are like more of the exemplars are classified faster. Thus, a sparrow is similar to many bird exemplars, so it is classified faster than a penguin, which is similar to a few bird exemplars. This is basically the same as the idea of family resemblance, described for prototypes, which states that “better” objects will have higher family resemblance.

Question 20

Exemplars

Answer 20

Specific instances or examples within a category that are used to represent and define that category

Question 21

Which Approach Works Better: Prototypes or Exemplars?

Answer 21

Which approach—prototypes or exemplars—provides a better description of how people use categories? One advantage of the exemplar approach is that by using real examples, it can more easily consider atypical cases such as flightless birds. Rather than comparing a penguin to an “average” bird, we remember that there are some birds that do not fly. This ability to consider individual cases means that the exemplar approach does not discard information that might be useful later. Thus, penguins, ostriches, and other birds that are not typical can be represented as exemplars, rather than becoming lost in the overall average that creates a prototype. The exemplar approach can also deal more easily with variable categories like games. Although it is difficult to imagine what the prototype might be for a category that contains rugby, video games, solitaire, pickle ball, and golf, the exemplar approach requires only that we remember some of these varying examples.

Some researchers have concluded that people may use both approaches. It has been proposed that as we initially learn about a category, we may average exemplars into a prototype; then, later in learning, some of the exemplar information becomes stronger (Keri et al., 2002; Malt, 1989). Thus, early in learning, we would be poor at taking into account “exceptions” such as ostriches or penguins, but later, exemplars for these cases would be added to the category. We know generally what dogs are—the prototype—but we know our own specific dog the best—an exemplar (Minda & Smith, 2001; Smith & Minda, 2000). A recent survey considering the virtues of both prototypes and exemplars ends with the following conclusion: “The two kinds of information work together to produce our rich store of conceptual knowledge allowing each kind of knowledge to explain the tasks that are most suited for it” (Murphy, 2016).

Question 22

hierarchical organization

Answer 22

Organization of categories in which larger, more general categories are divided into smaller, more specific categories. These smaller categories can, in turn, be divided into even more specific categories to create a number of levels.

One question cognitive psychologists have asked about this organization is whether there is a “basic” level that is more psychologically basic or important than other levels. The research we will describe indicates that although it is possible to demonstrate that there is a basic level of categories with special psychological properties, the basic level may not be the same for everyone.

Question 23

Rosch’s Approach: What’s Special About Basic Level Categories?

Answer 23

Her research distinguished three levels of categories:

(1)
the superordinate level , which we will call the global level (for example, “furniture”);

(2)
the basic level (for example, “table”); and

(3)
the subordinate level , which we will call the specific level (for example, “kitchen table”).

The following demonstration illustrates some characteristics of the different levels.

If you responded like the participants in the Rosch and colleagues’ (1976) experiment, who were given the same task, you listed only a few features that were common to all furniture but many features that were shared by all tables and by all kitchen tables. Rosch’s participants listed an average of 3 common features for the global level category “furniture,” 9 for basic level categories such as “table,” and 10.3 for specific level categories such as “kitchen table” .

Rosch proposed that the basic level is psychologically special because going above it (to global) results in a large loss of information (9 features at the basic versus 3 at the global level) and going below it (to specific) results in little gain of information (9 features versus 10.3).

Question 24

What names did you assign to each object?

Answer 24

When Rosch and colleagues (1976) did a similar experiment, they found that people tended to pick a basic-level name. They said guitar (basic level) rather than musical instrument (global) or rock guitar (specific), fish rather than animal or trout, and pants rather than clothing or jeans.

In another experiment, Rosch and colleagues showed participants a category label, such as car or vehicle, and then, after a brief delay, presented a picture. The participants’ task was to indicate, as rapidly as possible, whether the picture was a member of the category. The results showed that they accomplished this task more rapidly for basic level categories (such as car) than for global level categories (such as vehicle). Thus, they would respond “yes” more rapidly when the picture of an automobile was preceded by the word car than when the picture was preceded by the word vehicle.

Question 25

How Knowledge Can Affect Categorization

Answer 25

Rosch’s experiments, which were carried out on college undergraduates, showed that there is a category level, which she called “basic,” that reflects college undergraduates’ everyday experience. This has been demonstrated by many researchers in addition to Rosch. Thus, when J. D. Coley and colleagues (1997) asked Northwestern University undergraduates to name, as specifically as possible, 44 different plants on a walk around campus, 75 percent of the responses used labels like “tree,” rather than more specific labels like “oak.” However, instead of asking college undergraduates to name plants, what if Coley had taken a group of horticulturalists around campus? Do you think they would have said “tree” or “oak”? An experiment by James Tanaka and Marjorie Taylor (1991) asked a similar question for birds. They asked bird experts and novices to name pictures of objects. There were objects from many different categories (tools, clothing, flowers, etc.), but Tanaka and Taylor were interested in how the participants responded to the four bird pictures. The results (Figure 9.11) show that the experts responded by specifying the birds’ species (robin, sparrow, jay, or cardinal), but the novices responded by saying “bird.” Apparently, the experts had learned to pay attention to features of birds, of which the novices were unaware. Thus, to fully understand how people categorize objects, we need to consider not only the properties of the objects but also the learning and experience of the people perceiving those objects. From the result of Tanaka’s bird experiment, we can guess that a horticulturist walking around campus would be likely to label plants more specifically than people who had little specific knowledge about plants. In fact, members of the Guatemalan Itzaj culture, who live in close contact with their natural environment, call an oak tree an “oak,” not a “tree” (Coley et al., 1997). Thus, the level that is “special”—meaning that people tend to focus on it—is not the same for everyone. Generally, people with more expertise and familiarity with a particular category tend to focus on the more specific information that Rosch associated with the specific level. This makes sense because our ability to categorize is learned from experience and depends on which objects we typically encounter and what characteristics of these objects we pay attention to.

Question 26

We consider the following two ideas related to the network approach to categorization:

Answer 26

An early network approach, called the semantic network approach, that uses networks of connected concepts to explain how these concepts might be organized in the mind A more modern network approach called “connectionism” that describes how networks can be “trained” to recognize specific objects

Question 27

semantic network approach

Answer 27

A model that represents knowledge as a network of interconnected nodes. Each node represents a concept, and the connections between nodes represent the relationships between concepts.

Question 28

Introduction to Semantic Networks: Collins and Quillian’s Hierarchical Model

Answer 28

One of the first semantic network models was based on the pioneering work of Ross Quillian (1967, 1969), whose goal was to develop a computer model of human memory. We will describe Quillian’s approach by looking at a simplified version of his model proposed by Allan Collins and Quillian (1969). The network consists of nodes that are connected by links. Each node represents a category or concept, and concepts are placed in the network so that related concepts are connected. In addition, many properties are indicated for each concept. The links connecting the concepts indicate how they are related to each other in the mind. Thus, the model shown in Figure 9.13 indicates that there is an association in the mind between canary and bird, and between bird and animal (indicated by the dashes along the links in Figure 9.13). It is a hierarchical model because it consists of levels arranged so that more specific concepts, such as “canary” and “salmon,” are at the bottom, and more general concepts are at higher levels. Notice how this hierarchical model in 9.13 looks like the concept maps we discussed in Chapter 7. This similarity is not coincidental. A concept map represents a network of ideas with meaningful connections. Network models that explain how information is stored in the brain are very similar. Since the brain is likely organized in some form of semantic network to store concepts and ideas, it makes sense that the brain may prefer to learn, study, and store information from concept maps that resemble this organization.

Question 29

We can illustrate how a hierarchical network works, and how it proposes that knowledge about concepts is organized in the mind, by considering how we would retrieve the properties of canaries from the network.

Answer 29

We start by entering the network at the concept node for “canary.” At this node, we obtain the information that a canary can sing and is yellow. To access more information about “canary,” we move up the link and learn that a canary is a bird and that a bird has wings, can fly, and has feathers. Moving up another level, we find that a canary is also an animal, which has skin and can move, and finally, we reach the level of living things, which tells us it can grow and is living. You might wonder why we must travel from “canary” to “bird” to find out that a canary can fly. That information could have been placed at the canary node, and then we would know it right away. But Collins and Quillian proposed that including “can fly” at the node for every bird (canary, robin, vulture, etc.) was inefficient and would use up too much storage space. Thus, instead of indicating the properties “can fly” and “has feathers” for every kind of bird, these properties are placed at the node for “bird” because this property holds for most birds. This way of storing shared properties just once at a higher-level node is called cognitive economy .

Question 30

cognitive economy

Answer 30

The principle that the brain organizes information in a way that minimizes cognitive effort and redundancy, storing shared properties or features of concepts at the highest possible level of hierarchical structure to efficiently use mental resources. Although cognitive economy makes the network more efficient, it does create a problem; for instance, not all birds fly. To deal with this problem while still achieving the advantages of cognitive economy, Collins and Quillian added exceptions at lower nodes. For example, the node for “ostrich,” which is not shown in this network, would indicate the property “cannot fly.”

Question 31

How do the elements in this semantic network correspond to the actual operation of the brain?

Answer 31

The links and nodes we have been describing do not necessarily correspond to specific nerve fibers or locations in the brain. The Collins and Quillian model is not meant to mirror physiology but to indicate how concepts and their properties are associated in the mind, and to make predictions about how we retrieve properties associated with a concept.

Question 32

Putting aside any possible connection between semantic networks and actual physiology, we can ask how accurate the model’s predictions are.

Answer 32

One prediction is that the time it takes for a person to retrieve information about a concept should be determined by the distance that must be traveled through the network. Thus, the model predicts that when using the sentence verification technique, in which participants are asked to answer “yes” or “no” to statements about concepts (see Method: Sentence Verification Technique, page 279), it should take longer to answer “yes” to the statement “A canary is an animal” than to “A canary is a bird.” This prediction follows from the fact, indicated by the dashed lines in Figure 9.12, that it is necessary to travel along two links to get from “canary” to “animal” but only one to get to “bird.” Collins and Quillian (1969) tested this prediction by measuring the reaction time to several different statements and obtained the results shown in Figure 9.14. As predicted, statements that required further travel from “canary” resulted in longer reaction times.

Question 33

Spreading activation

Answer 33

Activity that spreads out along any link in a semantic network that is connected to an activated node. Another property of the theory, which leads to further predictions, is spreading activation. For example, moving through the network from “robin” to “bird” activates the node at “bird” and the link we use to get from robin to bird, as indicated by the colored arrow in Figure 9.15. But according to the idea of spreading activation, this activation also spreads to other nodes in the network, as indicated by the dashed lines. Thus, activating the canary-to-bird pathway activates additional concepts that are connected to “bird,” such as “animal” and other types of birds. The result of this spreading activation is that the additional concepts that receive this activation become “primed” and so can be retrieved more easily from memory.

Question 34

lexical decision task.

Answer 34

A procedure in which a person is asked to decide as quickly as possible whether a particular stimulus is a word or a nonword. The idea that spreading activation can influence priming was studied by David Meyer and Roger Schvaneveldt (1971) in a paper published shortly after Collins and Quillian’s model was proposed. They used a method called the lexical decision task. Meyer and Schvaneveldt used a variation of the lexical decision task by presenting participants with pairs of words, one above the other. The participants’ task was to press, as quickly as possible, the “yes” key when both items were words or the “no” key when at least one item in the pair was a nonword. Thus, pairs 1 and 2 would require a “no” response, and pairs 3 and 4 would require a “yes” response. The key variable in this experiment was the association between the pairs of real words. In some trials, the words were closely associated (like bread and wheat), and in some trials, they were weakly associated (chair and money). The result, shown in Figure 9.16, was that reaction time was faster when the two words were associated. Meyer and Schvaneveldt proposed that this might have occurred because retrieving one word from memory triggered a spread of activation to other nearby locations in a network. Because more activation would spread to words that were related, the response to the related words was faster than the response to unrelated words.

Question 35

Criticism of the Collins and Quillian Model

Answer 35

Although Collins and Quillian’s model was supported by the results of several experiments, such as their reaction time experiment (Figure 9.13) and Meyer and Schvaneveldt’s priming experiment, it did not take long for other researchers to call the theory into question. They pointed out that the theory could not explain the typicality effect, in which reaction times for statements about an object are faster for more typical members of a category than for less typical members (see page 279; Rips et al., 1973). Thus, the statement “A canary is a bird” is verified more quickly than “An ostrich is a bird,” but the model predicts equally fast reaction times because “canary” and “ostrich” are both one node away from “bird.” Researchers also questioned the concept of cognitive economy because of evidence that people may, in fact, store specific properties of concepts (like “has wings” for “canary”) right at the node for that concept (Conrad, 1972). In addition, Lance Rips and colleagues (1973) obtained sentence verification reaction time (RT) results such as the following: A pig is a mammal; RT = 1,476 milliseconds A pig is an animal; RT = 1,268 milliseconds “A pig is an animal” is verified more quickly, but as we can see from the network in Figure 9.17, the Collins and Quillian model predicts that “A pig is a mammal” should be verified more quickly because a link leads directly from “pig” to “mammal,” but we need to travel one link past the “mammal” node to get to “animal.” Sentence verification results such as these, plus the other criticisms of the theory, led researchers to look for alternative ways to using networks to describe how concepts are organized (Glass & Holyoak, 1975; Murphy et al., 2012) and eventually, in the 1980s, to the proposal of a new approach to networks, called connectionism.

Question 36

Connectionism

Answer 36

A network model of mental operation that proposes that concepts are represented in networks that are modeled after neural networks. This approach to describing the mental representation of concepts is also called the parallel distributed processing (PDP) approach. (propose that concepts are represented by activity that is distributed across a network.) This approach gained favor among many researchers because (1) it is inspired by how information is represented in the brain; and (2) it can explain several findings, including how concepts are learned and how damage to the brain affects people’s knowledge about concepts.

Question 37

connectionist network

Answer 37

The type of network proposed by the connectionist approach to the representation of concepts. Connectionist networks are based on neural networks but are not necessarily identical to them. One of the key properties of a connectionist network is that a specific category is represented by activity that is distributed over many units in the network. This contrasts with semantic networks, in which specific categories are represented at individual nodes. An example of a simple connectionist network is shown in Figure 9.18. The circles are units . These units are inspired by the neurons found in the brain. As we will discover, concepts and their properties are represented in the network by the pattern of activity across these units. 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 connection weights. A connection weight determines how signals sent from one unit either increase or decrease the activity of the next unit. These weights correspond to what happens at a synapse that transmits signals from one neuron to another. some synapses can transmit signals more effectively than others and therefore cause a high firing rate in the next neuron (Figure 7.11). Other synapses can cause a decrease in the firing 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. In the network of Figure 9.18, two of the input units are receiving stimuli. Activation of each of the hidden and output units is indicated by the shading, with darker shading indicating more activation. These differences in activation, and the pattern of activity they create, 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.

Question 38

How Are Concepts Represented in a Connectionist Network? Representing a Canary

Answer 38

Let’s first compare this model to the Collins and Quillian hierarchical model in Figure 9.13. The first thing to notice is that both models are dealing with the same concepts. Specific concepts, such as “canary” and “salmon,” shown in blue in Figure 9.13, are represented on the far left as concept items in Figure 9.18. 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 fly); and “has” (A canary has wings). But whereas the hierarchical network in Figure 9.13 represents these properties at the network’s nodes, connectionist networks indicate these properties by activity in the attribute units on the far right, and also by the pattern of activity in the representation and hidden units in the middle of the network. Figure 9.18 shows that when we activate the concept “canary” and a relation unit, can, activation spreads along the connections from “canary” and can so that some of the representation units are activated and some of the hidden units are activated. The connection weights, which are not shown, cause some units to be activated strongly and others more weakly, as indicated by the shading of the units. If the network is working properly, this activation in the hidden units activates the grow, move, fly, and sing property units. What is important about all of this activity is that the concept “canary” is represented by the pattern of activity in all of the units in the network.

Question 39

Training a network

Answer 39

According to the above description, the answer to “A canary can …” is represented in the network by activation of the property units plus the pattern of activation of the network’s representation and hidden units. But according to connectionism, this network had to be trained to achieve this result. We can appreciate the need for training by considering Figure 9.20, which indicates how the network might have responded before training had occurred. In the untrained network, stimulating the canary and can units sends activity out to the rest of the network, with the effect of this activation depending on the connection weights between the units. Let’s assume that in our untrained network, all the connection weights are 1.0. Because the connection weights are the same, activity spreads throughout the network, and property nodes such as flower, pine, and bark, which have nothing to do with canaries, are activated. For the network to operate properly, the connection weights must be adjusted so that activating the concept unit “canary” and the relation unit can activates only the property units grow, move, fly, and sing. This adjustment of weights is achieved by a learning process. The learning process occurs when the erroneous responses in the property units cause an error signal to be sent back through the network, by a process called back propagation (since the signals are being sent backward in the network starting from the property units). The error signals that are sent back to the hidden units and the representation units provide information about how the connection weights should be adjusted so that the correct property units will be activated. Although this “educated” network might work well for canaries, what happens when a robin flies by and alights on the branch of a pine tree? To be useful, this network needs to be able to represent not just canaries but also on other items like robins and pine trees. Thus, to create a network that can represent many different concepts, the network is not trained just on “canary.” Instead, presentations of “canary” are interleaved with presentations of “robin,” “pine tree,” and so on, with small changes in connection weights made after each presentation.

Question 40

To explain the idea behind activation and back propagation, let’s consider a behavioral example.

Answer 40

A young child is watching a robin sitting on a branch when suddenly the robin flies away. This simple observation, which strengthens the association between “robin” and can fly, would involve activation. However, if the child were to see a canary and say “robin,” the child’s parent might offer a correction and say, “That is a canary” and “Robins have red breasts.” The information provided by the parent is similar to the idea of feedback provided by back propagation. Thus, a child’s learning about concepts begins with little information and some incorrect ideas, which are slowly modified in response both to observation of the environment and to feedback from others. Similarly, the connectionist network’s learning about concepts begins with incorrect connection weights that result in the activation shown in Figure 9.20, which are slowly modified in response to error signals to create the correctly functioning network in Figure 9.19.

Question 41

Although our description has been based on one particular connectionist network, most networks have similar properties.

Answer 41

Connectionist networks are created by a learning process that shapes the networks so information about each concept is contained in the distributed pattern of activity across many units. Notice how different this operation of the connectionist network is from the operation of Collins and Quillian’s hierarchical network, in which concepts and their properties are represented by activation of different nodes. Representation in a connectionist network is not only far more complex, involving many more units for each concept, but it is also much more like what happens in the brain.

Question 42

Many researchers believe that the idea that knowledge is represented by distributed activity holds great promise because

Answer 42

of the resemblance between connectionist networks and the brain, and the fact that connectionist networks have been developed that can simulate normal cognitive functioning for processes such as language processing, memory, and cognitive development (Rogers & McClelland, 2004; Seidenberg & Zevin, 2006). The following results also support the idea of connectionism: 1. The operation of connectionist networks is not totally disrupted by damage. Because information in the network is distributed across many units, damage to the system does not completely disrupt its operation. This property, in which disruption of performance occurs only gradually as parts of the system are damaged, is called graceful degradation . It is similar to what often happens in actual cases of trauma to the brain, in which damage to the brain causes only a partial loss of functioning. Some researchers have suggested that studying the way networks respond to damage may suggest strategies for rehabilitation of human patients. 2. Connectionist networks can explain generalization of learning. Because similar concepts have similar patterns, training a system to recognize the properties of one concept (such as “canary”) also provides information about other, related concepts (such as “robin” or “sparrow”). This is similar to the way we actually learn about concepts because learning about canaries enables us to predict properties of different types of birds we’ve never seen.

Question 43

How Concepts Are Represented in the Brain We consider the following three ideas related to how categories are represented in the brain:

Answer 43

What the results of neuropsychological research tell us about where different categories are represented in the brain. How neuropsychological research has led to several different models that explain how categories are organized in the brain. What brain imaging research tells us about how and where different categories are represented in the brain?

Question 44

category-specific memory impairment

Answer 44

A result of brain damage in which the patient has trouble recognizing objects in a specific category. To explain why this selective impairment occurred, Warrington and Shallice considered properties that people use to distinguish between artifacts and living things. They noted that distinguishing living things depends on perceiving their sensory features. For example, distinguishing between a tiger and a leopard depends on perceiving stripes and spots. Artifacts, in contrast, are more likely to be distinguished by their function. For example, a screwdriver, chisel, and hammer are all tools but are used for different purposes (turning screws, scraping, and pounding nails)

Question 45

The Sensory-Functional Hypothesis

Answer 45

Explanation of how semantic information is represented in the brain that states that the ability to differentiate living things and artifacts depends on one system that distinguishes sensory attributes and another system that distinguishes function. While the S-F hypothesis explained the behavior of Warrington and Shallice’s patients, plus dozens of other patients, researchers began describing cases that could not be explained by this hypothesis. For example, Matthew Lambon Ralph and colleagues (1998) studied a patient who had a sensory deficit—she performed poorly on perceptual tests—yet she was better at identifying animals than artifacts, which is the opposite of what the S-F hypothesis predicts. In addition, some patients can identify mechanical devices even if they perform poorly for other types of artifacts. For example, Hoffman and Lambon Ralph (1998) describe patients who have poor comprehension of small artifacts like tools but better knowledge of larger artifacts, such as vehicles. Thus, “artifacts” are not a single homogeneous category as hypothesized by the S-F hypothesis. Results such as these have led many researchers to conclude that many effects of damage to the brain cannot be explained by the simple distinction between sensory and function

Question 46

multiple-factor approach

Answer 46

Seeking to describe how concepts are represented in the brain by searching for multiple factors that determine how concepts are divided up within a category. We can appreciate this approach by posing the following question: Assume that we start with many items selected from lists of different types of animals, plants, and artifacts. If you wanted to arrange them in terms of how similar they are to each other, how would you do it? You could arrange them by shape, but then items like a pencil, a screwdriver, a person’s finger, and a breakfast sausage might be grouped together. Or considering just color, you could end up placing fir trees, leprechauns, and Kermit the Frog together. Although members of specific categories do indeed share similar perceptual attributes, it is also clear that we need to use more than just one or two features when grouping objects in terms of similarity.

Question 47

experiment by Paul Hoffman and Matthew Lambon Ralph (2013)

Answer 47

used 160 items like the ones shown in Table 9.3a. The participants’ task was to rate each item on the features shown in Table 9.3b. For example, for the concept “door,” the participant would be asked, “How much do you associate door with a particular color (or form, or motion, etc.)?” Participants assigned a rating of 7 for “very strongly” to 1 for “not at all.” researchers picked several different features and had participants rate many items using these features. The results, shown in Figure 9.23, indicate that animals were more highly associated with motion and color compared to artifacts, and artifacts were more highly associated with performed actions (actions associated with using or interacting with an object). This result conforms to the S-F hypothesis, but when Hoffman and Lambon Ralph (1998) looked at the groupings more closely, they found some interesting results. Mechanical devices such as machines, vehicles, and musical instruments overlapped with both artifacts (involving performed actions) and animals (involving sound and motion). For example, musical instruments are associated with specific actions (how you play them), which goes with artifacts, and are also associated with sensory properties (their visual form and the sounds they create), which goes with animals. Thus, musical instruments and some mechanical devices occupy a middle ground between artifacts and living things, because they can involve both action knowledge and sensory attributes.

Question 48

Crowding

Answer 48

Animals tend to share many properties, such as eyes, legs, and the ability to move. This is relevant to the multiple-factor approach to the representation of concepts in the brain. In contrast, artifacts like cars and boats share fewer properties, other than that they are both vehicles. This has led some researchers to propose that patients who appear to have category-specific impairments, such as difficulty recognizing living things but not artifacts, do not really have a category-specific impairment at all. They propose that these patients have difficulty recognizing living things because they have difficulty distinguishing between items that share similar features. According to this idea, because animals tend to be more similar than artifacts, these patients find animals harder to recognize.

Question 49

The Semantic Category Approach

Answer 49

An approach to describing how semantic information is represented in the brain that proposes that there are specific neural circuits for some specific categories. According to Bradford Mahon and Alfonso Caramazza (2011), there are a limited number of categories that are innately determined because of their importance for survival. This idea is based on research that we described in Chapter 2, which identified regions of the brain that respond to specific types of stimuli such as faces, places, and bodies. While the semantic category approach focuses on regions of the brain that are specialized to respond to specific types of stimuli, it also emphasizes that the brain’s response to items from a particular category is distributed over several different cortical regions. Thus, identifying faces may be based on activity in the face region in the temporal lobe (see Chapter 2), but it also depends on activity in regions that respond to emotions, facial expressions, where the face is looking, and the face’s attractiveness. Similarly, the response to a hammer activates visual regions that respond to the hammer’s shape and color, but it also causes activity in regions that respond to how a hammer is used and to a hammer’s typical motions. This idea that some objects, like hammers, cause activity in regions of the brain associated with actions, brings us to the embodied approach to categorization.

Question 50

In addition, we described an experiment by Alex Huth and colleagues (2012),

Answer 50

which resulted in the map in Figure 2.21, showing where different categories are represented in the cortex. This “category map” was determined by measuring functional magnetic resonance imaging (fMRI) responses as participants were viewing films and determining how individual voxels responded to objects in the films. But semantic categories come into play not only when we look at a scene but also when we listen to someone speaking. Understanding spoken language involves not only knowing about concrete categories like living things, food, and places, but also about abstract concepts like feelings, values, and thoughts. To create a map based on spoken language, Huth and colleagues (2016) used a procedure similar to their earlier experiment, but instead of having their participants view films, they had them listen to more than two hours of stories from The Moth Radio Hour broadcast while in a brain scanner. Figure 9.25a shows a map that extends over a large area of the cortex, which indicates were specific words activate the cortex. Figure 9.25b zooms in to make some of the words easier to read. Figure 9.26 shows the cortex color-coded to indicate where different categories of words activate the cortex. For example, the light region in the back of the brain is activated by words associated with violence. The words that activated a single voxel in that region are indicated on the right. Another voxel, which is activated by words associated with visual qualities, is shown in the green region near the top of the brain. An interesting aspect of Huth’s results is that the maps were very similar for each of the seven participants.

Question 51

The Embodied Approach

Answer 51

Proposal that our knowledge of concepts is based on reactivation of sensory and motor processes that occur when we interact with an object. According to this idea, when a person uses a hammer, sensory regions are activated in response to the hammer’s size, shape, and color, and, in addition, motor regions are activated that are involved in carrying out actions involved in using a hammer. When we see a hammer or read the word hammer later, these sensory and motor regions are reactivated, and it is this information that represents the hammer (Barsalou, 2008). We can understand the basis of the embodied approach by returning to Chapter 3, where we described how perception and behavior interact, as when Elena reached across the table to pick up a cup of coffee. The important message behind that example was that even simple actions involve a back-and-forth interaction between pathways in the brain involved in perception and pathways involved in behavior (Almeida et al., 2014). Also in Chapter 3, we saw how mirror neurons in the prefrontal cortex fire both when a monkey performs an action and when the monkey sees the experimenter performing the same action (review pages 90–92; Figure 3.35). What do mirror neurons have to do with concepts? The link between perception (a neuron fires when watching the experimenter pick up the food) and motor responses (the same neuron fires when the monkey picks up the food) is central to the embodied approach’s proposal that thinking about concepts causes the activation of perceptual and motor regions associated with these concepts.

Question 52

Evidence for this link between perceptual and motor responses in the human brain is provided by an experiment by Olaf Hauk and colleagues (2004)

Answer 52

who measured participants’ brain activity using fMRI under two conditions: (1) as participants moved their right or left foot, left or right index finger, or togue; (2) as participants read “action words” such as kick (foot action), pick (finger or hand action), or lick (tongue action). The results show regions of the cortex activated by the actual movements and by reading the action words. The activation is more extensive for actual movements, but the activation caused by reading the words occurs in approximately the same regions of the brain. For example, leg words and leg movements elicit activity near the brain’s centerline, whereas arm words and finger movements elicit activity farther from the centerline. This correspondence between words related to specific parts of the body and the location of brain activity is called semantic somatotopy .

Question 53

Although there is convincing evidence linking concepts and activation of motor regions in the brain, some researchers question whether the embodied approach offers a complete explanation of how the brain processes concepts

Answer 53

For example, Frank Garcea and colleagues (2013) tested patient A.A., who had suffered a stroke that affected his ability to produce actions associated with various objects. Thus, when A.A. was asked to use hand motions to indicate how he would use objects such as a hammer, scissors, and a feather duster, he was impaired compared to normal control participants in producing these actions. According to the embodied approach, a person who has trouble producing actions associated with objects should have trouble recognizing the objects. A.A. was, however, able to identify pictures of the objects. Garcea and colleagues concluded from this result that the ability to represent motor activity associated with actions is not necessary for recognizing objects, as the embodied approach would predict. Another criticism of the embodied approach is that it is not well suited to explaining our knowledge of abstract concepts such as “democracy” or “truth.” However, proponents of the embodied approach have offered explanations in response to these criticisms

Question 54

Summarizing the Approaches

Answer 54

Our survey of how concepts are represented in the brain began with the sensory-functional approach, based on neuropsychological studies begun in the 1980s. But once it became clear that things were more complex than the distinction between sensory and functional, research led in different directions that resulted in more complex hypotheses One thing that all these approaches agree on is that information about concepts is distributed across many structures in the brain, with each approach emphasizing different types of information. The multiple-factor approach emphasizes the role of many different features and properties. The category-specific approach emphasizes specialized regions of the brain and networks connecting these regions, and the embodied approach emphasizes activity caused by the sensory and motor properties of objects. It is likely that, as research on concepts in the brain continues, the final answer will contain elements of each of these approaches.

Question 55

The Hub-and-Spoke Model

Answer 55

A model of semantic knowledge that proposes that areas of the brain that are associated with different functions are connected to the anterior temporal lobe, which integrates information from these areas. The ideas we have been discussing about how concepts are represented in the brain have been based largely on patients with category-specific memory impairments. However, there is another type of problem, called semantic dementia , which causes a general loss of knowledge for all concepts. Patients with semantic dementia tend to be equally deficient in identifying living things and artifacts. Evidence supporting the idea of a hub with spokes is that damage to one of the specialized brain regions (the spokes) can cause specific deficits, such as an inability to identify artifacts, but damage to the ATL (the hub) causes general deficits, as in semantic dementia (Lambon Ralph et al., 2017; Patterson et al., 2007). This difference between hub-and-spoke functions has also been demonstrated in participants with healthy brain function using a technique called transcranial magnetic stimulation (TMS) .