Running head: CATEGORIZATION AND DISCRIMINATION

Correspondences should be sent to: Dr. Robert Goldstone

Psychology Department

Indiana University

Bloomington, IN. 47405

812-855-4853

rgoldsto@ucs.indiana.edu

Four experiments investigated the influence of categorization training on perceptual discrimination. Ss were trained according to 1 of 4 different categorization regimes. Subsequent to category learning, Ss performed a Same/Different judgment task. Ss' sensitivities (d's) for discriminating between items that varied on category-(ir)relevant dimensions were measured. Evidence for acquired distinctiveness (increased perceptual sensitivity for items that are categorized differently) was obtained. One case of acquired equivalence (decreased perceptual sensitivity for items that are categorized together) was found for separable, but not integral, dimensions. Acquired equivalence within a categorization-relevant dimension was never found for either integral or separable dimensions. The relevance of the results for theories of perceptual learning, dimensional attention, categorical perception, and categorization are discussed.

Psychologists have long been intrigued by the possibility that the concepts that people learn influence their perceptual abilities. It may be that the way that people organize their world into categories alters the actual appearance of their world. The purpose of the research reported here is to investigate influences of concept learning on perception.

The notion that experience and expectations can influence perception can be traced back to the "New Look" movement of the 40's and 50's (Bruner & Postman, 1949). Evidence suggests that experts perceive structures in x-rays (Norman, Brooks, Coblentz, & Babcook, 1992), beers (Peron & Allen, 1988), and infant chickens (Biederman & Shiffrar, 1987) that are missed by novices. As the experts in these fields learn to distinguish between the concepts in their domain (types of fractures, brands of beer, or gender of chicken) they seem to acquire new ways of perceptually structuring the objects to be categorized.

This suggestion that categorization causes changes to perceptual abilities is not implicated in most traditional accounts of concept learning. In Bruner, Goodnow, and Austin's (1956) classic studies on concept learning, subjects saw flash cards with shapes, and were required to learn rules such as "If the flash card has a circle or something black then it belongs in Category A" on the basis of feedback provided by the experimenter. Although work in concept learning has come a long way since Bruner, Goodnow, and Austin (Estes, 1986; Kruschke, 1992; Nosofsky, 1986; Medin & Schaffer, 1978; Reed, 1972), vestiges of this earlier work are apparent in current research. Specifically, many researchers investigate concept learning using stimuli that have clear-cut dimensions with clearly different values on these dimensions. Although mandatory in many cases for reasons of experimental control and precision, such stimuli do not require subjects to perceptually learn new dimensions or finer discriminations. In the currently described concept learning tasks, subjects must make fine discriminations along dimensions, or isolate dimensions that are normally fused together. In both cases, the perceptual abilities required for the categorization task are not at a ceiling level before categorization training begins, and consequently, experience with categorization may drive perceptual learning.

Perceptual Learning

Although most concept learning work has not dealt with the development of new perceptual abilities from experience (but see Norman, Brooks, & Allen, 1989; Wisniewski & Medin, in press), this topic has been addressed in other literatures. Most influential perhaps has been Gibson's (1969) treatment of perceptual learning, the process by which there is "an increase in the ability to extract information from the environment, as a result of experience and practice with stimulation coming from it" (p. 3). Gibson provided several demonstrations that people can increase their perceptual sensitivity by categorizing or identifying stimuli. One type of perceptual learning, called "predifferentiation" (Gibson, 1991) or "preexposure effect," (Hall, 1991) entails heightened perceptual sensitivity following exposure to the tested materials. Simply preexposing subjects to stimuli often facilitates their later discriminations between the stimuli. For example, Gibson and Walk (1956) placed cut-out shapes in the cages of some rats, but not others. Although no response was required to the shapes, animals that were raised with the shapes in their cages were better able to learn a subsequent discrimination that used the shapes as relevant cues. Similar results have been found with human subjects (Goss, 1953; Vanderplas, Sanderson, & Vanderplas, 1964). Similarly, Gibson and Gibson (1955) showed that simple practice in identifying visual "scribbles" increased subjects' identification performance, even when no feedback was provided. The existence of predifferentiation effects indicates that differential rewards among cues to be discriminated are not required for facilitation of perceptual discrimination. Mere experience with the stimuli is sufficient.

However, there is also evidence that subjects become perceptually attuned to diagnostic physical features that facilitate discrimination between presented stimuli (e.g. Gardner & Wallach, 1965; Gibson & Gibson, 1955). For this class of experiments, the organization of the stimuli into categories has an influence on subsequent discrimination. Lawrence (1949) developed a theory of acquired distinctiveness of cues, according to which cues that are relevant for a task become generally distinctive. In one experiment, Lawrence trained rats on either a black/white or a rough/smooth discrimination. Rats received a reward for choosing one stimuli rather than another. Subsequently, rats were transferred to a discrimination in which, for example, when black shapes were presented, the rat was rewarded for a left response, and when white shapes were presented, the rat was rewarded for a right response. Rats learned this second discrimination better when they had been trained earlier to make a black/white discrimination. Stimuli also acquire "non-distinctiveness" (or "equivalence"); when cues are irrelevant for an earlier discrimination, there is a deleterious effect on subsequent discrimination learning with them (Waller, 1970). Both of these effects are commonplace in human subjects (Gibson, 1969).

Cautions from Perceptual Learning work.

It is important to distinguish transfer to two types of discrimination experiment: associative and perceptual. Lawrence's experiment pertains to associative discrimination; the task is to learn to associate a particular cue with a particular response. Perceptual discrimination experiments involve simply distinguishing between two perceptual cues. For example, in a perceptual same/different judgment task, subjects are shown a pair of stimuli and must decide whether the stimuli are exactly the same or not. Perceptual sensitivity in a same/different judgment task can be measured by comparing, in accordance with signal detection theory (Swets, Tanner, & Birdsall, 1961), the probability of correctly calling a pair "different" with the probability of incorrectly calling a pair of identical stimuli "different." Although most of the work testing the theory of acquired distinctiveness has used associative discrimination tasks, it is important to use perceptual discrimination tasks if the goal is to find an influence of stimulus categorization on perception per se. Findings of acquired distinctiveness that use associative discrimination do not necessarily require changes to perceptual abilities. In fact, Hall (1991) reviews the evidence used to support perceptual learning using associative discrimination paradigms and concludes that there is very little reason to postulate perceptual changes rather than the acquisition of associations between stimulus cues. For example, Lawrence's results can be accommodated if one assumes that stimulus cues that determine correct responses are more likely to be associated with other responses. However, simple associationist theories are less likely to be relevant in cases where perceptual sensitivity is measured by same/different judgments. One of the purposes of the reported experiments is to find evidence for perceptual, not simply associative, changes due to training.

Although same/different judgments will be used as the measure of perceptual sensitivity, the claim will not be made that they are strictly perceptual judgments, if "perceptual" is interpreted narrowly. Same/different judgments involve a memory component (Pisoni, 1973). Often, the two items that are compared are not presented simultaneously. Even if there is a only a short delay between items, a short-term memory encoding of the first presented item is required. It might even be argued that if the items are simultaneously presented, there is still a need for an item to encoded in memory before it can be compared to the other item if the items are spatially separated. Same/different judgments also involve attentional mechanisms. When viewing an item, subjects may attend to some aspects and ignore others. Contextual, instructional, and motivational variables may influence what stimulus aspects attract attention. Thus, although the same/different judgment task will be used here as a measure of perceptual sensitivity, higher-level cognitive functions such as memory and attention also have an influence on same/different judgments. In fact, one of the conclusions from the reported research will be that a clear distinction between sensory and cognitive processes is not tenable (see also Algom, 1992; Goldstone, 1993; Marks, 1992).

Categorical Perception

The largest bulk of evidence that is used to support the contention that concepts influence perception comes from the field of categorical perception (see Harnad, 1987 for reviews and recent research). According to the phenomenon of categorical perception, people are better able to distinguish between physically different stimuli when the stimuli come from different categories than when they come from the same category. The effect has been best documented with speech phoneme categories. For example, Liberman, Harris, Hoffman, and Griffith (1957) generated a set of vowel-consonant syllables going from /be/ to /de/ to /ge/ by varying the onset frequency of the second-formant transition of the initial consonant (the details are not important for the present purpose). The fourteen speech sounds were created by making equal physical spacings between neighboring sounds. Observers listened to three sounds - A followed by B followed by X, and indicated whether X was identical to A or B. Subjects performed this task more accurately when syllables A and B belonged to different phonetic categories than when they were physical variants of the same phoneme, even when the physical difference between A and B was equated. Liberman et al concluded that the phonemic categories possessed by an adult speaker of English influence the perceptual discriminations that can be made.

The degree to which categorical perception phenomena are learned rather than innate is not clear (Pastore, 1987; Rosen & Howell, 1987). On the one hand, it appears that discriminability in some regions of acoustical continua is higher than in other regions, irrespective of category structure. Phoneme categories may naturally use regions with intrinsically higher discriminability as boundaries (Kuhl & Miller, 1981; Stevens, 1981). Consistent with an innatist perspective are the results from Eimas (1974; Eimas, Siqueland, Jusczyk, and Vigorito, 1971) that show that infants of only four months show increased sensitivity to acoustical differences in the same region of a physical continua as do adults. Thus, there is evidence that suggests that people's increased sensitivity to acoustical differences that straddle category boundaries may either be innate or a property of the acoustical signal, rather than learned.

However, there is also evidence that the categorical perception effect is subject to learning (Logan, Lively, & Pisoni, 1991). Lane (1965) found categorical perception effects for laboratory-created materials that were placed in different categories by their labels, even though the categories did not correspond to a naturally occurring distinction. Burns and Ward (1978) found that expert but not novice musicians showed a categorical perception effect for pitch differences, suggesting that training was instrumental in creating differential sensitization along semitone boundaries. Finally, cross-cultural evidence suggests that the learning of a particular language influences the pattern of discriminability between speech sounds. In general, a sound difference that crosses the boundary between phonemes in a language will be more discriminable to speakers of that language than to speakers of a language in which the sound difference does not cross phonemic boundaries (Repp, 1984; Strange & Jenkins, 1978).

Another issue in categorical perception research concerns whether it is a general perceptual effect, or whether it is only found for language-related stimuli. Liberman, Harris, Kinney, and Lane (1961) originally argued that categorical perception is found for speech-like stimuli, but not for control stimuli that do not sound like speech (also see Miyawaki, Strange, Verbrugge, Liberman, Jenkins, & Fujimura, 1975). However, since then, other researchers (Cutting, 1982; Pisoni, 1977; also see Bornstein, 1987) have found categorical perception effects for non-speech materials.

The categorical nature of categorical perception has been called in doubt. Researchers (Pastore, 1987; Pisoni, 1977) have argued that discrimination for physical differences within a category is not at chance, as would be expected by a theory that stated that discrimination is based entirely on category membership. Massaro (1987; Masarro & Cohen, 1983) has developed a Fuzzy Logical Model of Perception (FLMP) that produces results that others have taken to be indicative of categorical perception, even though the model assumes completely continuous perceptual information. In specific, FLMP predicts sharp identification boundaries between categories. In FLMP, continuous perceptual information from different dimensions is integrated together, and classification of an item depends on the relative similarity of the perceptual information to each of the candidate categories. Thus, the simple presence of sharp boundaries between categories is not sufficient to conclude that perceptual dimensions are perceived categorically or even nonlinearly.

Research on attention to dimensions in categorization is also relevant for the current work. Virtually all of the work on categorical perception concerns differential patterns of perceptual sensitization within a dimension (for an exception, Massaro, 1987, considers how two dimensions interact to obtain effects similar to categorical perception). Much of the work in category learning concerns situations where multiple dimensions are relevant and varied.

One issue in the area of dimensional attention in categorization concerns whether entire dimensions are sensitized when they are relevant for a categorization (Nosofsky, 1986; Kruschke, 1992), or whether local regions of a dimension can be selectively sensitized. For example, consider applying Nosofsky's GCM model to objects that vary on size and hue. Suppose that objects that are 1 or 2 cms belong to Category 1, and objects that are 3 or 4 cms belong to Category 2. According to GCM, the "attentional weight" given to the size dimension will be greater than the attentional weight given to hue because of the relevance of size for the categorization. Attentional weights refer to the importance of a dimension in a categorization decision. No assumption is made that different attentional weights correspond to perceptual differences (Nosofsky, 1987). Because entire dimensions are weighted, the difference between 2 cm and 3cm objects will become particularly salient, but so too will the difference between 1 cm and 2 cm objects. That is, the difference between objects that fall in the same category but differ along a categorization-relevant dimension will become more important for determining categorization.

Aha and Goldstone (1990) predict that local regions of a dimension may become selectively attended. With their model, as applied to the above example, size categorization may selectively highlight the difference between 2 cm and 3 cm, without having much influence on the difference between 1 cm and 2 cm. Aha and Goldstone (1992) found empirical support for such a local selective attention effect. The design and partial results from one experiment are shown in Figure 1. Subjects categorized objects that varied on two dimensions: the size of a rectangle, and the position of a line within the rectangle. Six Category A and six Category B members were presented. After learning the categories, the subjects were shown new stimuli to categorize. The numbers in the grid show the percentage of time that a particular object, defined by its value on the two dimensions, was placed into Category B. The results indicate that subjects learn to selectively weight size when an object has a high value on size, and to selectively weight position of line when an object has a low value on size. Thus, people appear to be able to, at times, heighten the attention paid to just a local region of a dimension. It is not clear whether this attention is actually perceptual or simply a change to the categorization strategy.

One aim of the current research is to determine whether the perceptual changes that occur from category learning, if they occur at all, are limited to the particular local region that is the boundary between categories, or whether the perceptual changes generalize to other values on the categorization-relevant dimension. The answer to this question, in addition to being relevant to the above models of categorization, also bears on explanations of categorical perception. Typical explanations of categorical perception assume that local regions of a dimension can be selectively sensitized to a greater extent than other regions of the same dimension, in accordance with Aha and Goldstone's (1992) results. If entire dimensions must be attended without differential attention placed on select regions/values of a dimension, then a learning explanation of categorical perception is not possible.

____________________________

Insert Figure 1 about here

____________________________

An issue that arises with stimuli that vary on more than one dimension is whether dimensions compete for attention. In Nosofsky's GCM model of categorization, the sum of attentional weights is constrained to equal 1.0. Thus, if more attention is placed on one dimension, it must be removed from another dimension. On the other hand, it is possible that as the attention paid to one dimension increases, that the perception of other dimensions does not suffer. Again, the assumption of GCM does not necessarily extend to perceptual attention. One of the purposes of the current experiments is to determine the circumstances under which dimensions compete for perceptual attention.

Some research suggests that the degree to which dimensions compete for attention may depend on the nature of the dimensions tested. Garner (1974) distinguished between integral and separable dimensions. In general, with integral dimensions it is relatively difficult to attend to one dimension without attending to the other; with separable dimensions, this is relatively easy. Integral dimensions, such as the saturation and brightness of a color (Burns & Shepp, 1988), are argued to be psychologically "fused." Typically, variations along one integral dimension interfere with the processing of the other integral dimension (Garner, 1974). Categorization judgments of stimuli that vary on integral dimensions tend to made on the basis of overall (Burns & Shepp, 1988; Smith, 1979), Euclidean distance (Shepard, 1964) similarity. Conversely, separable dimensions, such as the brightness and size of a shape (Gottwald & Garner, 1975), are perceptually independent. Variations along one separable dimension cause no interference to the processing of the other dimension. Categorization judgments with separable dimensions tend to focus on particular dimensions, and to use city-block similarity.

Given this characterization, we might expect integral and separable dimensions to have different patterns of perceptual competition. It might be argued that if integral dimensions are used and Dimension X is relevant for a categorization, sensitization might spread over to Dimension Y. Integral dimensions are assumed to be similar/close to each other in an abstract space (Melara, 1992), and if the focus of attention is placed on one dimension, unless the focus of attention is very narrow, it will cover the other dimension as well. Assuming that separable dimensions are further separated, this view predicts less spread of sensitization for separable dimensions. On the other hand, it is also possible to predict the opposite pattern of results. During category learning it might be argued that subjects will learn to actively filter out or ignore irrelevant dimensions to the extent that they intrude on category decisions. Because integral dimensions intrude on each other more than do separable dimensions (Garner, 1974), it might be thought subjects will become particularly desensitized to irrelevant integral dimensions.

The current experiments address the above issues in perception and perceptual attention. The broadest question addressed is: Does categorization training alter perceptual judgments measured by a same/different task? If training does influence perceptual judgments, several more refined questions can be asked:

1. Do dimensions acquire equivalence or distinctiveness? A categorical perception effect, defined as elevated discriminability for items that straddle a category boundary relative to control items that fall into one category, could arise from acquired equivalence or distinctiveness. According to acquired equivalence, there is a decrease in perceptual sensitivity to differences that are not relevant for a categorization (Pearce & Hall, 1980). According to acquired distinctiveness, there is an increase in perceptual sensitivity to differences that are relevant for a categorization (Gibson, 1969; Miller & Dollard, 1946). Figure 2, from Pisoni (1991), distinguishes between these two theories. Panel 2A, depicting acquired similarity (aquired equivalence), shows that at one time (the white circles), all stimuli are well discriminated. At a later time (the black circles), stimuli that are at the boundary between two categories (the boundary is assumed to fall at Stimulus 4) retain their high discriminability while the other stimuli lose their discriminability. Panel 2B, depicting acquired distinctiveness, shows a final pattern of discriminability that is identical to Panel 2A, but achieves this result by elevating the discriminability of the boundary stimuli rather than by depressing the discriminability of the non-boundary stimuli.

____________________________

Insert Figure 2 about here

____________________________

In the domain of speech perception, claims have been made that acquired equivalence may underlie the categorical perception effect (Pisoni, 1991). Experiments have found infants to be able to make discriminations between speech sounds in a language that belong to the same phoneme category in adult speakers of the language (Eimas, Miller, & Jusczyk, 1987; Eimas, Siqueland, Jusczyk, & Vigorito, 1971). Others have found that very young infants (2 mo. old) show sensitivity to differences between speech sounds that they lose by the age of 10 mo. (Werker & Tees, 1984). However, one should be wary of the conclusion from these studies that acquired equivalence underlies categorical perception. First, the strategy of comparing adult to infant performance is far from perfect because it is difficult to test adults and infants with similar methods, adults and infants differ in many ways besides their experience with speech, and obtaining accurate measures of infants' perceptual abilities is subject to many potential obstacles. Although infants show above chance ability to discriminate sounds within a phoneme category, so do adults (Pisoni, 1973). Thus, it would be important to quantitatively compare perceptual sensitivities between adults and infants, but this is pragmatically precluded due to difficulties in testing infants with standard discrimination paradigms. In the currently reported experiments, a cleaner comparison is achieved by presenting different groups from a single population (undergraduate students at Indiana University) with the relevant categories.

____________________________

Insert Figure 3 about here

____________________________

Although many theories have proposed either acquired equivalence or distinctiveness, Hall (1991) notes that few if any experiments have had an appropriate design to tease these two processes apart. Figure 3 abstractly shows a set of stimuli that can distinguish between these processes. Subjects are given different categorization rules; for one group, size is relevant; for another group, brightness is relevant; another group receives no categorization training. For example, a subject assigned to the "size categorization" group would receive feedback indicating that any object with a size dimension value of 1 or 2 belongs to Category A, and any object with a size value of 3 or 4 belongs to Category B. After prolonged categorization training, subjects are presented with pairs of adjacent objects (or with the identical object repeated twice) and are asked to indicate whether the objects are exactly identical or not.

It is possible to obtain evidence for acquired equivalence along a categorization-relevant or -irrelevant dimension. Evidence for acquired equivalence of a categorization-irrelevant dimension exists, for example, if the size categorizers are worse at making brightness discriminations than are subjects who had no categorization training at all. Evidence for acquired distinctiveness of a categorization-irrelevant dimension exists if the size categorizers are better at making size discriminations than are "no categorization" subjects.

Evidence for acquired equivalence of a categorization-relevant dimension exists if the size categorizers are worse at making size discriminations between objects with size values of 1 and 2 than are the "no categorization" control subjects. That is, acquired equivalence of a categorization-relevant dimension occurs if discrimination among values of Dimension X is impaired when categorization depends on Dimension X and the values being discriminated belong to the same category.

By comparing categorization groups to a control group that performs no categorization, acquired equivalence and distinctiveness are teased apart. Thus, it is possible to gather evidence for either acquired equivalence or acquired distinctiveness, both, or neither.

2. Must entire dimensions be perceptually sensitized or can regions within a dimension be sensitized? If entire dimensions must be perceptually sensitized, then acquired equivalence in a categorization-relevant dimension should not be found. In fact, the opposite effect would be predicted. Size categorizers' ability to distinguish between size values of 1 and 2 (in Figure 3) should exceed that of subjects with no categorization training, even though size values of 1 and 2 both belong to Category A for size categorizers. If particular regions of a dimension can be sensitized, then size categorizers should be able better discriminate between values 2 and 3 than they can distinguish between values 1 and 2. It is possible that evidence could be obtained in favor of both general sensitization of a dimension and specific sensitization of one region, because the former hypothesis involves comparison to the control group, while the latter claim involves comparing sensitivities within a categorization condition.

3. Does the degree of competition between dimensions for perceptual attention vary with the integrality/separability of the dimensions? Competition between dimensions can be measured in two ways. First, the extent or presence of acquired equivalence along a dimension that is irrelevant for categorization may be related to the ability of subjects to attend to both the relevant and irrelevant dimensions. if acquired equivalence along the irrelevant dimension occurs, then this can be taken as evidence that the requirement to attend a relevant dimension causes the other dimension to become desensitized.

Competition between dimensions can also be measured by including a fourth category training condition in which subjects must use information from both dimensions in order to make the categorization. In this condition, four categories are learned, one for each quadrant of four stimuli in Figure 3. This condition can be compared to the other categorization conditions to determine whether attention, when divided across two dimensions, results in less perceptual sensitization than when it is focused on one dimension. Such competition would exist if, for example, size categorizers are better able to make size discriminations than are subjects who must use both size and brightness to categorize. If such competition exists, the refined question can be asked: Is this competition greater for integral or separable dimensions?

The stimuli for Experiment 2 are shown in Figure 3. Sixteen squares are constructed by factorially combining four values on two dimensions (size and brightness). In Experiment 1, the psychological differences between adjacent stimuli in Figure 2 are roughly equated. This is not strictly necessary in order to test perceptual learning effects because learning is assessed by comparing the perceptual sensitivities of a categorization group to the sensitivities of a control group. By this technique, discriminability differences between pairs of items should not affect the measure of learning as long as it influences both the experimental and control groups equally. However, pragmatically, it is useful to scale the stimuli because it increases the sensitivity of the experiment.

Consequently, Experiment 1 is a preliminary "stimulus customization" experiment conducted in order to determine the correct appearance for each of the 16 squares in Figure 3. On some of the trials, two identical squares are shown; on the other trials, two adjacent squares are shown. Subjects respond as to whether the two stimuli are exactly the same or different in their sizes or colors. When the squares are actually different but the subject responds "Same," the square with the larger value on the varying dimension increases its value on this dimension by a small amount. With this method, the next time that the same two squares are compared, it will be somewhat easier to distinguish them. When the squares are different and the subject responds "Different," the square with the larger dimension value decreases its value on this dimension by the same small amount, making it somewhat more difficult to distinguish the squares on subsequent trials. After many adjustment trials, the set of squares will be scaled such that each pair of adjacent squares is approximately equally discriminable.

Method

Subjects. Forty-five undergraduate students from Indiana University served as subjects in order to fulfill a course requirement.

Materials. Sixteen squares that varied in their size and brightness were displayed on Macintosh IIsi screens. The sixteen squares were obtained by factorially combining four values of size with four values of brightness. The exact values for size and brightness were varied during the procedure.

Procedure. Subjects were instructed that they would see a pair of squares on the screen that might or might not slightly differ on their size or the brightness of their color. Subjects were instructed to press the "S" key on the keyboard if they believed the two squares to be physically identical, and to press the "D" key if they believed the two squares to differ along either dimension. They were also told that the discriminations that they would be required to make would become more difficult as the experiment continued.

The two squares were presented successively on the screen. The exact vertical and horizontal locations of the two squares was randomized, under the constraint that the first presented square occupied the left portion of the screen, and the second presented square occupied the right portion of the screen. No square was shown within a 5 cm vertical column in the center of the screen. At the beginning of a trial, the first square was displayed for 1000 msec. The square was then removed, and a blank screen was presented for 33 msec. The second square was then displayed for 1000 msec. The second square was then replaced with a blank screen until the subject made a "S" or "D" response. The interval between trials was 1500 msec. Subjects made 576 judgments in all.

The 576 judgments were divided into 18 blocks. Each block consisted of 8 "same" trials and 24 "different" trials, randomly ordered. For the "same" trials, one of the sixteen squares abstractly shown in Figure 3 was shown to subjects, and was repeated as the second square. For the 24 "different" trials, two vertically or horizontally adjacent squares from Figure 3 were selected as the squares to be compared. There are 24 different pairs of immediately adjacent squares. Diagonally adjacent squares were not tested.

The distances between adjacent squares were "customized" for each subject by changing individual squares' appearances. No changes were made on "same" trials. On "different" trials, if a subject correctly responded "different," the larger/brighter square was assigned a slightly smaller size/brightness value. If a subject incorrectly responded "same" when the squares were different, the larger/brighter square was assigned a slightly larger size/brightness value. In this manner, subjects will eventually have an accuracy of 50% on "different" trials. However, subjects accuracy on "same" trials will be higher, yielding d' scores that are above zero. For brightness changes, the constant adjustment increment was 1200 brightness units in Macintosh's hue-saturation-brightness representation for colors (the final standard CIE color coordinates and lengths in cm will be presented later). For size changes, the constant adjustment increment was 0.07 cm.

In addition to changing the size/brightness of the larger/brighter square, it is also necessary to adjust the sizes of all of the squares that are brighter/larger than this square. If this were not done, then making the square more similar to a smaller/darker square would also make the square much less similar to larger/brighter squares. Consequently, the same absolute adjustment (in size or brightness units) that is made to the larger/brighter of the compared squares is also made to all squares that have still larger/brighter values on the dimension of variation between the compared squares. With this method, the absolute differences between other adjacent pairs of objects remains fixed. Initial dimension values for the sixteen squares were set to plausible, fairly discriminable values.

Results

Subjects required about 55 minutes in order to complete the same/different judgments. The average values along the size and brightness dimension for each of the 16 squares are shown in Table 1. The values indicate a slight interaction between size and brightness. In particular, the larger the object, the less different brightness values needed to be in order to yield the same discriminability.

_______________________

Insert Table 1 about here

_______________________

For the purposes of Experiment 2, it is necessary to calculate dimension values as if there were no dimension interactions. The height (and width) of the four size values in centimeters are: 2.85, 3.12, 3.45, and 3.79. These values are roughly consistent with Weber's law; bigger absolute differences are required to produce the equal psychological difference as the magnitude of the stimuli increases. The CIE (1976 model) color coordinates and luminance (in candelas per meter squared) as measured by a Spectra Scan 714 chromometer were: {X=.5463, Y=.3696, luminance = 21.62}, {X=.5405, Y=.3730, lum.=20.09}, {X=.53, Y=.3737, lum.=18.79}, and {X=.5266,Y=.3764, lum.=16.17}.

Discussion. It is not a complete surprise that brightness and size values should interact such that finer discriminations in brightness are easier to make when the sizes of the compared objects are relatively large. Although size and brightness are separable dimensions in that there is no interference of one dimension on another in a speeded sorting task (Gottwald & Garner, 1975), they fail other tests of separable dimensions. For example, Biederman & Checkovsky (1970) found that size and brightness produced a redundancy gain. Subjects were faster to make judgments about materials that varied in a correlated fashion on size and brightness than judgments about materials that varied only on size or brightness. The results support the notion that different types of stimulus integrality are experimentally dissociated. The discriminability of differences on one dimension is influenced by the value along another dimension, even though previous results indicate that attention can be selectively focused on only one of the two dimensions. Perceptual interactions between dimensions do not necessarily entail a failure to selectively attend (see also Ashby & Townsend, 1986)

In Experiment 2, stimuli that are horizontally or vertically adjacent (in Figure 3) only vary on one dimension. Two squares that are horizontally adjacent must vary only on size and not brightness. If this condition were violated, it would be impossible to tell whether size or brightness differences were used to make a particular perceptual discrimination. Thus, the brightness and size values obtained from Experiment 1, averaging out the slight dimensional interaction, will be used in Experiment 2.

Experiment 2 addresses the three questions raised at the end of the introduction using the stimuli scaled in Experiment 1. The first part of Experiment 2 involves training subjects to make a single- or a double-dimension categorization. The second part measures subjects' sensitivity at making a perceptual same/different judgment.

Method

Subjects. Eighty-four undergraduate students from Indiana University served as subjects in order to fulfill a course requirement. The subjects were evenly split into the four conditions.

Materials. The sixteen squares scaled in Experiment 1 were used as stimuli. The stimuli formed a 4X4 matrix (see Figure 3) with every brightness level combined with every size level. The scaling results, after averaging across dimension values to eliminate dimension interactions, guarantees that horizontally and vertically adjacent squares will differ on only their size or brightness, and that each pair of adjacent squares will be approximately equally discriminable for control subjects.

Procedure. Subjects were given an approximately 60 minute categorization task (except for the control group) followed by a forty minute perceptual discrimination task. Both tasks were completed in the same session.

There were four categorization conditions: size categorizers, brightness categorizers, size and brightness categorizers, and controls. The control subjects did not undergo any categorization training. For the size categorizers, the squares in the left two columns of Figure 3 were assigned to Category A, and the squares in the right two columns were assigned to Category B. For the brightness categorizers, the squares in the upper two rows were assigned to Category A, while the squares in the lower two rows were assigned to Category B. For the size and brightness categorizers, four categories, corresponding to the four quadrants (upper-left, upper-right, lower-left, and lower-right), each included four squares.

In the categorization training, 20 repetitions of the 16 squares were shown. On an individual trial, a square was shown in a randomly generated location on the screen. The square remained on the screen until the subject pressed a key corresponding to their guess as to the square's category. Category responses were made by pressing the keys "1," "2," "3," and "4"; only the first two responses were required for size categorizers and brightness categorizers. After a response was made, feedback was given as to the correctness of the response, and the correct category label was displayed. After 1.5 sec, the screen was erased, and after another 1 sec, the next trial began. Subjects were instructed that the squares would vary on brightness and size.

All four categorization training groups received the identical subsequent discrimination experiment. Subjects were shown pairs of adjacent squares (or the identical square repeated twice) and responded either "same" or "different." The procedure exactly followed the procedure used in Experiment 1, with two exceptions. First, the initial values for the dimensions were set to the values determined in Experiment 1, rather than arbitrary initial values. Second, there was no modification of the squares' size and brightness values. The number and distribution of "same" and "different" trials was identical to Experiment 1.

Results

The most important results are related to the relative ability of each of the four categorization groups to discriminate between different pairs of squares in the same/different judgment portion of the experiment. The measure d' is computed as an indicator of perceptual sensitivity (Swets, Tanner, & Birdsall, 1961) that is based on subjects' ability to correctly discriminate between different stimuli (subjects' probability of responding "different" when different squares are presented), adjusted by subjects' "false alarm" rate (subjects' probability of responding "different" when identical squares are presented). By comparing the perceptual sensitivity (d') of subjects who undergo category learning with the d' of those subjects who do not, we can assess the nature of the influence that category training has on perceptual "same/different" judgments.

____________________________

Insert Table 2 about here

____________________________

The results from all four categorization training conditions are shown in Table 2. Correct response percentages are shown. For example, when the square in the upper left hand corner of Figure 3 is displayed twice in the size categorization condition, subjects correctly respond "same" 82% of the time. When these same subjects are shown the upper-left hand square compared to the square immediately to its right, they correctly respond "different" on 44% of the trials, yielding an average d' of 0.78. Although a correct response rate of 44% might seem to be less accurate than predicted by chance, the 84% "correct same" response rate shows a tendency for subjects to respond "same" more than "different." In fact, the greatest tendency to respond "same" over "different" is found in the "no categorization" control condition, followed by the size categorizers and brightness categorizers which did not significantly differ, followed by the size and brightness categorizers (F(3,20) = 7.4, p < .01, Fisher's post-hoc PLSD p < .01).

Many of the important questions involve the comparison of a categorization group's sensitivity (d') to the control group's sensitivity. Consequently, Figures 4, 5, and 6 show the difference between d' scores of subjects that are trained to make categorizations and d' scores of subjects that receive no categorization training. There is a rectangle for each of the 24 pairwise comparisons of squares. A black rectangle indicates a greater d' for the categorization condition than the control. A white rectangle indicates the opposite. The size of the rectangle indicates the absolute magnitude of the difference between the d's.

____________________________

Insert Figure 4 about here

____________________________

____________________________

Insert Figure 5 about here

____________________________

____________________________

Insert Figure 6 about here

____________________________

Acquired Distinctiveness. Dimensions acquire distinctiveness if perceptual discriminations along a categorization-relevant dimension are better for the categorization condition than for the control condition. Separate tests were conducted for size categorizers and brightness categorizers. For size categorizers, there are four comparisons that would unambiguously become sensitized if relevant dimensions acquire distinctiveness - the four comparisons that involve a square with a size value of 2 and a square with a size value of 3. In Figure 4, these four comparisons occupy the center column of horizontally extended rectangles. These four comparisons were averaged together for each subject in the size categorization and the control conditions. Overall, the size categorizers' d' for these four comparisons was greater than the control subjects' d's for the same comparisons (unpaired t(40)=3.6, p < .05). This difference can be clearly seen in Figure 4 from the fact that all four of the d' differences for the center size comparisons are large (wide) and positive (black).

The equivalent test for acquired distinctiveness for brightness categorizers compares brightness values of 2 and 3. These comparisons are found in the center row of vertically extended rectangles in Figure 5. The brightness categorizers' d' was greater than the control subjects' d's for these comparisons (t(40)=2.9, p<.05). Again, this is evident in Figure 5 from the long black rectangles along the center row. Thus, for both categorization groups, there is evidence for acquired distinctiveness between values of the categorization-relevant dimension that belong to different categories.

Acquired Equivalence of the Irrelevant Dimension. Acquired equivalence along a dimension that is irrelevant for a categorization occurs if discriminations along such a dimension are worse than they are for the control subjects. This occurs if, for example, size categorizers have a lower d' for trials that differ on brightness than do control subjects. Again, this hypothesis can be tested separately for the two single dimension categorization conditions. For size categorization subjects, there are 12 square comparisons that involve differences in brightness (the 12 vertically extended rectangles in Figure 4). Averaging over these comparisons, there is no significant difference between d' scores for the control and size categorization conditions (unpaired t(40) = 0.85, p > 0.1).

The analogous test of acquired equivalence in brightness categorizers involves the 12 horizontally extended rectangles in Figure 5. These data show a significant difference between the brightness categorizers and the control subjects (unpaired t(40)=2.3, p<.05). The d's are significantly greater for control subjects than for brightness categorizers, as shown by the preponderance of white horizontally extended rectangles in Figure 5.

Acquired Equivalence Within a Relevant Dimension. Acquired equivalence within a categorization-relevant dimension can occur if squares that have different dimension values on this dimension but belong to the same category become less discriminable as a result of categorization training. Again, this can be tested separately for both categorization conditions. For the size categorization condition, 8 relevant square comparisons are collapsed together - the left and right columns of horizontally extended rectangles in Figure 4. These rectangles span squares that differ on their sizes, but belong to the same category. Overall, the d' associated with these 8 squares is greater for the size categorization condition than it is for the control group (t(40) = 2.6, p < .05). This result is apparent in Figure 4 in that all but one of the rectangles in the first and third columns are black. This significant difference is in the opposite direction as hypothesized by acquired equivalence within a relevant dimension.

The analogous comparison for the brightness condition involves the 8 vertically extended rectangles that form the top and bottom rows of Figure 5. The d's associated with these comparisons are (marginally) significantly greater for the brightness condition than for the control group (t(40)=1.87, p=0.07). In Figure 5, 6 out of the 8 d's are greater for the brightness condition, as shown by the black rectangles.

Local Sensitization of a Dimension. Given that acquired equivalence was not found within the categorization-relevant dimension, it may or may not be the case that all values on a categorization-relevant dimension are equally sensitized. To test this, discriminations between the squares that belong to different categories can be compared to discriminations between squares that belong to the same category but vary on the categorization-relevant dimension. However, even though the psychological scaling conducted in Experiment 1 roughly equated the psychological differences between adjacent squares, the data in Table 2 for the control subjects indicate that the scaling was not perfect; for the control subjects, some pairs of squares are more discriminable than others. Thus, although the test for local versus dimension-wide sensitization is ideally a within-condition test, the results described here take into account the slightly different context-free discriminabilities of the pairs of squares. This is done by comparing two differences - the difference between control and categorization conditions on relevant dimension differences that straddle categories, and the difference between the two conditions on relevant dimension differences that remain in one category.

To test local sensitization for size categorizers, two sets of comparisons were formed. One set (the "critical value" set) contained the four comparisons that paired a square with a size value of 2 with a square with a size value of 3 (the middle column of horizontally extended rectangles in Figure 4). The other set (the "noncritical value" set) contained the other 8 eight comparisons involving squares that differed in their sizes. The d' scores for the size categorizers for each set were adjusted by subtracting from these scores the respective d' scores from the control condition. The adjusted d' scores for the critical set were significantly greater than the adjusted d' scores for the noncritical set (t(40) = 2.6, p < .05). In Figure 4, this is apparent in the relatively wider rectangles in the middle column than in the left and right columns.

The analogous sets of critical and noncritical comparisons were compared for the brightness categorizers. Again, the adjusted d' scores for the critical set were significantly greater than the adjusted d' scores for the noncritical set (t(40)=3.5, p < .05). This strong effect is shown clearly in Figure 4. The middle column of vertically extended rectangles contains very tall, black rectangles. The upper and lower columns are much shorter, and are not uniformly black.

Attentional competition. The final major question that was addressed by the results of Experiment 2 concerns the competition of dimensions for perceptual sensitization/attention. The specific question was: Does sensitization produced by categorization of Dimension X diminish if categorization by Dimension Y is also required? Once again, this question can be tested individually with size and brightness each occupying the Dimension X slot. For the size dimension, the ability of size categorizers to make size discriminations is compared to the ability of subjects who need to categorize on the basis of both size and brightness to make size discriminations. The overall d' for size discriminations by size categorizers is greater than the d' for the same items by the size and brightness categorizers (t(40) = 3.0, p<.05). This result can be seen by comparing the rectangles spanning horizontally adjacent squares in Figures 4 and 6. These rectangles are somewhat wider in Figure 4 than they are in Figure 6.

The analogous test of dimensional competition for brightness compares the 12 brightness discriminations for brightness categorizers to the same brightness discriminations for the size-and-brightness categorizers. Again, the overall d' is greater for the brightness condition (t(40)=2.4, p<.05).

The size-and-brightness categorizers develop sensitivities to both size and brightness (as evidenced by the prevalence of black rectangles in Figure 6). The overall d' for this group is higher than the d' for the control subjects (t(40)=3.4, p<.05). Still, the d's are not as high as they are for the single-dimension categorization conditions along the categorization-relevant dimension.

Other Results. Despite the fact that subjects did not receive feedback in the same/different judgment portion of the experiment, their perceptual sensitivities improved over the course of testing. There was a positive correlation of r=0.05, p<.05 between trial number and d' score. This is consistent with Gibson and Gibson's (1955) finding that false "old" judgments in an identification judgment decreased during practice without feedback, although the current result is somewhat more surprising in that the d' measure is an unbiased measure of perceptual sensitivity, and thus cannot be attributed to changes in decisional strategies.

Subjects in the size-and-brightness condition took longer to learn than the other two categorization conditions (F(2, 20) = 5.2, p<.05). In order to achieve a categorization accuracy of 90%, the size, brightness, and "size and brightness" groups required an average of 134, 113, and 217 trials respectively.

Discussion

At the broadest level, the results confirm that experience in categorization influences later perceptual sensitivity, as measured by same/different judgments. The clearest support for this comes from the acquired distinctiveness that occurs along the critical values of a dimension that are relevant for categorization. Strong acquired distinctiveness was found for both size and brightness.

The evidence for acquired equivalence was somewhat more complex. There was evidence on one dimension, but not the other, for acquired equivalence when a dimension is irrelevant for categorization. When size is irrelevant for the categorization (for brightness categorizers), size discriminations become desensitized, relative to the control group. This provides a demonstration of a situation in which subjects would have done better in a perceptual task if they had never experienced the materials before in a categorization experiment, or if they had been able to ignore their previous exposure. Neither acquired equivalence nor distinctiveness was found along the brightness dimension for size categorizers. Clearly there is a difference between how these dimensions are processed, even though scaling was done to make differences on the dimensions equally salient. Future work will be necessary in order to determine why the two dimensions do not equally acquire equivalence.

The two dimensions do, however, behave similarly as far as acquired equivalence along non-critical values of a relevant dimension. In particular, the opposite of acquired equivalence is found for both dimensions. For example, although size values of 1 and 2 both belong to Category A for size categorizers, size categorization training produces sensitization to the difference between these values.

Although the opposite of an acquired equivalence effect within the categorization-relevant dimension is consistent with dimension-wide sensitization, there is other evidence for a more local, value-specific sensitization. For both size and brightness dimensions, sensitization due to categorization is most pronounced when dimension values are chosen that straddle two categories. This local, value-specific sensitization is more clear for brightness than it is for size.

The results from Experiment 2 also indicate a particular way that dimensions compete with each other for attention. For both size and brightness, less perceptual sensitization occurs when two dimensions must be attended for categorization than when only one must be attended. This result is important for two reasons. First, it suggests that the attentional competition between dimensions that has been hypothesized in models of categorization (e.g. Nosofsky, 1986) is not simply a strategic selective weighting of dimensions for categorization. Rather, it also has a non-volitional component that emerges even when tested in a perceptual task with no logical relation to the categorization task (see General Discussion).

Second, the result opposes one possible non-perceptual explanation of the differences that have been found between categorization conditions. It might be argued that subjects are better at making size discriminations after size categorization because during categorization they learn that even a very small size difference counts as a difference. Thus, categorization training might increase only one's willingness to base a decision on a dimension if it is relevant for categorization, but might not change one's ability to spot differences on the dimension. The comparison of the "size and brightness" condition to the "size" condition makes this approach less plausible. For both of these groups, size is a clearly relevant dimension for categorization. Thus, there is evidence for two groups having different perceptual sensitivities along a dimension even when both groups know that the dimension is relevant, and are calibrated to the correct magnitude of value differences along the dimension.

Experiments 3 and 4 provide a replication to Experiments 1 and 2 using different dimensions. It is desirable to test new dimensions for two reasons. First, one of the results from Experiment 2 was that although most conclusions were similar for the two dimensions tested, there were also some differences. Additional dimensions must be tested to gain security in the external validity of the results from Experiment 2. Second and more importantly, it is informative to compare dimensions that vary in their integrality. While size and brightness are fairly separable dimensions, results may differ with integral dimensions. In particular, there are reasons (see Introduction) for predicting either more or less acquired equivalence and attentional competition between integral dimensions than was obtained with separable dimensions. Thus, in Experiments 3 and 4, stimuli vary on brightness and saturation, two color dimensions that substantial evidence has shown to be integral.

Experiment 3 is the analog of Experiment 1 in that it provides the scaling data for Experiment 4. Experiment 3 provides 16 stimuli that vary on saturation and brightness such that adjacent stimuli will be approximately equally discriminable for the control subjects that do not receive any categorization training.

Method

Subjects. Forty-two undergraduate students from Indiana University served as subjects in order to fulfill a course requirement.

Materials. Sixteen squares that varied in their saturation and brightness were displayed on Macintosh IIsi screens. The sixteen squares were obtained by factorially combining four values of saturation with four values of brightness. All squares were 6 cm2.

Procedure. The same/different judgment procedure from Experiment 1 was used, with only the slight variations described here. Initial dimension values for the sixteen squares were set to plausible, fairly discriminable values. For saturation, the constant increment/decrement to the more saturated square when incorrect/correct responses on "different" trials were made was 527 units. For brightness, the increment/decrement was 1462 units.

Results and Discussion

The average values along the saturation and brightness dimension for each of the 16 squares are shown in Table 2. The values indicate very little interaction between saturation and brightness. That is, the physical difference in saturation values required to make a discriminable difference was fairly constant across different values of brightness.

For Experiment 3, it is necessary to calculate dimension values as if there were no dimension interactions. The CIE 1976 color coordinates were determined by Spectra Scan 714 chromometer. The X and Y coordinates for all 16 squares are approximately 0.45 and 0.35 respectively. The actual luminance and purity values are shown in Table 3.

________________________

Insert Table 3 about here

________________________

Experiment 4 is a replication of Experiment 2 using integral instead of separable dimensions. Three issues are explored by varying the integrality of the materials. First, the extent to which perceptual attention is naturally and non-strategically placed on integral dimensions is explored. Previous researchers have shown that even with integral materials, categorization decisions can be made to selectively weight one dimension over another (Foard & Kemler Nelson, 1984; Nosofsky, 1987). For example, if saturation but not brightness is relevant for a categorization, subjects can learn this categorization rule with sufficient practice. However, it is still an open question whether this categorization regime causes a change to subjects' actual perceptual processes. Some researchers speculate that during the course of experience or maturation, dimensions that were once originally fused become differentiated (Smith, Carey, & Wiser, 1985; L. B. Smith, 1979, 1989). It is conceivable that if saturation but not brightness (or vice versa) is relevant for a categorization, that saturation would become differentiated from brightness and selectively sensitized. If this occurs, then saturation but not brightness might show sensitization due to saturation categorization.

Second, integral dimensions are used to test whether acquired equivalence along a categorization-irrelevant dimension is more or less prominent with fused dimensions. As discussed in the introduction, the answer to this question has a bearing on whether subjects learn to actively filter out irrelevant dimensions that intrude on categorization (yielding greater acquired equivalence for integral dimensions), or whether irrelevant dimensions become less influential when attention shifts toward relevant dimensions that are far removed (yielding greater acquired equivalence for separable dimensions).

Third, integral dimensions also test predictions about competition between dimensions. In Experiment 2, there was evidence of competition from the comparison of two-dimension and one-dimension categorization conditions. However, if it is easier to attend to two dimensions simultaneously if they are integral (Garner, 1974; Melara, 1992) then there may be less evidence for competition with saturation and brightness.

Method

Subjects. Eighty undergraduate students from Indiana University served as subjects in order to fulfill a course requirement. The subjects were evenly split into the four conditions.

Materials. The sixteen squares scaled in Experiment 3 were used as stimuli. The stimuli form a 4X4 matrix similar to that shown in Figure 3 with every brightness level combined with every saturation.

Procedure. The procedure was identical to that used in Experiment 2 except as noted here. There were four categorization conditions: saturation categorizers, brightness categorizers, saturation and brightness categorizers, and controls (no categorization training). Before the categorization training, subjects were instructed that the presented squares would have slightly different colors. After categorization training, all groups were given the same same/different judgment task. Subjects were told that they should respond "different" if the presented squares had even slightly different colors.

Results

____________________________

Insert Table 3 about here

____________________________

The d' results from all four categorization training conditions are shown in Table 3. As with Experiment 2, there was a bias for subjects in the control condition to respond "same" more frequently than in the other conditions, and a bias for subjects in the "saturation and brightness" categorization condition to respond "same" less frequently than in the other conditions (F(3, 19) = 4.5, p < 0.05, post-hoc differences significant at p < .05 by Fisher's PLSD).

____________________________

Insert Figure 7 about here

____________________________

____________________________

Insert Figure 8 about here

____________________________

____________________________

Insert Figure 9 about here

____________________________

Acquired Distinctiveness. Dimensions acquire distinctiveness if perceptual discriminations along a categorization-relevant dimension are better for the categorization condition than for the control condition. Separate tests were conducted for saturation categorizers and brightness categorizers. For saturation categorizers, there are four comparisons that would unambiguously become sensitized if relevant dimensions acquire distinctiveness - the four comparisons that involve a square with a saturation value of 2 and a square with a saturation value of 3. In Figure 7, these four comparisons occupy the center column of horizontally extended rectangles. These four comparisons were averaged together for each subject in the saturation categorization and the control conditions. Overall, the saturation categorizers' d' for these four comparisons was greater than the control subjects' d's for the same comparisons (unpaired t(38)=3.2, p < .05). This difference can be clearly seen in Figure 7 from the fact that all four of the d' differences for the center size comparisons are large (wide) and positive (black).

The equivalent test for acquired distinctiveness for brightness categorizers compares brightness values of 2 and 3. These comparisons are found in the center row of vertically extended rectangles in Figure 8. The brightness categorizers' d' was greater than the control subjects' d's for these comparisons (t(38)=3,4, p<.05). This is evident in Figure 8 from the long black rectangles along the center row. Thus, for both categorization groups, there is evidence of acquired distinctiveness between values of the categorization-relevant dimension that belong to different categories.

Acquired Equivalence of the Irrelevant Dimension. Acquired equivalence along a dimension that is irrelevant for categorization occurs if discriminations along such a dimension are worse than they are for the control subjects. This occurs if, for example, saturation categorizers have a lower d' for trials that differ on brightness than do control subjects. For saturation categorization subjects, there are 12 square comparisons that involve differences in brightness (the 12 vertically extended rectangles in Figure 7). Averaging over these comparisons, there is a significant difference between d' scores for the control and saturation categorization conditions (unpaired t(38) = 2.7, p <.05). This result is in the opposite direction as predicted by acquired equivalence. As evidenced by the black vertically extended rectangles, subjects in the saturation group are more sensitive than the control group at detecting brightness differences.

For the analogous test (the 12 horizontal rectangles in Figure 8) of acquired equivalence in brightness categorizers, there is a significant difference between the brightness categorizers and the control subjects (unpaired t(38)=2.4, p<.05) with the brightness categorizers producing higher d's. Again, this is the opposite of the pattern predicted by acquired equivalence. This is evident in Figure 8 from the preponderance of black horizontal rectangles spanning squares that differ in their saturations.

Acquired Equivalence Within a Relevant Dimension. Acquired equivalence within a categorization-relevant dimension occurs if squares that have different dimension values on a categorization-relevant dimension but belong to the same category become less discriminable as a result of categorization training. For the saturation categorization condition, 8 relevant square comparisons are collapsed together - the left and right columns of horizontal rectangles in Figure 7. These rectangles span squares that differ on their saturations, but belong to the same category. Overall, the d' associated with these 8 squares is greater for the saturation categorization condition than it is for the control group (t(38) = 2.8, p < .05). This significant difference is in the opposite direction as hypothesized by acquired equivalence within a relevant dimension.

The analogous comparison for the brightness condition involves the 8 vertically extended rectangles that form the top and bottom rows of Figure 8. The d's associated with these comparisons are greater for the brightness condition than for the control group (t(38)=2.1, p<.05), in opposition to acquired equivalence with a relevant dimension.

Local Sensitization of a Dimension. Given that acquired equivalence was not found within the categorization-relevant dimension, it may or may not be the case that all values on a categorization-relevant dimension are equally sensitized. To test this, discriminations between the squares that belong to different categories can be compared to discriminations between squares that belong in the same category but that vary on the categorization-relevant dimension. This is done by comparing two differences - the difference between control and categorization conditions on relevant dimension differences that straddle categories, and the difference between the two conditions on relevant dimension differences that remain in one category.

To test local sensitization for saturation categorizers, two sets of comparisons were formed. One set (the "critical value" set) contained the four comparisons that paired a square with a saturation value of 2 with a square with a saturation value of 3 (the middle column of horizontally extended rectangles in Figure 7). The other set (the "noncritical value" set) contained the other 8 eight comparisons involving squares that differed in their saturations. The d' scores of the saturation categorizers for each set were adjusted by subtracting the respective d' scores from the control condition. The adjusted d' scores for the critical set were (marginally) significantly greater than the adjusted d' scores for the noncritical set (t(38) = 1.74, p = .083). In Figure 7, this is apparent in the relatively wider rectangles in the middle column than in the left and right columns.

The analogous sets of critical and noncritical comparisons were compared for the brightness categorizers. The adjusted d' scores for the critical set were significantly greater than the adjusted d' scores for the noncritical set (t(38)=2.2, p < .05). The middle column of vertically extended rectangles in Figure 8 contains relatively tall, black rectangles. The upper and lower columns are somewhat shorter.

Attentional competition. Does sensitization produced by categorization on Dimension X diminish if categorization by Dimension Y is also required? For the saturation dimension, the ability of saturation categorizers to make saturation discriminations is compared to the ability of subjects who need to categorize on the basis of both saturation and brightness to make saturation discriminations. The overall d' for saturation discriminations by saturation categorizers was not significantly greater than the d' for the same items by the size and brightness categorizers (t(38) = 1.4, p=0.17). The overall d' for brightness differences was marginally greater for the brightness condition (t(40)=2.0, p=.05) than it was for the "saturation and brightness" condition.

Selective Sensitization of a Dimension. Thus far, the evidence has suggested sensitization of the irrelevant dimension, rather than acquired equivalence. Given this, a second question can be asked: If Dimension X is relevant for categorization, is there equal sensitization of Dimensions X and Y? To answer this question, we can compare the d's along the saturation (or brightness) dimension for the two single-dimension categorization conditions. The d's for saturation differences were greater for saturation categorizers than for brightness categorizers (t(38)=3.4, p<.05). The horizontally extended rectangles are longer in Figure 7 than Figure 8. Likewise, the d's for brightness differences were greater for brightness categorizers than for saturation categorizers (t(38)=3.1, p < .05). Thus, although brightness is sensitized for saturation categorizers (and vice versa), there is still more sensitization of the dimension that is relevant for categorization.

Other Results. As was found in Experiment 2, subjects' perceptual sensitivities improved over the course of testing. There was a positive correlation of r=0.07, p<.05 between trial number in the same/different portion of the experiment and d' score.

The categorization portion of the experiment was harder than the categorization in Experiment 2. In order to achieve a categorization accuracy of 90%, the saturation, brightness, and "saturation and brightness" groups required an average of 167, 173, and 212 trials respectively.

Discussion

The results from Experiment 4 share significant similarities and differences with the results from Experiment 2. The current results with integral dimensions replicate several of the results found with separable dimensions. First, acquired distinctiveness effects were always found. Categorization yielded selective sensitization along dimension values that determined the categories. Second, no evidence for acquired equivalence within a categorization-relevant dimension was found. In fact, the results indicated that dimension differences along a relevant dimension that did not cross category lines were still sensitized by categorization training. Third, there was evidence for local sensitization of particular dimension values. Although entire saturation and brightness dimensions were sensitized, there was also selective sensitization of dimension values that fell on the boundaries between categories.

There were also two important differences in the results between Experiments 2 and 4. First, Experiment 4 showed the opposite of acquired equivalence along the categorization-irrelevant dimension, for both saturation and brightness. Second, far less competition between dimensions was found in Experiment 4 than in Experiment 2. Requiring subjects to use both saturation and brightness during the categorization portion of Experiment 4 did not attenuate sensitization on either dimension as much as it did in Experiment 2. This result is consistent with integral dimensions being easy to attend simultaneously. The claim is that perceptual sensitization occurs when subjects learn to perceptually attend to a dimension during categorization training.

At the broadest level, the experiments support the contention that experience in acquiring new categories can alter perceptual sensitivity. Training subjects on different simple categorization rules resulted in different abilities to make perceptual discriminations (subject to the earlier caveat that same/different judgments probably involve memory and attentional components in addition to purely sensory processes). In addition to this broad conclusion, the experiments also identify a number of specific details about the particular manner in which categorization exerts its influence.

Acquired Equivalence and Distinctiveness

In both experiments, there was consistent support for acquired distinctiveness. For all three dimensions tested (saturation, brightness, and size), dimension differences that cross category boundaries become sensitized with training. This effect is predicted by several theories of associative and perceptual learning (Lawrence, 1949; Gibson, 1969; Miller & Dollard, 1946).

It is particularly informative that acquired distinctiveness was found for integral dimensions. Although saturation and brightness behave as psychologically "fused" dimensions by many measures, categorization training acted to separate/differentiate the dimensions. For example, although saturation is sensitized somewhat when categorization is based on brightness, brightness becomes more sensitized than saturation. Thus, both Experiments 2 and 4 provide evidence that categorization can change subjects' perceptual abilities. Experiment 4 goes further, suggesting that categorization training can cause previously undifferentiated/fused dimensions to become differentiated.

There was similarly consistent evidence against the variety of acquired equivalence that occurs within a categorization-relevant dimension. For all three dimensions, the evidence suggests that sensitization to stimuli differences occurs even if the stimuli belong in the same category, as long as the dimension that the stimuli vary on is generally relevant for categorization. This result is at odds with acquired equivalence explanations of categorical perception. If categorical perception occurs by a process of acquired equivalence, then dimension values that belong to a categorization-relevant dimension must become desensitized with time. However, the current experimental results, to the extent that they can be generalized to speech perception, argue that this process of learning to neglect differences along a categorization-relevant dimension does not occur. Instead, acquired distinctiveness seems to be the general method for establishing selectively heightened sensitivities within a dimension.

The results with regard to acquired equivalence along an irrelevant dimension were more complicated. With the separable dimensions of size and brightness, acquired equivalence was found for size, but not for brightness. That is, when brightness determined categorization, subjects actually decreased their sensitivity at spotting size differences relative to the control group. When size determined categorization, there was neither cost nor benefit for brightness discriminations relative to the "no categorization" control condition. This latter null effect may mask predifferentiation and acquired equivalence effects that cancel out one another. Because effects of predifferentiation would attenuate any effects of acquired equivalence, the prevalence of acquired equivalence may be underestimated by the current experiments.

Acquired equivalence along the irrelevant dimension was found for one dimension when separable dimensions were tested, but was not found with integral dimensions. In fact, the opposite of acquired equivalence was found. Training with brightness facilitated saturation discriminations, and vice versa. While this is consistent with predifferentation, predifferentiation does not explain why this strong bi-directional facilitation is found with integral but not separable dimensions. An account that resorts to differences between individual dimensions is not satisfactory because one of the dimensions, brightness, was the same in the two situations. An explanation that takes into account the relation between the two dimensions seems necessary. Such an account is provided in a later section.

Local Sensitization of Dimensions

Although apparently contradictory, evidence in favor of both dimension-wide and more local sensitization were found. The evidence against acquired equivalence within a categorization-relevant dimension is evidence in favor of dimension-wide sensitization. For example, when size values 1 and 2 were assigned to Category A, and size values 3 and 4 were assigned to Category B, there was sensitization to the difference between values 1 and 2 on size relative to the control group. Both saturation and brightness in Experiment 4 showed significant results in the same direction. These results suggest that subjects learn to attend not simply to particular values on a dimension, but also learn to attend to the dimension itself (Zeamon & House, 1963).

However, there is also evidence for local sensitization. Specifically, although the difference between values 1 and 2 were sensitized, the value difference that was critical for categorization, the difference between values 2 and 3, was sensitized even more. This was found for all three dimensions that were tested. This local sensitization indicates that the perception of particular parts of dimensions can be selectively enhanced when useful for categorization. If the basic result of acquired distinctiveness due to categorization training attests to the flexibility of the perceptual system, the additional results of selective acquired distinctiveness within a dimension attests to even greater flexibility.

Competition Between Dimensions and Integrality/Separability

The current experiments provide two sources of evidence for competition between dimensions. The first evidence comes from comparing perceptual sensitization produced by training with a one-dimension categorization rule with the sensitization produced by a two-dimension categorization rule. If Dimension X is more sensitized when categorization depends only on Dimension X than when it depends on Dimensions X and Y, then this is some evidence for competition between dimensions.

By this measure of dimensional competition, there is evidence for strong competition between dimensions in Experiment 2 and evidence for some competition in Experiment 4. The extent of the competition appears to be somewhat greater in Experiment 2 than in Experiment 4. In Experiment 2, the sensitization of either dimension suffered from a categorization rule that required attention to both dimensions. In Experiment 4, only brightness sensitization was marginally affected. A major difference between these experiments is that integral dimensions were used in Experiment 4 while separable dimensions were used in Experiment 2. Thus, by this measure, separable dimensions seem to compete with each other for dimensional attention more than do integral dimensions. These results fit with Garner's (1974) conception of integral dimensions as being easy to attend simultaneously relative to separable dimensions. The results augment previous support for Garner's distinction by showing that this competition for attention has an influence on a later task, and that it has a relatively low-level perceptual influence.

The second measure of dimensional competition gives similar results. The second measure examines the degree of acquired equivalence/distinctiveness along the categorization-irrelevant dimension. The claim is that if two dimensions are naturally attended simultaneously, then requiring attention to be placed on one dimension should not necessarily cause another dimension to be processed less. However, if the dimensions cannot be simultaneously attended, then desensitization (acquired equivalence) of competing dimensions is expected when attention is required elsewhere. As earlier reviewed, the opposite of acquired equivalence was found for categorization-irrelevant integral dimensions. One case of acquired equivalence (and one null effect) was found for categorization-irrelevant separable dimensions. This is consistent with the view that integral dimensions are easily and naturally attended simultaneously.

There are at least two accounts for how the opposite of acquired equivalence can be obtained for irrelevant dimensions. According to predifferentiation (Gibson, 1969), distinctiveness is acquired simply by exposure to the stimuli that are later included in the same/different task. However, predifferentiation does not explain why acquired equivalence was found for separable dimensions. Another explanation can be given in terms of diffusion of attention. Sensitization of an irrelevant dimension is hypothesized if attention is focused on the categorization-relevant dimension, but also spreads to cover similar dimensions (also see Melara, 1992). More diffusion would be expected to occur for integral than for separable dimensions because of the greater similarity of integral dimensions.

There are also two accounts for the occurrence of acquired equivalence within an irrelevant dimension. According to one account, acquired equivalence occurs when a person actively learns to filter out intrusive dimensions. According to the other account, acquired equivalence occurs when attention, which is usually distributed over many dimensions, becomes focused on one dimension, and thereby displaces attention from irrelevant dimensions. The current results with respect to integral/separable dimensions are more consistent with the latter account. According to the latter account, separable dimensions are less similar/further removed from each other than are integral dimensions, and consequently, when attention is focused on one separable dimension, it must be removed from the other separable dimension. At least in this instance, irrelevant dimensions seem to be ignored because other dissimilar dimensions are relevant, rather than because they are actively filtered out (James, 1890/1950; Neisser, 1976).

Conclusion

Although there has been considerable speculation about the relation between conceptual structure and perception (Bruner & Postman, 1949; Goldstone, in press; Whorf, 1941), there have been few attempts to obtain rigorous experimental support for concept/percept interactions in controlled laboratory conditions. By training otherwise equivalent subjects on different categories, it was possible to single out category differences as the grounds for the perceptual differences that were observed. The particular results obtained generally indicate that categorization training does influence perceptual sensitivity on later tasks. In particular, acquired distinctiveness effects were found with all tested dimensions; no evidence for acquired equivalence for same-category values within a relevant dimension was found; evidence both for dimension-wide sensitization and for more local value-specific sensitization was found; and competition between dimensions for sensitization was found, and was stronger for separable than for integral dimensions.

In sum, the results suggest that categorization training, in addition to requiring subjects to determine a categorization rule, can also produce perceptual learning that lasts beyond the categorization task. One productive and influential approach to cognition maintains that categorization, and higher-level cognitive processes in general, operate on the output of lower-level perceptual processing. That is, our perceptual systems provide us with a set of fixed features. These features are the inputs to higher-level cognitive processes. However, the current experiments may illustrate a situation in which the higher-level cognitive process being executed has an influence on the lower-level features that are used. In addition to categorization being based on featural/dimensional descriptions of objects, it also appears that the categorization process partially forms the descriptions that are used.

Table 1. Averaged stimulus values obtained from Experiment 1.

Size Value

Brightness Value Ø	1	2	3	4
4	Size= 2.85 Brightness= 21.62	3.19 21.62	3.59 21.62	4.02 21.62
3	2.85 19.61	3.16 19.95	3.49 20.36	3.83 20.43
2	2.85 18.12	3.08 18.61	3.38 19.1	3.65 19.33
1	2.85 13/6	3.09 15.67	3.38 17.21	3.69 18.16

Note: Size values are expressed in height and width centimeters. Brightness values are expressed in candelas/meter2.

Table 2. Discrimination performance for four training conditions in Experiment 2.

Size Values

Brightness Values	Training condition	1&1 P(C)	1&2 P(C), d'	2&2 P(C)	2&3 P(C), d'	3&3 P(C)	3&4 P(C), d'	4&4 P(C)
4&4 P(C)	S B No S&B	82 86 84 74	44, 0.78 33, 0.59 36, 0.64 49, 0.69	84 84 85 76	52, 0.98 38, 0.67 40, 0.73 54, 0.81	81 83 83 77	41, 0.71 34, 0.56 36, 0.65 43, 0.60	82 84 82 77
3&4 P(C), d'	S B No S&B	47, 0.83 44, 0.84 40, 0.80 54, 0.81		47, 0.87 54, 0.94 47, 0.92 58, 0.93		55, 1.05 51, 0.99 52, 1.02 64, 1.07		53, 1.00 58, 1.11 53, 1.07 60, 0.97
3&3 P(C)	S B No S&B	81 82 84 79	47, 0.86 33, 0.53 33, 0.56 45, 0.59	82 85 86 73	55, 1.07 35, 0.59 35, 0.61 62, 1.0	81 81 86 73	38, 0.63 33, 0.53 33, 0.56 45, 0.58	83 83 87 76
2&3 P(C), d'	S B No S&B	46, 0.83 50, 0.97 45, 0.87 63, 1.03		50, 0.94 56, 1.12 45, 0.87 65, 1.09		52, 0.99 59, 1.20 48, 0.94 64, 1.06		55, 1.06 61, 1.25 49, 0.97 68, 1.19
2&2 P(C)	S B No S&B	81 83 86 75	46, 0.83 34, 0.56 37, 0.67 50, 0.71	80 86 84 74	51, 0.96 34, 0.56 32, 0.54 53, 0.79	82 85 83 75	38, 0.64 29, 0.42 37, 0.67 45, 0.58	81 83 85 76
1&2 P(C), d'	S B No S&B	46, 0.83 45, 0.84 45, 0.87 57, 0.89		46, 0.83 48, 0.92 46, 0.90 57, 0.88		49, 0.95 54, 1.04 50, 0.97 61, 0.98		48, 0.91 52, 0.95 48, 0.93 57, 0.90
1&1 P(C)	S B No S&B	80 82 87 76	46, 0.83 34, 0.56 36, 0.64 56, 0.67	83 82 85 79	54, 1.04 36, 0.61 40, 0.74 55, 0.83	83 85 85 78	40, 0.68 33, 0.53 36, 0.64 48, 0.69	82 83 86 76

Note: Each of the two presented squares has a brightness and size value. The values for the two squarse on a dimension are combined by the "&" symbol. P(C) = percentage of correct judgments. When the presented squares have identical values on both dimensions (e.g. values 1&1 on brightness and 4&4 on size), the correct response is "same." When the squares differ on either brightness or size, the correct response is "different." S = size categorization condition. B = brightness categorization condition. No = no categorization training. S&B = size and brightness categorization.Table 3. Averaged stimulus values obtained from Experiment 3.

Saturation Value

Brightness Value Ø	1	2	3	4
4	Luminance= 19.57 Purity= 70.37%	23.85 52.62	29.08 35.71	36.66 26.19
3	17.7 71.53	21.89 54.81	25.82 38.4	31.94 28.27
2	16.74 71.21	18.09 55.7	22.05 39.73	26.84 28.54
1	15.26 70.29	15.27 56.25	19.89 42.31	23 29.71

Note: Luminance values are expressed in candelas/meter2.

Table 4. Discrimination performance for four training conditions in Experiment 4.

Saturation Values

Brightness Values	Training condition	1&1 P(C)	1&2 P(C), d'	2&2 P(C)	2&3 P(C), d'	3&3 P(C)	3&4 P(C), d'	4&4 P(C)
4&4 P(C)	S B No S&B	82 84 83 80	61, 1.23 56, 1.08 54, 1.04 62, 1.19	83 83 86 81	68, 1.35 55, 1.09 53, 1.06 66, 1.29	80 81 85 83	64, 1.28 56, 1.10 53, 1.02 61, 1.16	80 82 83 82
3&4 P(C), d'	S B No S&B	52, 0.97 57, 1.12 43, 0.78 59, 1.11		56, 1.04 54, 1.05 48, 0.91 61, 1.16		56, 1.06 58, 1.16 57, 1.14 57, 1.08		55, 1.05 55, 1.11 46, 0.86 49, 0.88
3&3 P(C)	S B No S&B	83 84 85 82	58, 1.12 55, 1.07 49, 0.93 57, 1.06	84 85 83 82	65, 1.26 56, 1.10 51, 0.99 63, 1.21	84 83 84 79	62 , 1.23 56, 1.10 51, 0.98 58, 1.08	83 82 82 81
2&3 P(C), d'	S B No S&B	55, 1.08 61, 1.23 49, 0.93 60, 1.13		54, 1.02 62, 1.25 51, 0.98 63, 1.21		55, 1.09 65, 1.33 51, 1.03 64, 1.24		52, 0.98 66, 1.36 53, 1.03 63, 1.21
2&2 P(C)	S B No S&B	82 82 86 84	58, 1.16 56, 1.10 48, 0.90 57, 1.05	81 82 82 80	64, 1.35 55, 1.09 52, 1.00 64, 1.24	82 83 83 82	61, 1.23 56, 1.08 53, 1.03 59, 1.11	81 84 82 83
1&2 P(C), d'	S B No S&B	53, 1.03 55, 1.10 47, 0.88 55, 1.00		54, 1.06 56, 1.07 47, 0.89 56, 1.03		54, 1.03 55, 1.10 53, 0.96 56, 1.03		57, 1.06 57, 1.12 51, 0.98 55, 1.03
1&1 P(C)	S B No S&B	83 83 85 81	59, 1.14 53, 1.10 53, 1.03 61, 1.16	81 84 83 80	62, 1.18 54, 1.02 50, 0.95 60, 1.13	82 81 82 81	59, 1.07 50, 0.90 49, 0.93 58, 1.08	82 83 84 79

Note: Each of the two presented squares has a brightness and saturation value. The values for the two squarse on a dimension are combined by the "&" symbol. P(C) = percentage of correct judgments. When the presented squares have identical values on both dimensions (e.g. values 1&1 on brightness and 4&4 on saturation), the correct response is "same." When the squares differ on either brightness or saturation, the correct response is "different." S = saturation categorization condition. B = brightness categorization condition. No = no categorization training. S&B = saturation and brightness categorization.

Aha, D. W., & Goldstone, R. L. (1990). Learning attribute relevance in context in instance-based learning algorithms. Proceedings of the Twelfth Annual Conference of the Cognitive Science Society. (pp. 141-148). Hillsdale, New Jersey: Lawrence Erlbaum Associates.

Aha, D. W., and Goldstone, R. L. (1992). Concept learning and flexible weighting. Proceedings of the Fourteenth Annual Conference of the Cognitive Science Society. (pp. 534-539). Hillsdale, New Jersey: Lawrence Erlbaum Associates.

Algom, D. (1992). Psychophysical approaches to cognition. Amsterdam: North Holland.

Andrews, J., Livingston, K., Harnad, S., & Fisher, U. (unpublished manuscript). Learned categorical perception in human subjects: Implications for symbol grounding.

Ashby, F. G., & Lee, W. W. (1991). Predicting similarity and categorization from identification. Journal of Experimental Psychology: General, 120, 150-172.

Ashby, F. G., & Perrin, N. A. (1988). Toward a unified theory of similarity and recognition. Psychological Review, 95, 124-150.

Ashby, F. G., & Townsend, J. T. (1986). Varieties of perceptual independence. Psychological Review, 93, 154-179.

Biederman, I., & Checkovsky, S. F. (1970). Processing redundant information. Journal of Experimental Psychology, 83, 486-490.

Biederman, I., & Shiffrar, M. M. (1987). Sexing day-old chicks: A case study and expert systems analysis of a difficult perceptual-learning task. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13, 640-645.

Bornstein, M. H. (1987). Perceptual categories in vision and audition. in S. Harnad (Ed.) Categorical Perception: The groundwork of cognition. Cambridge University Press: Cambridge (pp. 287 - 300).

Bruner, J. A., & Postman, L. (1949). Perception, conception, and behavior. Journal of Personality, 18, 14-31.

Bruner, J. S., Goodnow, J. J., & Austin, G. A. (1956). A study of thinking. New York: Wiley.

Burns, B., & Shepp, B. E. (1988). Dimensional interactions and the structure of psychological space: The representation of hue, saturation, and brightness. Perception and Psychophysics, 43, 494-507.

Burns, E. M., & Ward, W. D. (1978). Categorical perception - phenomenon or epiphenomenon: Evidence from experiments in the perception of melodic musical intervals. Journal of the Acoustical Society of America, 63, 456-468.

Cutting, J. E., (1982)). Plucks and bows are categorically perceived, sometimes. Perception & Psychophysics, 31, 462-476.

Duncan, J. (1984). Selective attention and the organization of visual information. Journal of Experimental Psychology: General, 113, 501-517.

Eimas, P. D. (1974). Auditory and phonetic coding of the cues for speech: Discrimination of the [r-l] distinction by young infants. Perception & Psychophysics, 18, 341-347.

Eimas, P. D., Miller, J. L., & Jusczyk, P. W. (1987). On infant speech perception and the acquisition of language. in S. Harnad (Ed.) Categorical perception. (pp. 161-195). Cambridge University Press: Cambridge.

Eimas, P. D., Siqueland, E. R., Jusczyk, P. W., & Vigorito, J. (1971). Speech perception in infants. Science, 171, 303-306.

Estes, W.K. (1986). Array models for category learning. Cognitive Psychology, 18, 500-549.

Fechner, G. T. (1966). Elements of psychophysics. (H. E. Alder, Trans.). New York: Holt, Rinehart, & Winston. (Original work published 1860).

Foard, C.F., & Kemler Nelson, D.G. (1984). Holistic and analytic modes of processing: The multiple determinants of perceptual analysis. Journal of Experimental Psychology: General, 113, 94-111.

Garner, W. R. (1974). The processing of information and structure. Hillsdale, NJ: Erlbaum.

Gibson, E. J. (1991). An odyssey in learning and perception. MIT Press: Cambridge.

Gibson, E. J. (1969). Principles of perceptual learning and development. Appleton-Centure-Crofts, New York.

Gibson, E. J., & Walk, R. D. (1956). The effect of prolonged exposure to visually presented patterns on learning to discriminate them. Journal of Comparative and Physiological Psychology, 49, 239-242.

Gibson, J. J., & Gibson, E. J. (1955). Perceptual learning: Differentiation or enrichment? Psychological Review, 62, 32-41.

Goldstone, R. L. (in press). The role of similarity in categorization: Providing a groundwork. Cognition.

Goldstone, R.L. (1993). Feature distribution and biased estimation of visual displays. Journal of Experiment Psychology: Human Perception and Performance, 19, 564-579.

Goss, A. E. (1953). Transfer as a function of type and amount of preliminary experience with task stimuli. Journal of Experimental Psychology, 46, 419-428.

Gottwald, R. L., & Garner, W. R. (1975). Filtering and condensation tasks with integral and separable dimensions. Perception & Psychophysics, 18, 26-28.

Hall, G. (1991). Perceptual and associative learning. Clarendon Press: Oxford.

Harnad, S. (1987). Categorical perception. Cambridge University Press: Cambridge.

James, W. (1890/1950). Principles of Psychology. New York: Holt.

Kemler, D. G., & Smith, L. B. (1978). Is there a developmental trend from integrality to separability in perception? Journal of Experimental Child Psychology, 26, 460-507.

Kruschke, J. K. (1992). ALCOVE: An exemplar-based connectionist model of category learning. Psychological Review, 99, 22-44.

Kuhl, P. K., & Miller, J. D. (1978). Speech perception by the chinchilla: Identification functions for synthetic VOT stimuli. Journal of the Acoustical Society of America, 63, 905-917.

Lane, H. (1964). Motor theory of speech perception: A critical review. Psychological Review, 72, 275-309.

Liberman, A. M., Harris, K. S., Hoffman, H. S., & Griffith, B.C. (1957). The discrimination of speech sounds within and across phoneme boundaries. Journal of Experimental Psychology, 54, 358-368.

Liberman, A. M., Harris, K. S., Kinney, J. A., & Lane, H. (1961). The discrimination of relative onset time of the components of certain speech and nonspeech patterns. Journal of Experimental psychology, 61, 379-388.

Logan, J. S., Lively, S. E., & Pisoni, D. B. (1991). Training Japanese listeners to identify /r/ and /l: A first report. Journal of the Acoustical Society of America, 89, 874-886.

Marks, L. E. (1992). "What thin partitions sense from thought divide": Toward a new cognitive psychophysics. in D. Algom (ed.), Psychophysical approaches to cognition. Amsterdam: North Holland (pp. 115-186).

Massaro, D. W. (1987). Categorical partition: A fuzzy-logical model of categorization behavior. in S. Harnad (Ed.) Categorical perception. (pp. 254-283). Cambridge University Press: Cambridge.

Massaro, D. W., & Cohen, M. M. (1983). Evaluation and integration of visual and auditory information in speech perception. Journal of Experimental Psychology: Human Perception and Performance, 9, 753-771.

Medin, D. L., & Schaffer, M. M. (1978). A context theory of classification learning. Psychological Review, 85, 207-238.

Melara, R. D. (1992). The concept of perceptual similarity: From psychophysics to cognitive psychology. in D. Algom's (Ed.) Psychophysical approaches to cognition. Amsterdam: North Holland.

Miller, N. E., & Dollard, J. (1941). Social learning and imitation. Yale University Press.

Miyawaki, K., Strange, W., Verbrugge, R., Liberman, A. M., Jenkins, J. J., & Fujimura, O. (1975). An effect of linguistic experience: The discrimination of [r] and [l] by native speakers of Japanese and English. Perception & Psychophysics, 18, 331-340.

Neisser, U. R., (1976). Cognition and reality: Principles and implications of cognitive psychology. San Francisco: Freeman.

Norman, G. R., Brooks, L. R., & Allen, S. W. (1989). Recall by expert medical practitioners as a record of processing attention. Journal of Experimental Psychology: Learning, Memory, & Cognition, 15, 1166-1174.

Norman, G. R., Brooks, L. R., Coblentz, C. L., & Babcook, C. J. (1992). The correlation of feature identification and category judgments in diagnostic radiology. Special Issue: Memory and cognition applied. Memory & Cognition, 20, 344-355.

Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of Experimental Psychology: General, 115, 39-57.

Nosofsky, R. M. (1987). Attention and learning processes in the identification and categorization of integral stimuli. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13, 87-108.

Nosofsky, R. M. (1991). Typicality in logically defined categories: Exemplar-similarity versus rule instantiation. Memory & Cognition, 19, 131-150.

Pastore, R. E. (1987). Categorical perception: Some psychophysical models. in S. Harnad (Ed.) Categorical perception. (pp. 29-52). Cambridge University Press: Cambridge.

Pearce, J. M., & Hall, G. (1980). A model for Pavlovian learning: Variations in the effectiveness of conditioned but not of unconditioned stimuli. Psychological Review, 87, 532-552.

Peron, R. M., & Allen, G. L. (1988). Attempts to train novices for beer flavor discrimination: A matter of taste. The Journal Of General Psychology, 115, 403-418.

Pisoni, D. B. (1991). Modes of processing speech and nonspeech signals. In I. G. Mattingly and M. Studdert-Kennedy (Eds.) Modularity and the Motor Theory of Speech Perception (pp. 225-238). Hillsdale, N.J.: Lawrence Erlbaum Associates.

Pisoni, D. B. (1977). Identification and discrimination of the relative onset time of two-component tones: Implications for voicing perception in stops. Journal of the Acoustical Society of America, 61, 1352-1361.

Pisoni, D. B., (1973). Auditory and phonetic memory codes in the discrimination of consonants and vowels. Perception & Psychophysics, 13, 253-260.

Postman, L. (1955). Association theory and perceptual learning. Psychological Review, 62, 438-446.

Reed, S. K. (1972). Pattern recognition and categorization. Cognitive Psychology, 3, 382-407.

Repp, B. H. (1984). Categorical perception: Issues, methods, findings. Speech and language: Advances in basic research and practice, 10, 243-335.

Rosen, S., & Howell, P. (1987). Auditory, articulatory, and learning explanations of categorical perception in speech. in S. Harnad (Ed.) Categorical perception. (pp. 455-490). Cambridge University Press: Cambridge.

Shepard, R. N. (1964). Attention and the metric structure of the stimulus space. Journal of Mathematical Psychology, 1, 54-87.

Smith, C., Carey, S., & Wiser, M. (1985). On differentiation: A case study of the development of the concepts of size, weight, and density. Cognition, 21, 177-237.

Smith, L. B. (1979). A continuum of dimensional separability. Perception & Psychophysics, 25, 285-291.

Smith, L. B. (1989). From global similarity to kinds of similarity: The construction of dimensions in development. In S. Vosniadou and A. Ortony (Eds.), Similarity and analogical reasoning. Cambridge: Cambridge University Press.

Stevens, K. N. (1981). Constraints imposed by the auditory system on the properties used to classify speech sounds: Data from phonology, acoustics, and psycho-acoustics. In T. F. Myers, J. Laver, & J. Anderson (Eds.), The cognitive representation of speech. Amsterdam: North Holland.

Stevens, S. S. (1976). Psychophysics. New York: Wiley.

Strange, W., & Jenkins, J. J. (1978). Role of linguistic experience in the perception of speech. In R. D. Walk & H. L. Pick, Jr. (Eds.) Perception and experience (pp. 125-169). New York: Plenum Press.

Swets, J. A., Tanner, W. P, & Birdsall, T. G. (1961). Decision process in perception. Psychological Review, 68, 301-340.

Vanderplas, J. M., Sanderson, W. A., & Vanderplas, J. N. (1964). Some task-related determinants of transfer in perceptual learning. Perceptual and Motor Skills, 18, 71-80.

Waller, T. G. (1970). Effect of irrelevant cues on discrimination acquisition and transfer in rats. Journal of Comparative and Physiological Psychology, 73, 477-480.

Werker, J. F., & Tees, R. C. (1984). Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. Infant Behavior & Development, 7, 49-63.

Whorf, B. L. (1941). Languages and logic. in J. B. Carroll (ed.) Language, Thought, and Reality: Selected papers of Benjamin Lee Whorf. MIT Press (1956), Cambridge, Mass. (pp. 233-245).

Wisniewski, E. J., & Medin, D. L. (in press). On the interaction of theory and ata in concept learning. Cognitive Science.

Zeamon, D., & House, B. J. (1963). The role of attention in retardate discrimination learning. In N. R. Ellis (Ed.), Handbook of mental deficiency. New York: McGraw-Hill.

Correspondences should be addressed to Robert Goldstone, Psychology Department, Indiana University, Bloomington, IN. 47405. The author's electronic mail address is rgoldsto@ucs.indiana.edu.

Many extremely useful comments and suggestions were provided by Dedre Gentner, Richard Ivry, Earl Hunt, John Kruschke, Dominic Massaro, Douglas Medin, Robert Nosofsky, David Pisoni, Richard Shiffrin, and Linda Smith. The author wishes to thank Audrey Bancroft, Matthew Bennett, Ryan Cobine, Leslie Darrow, Marci Days, Daniel Lipson, Holly Mashburn, Leslie Masterson, Maureen McMahon, and James Wiltz for assistance in conducting the experiments. The research was funded by a Biomedical Research Grant from the National Institute of Mental Health.

Figure 1. Categorization data from Aha and Goldstone (1992). Numbers indicate the percentage of Category A classifications for each specified stimulus. "A"s and "B"s represent the twelve original training stimuli. The results indicate that subjects tend to classify on the basis of line position for stimuli with small sizes, but tend to classify on the basis of size for stimuli with larger sizes.

Figure 2. Demonstrations of acquired equivalence (acquired similarity) and acquired distinctiveness. The category boundary is assumed to occur at stimulus number 4. The white circles show discrimination performance before training. The black circles show discrimination performance after training. Figure reprinted with author's permission (Pisoni, 1991).

Figure 3. Stimuli used in Experiment 2. Sixteen squares are constructed by combining four values of brightness with four values of size factorially.

Figure 4. (Size categorizers' d') - (No categorizers' d'). This figure shows the gain in perceptual sensitivity that is due to size categorization training. A black rectangle indicates a positive difference. A white rectangle indicates a negative difference. The size of the rectangle indicates the absolute magnitude of the difference. Rectangles are placed between the two squares that are being discriminated.

Figure 5. (Brightness categorizers' d') - (No categorizers' d'). This figure shows the gain in perceptual sensitivity that is due to brightness categorization training.

Figure 6. (Size & brightness categorizers' d') - (No categorizers' d'). This figure shows the gain in perceptual sensitivity that is due to categorization training that requires attention to both size and brightness.

Figure 7. (Saturation categorizers' d') - (No categorizers' d'). This figure shows the gain in perceptual sensitivity that is due to saturation categorization training.

Figure 8. (Brightness categorizers' d') - (No categorizers' d'). This figure shows the gain in perceptual sensitivity that is due to brightness categorization training.

Figure 9. (Saturation & brightness categorizers' d') - (No categorizers' d'). This figure shows the gain in perceptual sensitivity that is due to categorization training that requires attention to both saturation and brightness.