Avian Visual Cognition

Categorization & Acquired Equivalence 

Peter J. Urcuioli 
Department of Psychology, Purdue University 


 

Like humans, pigeons exhibit the ability to categorize different objects on the basis of physical resemblance (i.e., "looking alike") and on the basis of having learned that some objects share a common function or association (e.g., producing the same response). In this chapter, I review some of the avian research on categorization and acquired equivalence. The review includes a discussion of the procedures used to train categorical discriminations and, more importantly, to subsequently confirm that what pigeons have learned is something that transcends their specific training experiences. The data obtained from these procedures confirm that pigeons and other avians readily exhibit one of the hallmarks of categorization and acquired equivalence: the ability to immediately generalize what they were explicitly taught to new examples and/or from one member of a group to another. I also describe a long-standing theoretical account of how pigeons accomplish transfer of what they've learned about some members of a learned equivalence class to other members, as well as some recent data suggesting that the mechanisms of acquired equivalence are richer and more varied than previously thought.

 

Chapter Outline & Navigation

I.     Introduction

II.   Categorization by Appearance

III.  Categorization by Association

IV.  Questions and Issues

V.   Psychological Mechanism

VI.  Conclusions and Summary

VII. References

I. Introduction

The purpose of my chapter is to give readers an appreciation of the fact that a well-developed characteristic of our own behavior - the ability to group different objects or stimuli together, to treat them as though they were equivalent - is not unique to our species. This behavioral characteristic, known variously as "categorization" or "conceptualization", is easy to take for granted because for human adults, it is certainly second nature. But the psychological processes responsible for this remarkable ability may not be entirely obvious or may even be somewhat different from what we might think. For instance, it is reasonable to believe that human language provides the groundwork for categorization given that we typically provide the same name for many things that we group together (e.g., "fruit"). If language were truly necessary for categorization, however, it would imply that animals without language would be incapable of grouping together different objects they encounter into the same class. But, as you will see, this implication is demonstrably false. Consequently, other processes besides language must be involved in categorization. Indeed, it may be that the common names given to things we treat as equivalent are mostly a reflection of, rather than just a foundation for,  these processes. 

The research literature contains many demonstrations of categorization and conceptualization in non-human animals and a substantial proportion of these involve pigeons and other avian species as experimental subjects (see, for example, the chapters by Cook, (2001), Huber (2001), Kirkpatrick (2001), and Young & Wasserman, (2001)). So, in the spirit of this cyberbook and my own particular research interests, I will be focusing my discussion on the avian work. 

There are two major points or "take-home" messages I hope to convey. The first is that pigeons, like humans, readily group together things that look alike. In other words, pigeons react similarly to objects that perceptually resemble one another just as we react similarly to the coherent groups that we call "cars", "people", "trees", etc. Of course, pigeons (unlike humans) cannot supply a common language label to the members of these groups, but they can be taught to do something very much analogous. Indeed, their version of "common names" is one way that we know that pigeons, too, treat complex stimuli that are similar (but not identical) in appearance as members of an equivalent class of objects. "Looking alike" or perceptual similarity, then, is one basis for pigeon categorization, and it supports the development of the pigeon version of common names. The issue of what constitutes perceptual similarity is rather involved (see the chapters by Blough (2001) and by Huber (2001)), so I will skirt this issue and focus instead on the behavioral consequences of providing the same name to things that look alike.

Learning common names or responses to a set of stimuli is not merely diagnostic of categorization. It is noteworthy for another reason: those responses or names may themselves help to hold the group members together more firmly (Catania, 1996). For example, our category "person" may cohere even more strongly than the cohesion generated by perceptual similarity alone precisely because we provide a common label for all objects in that category. Similarly, the "person" category might be strengthened because we've learned through experience that its members share the common attribute of answering us when we speak to them.

This, then, brings me to my second, and perhaps more important, point. The common associations shared by different objects promote the development of an acquired equivalence between them. More specifically, pigeons (like humans) group together objects on the basis of associative similarity, and this occurs even when the objects do not look at all alike (i.e., are perceptually dissimilar). A human example of this type of categorization would be the class formed by cars, trains, airplanes, boats, bicycles, and mopeds. Although very different in appearance, we recognize all of these things as members of a group ("modes of transportation") bound together by their common function of moving people from one place to another. Thus, common behavioral functions or associations are another basis for categorization. Such acquired equivalence -  what some psychologists (e.g., Lea, 1984) refer to as "conceptualization" - is a major topic for discussion in my chapter because it transcends mere "looks" and because it, too, can be observed in pigeons' behavior.

After describing some techniques for revealing categorization and acquired equivalence in pigeons and related species and the data that these techniques have yielded, I will address how "bird brains" might manage such remarkable behavioral feats. Here, I'll get into some psychological nitty-gritty - namely, the theorizing about possible mechanisms that might produce acquired equivalence effects in creatures without language. I'll end by describing how these hypothesized mechanisms can be evaluated experimentally and what actual experimental tests have revealed about them.

II. Categorization by Appearance

One way that pigeons and humans categorize events is on the basis of how similar they are in appearance to each other. In short, things that look alike tend to be treated alike. The most straightforward example of how perceptual similarity yields behavioral similarity is the common name we typically assign to groups of objects, such as people, cars, flowers, etc. We can, of course, perceptually differentiate between group members (viz., we can tell one person from another). Nonetheless, their common perceptual features are a major factor in our categorical discrimination (Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1976), and this ability produces tremendous economy in behavior (Younger & Cohen, 1985). For example, we do not have to learn how to react individually to each and every person but, instead, can generalize (quite appropriately) the behavior we've learned across different individuals. Thus, we know that speaking to one member of the group "people" - even someone we've never previously encountered (and who is recognized as such) - will most likely produce a reply in just the same way that we've actually experienced in our interactions with other people. Conversely, we know that speaking to members of the group "fish" will not produce the same effect.

This sort of discriminative behavior (i.e., generalization within a group and discriminating between groups) and the generalization of this behavior to new instances, has been elegantly documented and analyzed in pigeons. Herrnstein and Loveland (1964), for example, trained pigeons to peck for food at any of a number of photographic pictures that contained a person or people somewhere in the picture. Figure 1 shows a series  of "people" slides used in some of Herrnstein's work. Click Here To View Figure 1Just as is typical in human experience, there was wide variation in the appearance of the people shown in these slides - e.g., their number, orientation, size, color of their clothing, etc. Other slides that did not contain a person in them were also shown to the pigeons; when these appeared, pecking at them did not produce food. Examples of "non-people" slides can also be seen in Figure 1. With training, pigeons soon came to peck quickly and rapidly at the "people" slides and not to peck at the "non-people" slides. Moreover, this learned discriminative behavior of pecking versus not-pecking at the various slides in the "people" versus "non-people" categories, respectively. also generalized immediately to pictures that the pigeons had never seen in training. In other words, when shown a new slide containing a person, pigeons pecked at it, despite never having been reinforced with food for pecking at that slide. Similarly, when shown a new slide without a person in it, the pigeons either did not peck at it or did so much less frequently than to a "people" slide, again despite never having experienced non-reinforcement with that particular slide. 

These latter results represent very crucial data for inferring categorization: the pigeon's explicitly taught behavior generalized to new instances of the "people" and "non-people" groups. In other words, they behaved like we would if asked to say what type of object something was even though we had never seen that particular object before. Our ability to generalize the name/label we've learned for a group to novel instances/objects is one of the litmus tests for categorization (Herrnstein, 1990; Keller & Schoenfeld, 1950; Roberts, 1996).

An even closer analog to human categorization by appearance was developed more recently  by Wasserman and his colleagues (e.g. Bhatt, Wasserman, Reynolds, & Knauss, 1988). Wasserman aptly refers to their task as the pigeon version of the "name game", which parents often use to teach their young children the names of different objects from a number of perceptually distinct categories. On each training trial in the pigeon version of the name game, hungry pigeons were shown a photographic slide on a center viewing screen that depicted one of a variety of examples drawn from four different categories: cats, chairs, cars, and flowers. After pecking at the slide appearing on the center screen, pigeons could then obtain food by "naming/labeling" it appropriately. This Click Here to view Figure 2 meant pecking one of four different keys located at the vertices of the viewing screen. Figure 2 shows the apparatusClick here to view Figure 3 used to display the pictures to the pigeons along with the location of the different response options, and Figure 3 shows some of the slides used by Bhatt et al. (1988) in their studies.  Thus, if a picture of a flower appeared on the center screen, pecking the top left key produced food; if a picture of a car appeared, pecking the top right key produced food; and so on.

Like the pigeons in Herrnstein and Loveland (1964) study, these pigeons readily learned to correctly categorize the various pictures shown to them in each training session. In this case, they quickly learned to confine their pecks to one key when shown any instance of a cat, to another key when shown any instance of a flower, and so on. The consistency in where the pigeons pecked as a function of category (like the consistency in the names we would provide) suggested that the pigeons had grouped the various slides together by appearance. Things that looked alike were treated alike, in other words. Click here to view  Figure 4 More importantly, the four different, explicitly taught pigeon "names" (viz., location responses) generalized immediately and appropriately to novel instances of each group that were later shown to them during a subsequent generalization test (see Figure 4). I say "more importantly" because one might dismiss the data obtained during initial training by claiming that the pigeons simply did what they were taught, independently of  any similarities (or dissimilarities) each stimulus may have had to the other training stimuli. The generalization test results obtained with the never-before-seen stimuli make that alternative explanation less plausible.

Still, how confident can we be that these performances represent the same sort of categorization-by-appearance that humans exhibit?  For instance, do pigeons really "see" perceptually coherent groups of objects, as implied by the idea that perceptual similarity underlies their categorization performances?  These are difficult questions to answer definitively, but other research (e.g., Astley & Wasserman, 1992; Cole & Honig, 1994) strongly suggests that the processes underlying the pigeons' performance are indeed perceptually based. In other words, the consistency in the pigeons' categorical responses is not merely an "accident" of learning (and remembering) what response to make to each particular picture in order to get food. To distinguish between "accident" and true categorization, researchers have compared pigeons' performances on the task I've just described with the performances of other pigeons that have been shown exactly the same set of photographic slides but have been taught a "pseudo-category" task. In this task (Wasserman, Kiedinger, & Bhatt, 1988, Exp. 2), each picture is again associated with one of the four possible responses but different pictures within a particular group (e.g., different cats or different flowers) are associated with Click here to view Figure 5different responses. For example, the pigeon might receive food for pecking the top left key after seeing one of the flower pictures, but the bottom right key after seeing a picture of a different flower, and so on. Pigeons learn this task, too, but they do so much more slowly and they are not as accurate in their choices as the "true category" pigeons as shown in Figure 5.

Other evidence that the pigeons' choices reflect the perceptual similarities among the pictures in each class comes from Wasserman et al. (1988, Exp. 1) and from Astley and Wasserman (1992). In the Wasserman et al. study, pigeons learned the four-choice task as depicted in Figure 3 but with only two (instead of four) groups of stimuli  - e.g., with only flowers and cats. Pigeons had to respond to one location (e.g., the top left key) to get food after seeing some of the flowers, but had to respond to another location (e.g., the bottom right key) after seeing other flowers. Similarly, different reinforced choices (e.g., top right versus bottom left) were associated with different subgroups of cats. The pigeons were able to learn this task indicating that they could discriminate between different flowers and between different cats. Of greater interest were the types of errors the pigeons made when they misclassified a slide. When pigeons pecked a wrong location following a particular flower (or cat), the wrong location they pecked most often was the correct location for the other subgroup of flowers (or cats). In other words, pigeons were more likely to confuse the items within a category than the items from different categories.

We can also ask whether the pigeons' ability to categorize different objects in a group is something they've learned or is simply a reflection of their biological/ perceptual apparatus. The answer to this nature-nurture question is similar to those for many other nature-nurture questions: Most likely, both. For instance, the category-pseudocategory difference shows that the pigeon's visual system, too, perceives objects that look alike and those that don't ("nature") and, as result, they can be readily taught to respond in consistent ways on the basis of visual appearance. However, the ability to transfer those responses to new instances appears to depend on learning ("nurture") - in particular, on learning to respond similarly to objects that are similar but not identical. In other words, pigeons (like humans) must learn that the inherent variation among the features exhibited by group members is behaviorally unimportant or inconsequential.

The contribution of the learning ("nurture") component can be gauged by comparing how well pigeons transfer their responses to new exemplars after having learned to categorize 1 instance from each group versus 12 instances (Wasserman & Bhatt, 1992). After training with only 1 instance/group, pigeons do not respond correctly to new examples (i.e., the accuracy of their choices is no better than chance). On the other hand, they are very accurate after training with 12 pictures/group.

This latter finding is important for another reason to which I alluded earlier. Specifically, even when it appears as though categorization/equivalence is perceptually based (i.e., due to "nature"), chances are that the category is also held together by the common behavior conditioned to the group members. Common behavior is only possible, however, when pigeons and humans are exposed to more than one member/group. But that situation is undoubtedly the typical one. Thus, I would argue that these common associations enhance the integrity of the group. Moreover, we know that common associations per se will generate an acquired equivalence, even when the members of the group do not look at all alike. It is that phenomenon to which I now turn.

III. Categorization by Association

Humans are adept at grouping together objects that have common functional attributes. Examples of categories that fit this description (so-called "superordinate" categories) include articles of clothing, modes of transportation, and animals. Some authors (e.g., Lea, 1984) have argued that this form of categorization is more in line with what we mean by the term "conceptualization" because it carries with it the connotation that categorization often transcends mere appearances. Are pigeons also capable of conceptualization, despite the absence of language that could potentially provide mediators to link such disparate objects together?  They are indeed. Furthermore, their abilities at categorization by association can be observed in a variety of different learning situations.

Revealing this sort of grouping between objects sharing a common association makes use of the transfer (or transfer-of-control) technique. The rationale behind this technique is that new things learned specifically about some members of a group should transfer without additional explicit training to other members of the same group. Stated otherwise, if a common association shared by different objects produces an acquired equivalence between them, then those objects should also be interchangeable with one another in new situations.

One example of categorization by association can be seen after differential outcome training, a procedure described in the chapter by Grant and Kelly (2001). Such training is often conducted in the context of delayed matching-to-sample, so a brief description of a simple version of this task is in order. On each trial of delayed matching, pigeons see one of two different sample stimuli on the center key of a three-key display. After viewing and/or pecking the sample stimulus, two comparison or choice stimuli appear simultaneously on the adjacent side keys. Pecking one of the two simultaneously displayed comparisons is reinforced on trials beginning with one sample, whereas pecking the other comparison is reinforced on trials beginning with the other sample. An "incorrect" comparison choice is never reinforced. Successive trials Click Here for Animation are separated from one another by a blank period called the intertrial interval. An animated illustration of "symbolic" delayed matching, in which the comparison alternatives are different stimuli than the samples that precede them, can be viewed by clicking here. Note that in this animation, the pigeon receives the same reinforcer - food - for pecking the vertical comparison stimulus after the red sample and for pecking the horizontal comparison stimulus after the green sample.

We can transform this procedure into a differential outcome task simply by providing different reinforcers for the two correct comparison choices. For instance, if pigeons were both hungry and thirsty, they might receive food for pecking the vertical comparison after the red sample and water for pecking the horizontal comparison after the green sample (Brodigan & Peterson, 1976). In other versions of differential outcome training, different types of food are delivered for each correct choice response (Edwards, Jagielo, Zentall, & Hogan, 1982) or food is delivered with a higher probability for one response than for the other (Urcuioli, 1990). In some of my research (Urcuioli, 1990), hungry pigeons received food in a lighted food hopper for making a correct choice response on one trial type (e.g., red sample/vertical stripes) versus the lighted hopper without food for making a correct response on the other trial type (e.g., green sample/horizontal stripes). The lighted hopper by itself serves as a conditioned reinforcer, meaning that it is rewarding because of its learned association with food. In any event, the lighted hopper is a different outcome than when food accompanies it.

We might reasonably ask, then, whether the hue samples and the stripe comparisons that "share"  the same reward become equivalent. Unfortunately, the answer to this particular question is uncertain because of complications arising from the fact that these stimuli appear in different locations (i.e. the center key vs. the adjoining side keys). It turns out that pigeons treat the same stimulus appearing in different locations as if it were a different stimulus (Lionello & Click here to view Figure 6 Urcuioli, 1998), so devising a valid test for equivalence between samples and comparisons that are associated with the same reinforcer is extraordinarily difficult. However, we can show that different center-key (sample) stimuli associated with the same reward become equivalent (Edwards et al., 1982). Here's how this demonstration is accomplished (See Figure 6).

Initially, pigeons learn a red-sample/vertical-stripes and green-sample/horizontal-stripes differential-outcome delayed matching task, as just described. Afterwards, they learn two new associations that involve new stimuli appearing on the center key (circle and dot) and the two familiar outcomes (food and lighted hopper). For example, they might receive food shortly after seeing after the circle on the center key but just the lighted hopper after seeing the dot. After this second phase of training, then, pigeons will have been exposed to pairs of different center-key stimuli that have the same reward association: namely, red and circle (both associated with food), and green and dot (both associated with the lighted hopper only). We can now ask if those common reward associations have made the members of each of these pairs equivalent. In other words, do the pigeons treat the red and circle as if they belonged to one category, and likewise green and dot?

If they do, then red and circle, and green and dot, should be interchangeable with one another in novel situations. We can easily conduct a "novel-situation" transfer test by returning to delayed matching but with the following modification. Rather than presenting red and green as samples, present the circle and dot instead. Then, when the vertical and horizontal stripes appear on the side keys, observe which of these two comparisons the birds peck following each "novel" sample. Note that although the vertical and horizontal comparisons are familiar from initial training, they've never before appeared after the circle and the dot. Nonetheless, assuming that red and circle have become equivalent (and, likewise, green and dot), Click Here To View Figure 7pigeons should peck the vertical stripes after seeing the circle (because that choice had been correct after red) and the horizontal stripes after seeing the dot (because that choice had been correct after green). That's precisely how pigeons behave, thus illustrating categorization by association. Figure 7  shows transfer results from two studies that used a design like that depicted in Figure 6. Urcuioli (1990) actually used food and a lighted hopper as the two outcomes for correct choice; Peterson (1984) used food and the brief sounding of a tone. The blue bars show pigeons' accuracy with the originally trained samples; the red bars show their accuracy on the very first test session with the "novel" stimuli now appearing as samples. With only two possible choices (i.e., between vertical and horizontal comparisons), chance performance = 50% correct. Clearly, birds in both of these studies were well above chance in their test performances and, in Urcuioli (1990), chose as accurately with the "novel" samples as with the explicitly trained ones.

The difference between this sort of test and the generalization tests run in studies of categorization by appearance  is that the stimuli used to conduct the novel-situation test are not truly novel. The circle and the dot center-key stimuli were previously seen by the pigeons, but in a different context. The transfer test with these stimuli simply assesses whether or not pigeons have learned "what's true of one member of a group is true for all."  In other words, the transfer test essentially asks if prior learning to choose vertical after presentation of one member of the "food" group (red), and horizontal after presentation of one member of the "lighted hopper" group (green), generalizes to other (familiar) members of these groups (i.e., circle and dot, respectively). Lea (1984) has called this "instance-to-category generalization" - the behavior learned to one instance of a category generalizes to other instances of that same category.

Another example of categorization by association in pigeons occurs after training on a task called Click here to view Figure 8 "many-to-one matching-to-sample". It is similar to the matching task just described in that pigeons see center-key (sample) stimuli presented individually on each trial followed by a pair of choice or comparison alternatives on the adjacent side keys. In the simplest many-to-one procedure, there are still two comparison alternatives but four (rather than two) possible sample stimuli. Pecking one comparison is reinforced after two of the sample stimuli, and pecking the other comparison is reinforced after each of the remaining two samples. The common association in this task, then, is the correct comparison choice shared by pairs of sample stimuli, as shown in the top panel of Figure 8.

Given the associative similarity between the red and circle samples (i.e., both are cues to choose vertical) and between the green and the dot samples (i.e., both are cues to choose horizontal), have pigeons learned that the members of each sample pair are equivalent?  We can again answer this question by conducting a transfer test to see if what's true of one instance of each category is true of all (viz., instance-to-category generalization). For example, if we now explicitly teach pigeons to make a new choice response to one member of each sample pair, will that new response transfer to the other member of each pair?

To find out, we train pigeons (after their many-to-one experience) on another delayed matching task involving two ofClick here for MTS Demonstration the original four sample stimuli (e.g., red and green) and two new comparison choices (e.g., blue and yellow). In this "reassignment" task, we might arrange for pigeons to obtain food for pecking, say, blue after seeing the red sample and yellow after seeing the green sample, as shown in the middle panel of Figure 8. (Click here for an interactive demonstration of this experiment) After teaching the birds these new choices, we then ask if pigeons will transfer these newly learned choices to the remaining (viz., the circle and dot) samples. Specifically, will pigeons now peck blue after seeing the circle sample and yellow after seeing the dot sample (bottom panel of Figure 8), even though they were never explicitly taught to do so?  If so, this would be evidence of an acquired equivalence between the circle and red, and between the dot and green, based upon the common comparison associations shared by these samples during many-to-one training.

Actual tests of this nature have been conducted in a couple of different ways. One has been reinforce whatever choice (blue or yellow) the pigeon makes following each sample in the test phase (nondifferential reinforcement). This way, how the pigeon performs over the course of a test session is not biased in favor of certain sample-choice combinations. By using nondifferential reinforcement in testing, any preference that pigeons do show Click here to view  Figure 9 for choosing blue after the circle sample and yellow after the dot sample is strong evidence of an acquired sample equivalence that had developed during initial, many-to-one training. Wasserman, DeVolder, and Coppage (1992) have run this type of transfer test and found clear preferences of just this sort. Some of their data appear in Figure 9.

Another test procedure run in my laboratory, provides differential reinforcement in the test session: food is delivered only for one particular choice after each sample (the alternative choice being nonreinforced). For instance, pigeons might receive food for pecking blue (but not yellow) after the circle sample and for pecking yellow (but not blue) comparison after the dot sample. As just described, these reinforced choices would be consistent with any acquired equivalence that may have developed between the red and circle, and between the green and dot, samples given that pigeons had already learned to peck blue after red and yellow after green. Such "consistent" contingencies, then, should promote very accurate performances in testing.

To avoid being misled about what that result, if obtained, would mean - e.g., to avoid the possibility that "consistent" pigeons are accurate in testing simply because they very quickly learn what goes with what within the test session itself - another group of pigeons is tested with the opposite contingencies:  contingencies inconsistent with the choices that should occur if the form and hue samples are equivalent. In the inconsistent condition, pigeons would receive food for pecking yellow (but not blue) after the circle sample and for pecking blue (but not yellow) after the dot sample. In contrast to the consistent birds, they should perform very poorly if the hues and forms are equivalent. 

Actual data from a consistent versus inconsistent test assessment are shown in Figure 10 which plots the percentage of the initial 16 transfer trials in which individual birds in each condition chose the reinforced comparison. As predicted on the basis of Click here to view Figure 10 acquired equivalence, the consistent birds were generally very accurate, matching well above the level expected by chance (50%) alone, whereas the inconsistent birds were quite inaccurate, matching at or below the level expected by chance. The substantial difference between conditions shows, without a doubt, that the birds' choices in testing were not simply the result of quickly learning new sample-choice relations. (If they were, the average performances in these two test conditions should be identical.)  Instead, the data provide solid evidence that pigeons categorize by association. Pigeons had learned in many-to-one training, in other words, that the red and circle samples were members of the same category as were the red and green dot samples - categories formed on the basis of their common association with a particular comparison choice.

These are not the only techniques for demonstrating categorization by association (e.g., see Honey & Hall, 1989, and Kaiser, Sherburne, Steirn, & Zentall, 1997) but they are common and have been used  in similar studies with humans (Spradlin, Cotter, & Baxley, 1973). An interesting alternative assay, however, was recently described by Manabe, Kawashima, and Staddon (1995) in a study with a different avian species - the budgerigar. They taught budgies to produce the comparison stimuli in matching-to-sample by making one of two different vocal calls, a high or a low call, depending on which sample stimulus (red or green) appeared on the center key. Pecking one of the two side-key comparisons (vertical or horizontal lines) then produced food after one sample, whereas pecking the alternative comparison produced food after the other sample. After the budgies had learned to make the correct call to each sample and to peck the correct set of stripes after each sample, Manabe et al. added two more samples to the procedure (viz., vertical and horizontal stripes on the center key). This addition created a many-to-one task: now, two different samples (one  hue and one  line) shared the same reinforced choice response. The unique twist to the Manabe et al. procedure was that the budgies could produce the  comparisons on trials with the new samples by making any vocal call. (They were still required, however, to call high or low to the hue samples.) 

As training on the many-to-one task progressed, the budgies learned which comparison to choose after each of the newly introduced line samples. But the more interesting  result was how they reacted to the new samples themselves as learning about the correct choices proceeded. Specifically, the budgies began to emit high and low calls to the vertical and horizontal samples and, furthermore, these calls  "matched" the calls they made to the hue samples associated with the same comparison choice. In other words, if the budgies were required to make a high call to the red sample, they began to make a high call to the line sample that was associated with the same comparison choice as red. Likewise, if they had to make a low call to the green sample, they began to make a low call to the line sample that was associated  with the same choice. Again, this occurred despite the fact that any call made in the presence of the line samples Click here to see and hear video of this "equivalence" behavior would produce the comparisons. An example of a budgie making high and low calls to red and green samples and then matching these samples to red and green comparisons is shown in the following video clip. This clip does not correspond exactly to the delayed matching procedure I've just described (viz., the comparisons are hues instead of lines) butClick here to view Figure 11 it does clearly illustrate examples of the two different calls that were explicitly  conditioned to the hue samples and the matching procedure in general. The choice and call data during many-to-one training for one subject (S6)  in Manabe et al.'s study appear in Figure 11. This figure shows the accuracy of correct comparison choice on trials with the newly introduced samples over successive training sessions and the proportion of calls to these samples that were consistent with those made to the other (hue) samples with the same comparison association. 

These fascinating and provocative results have another possible (and plausible) interpretation (Saunders & Williams, 1998) which, unfortunately, is too involved to explain here in a brief and simple fashion. Nonetheless, the Manabe et al. (1995) results may yet  represent another example of categorization by association. Different-looking samples associated with the same comparison ostensibly became members of the same group, and that acquired equivalence became apparent in how the birds behaved to the samples themselves. In other words, the behavior explicitly taught to some samples - the high and low calls - transferred to the other samples sharing the same comparison association. Assuming the acquired equivalence interpretation to be correct, this effect would, once again, represent instance-to-category generalization:  what was true of one member of each category (i.e., making a high or low call) became true of other members of the same category (resulting in "matching" calls). 

IV. Questions and Issues

There can be little doubt that pigeons are able to categorize different things they see on the basis of similarity in appearance and similarity in function. These capabilities show a higher level of functioning than most people might normally attribute to them and, from the standpoint of psychological research and theorizing, show that categorization and acquired equivalence does not require human language. But the generalization and transfer data that demonstrate the pigeon's categorization abilities raise a few questions.

For example, if pigeons truly treat different objects as belonging to the same category - either because of their appearance or their function - then why are their performances during the generalization or transfer tests not "perfect" or at least indistinguishable from their performances with the training stimuli?  One answer to this question is that "equivalent" does not mean "equal/identical to". For instance, the fact that pigeons can correctly sort new instances of flowers, chairs, cats, etc. into different, previously learned response categories does not mean that they are unable to distinguish between these new instances and the ones used in training. In fact, the "subgroup" categorization study (Wasserman et al.,1988, Exp. 1) discussed earlier shows that pigeons can make those distinctions. This is important because if pigeons could not discriminate between a new instance and a familiar one, then the generalization test results from these studies would be trivial (viz., they would simply show that pigeons continue to do what they had been explicitly taught to do with each given stimulus). A truly novel stimulus, then, has some attribute(s) that is (are) detectably different from the stimuli with which pigeons have been trained. Consequently, if the "particulars" of the explicitly trained category members also influence the bird's behavior, then new particulars would be expected to cause some disruption in performance, thus yielding less than perfect test performance.

The same argument can be made for the transfer data from categorization-by-association studies. An analogy to consider is that even though we treat shoes and ties as belonging to a particular category (articles of clothing), we do not treat them identically in all contexts and situations. Equivalent implies treating these objects the same with respect to a particular function (e.g., we wear them), but not necessarily with respect to all functions (e.g., the part of the body on which we put them). Similarly, pigeons that learn to choose vertical (as opposed to horizontal) after seeing either a red or a circle sample might treat those samples as interchangeable (equivalent) within that context. When later taught to choose a new (e.g., blue) comparison after seeing just the red sample, they might learn (among other things) that blue is what to peck after that particular stimulus. Consequently, when later shown the circle and asked to choose between blue and yellow, they may not consistently choose blue because they recognize that the circle is not red. Here, too, the particular physical properties of each sample may exert some control over the birds' behavior, such that a change in those properties should yield less-than-perfect transfer despite acquired equivalence.

A second question might be:  Is it truly the case that acquired equivalence requires a common association in training?  The answer to that question is "yes". For example, in the differential outcome paradigm, transfer of delayed matching across samples does not occur if those samples do not share the same reward association (e.g., Peterson, 1984; Urcuioli 1990). Similarly, in the common comparison (many-to-one) task, the transfer effects indicative of acquired equivalence do not materialize if pigeons learn exactly the same set of sample-choice relations in training but learn them in such a way that the common associations are never experienced concurrently (Urcuioli, DeMarse, & Zentall, 1995). 

This "non-concurrent" type of training can be seen in the bottom half of Figure 12 in the row Click here to view Figure 12 designated "One-to-many". The top half of Figure 12 shows the concurrently trained common associations (in italics) that are part of many-to-one delayed matching. [The entire top row writes out the training contingencies and the transfer test that are depicted visually in Figure 8.] Note especially that the two (initial and reassignment) training phases for the One-to-many condition involve learning the same set of sample-choice relationships as in the Many-to-one condition. Likewise, the transfer test in both conditions is identical. More importantly, however,  note that the samples associated with a common comparison stimulus (e.g., red and circle, and green and dot) appear in separate phases of training in the One-to-many condition (italicized) - this is what I mean by "non-concurrent". That seemingly minor difference has a major impact,  however, on the results of the transfer test used to assess acquired equivalence.

Specifically, the consistent versus inconsistent differences Click here to view Figure 13 in test performances typically seen following many-to-one training are absent following one-to-many training. This can be seen in the right panel of Figure 13 which shows choice accuracies over the first 20 test trials for consistently and inconsistently tested pigeons that had been exposed to the one-to-many training procedure (Urcuioli et al., 1995, Exp. 2). Accuracies look very much alike in both test conditions. By contrast, the left panel shows the corresponding data from birds initially trained on many-to-one matching, the concurrently trained, common comparison association task. A clear and substantial consistent-inconsistent difference in transfer is once again obtained.

V. Psychological Mechanism

It's one thing to show that training pigeons in certain ways yields an acquired equivalence. It's another thing to explain how the transfer indicative of equivalence comes about. What mechanism allows pigeons that have been trained on many-to-one delayed matching to transfer the new choice responses explicitly learned to one pair of samples to the other pair? 

This general issue was raised 60 years ago by the famous learning psychologist, Clark Hull, in a paper entitled "The problem of stimulus equivalence in behavior theory". In the introduction of that paper, Hull quickly identified the problem discussed in this section:  "How can we account for the fact that a stimulus will sometimes evoke a reaction to which it has never been conditioned?" (Hull, 1939, p. 9). If we substitute "sample stimulus" in place of the word "stimulus" and "choice response" in place of "a reaction", we have much the same question I posed at the end of the preceding paragraph. Let's review the example of transfer following many-to-one matching (top half of Figure 12) to see the issue.

Pigeons initially learn to match the red and circle samples to a vertical choice and the green and dot samples to a horizontal choice. They then learn to match one sample from each associative "set" to new comparisons: the red sample to blue and the green sample to yellow. We then find that pigeons can now match the circle sample to blue and the dot sample to yellow, reactions that had never been conditioned. How do they do it?  Obviously, pigeons treated the red and circle samples as equivalent and the green and dot samples, too. But what does "treating them as equivalent" entail?

For instance, does it mean that the circle begins to look like red or vice versa?  This seems unlikely. Alternatively, do both the circle and red samples remind the bird of the common choice (i.e., vertical) it has to make after each of them to get food?  Interestingly, something very similar to this notion was Hull's explanation of equivalence effects. Were he to analyze the categorization-by-association effect, Hull would argue that pigeons taught many-to-one matching to high levels of accuracy anticipate which choice they will make when they see each sample stimulus. In other words, pigeons anticipate vertical upon seeing the red or the circle sample and they anticipate horizontal upon seeing the green or dot sample. The notion of anticipation is discussed and illustrated in Grant and Kelly (2001) and is popular for explaining differential outcome and certain types of memory effects.

The diagram in Figure14 shows how these hypothesized anticipatory reactions could produce Click here to view Figure 14 transfer of  blue-yellow choices from one set of (explicitly trained) samples to the other ("untrained") set. During reassignment training that follows many-to-one matching and when pigeons learn to peck blue after red and yellow after green, seeing the red sample generates the anticipatory reaction of "vertical" (from the many-to-one task) and seeing the green sample generates the anticipatory reaction of "horizontal". Consequently, pigeons are also learning to choose blue when anticipating vertical and to choose yellow when anticipating horizontal. The result, then, is that birds should now be able to match any sample generating a "vertical" or "horizontal" anticipation to blue and yellow. This means that the circle and dot should be effective samples because they too remind the birds of "vertical" and "horizontal", respectively. Hull called this "secondary generalization"; others have termed it "mediated generalization":  blue-yellow choices generalize from samples to which they've been explicitly conditioned (red and green) to samples to which they've not been conditioned (circle and dot) because each pair of samples produces common mediators (the anticipation of "vertical" and "horizontal"). 

Although this is a perfectly plausible explanation of categorization-by-association in pigeons, is it accurate?  One way to find out is to devise an independent test of the mediated generalization hypothesis by, first, deriving a prediction based on the assumption of anticipation and, second, running an experiment to see if the prediction holds.

I have recently completed just such an experiment. The entire experiment involved three groups of pigeons, but only two need to be described. One group (Group First) was initially trained on many-to-one matching and were then taught during subsequent reassignment training to make new choice responses to one pair of samples. Finally, they were tested for acquired equivalence using the remaining samples and the new choice responses. This is the usual training and testing sequence I've previously described for the many-to-one condition and that's illustrated in the top row of Figure 12. These pigeons, then, were expected to show the usual transfer effects: accurate performances if tested with "consistent" contingencies, but inaccurate performances if tested with "inconsistent" contingencies. Such findings would replicate my previous findings and, of course, would be in line with the mediated generalization account.

The other group (Group Second) in this experiment was the more interesting one. These pigeons learned the same two training tasks as Group First (i.e., many-to-one matching and reassignment) but in reverse order. In other words, Group Second learned many-to-one matching after they had already learned to match two of the samples to the comparisons that would later appear in testing. Otherwise, what they learned in the two phases of training and the contingencies (consistent or inconsistent) to which they were exposed during the transfer test were exactly the same as in Group First. [Imagine the order of initial and reassignment training for the Many-to-one condition shown in Figure 12 being switched - that's Group Second.]

According to the mediated generalization (anticipation) explanation, Group Second should not show transfer indicative of acquired equivalence. In other words, their performances during the initial trials of the transfer test should be close to chance (50%) and should not differ, on average, as a function of whether the test involves consistent or inconsistent relations. The reasoning behind these predictions is that what pigeons might learn to anticipate upon seeing the samples in many-to-one matching (i.e., vertical and horizontal) won't become connected to the other comparison stimuli (blue and yellow) that appear in the other phase of training and, later,  in the transfer test. Why not?  Because birds in Group Second will have already learned to match with those other comparisons before they learn any common anticipatory reactions to the samples in many-to-one matching (and, thus, to the two samples originally matched to those other comparisons). 

Click here to view Figure 15The results from this experiment, however, did not support this prediction. There was clear evidence of transfer (i.e., of acquired equivalence) in Group Second as well as in Group First. Figure 15 shows some of the first-session transfer data.  The blue bars show choice accuracy over the first 20 transfer trials for birds tested in the consistent condition; orange bars show choice accuracy for birds tested in the inconsistent condition. The left and right panels plot data from six birds each in Groups First and Second, respectively. For both groups, "consistent" pigeons (blue bars) matched well above the level expected by chance, whereas "inconsistent" pigeons (orange bars) matched at or below the level expected by chance.

What do these results mean? To me, they suggest that if pigeons use a mediational strategy to guide their performances during transfer, then the mediator - the common reaction that ties  samples previously associated with the same comparison choice together - can be something other than the anticipation of those choices. In other words, even though the comparison anticipation could serve as a mediator to produce the test results observed in Group First, Group Second's results indicate that pigeons probably have other means for grouping together the commonly associated, many-to-one samples . 

At this point, what those "other means" might be can only be surmised. One possibility is that one sample serves as a reminder or a retrieval cue for the other sample with which it shares a common association. For instance, given that the red and circle sample were both cues to choose vertical, perhaps birds are reminded of red when they see the circle, and/or vice versa. Researchers call this a "retrospective" process as opposed to the anticipatory or "prospective" process already described, and distinction that is nicely illustrated by animation 1 (retrospection) and animation 2 (prospection) from Grant and Kelly (2000).

Figure 16 illustrates how such a retrospective process might work for Group Second. Some of the pigeons in the consistent Click here to view Figure 16 condition were required, during their transfer test, to match the circle sample to blue and the dot sample to yellow. If the circle sample reminded these birds of "red" and the dot sample reminded them of "green" (i.e., of the hue samples associated with the same comparisons as the circle and dot in many-to-one matching), then this would closely correspond to what these pigeons actually learned during initial training: choose blue after red and yellow after green. This correspondence is highlighted in Figure 16 by comparing the connections between blue and yellow and the colored words. If being reminded of something ("red"and "green") is similar to the "real thing" (red and green), then the consistent birds should be quite accurate during testing (and the inconsistent birds quite inaccurate), as observed. This is, of course, speculative although no more so that postulating anticipatory reactions.

VI. Conclusion and Summary

Trying to pin down the psychological mechanisms responsible for these and other findings in the categorization and acquired equivalence literature represents some of the interesting and challenging questions in this line of research. Whatever the mechanisms, however, they are clearly not language-based and they permit a type of "cognitive" behavior that we so readily take for granted in ourselves. 

I should mention at this point, I suppose, that I am not adopting the position (nor are others in this line of work) that the extent of, and the processes underlying, categorization and acquired equivalence are no different in pigeons than they are in humans. To the contrary, researchers are keenly aware of the fact that there are equivalence phenomena readily exhibited by humans that have little, if any, known counterpart in pigeons and other animals (e.g., Saunders, Williams, & Spradlin, 1996; Zentall, 1998). Perhaps down the road, this observed "gap" in known conceptual abilities will shrink as we become better at refining our techniques, and at developing new techniques, to study avian cognition. But even if  more discriminating research in the future fails to shrink the gap, this will not detract at all from the significance of the demonstrable abilities I have briefly reviewed in this chapter. After all, what we see in the pigeon's behavior may be a reflection of the origins of our own, more extensive abilities, especially when we consider that we too are largely the product of our experiences. The continued study of avian cognition thus holds considerable promise in providing additional, valuable insights into some of the basic behavioral processes of categorization, acquired equivalence, and other cognitive phenomena.

VII. References

Astley, S., & Wasserman, E. A. (1992). Categorical discrimination and generalization in pigeons: All negative stimuli are not created equal. Journal of Experimental Psychology: Animal Behavior Processes, 18, 193-207.

Bhatt, R. S., Wasserman, E. A., Reynolds, W. F., Jr., & Knauss, K. S. (1988). Conceptual behavior in pigeons: Categorization of both familiar and novel examples from four classes of natural and artificial stimuli. Journal of Experimental Psychology: Animal Behavior Processes, 14, 219-234.

Blough, D. S. (2001). The perception of similarity. In R. G. Cook  (Ed.), Avian visual cognition [On-line]. Available: pigeon.psy.tufts.edu/avc/dblough/

 Brodigan, D. L., & Peterson, G. B. (1976). Two-choice conditional discrimination performance of pigeons as a function of reward expectancy, prechoice delay, and domesticity. Animal Learning & Behavior, 4, 121-124.

Catania, A. C. (1996). On the origins of behavior structure. In T. R. Zentall & P. M. Smeets (Eds.), Stimulus class formation in humans and animals (pp. 3-12). Amsterdam, NL: Elsevier.

Cole, P. D., & Honig, W. K. (1994). Transfer of a discrimination by pigeons (Columba livia) between pictured locations and the represented environments. Journal of Comparative Psychology, 108, 189-198.

Cook R. G. (2001). Hierarchical stimulus processing in pigeons.  In R. G. Cook  (Ed.), Avian visual cognition [On-line]. Available: pigeon.psy.tufts.edu/avc/cook/

 Edwards, C. A., Jagielo, J. A., Zentall, T. R., & Hogan, D. E. (1982). Acquired equivalence and distinctiveness in matching-to-sample by pigeons: Mediation by reinforcer-specific expectancies. Journal of Experimental Psychology: Animal Behavior Processes, 8, 244-259.

Grant, D. S. & Kelly, R. (2001). Anticipation and short-term retention in pigeons. In R. G. Cook  (Ed.), Avian visual cognition [On-line]. Available: pigeon.psy.tufts.edu/avc/grant/

 Herrnstein, R. J. (1990). Levels of stimulus control: A functional approach Cognition, 37, 133-166.

 Herrnstein, R. J., & Loveland, D. H. (1964). Complex visual concept in the pigeon. Science, 146, 549-551.

 Herrnstein, R. J., Loveland, D. H., & Cable, C. (1976). Natural concepts in pigeonsJournal of Experimental Psychology: Animal Behavior Processes, 2, 285-302.

 Honey, R. C., & Hall, G. (1989). The acquired equivalence and distinctiveness of cuesJournal of Experimental Psychology: Animal Behavior Processes, 15, 338-346.

Huber, L. (2001). Visual categorization in pigeons. In R. G. Cook  (Ed.), Avian visual cognition [On-line]. Available: pigeon.psy.tufts.edu/avc/huber/

 Hull, C. L. (1939). The problem of stimulus equivalence in behavior theory. Psychological Review, 46, 9-30.

Kaiser, D. H., Sherburne, L. M., Steirn, J. N., & Zentall, T. R. (1997). Perceptual learning in pigeons: Decreased ability to discriminate samples mapped onto the same comparison in many-to-one matching. Psychonomic Bulletin & Review, 4, 378-381.

Keller, F. S., & Schoenfeld, W. N. (1950).  Principles of psychology.  New York, NY:  Appleton-Century-Crofts.

Kirkpatrick, K. (2001). Object recognition. In R. G. Cook  (Ed.), Avian visual cognition [On-line]. Available: pigeon.psy.tufts.edu/avc/kirkpatrick/

Lea, S. E. G. (1984). In what sense do pigeons learn concepts? In H. L. Roitblat, T. G. Bever, & H. S. Terrace (Eds.), Animal cognition (pp. 263-276). Hillsdale, NJ: Lawrence Erlbaum.

Lionello, K. M., & Urcuioli, P. J. (1998). Control by sample location in pigeons' matching-to-sampleJournal of the Experimental Analysis of Behavior, 70, 235-251.

Manabe, K., Kawashima, T., & Staddon, J. E. R. (1995). Differential vocalization in budgerigars: Towards an experimental analysis of namingJournal of the Experimental Analysis of Behavior, 63, 111-126.

Peterson, G. B. (1984). How expectancies guide behavior. In H. L. Roitblat, T. G. Bever, and H. S. Terrace (Eds.), Animal cognition (pp. 135-148). Hillsdale, NJ: Lawrence Erlbaum.

Roberts, W. A. (1996). Stimulus generalization and hierarchical structure in categorization by animals. In T. R. Zentall & P. M. Smeets (Eds.), Stimulus class formation in humans and animals (pp. 35-54). Amsterdam, NL:  Elsevier.

Rosch, E., Mervis, C. B., Gray, W. D., Johnson, D. M., & Boyes-Braem, P. (1976). Basic objects in natural categories. Cognitive Psychology, 9, 382-439.

Saunders, K. J., & Williams, D. C. (1998). Do parakeets exhibit derived stimulus control? Some thoughts on experimental control procedures. Journal of the Experimental Analysis of Behavior, 70, 321-324.

Saunders, K. J., Williams, D. C., & Spradlin, J. E. (1996). Derived stimulus control: Are there differences among procedures and processes? In T. R. Zentall & P. M. Smeets (Eds.), Stimulus class formation in humans and animals (pp. 93-109). Amsterdam, NL:  Elsevier.

Spradlin, J. E., Cotter, V. W., & Baxley, N. (1973). Establishing a conditional discrimination without direct training: A study of transfer with retarded adolescents. American Journal of Mental Deficiency, 77, 556-566.

Urcuioli, P. J. (1990). Some relationships between outcome expectancies and sample stimuli in pigeons’ delayed matching. Animal Learning & Behavior, 18, 302-314.

Urcuioli, P. J. (1996). Acquired equivalences and mediated generalization in pigeon’s matching-to-sample. In T. R. Zentall & P. M. Smeets (Eds.), Stimulus class formation in humans and animals (pp. 55-70). Amsterdam, NL: Elsevier.

Urcuioli, P. J., Zentall, T. R., & DeMarse, T. (1995). Transfer to derived sample-comparison relations by pigeons following many-to-one versus one-to-many matching with identical training relations. Quarterly Journal of Experimental Psychology, 48, 158-178.

Urcuioli, P. J., Zentall, T. R., Jackson-Smith, P., & Steirn, J. N. (1989). Evidence for common coding in many-to-one matching: Retention, intertrial interference, and transfer. Journal of Experimental Psychology: Animal Behavior Processes, 15, 264-273.

Wasserman, E. A., & Bhatt, R. S. (1992). Conceptualization of natural and artificial stimuli by pigeons. In W. K. Honig & J. G. Fetterman (Eds.), Cognitive aspects of stimulus control (pp. 203-223). Hillsdale, NJ: Lawrence Erlbaum.

Wasserman, E. A., DeVolder, C. L., & Coppage, D. J. (1992). Nonsimilarity-based conceptualization in pigeons. Psychological Science, 3, 374-379.

Wasserman, E. A., Kiedinger, R. E., & Bhatt, R. S. (1988). Conceptual behavior in pigeons: Categories, subcategories, and pseudocategories.  Journal of Experimental Psychology: Animal Behavior Processes, 14, 235-246.

Younger, B. A., & Cohen, L. B. (1985). How infants form categories. In G. H. Bower (Ed.), The Psychology of Learning and Motivation (pp. 211-247). New York, NY: Academic Press.

Young, M. E. & Wasserman, E.A. (2001). Stimulus control in complex arrays. In R. G. Cook  (Ed.), Avian visual cognition [On-line]. Available: pigeon.psy.tufts.edu/avc/young/

Zentall, T. R. (1998). Symbolic representation in animals: Emergent stimulus relations in conditional discrimination learning. Animal Learning & Behavior, 26, 363-377.

Preparation of this chapter and some of the research described in it were supported by a research grant from the National Institutes of Mental Health (MH56487) to Peter J. Urcuioli. The author thanks Will Vaughan, Jr. for providing the slides used by Herrnstein in his visual categorization studies and, especially, Bob Cook whose extraordinary and unselfish efforts brought this chapter and this book to fruition. 

This chapter is dedicated to the memory of Werner K. Honig, the author's mentor and one of the inspirational founders of the field of avian cognition.