Avian Visual Cognition


Learning Strategies in
Matching to Sample

Anthony Wright
University of Texas Health Science Center

avcrule.gif (935 bytes)

 

 

 

 

 

 

The topic of this chapter is the different ways or strategies that pigeons learn matching to sample. Some of these strategies, or types of learning, lead to concept learning, whereas others do not. Concept learning is based upon learning the relationship (matching) between the correct choice and the sample stimulus. Failures of concept learning have been previously attributed to learning separate if-then rules for each sample stimulus. In each of two similar experiments, a special test for if-then rule learning along with a transfer test for concept learning was administered to four groups of pigeons. The groups were distinguished by different fixed-ratio response requirements (0, 1, 10, or 20 pecks) to the sample stimulus on each trial. Three cartoon stimuli were used to train matching to sample (duck, apple, and grapes). As with any three stimuli, there were 12 unique displays (e.g. duck, duck, apple) when the right vs. left positions of the choice stimuli are taken into account. Pigeons were trained on 6 of these 12 displays. The other 6 displays were reserved for testing if-then rule following acquisition. There was no substantial evidence for if-then rule learning in these experiments. Pigeons responding 0 or 1 time to the sample (FR0 & FR1 groups) apparently learned the configural pattern of each display. These pigeons seemed to learn each display as a unique gestalt and to go right or left for each display. Pigeons responding 20 times to the sample showed complete concept learning. These pigeons performed as accurately with novel stimuli as with training stimuli. Sample responding by these pigeons apparently directed attention to the elements (e.g., duck) of each display and to relationships (e.g., the identity relationship between the sample and matching comparison) among display elements. Pigeons responding 10 times to the sample were a transition group between these two types of learning. These pigeons showed moderate amounts of concept learning and moderate amounts of configural learning. This transition group indicates that different ways in learning matching to sample by pigeons may occur at the same time. Additional tests of these different types of learning are presented and discussed.

avcrule.gif (935 bytes)

Chapter Outline & Navigation 

I.     Introduction
         Concept learning
           MTS with 2 or 152 stimuli
           The divided-set experiment

II.    General Procedures

III.   Results
         Predictions
           Acquisition results
           Transfer results

IV.   Discussion
           Configural learning
           Relational learning

V.    References

   I. Introduction

The focus of this chapter is on the mechanisms, processes, or strategies by which pigeons learn matching to sample (MTS). In a typical MTS task, a pigeon is presented with a sample stimulus. It pecks the sample and then is presented with two comparison stimuli. One comparison stimulus matches the sample and the other does not match. If it chooses (pecks) the matching comparison, then it is rewarded. It turns out that there are different ways to learn the MTS task.

One way to learn MTS is to compare each comparison stimulus to the sample stimulus and choose the one that matches the sample. This is relational learning. From a description of the task, this is such an obvious way to learn MTS that it was almost implicitly assumed that this was THE way. But descriptions can be deceiving. This is not the only way that pigeons can learn MTS. Indeed, it is not even the preferred way that pigeons learn MTS.

The preferred way that pigeons learn MTS, so it seems, is to learn the configural pattern of the whole display. The display is made up of the sample stimulus and two comparison stimuli. Each display is learned as a unique pattern, similar to an artist’s collage. A correct response to the left side or right side of each unique pattern is learned. If there are not too many displays, pigeons can easily learn the unique pattern of each display.

There is also another way that pigeons might learn MTS. This way is to learn an if…then rule for each sample stimulus (e.g., if sample "duck", then choose comparison "duck"). In operant conditioning terms this is learning a behavioral chain for each sample or conditional discriminative stimulus (Skinner, 1950).

In this chapter, evidence will be presented in support of each of these ways of learning MTS. Although the evidence for if…then rule learning in the MTS experiments presented in this chapter is sparse, the original purpose of these experiments was to test for if…then rule learning. At the time, the accepted explanation for why pigeons did not learn the matching concept in MTS experiments was that they learned if…then rules.

Results from two experiments are presented and discussed. The experimental manipulation that determines which way the task was learned in these experiments was the sample-response requirement. Some of the results from one of these experiments (Experiment B) have been previously published (Wright, 1997). The results from these two experiments were similar and I feel that they are mutually supportive. Before presenting these experiments, a few words will be said about concept learning and learning strategies, generally.

Concept learning

Abstract concept learning is different from other types of learning in its range and degree of generalization. An abstract concept is a rule for accurately dealing with new stimuli. An abstract concept transcends the stimuli used to train it. With humans, problem solving often depends upon concept learning (e.g., Nisbett, Fong, Lehman, & Cheng, 1987), as do many of the basic skills during development (e.g., Daehler & Bukatko, 1985).

Abstract concepts are to be distinguished from so-called "natural" concepts where, for example, pigeons learn a category of person, water, or tree (Herrnstein, Loveland, & Cable, 1976). A more appropriate label for natural-concept learning might be category learning. In any case, the training, testing, and issues are often different. The issues of what is learned in category learning often involves whether subjects learn exemplars, features, or prototypes of the stimuli (e.g., Chase and Heinemann, 2001;  Huber, 2001). On the other hand, abstract concept learning is based on a higher-order rule that transcends any features of the training stimuli. The test for abstract concept learning involves stimuli, which may have no features in common with the training stimuli.

Abstract concepts may appear complex in some applications (e.g., mathematical equations for conservation of energy), but nevertheless they are usually reducible to some simple relationship such as identity or equivalence. A child's cognitive development has been thought to progress through stages where, among other things, the basic identity concept is progressively expanded to include equivalence (conservation) of substance (e.g., matching lumps of clay), length (e.g., matching lines), number (e.g., matching equal numbers irrespective of spatial distribution) and volume (e.g., matching volumes irrespective of container dimensions) (e.g., Daehler & Bukatko, 1985; Piaget & Inhelder, 1966/1969).

MTS with 2 or 152 stimuli

A popular task for testing abstract concept learning and other cognitive processes of animals has been matching to sample. A considerable amount has been written about how some species are capable of learning the matching concept and others are not (e.g., D’Amato, Salmon, & Colombo, 1985; Premack, 1978, 1983). Such hypotheses were supported, if not stimulated, by a string of failures to find concept learning by (e.g., Berryman, Cumming, Cohen, & Johnson, 1965; Cumming, Berryman, & Cohen, 1965; Farthing & Opuda, 1974; Holmes, 1979; Santi, 1978, 1982). It seems clear now why these experiments probably failed to find pigeon concept learning. Pigeons have more recently been shown capable of learning the matching concept by using a large set of training stimuli (Wright, Cook, Rivera, Sands, & Delius, 1988). In the Wright et al. (1988) experiment, two groups of pigeons were trained and tested. One group (2-stimuli) was trained with two cartoon stimuli (duck and apple). Another group (trial-unique) was trained with 152 cartoon stimuli, drawn in a way similar to the duck and apple. The 152 stimuli were shown once each daily session to the pigeons and scrambled between days (Click the link to see the entire stimulus set). Both groups learned the task. The 2-stimuli group learned the task in an average of 25 sessions (76 trials per session). TheClick here to see the trial-unique condition trial-unique group learned the task in an average of 360 daily sessions. Clearly, the trial-unique group required a lot more training to learn the task than those trained with two stimuli, indicating that something different was being learned. A difference in what was learned was revealed by the transfer tests. Novel stimuli were used in these transfer tests. The novel stimuli were each tested only once, by definition of novelty. When the pigeons were correct on novel-stimulus test trials, they were rewarded. The 2-stimuli group showed very little transfer; it was not different (statistically) from chance performance. The trial-unique group, on the other hand, showed complete transfer and hence complete concept learning; transfer performance (80% correct) was not (statistically) different from training-trial performance (also 80% correct). This study showed that pigeons have the cognitive ability to learn abstract concepts--an abstract concept of matching to sample-- in this case.

The Wright et al. (1988) study raised the question of what was being learned by the pigeons (2-stimulus group) that did not learn the concept. Pigeons in the 2-stimulus group clearly learned the task. Indeed, they learned it as well as those subjects (trial-unique group) that learned the concept. The conventional wisdom at that time was that pigeons learned MTS by learning a series of stimulus-specific behavioral chains (Skinner, 1950), equivalent to sample-specific if-then rules (Carter & Werner, 1978). Chaining has been widely used to describe how animals learn a sequence of responses to a sequence of stimuli leading up to reinforcement. If-then rules have been widely used to describe how humans make decisions and solve problems (e.g., Nisbett, 1993). In the experiments of this article, the sample-specific if-then rules with the duck, apple, and grapes stimuli would be: "If sample Duck, then choose comparison Duck", "If sample Apple, then choose comparison Apple," and "If sample Grapes, then choose comparison Grapes." The hypothesis of if-then rule learning explains the lack of pigeon concept learning because a new if-then rule must be learned for each novel stimulus. Support for the hypothesis of if-then learning came from demonstrations showing that pigeons learned symbolic-matching (arbitrary sample-comparison relationship) as rapidly as identity-matching (where a concept could enhance learning) (Carter & Eckerman, 1975; Carter & Werner, 1978). Opposition to the if-then learning hypothesis came from demonstrations showing that pigeons had a slight advantage for new learning if the training relationship (e.g., matching) was preserved rather than changed (e.g., from matching to oddity) (e.g., Zentall & Hogan, 1978). But even this indirect evidence in favor of concept learning, and against if-then rule learning, was shown to result from factors other than the concept (Wilson, Mackintosh, & Boakes, 1985).

The divided-set experiment

The experiments of this article arose from the desire to test whether or not pigeons learned the MTS task via if…then rules. The basic idea was to use three training stimuli (e.g., duck, apple, and grapes) instead of just two. Then the 12 displays resulting from these three stimuli could be divided into two equivalent subsets as shown in Figure 1. The roles of each stimulus were counterbalanced in these two subsets. Each stimulus was represented twice as a sample, was correct once on the left and once on the right, and was incorrect once on the left and once on the right. One subset was used for training and the other subset was used for testing. The idea was that if, if…then rules were learned, then the pigeons should be as accurate on the six testing displays (i.e., those not used in training) as with the six displays used in training. Additionally, the same three elements would be used in all displays so that there could be no novelty-averse reaction to disrupt transfer or confound the interpretation of the results.

Experiment A began as a pilot experiment. The first pigeons were trained in a typical MTS procedure of requiring one peck, fixed ratio of one (FR1), to the sample to produce the comparison stimuli. After they learned the task with six training displays, they were tested with the six testing (untrained) displays, composed of the same three elements, duck, apple, and grapes. I expected good performance with these novel displays, a result that would have demonstrated if…then rule learning. I was absolutely shocked that these pigeons were at chance performance with these ‘untrained’ displays. They were also at chance performance with novel stimuli, a somewhat less surprising result given the history of a lack of concept learning by pigeons in MTS tasks. This result of no transfer to the testing (untrained) displays was very puzzling because if they did not learn the task relationally (i.e., no concept learning), and did not learn the task via if…then rules, then how did they learn?

A second pilot experiment was conducted. Two more pigeons were then trained on FR10 as the sample-response requirement. Previous experiments had shown that sample responding could produce faster acquisition (e.g., Sacks, Kamil, & Mack, 1972). Possibly this faster acquisition in the Sacks et al., (1972) experiment was indicative of a different type of learning. The pigeons trained with a FR10 sample response requirement did learn the MTS task somewhat more quickly than the FR1 pigeons, and these pigeons did show some partial transfer to untrained displays and to novel stimuli. This was change in the right direction. The next obvious manipulation was to increase the FR even more. Two new pigeons were trained on FR20 and tested. These FR20 pigeons now showed complete transfer to the untrained displays and complete transfer to novel stimuli. Thus, there appeared to be a continuum of transfer. The degree of transfer seemed to depend upon the FR value. In order to explore the other side of this continuum, two pigeons were trained on FR0. FR0 is an unusual MTS procedure. The trial began with the whole display presented all at once. The pigeon simply chose (pecked) one or the other comparison stimulus. It was not clear (at least to me) whether pigeons could learn MTS without responding to the sample. But they did learn, although they required more training. They did not show transfer to untrained displays or to novel stimuli. The four groups were then expanded to include six pigeons in each group.

During training and testing in Experiment A, the six training displays (i.e., trial types) were randomly selected. Although the number of presentations of each display was similar at the end of the experiment, there were differences within sessions. In order to equate the number of presentations of each display type within sessions and thereby control this variable, Experiment B was conducted. In Experiment B there were 84 training trials per daily session. Each display was presented 14 times during each training and testing session. Test trials were added, rather than substituted, in order to maintain equal presentation frequencies during testing. Experiment B also employed a slightly stricter acquisition criterion. The pigeons had to be 70% correct or better on all six trial types for two consecutive sessions, instead of 80% correct overall on a single session as in Experiment A.

II. General Procedures

The stimuli were full-color cartoons (computer drawn and presented). The training stimuli were 3 cartoons (duck, apple, and grapes) as shown in Figure 1. The sample was the center cartoon in each display. The choice or comparison stimuli were to either side of the sample. Stimuli were presented from the floor of the chamber by orienting the monitor so that it was tipped upward at a slight angle (4 deg.) toward the subject. Thus, the pigeons pecked down at the stimuli as they might toward objects and seeds on the ground. (See Wright et al., 1988 for more details).

On each trial, the center or sample stimulus appeared first (except for Group 0 where all 3 stimuli appeared simultaneously). The pigeon pecked the sample (number of required pecks varied for each group) and then the comparison stimuli appeared on either side of the sample. If the pigeon pecked (chose) the comparison stimulus that matched the sample, then the incorrect comparison was turned off and 4-5 wheat seeds (mean = 4.4, SD =1.5) were placed on top of the correct comparison stimulus. The sample was turned off 1 s after reward was delivered and the correct comparison remained on for 8 s while the pigeon ate the seeds. If the pigeon pecked the incorrect comparison, the incorrect comparison stimulus was turned off immediately, the sample remained on for 1 s, and the correct comparison remained on for 2 s. Responses to these stimuli that remained following incorrect choices had no (programmed) consequence. These responses were recorded but seldom occurred. No reward was given on a trial where an incorrect choice was made. Following incorrect choices, the chamber was dark (punishment) for 8 s following the offset of the correct comparison stimuli and then the trial was repeated (correction procedure). (Correction trial results are not included in any of the analyses presented in this chapter.) A 15-s inter-trial interval followed the 8-s eating time or the 8-s punishment time.

Pigeons were trained in one of four groups. The difference among groups was the number of pecks to the sample before the choice/comparison stimuli were presented: 0, 1, 10, or 20 pecks for Group 0, Group 1, Group 10, and Group 20, respectively. Although sample responding has not been explicitly investigated with regard to concept learning, it has been shown to enhance matching learning (e.g., Sacks, Kamil, & Mack, 1972) and some indications of partial concept learning have come from studies requiring several sample responses (e.g., Lombardi, Fachinelli, & Delius, 1984; Zentall, Edwards, Moore, & Hogan, 1981). Some intermediate training was required for Groups 10 and 20 before they would respond reliably on their final FR values. The sample FR was increased from 1 to the final value of 10 or 20 (depending upon the group) over several sessions. Increases were titrated as rapidly as possible without causing pauses (ratio strain) due to extinction. Click here for an interactive demonstration of the FR 20 condition.

Figure 1 shows the 12 possible matching displays divided into 2 counterbalanced sets: a training display set and a testing display set. This training/testing regime was used to test for if…then rule learning. In each display subset, the three cartoon stimuli were counterbalanced in terms of the role that they played within each display set of Figure 1. Each cartoon appeared twice as the sample stimulus, twice as the correct comparison stimulus (once on the left and once on the right), and twice as the incorrect comparison stimulus (once on the left and once on the right). The subset in the upper portion of Figure 1 was the training subset for 4 pigeons in each group of Experiment A and 2 pigeons in each group of Experiment B. The subset in the lower portion of Figure 1 was the training subset for 2 pigeons in each group of Experiment A and 2 pigeons in each group of Experiment B.

Pigeons were trained 5-6 days per week. The pigeons in Experiment A were trained on 76 trials daily until (mean) performance was 80% correct or better on a single session. The pigeons in Experiment B were trained on 84 trials daily until performance was at least 70% correct on all six training displays for two consecutive sessions. The criterion used in Experiment B turned out to be slightly more stringent than that used in Experiment A. None of the pigeons in Experiment A would have met the criterion of Experiment B. Thus, all the pigeons in Experiment A would have required more training had this more stringent criterion been used.

On the session following criterion performance in both experiments, the correction procedure was removed and testing begun. There were 10 consecutive daily test sessions with the six testing (untrained) displays. In Experiment A, each test session included 1 trial of each of the 6 testing displays pseudorandomly intermixed with 70 training trials for a total of 76 trials. In Experiment B, each test session included 1 trial of each of the 6 testing displays pseudorandomly intermixed with 84 training trials for a total of 90 trials.

A novel-stimulus test (concept-learning) test was conducted immediately following the if-then rule test. Novel cartoons (objects and animals) were only tested once to eliminate any possibility that rapid learning might confound the measure of concept learning (Premack, 1978, 1983; Wright et al., 1988). There were five consecutive daily test sessions. In Experiment A, each test session consisted of 10 novel-stimulus test trials pseudorandomly intermixed with 66 training trials for a total of 76 trials. In Experiment B, each test session consisted of 10 novel-stimulus test trials pseudorandomly intermixed with 84 training trials for a total of 94 trials. In both experiments there were a total of 50 novel-stimulus test trials and 100 novel stimuli tested (see Wright, et al., 1988, for examples). Stimulus combinations and sequences were varied and counterbalanced within the limits of the experiment. The novel cartoons varied greatly in form and color from each other and from the training cartoons.

III. Results

Predictions

One requirement for evidence in support if-then rule learning would as good performance with the testing displays as with training displays. This is to be expected because the stimulus elements (e.g., duck, apple, grapes) were identical in both cases. If the pigeons had learned if-duck, then choose "duck", they should have no trouble with the testing or untrained displays. Furthermore, performance with novel stimuli should be at chance (i.e., no concept learning) for clear case for if-then rule learning.

Evidence in support of concept learning would be good performance on the novel-stimulus test. A clear case for concept learning would be as accurate performance with novel stimuli as with training displays. Stimuli in the novel test were only tested once (by definition of novelty). Thus, any good performance with the novel stimuli could not be due to a history of reinforcement (i.e., learning). To perform accurately with the novel stimuli the pigeons would have to know the rule (i.e., the matching concept), choose the comparison stimulus that matches the sample stimulus.

There is a third possible learning strategy, configural-pattern learning. Evidence in support of configural learning from this experiment is more or less ‘default’ evidence. After learning the MTS task, if the pigeons are inaccurate on both tests, that is, inaccurate with the untrained displays and with novel stimuli, then this pattern of results points (by default) to configural-pattern learning.

Acquisition Results

The individual acquisition results for the four groups of ExperimentsClick here to view figure 3 A and B are shown in Figure 2 and Figure 3, Click here to view figure 2 respectively. The subjects within each group performed similar to one another in their acquisition of MTS.  There is a trend, in both experiments, for acquisition to be slightly more rapid as the FR increased. These group differences are more clearly shown in Figure 4 by the mean acquisition functions for the two experiments. The separation is  somewhat greater in Experiment A than in Experiment B, at least between groups FR20 and FR0. The comparison between experiments is more clearly shown in Figure 5. The mean functions for groups with the same FR are shown in each of the four panels. There is very little difference between the FR0 groups. Somewhat more training was required for the FR1 group of Experiment B [FR1(B)] than for the FR1 group of Experiment A [FR1(A)]. This Click here to view figure 4 training difference increases somewhat withClick here to view figure 5 increases in the FR. Only in the case of FR20 does there appear to be any substantial difference between experiments in terms of acquisition rate. The FR20 group of Experiment A shows an acquisition advantage over that of Experiment B. This difference was not due to any one subject, for example, one slow learner in Experiment B. For Experiment A the number of trials-to-acquisition were 1368, 1444, 1596, 1596, 1672, and 1976. For Experiment B, the number of trials-to-acquisition were 1764, 2016, 2100, and 3444.

The reason for this acquisition advantage of FR20(A) over FR20(B) is unclear. Possibly fewer trials per session (76 vs. 84) in Experiment A might have had a distributed-practice effect that enhanced performance. If so, that would raise the question of why other groups of Experiment A do not show similar distributed-practice enhancements. Perhaps the FR20 groups learned the task differently than other groups, and this learning was more susceptible to this small distributed-practice effect than the other groups. What the different groups learned is shown from the transfer tests in the next section.

Transfer Results

Transfer test results are shown for both experiments in Figure 6. Results fromClick here to view figure 6 Experiment A are shown in the upper panel and from Experiment B in the lower panel. The baseline (training display) performance from the test with ‘untrained’ displays was similar and not systematically different from the baseline performance from the test with novel stimuli. Therefore, the training-trial performances during the two tests were combined for each group.

The degree of transfer varied considerably across groups, but was similar between experiments. Both experiments show increasing transfer with increasing sample FR. There was little or no transfer for the FR0 and FR1 groups in either experiment. For the FR10 groups of both experiments, novel-stimulus transfer was intermediate between baseline and chance performance. For the FR20 groups, novel-stimulus transfer further increased. In Experiment B, the FR20 group showed complete concept learning. That is, this group’s novel-stimulus transfer was equivalent to its baseline (training-trial) performance. Although novel-stimulus transfer by the FR20 group of Experiment A was slightly less that its baseline, this difference may have resulted from the somewhat less stringent acquisition criterion and consequently less training.

By in large, the degree of transfer to the untrained displays was similar to the degree of transfer to novel stimuli for all groups. This result implicates relational learning being responsible for both the transfer to untrained displays as well as transfer to novel stimuli, and is evidence against if…then rule learning. The one possible exception was the FR10 group of Experiment A. For this group, transfer to the untrained displays was similar to baseline performance with somewhat less transfer to novel stimuli. Had transfer to novel stimuli been at chance performance then this group would have shown a pattern of results in keeping with the predictions of if...then rule learning. This intermediate novel-stimulus transfer was shown by all pigeons. It was not the result of some pigeons showing chance performance and others showing similar-to-baseline performance. The FR10 groups of both experiments showed a blend of different learning processes or strategies. There is some relational-concept learning as shown by partial transfer to novel stimuli. The difference between baseline performance and concept learning (novel-stimulus transfer) is in all probability due to configural-pattern learning, particularly the FR10 group of Experiment B. In general, the evidence for any if…then rule learning in either of these experiments is scant.

IV. Discussion

Configural Learning

Pigeons that did not learn the MTS task relationally (e.g., FR0 and FR1 groups) appear to have learned the task configurally. That is, they learned the gestalt of the three elements (e.g., duck, duck, apple) of the display. The three stimuli of the display form a unique configural pattern. Even pigeons learning delayed MTS showed configural-pattern learning (unpublished pilot experiment). To learn the MTS task, all they needed to do was to memorize each configural pattern and learn whether to go to the right or left for each pattern. Pigeons are very good at memorizing patterns. Photographs are patterns. Pigeons can memorize sets of at least 320 photographs and possibly more (Vaughan & Greene, 1983). Pigeons are predisposed to memorizing as opposed to relational learning, relative to primates such as rhesus monkeys (Wright, Santiago, Sands, & Urcuioli, 1984; Wright, Cook, & Kendrick, 1989).

Pigeons that learned configurally (e.g., FR0 and FR1 groups) made end-runs around relational learning. These pigeons transformed what was assumed to be a relational task, into an item-specific task (See also Huber, 2001 and Pearce, 1987 for configural and item-specific learning in go/no-go tasks). They memorized the specific response (right or left in this case) for each of the six unique displays. A bias of pigeons toward item-specific learning--rather than any limitation on their cognitive abilities for abstract concept learning--may explain why there have been indications of at least partial concept learning by pigeons (e.g., Holmes, 1979; Lombardi et al., 1984; Pisacreta, Redwood, & Witt, 1984; Urcuioli & Nevin, 1975; Zentall et al., 1981).

Display manipulations. Further evidence for item-specific configural-pattern learning was shown by manipulations of certain elements (e.g., duck) of the display (unpublished). On these special tests, some aspects of the training displays were altered. In one such test, the duck was turned upside down. With this alteration, the pigeons performed at chance. This was true whether the upside-down duck was the sample and correct comparison or whether it was the incorrect comparison. Turning the duck upside down apparently altered the configuration of the display for these pigeons. Pigeons that had learned the concept (FR20 group) had no problem with these altered ‘training’ displays. Other alterations of the individual elements (e.g., changing the color of the duck’s beak, or reversing its direction) had a much smaller effect on performance and thus probably had little effect on the overall perception of the display configuration. 

Learning MTS with the entire (configural) display as a sample. Other evidence of configural pattern learning comes from other pigeons trained to use the whole display as a sample stimulus (unpublished). The rationale here was that if the pigeons trained on FR0 or FR1 could memorize the whole display, then this display ought to be usable as a unitary (sample) stimulus. For this training, pigeons were presented with one of the four displays composed from the "duck" and "apple" cartoons (e.g., duck, duck, apple). A peck to the display in the middle position (the sample in other experiments) resulted in presentation of two choice stimuli, "candle" and "bicycle" in the side positions. Right and left positions of candle and bicycle varied quasirandomly following presentation of the duck-duck-apple display. With this display, candle was the correct choice. When the "mirror-image" display, apple-duck-duck, was presented then the bicycle was the correct choice. According to a similar design, the cartoons "lobster" and "book" followed presentations with apple-apple-duck and duck-apple-apple displays. Two pigeons learned this symbolic MTS task demonstrating that the whole configural display can be learned as a unitary stimulus by pigeons.

Relational Learning

The groups of pigeons that made the most sample responses (Groups 10 & 20) clearly learned more than item-specific responses to configural displays. These pigeons learned the task relationally, either partially in the case of the FR10 group, or completely in the case of the FR20 group. Relationally, in this case, means that each comparison stimulus was "compared" to the sample stimulus, and an identity judgment was made based on this comparison. Relational learning is a requirement for accurate performance with novel stimuli. From the standpoint of comparing the cognitive abilities of different species, it is clear that pigeons do have the cognitive ability (and intelligence) to learn the higher-order, abstract concept of "matching."

The greater sample responding by the FR20 group was likely to have altered the pigeon’s perception of the display by breaking up the configural pattern into individual elements (duck, apple, and grapes). Responding many times to the sample may have focused the pigeon’s attention on this element of the display. When the pigeon then made a response to the correct comparison stimulus, then it might have noticed that the comparison element was the same as the sample.

Training MTS with a FR40 sample-response requirement. The FR20 sample response requirement in Experiment B resulted in rapid and complete concept learning. This result and the trend across the other groups raise the question about what would happen with an even larger FR (e.g., FR40). Two experimentally naďve pigeons were trained with a FR40 sampleClick here to view figure 7 response requirement. Other aspects of the procedure were similar to Experiment B (e.g., FR20 group). Figure 7 shows that P6531 learned the task (70% correct with all training displays for two consecutive sessions) in 23 sessions (84-trials per session). P4019 learned in 21 sessions. This learning was not too different from the learning by the FR20 pigeons of Experiment B: P279—21 sessions, P465—24 sessions, P700—25 sessions, and P1146—41 sessions. What was different was the lack of complete concept learning: P6531—70%, P4019—66%. This is considerably less than the 80% novel-stimulus transfer by the FR20 pigeons of Experiment B. In the case of P6531 (but not P4019), even baseline performance dropped during testing: 77.6% correct during the testing-display test and 73% during the novel-stimuli test. Why concept learning was less complete with FR40 than FR20 is unclear.

One possibility is that the overall relationship between transfer and FR is an inverted "U" shaped function. An inverted U-shaped function might be due to the conditioning properties of ratio schedules. Ratio schedules tend to establish a chaining of responses, such that the response itself becomes the stimulus for more responding. The initiation of responding may be the visual stimulus, but after responding is initiated, stimulus control shifts to responding. The degree to which stimulus control is usurped from the sample stimulus might be a function of the size of the FR. Such FR response control might become particularly pronounced with FR40. Perhaps a better schedule for the sample-response requirement would be a variable-interval (VI) schedule. A VI schedule would tend to promote more stimulus control to the sample stimulus and take control away from responding. A disadvantage of a VI schedule is that the numbers of responses would vary from subject to subject and from trial to trial. A blend of VI and FR schedules (e.g., a conjunctive VI- FR schedule) might prove to be a good compromise. Under such a schedule the pigeon would have to fulfill both a ratio and an interval requirement. Responding during the interval would count towards the ratio requirement.

Another possible explanation for the lack of complete concept learning by the FR40 pigeons might be that transfer was conducted prematurely. Additional training might have been necessary for transfer to be complete following FR40 training. In the present experiment, the FR was gradually increased to FR40 over the first 17 sessions. There were only 5 or 7 sessions (depending on the subject) with FR40 before transfer began. Perhaps more training sessions were needed for the pigeons to settle into the routine of responding 40 times to each sample prior to being tested with new stimuli (transfer). In order to test this possibility, pigeons could be trained to criterion and then additional sessions (e.g., 10) conducted. If transfer were complete under such training conditions, then this result would show that the lack of complete transfer was not due to response control by the ratio schedule, but instead was due to incomplete relational learning.

Training-trial performance was excellent for both pigeons at criterion--80% for P4019 and 86% for P6531--yet some of this good performance (especially in the case of P4019) appears to be due to configural-pattern learning. P4019 was 82% correct with the training displays during both transfer tests and only 66% correct with the novel stimuli. (This pigeon was only 73% correct with the "untrained" displays which argues against if…then rule learning.) Superior training-display performance over novel-stimuli performance means that there was some configural-pattern learning. Possibly these pigeons did learn the configural patterns of the displays and their relational performance was only beginning to develop, perhaps aided by the configural-pattern learning. Since there was partial relational learning (66%), this type of learning might have been in the process of developing and might have yielded complete concept learning with a little more training.

Possibility of multiple learning strategies. Pigeons in the FR20 groups of Experiments A and B clearly learned the task relationally; otherwise they could not have performed well with novel stimuli. But relational learning does not in itself preclude configural learning. The FR20 pigeons could have learned the task configurally, as well as relationally. These pigeons could have, for example, begun by learning the task configurally and this configural learning might then have enhanced relational learning. Configural-pattern learning appears to be the preferred way for pigeons to learn the MTS task. Such learning might provide a "crutch" for the less preferred, and possibly more difficult, relational learning for pigeons. Such a possibility makes a certain amount of sense. Subjects might first learn to accurately perform the task in the easiest possible manner, for example, by memorizing examples. Then they might figure out the rules and relationships in the process of making correct responses. Such a possibility may be similar to memorizing language phrases before learning language grammar. There are several pieces of evidence that point to this possibility of first learning the task configurally and then using this configural learning to aid and enhance relational learning.

One piece of evidence that pigeons might learn the MTS task both configurally as well as relationally is the intermediate transfer results for the FR10 group in Experiment B. This group showed intermediate transfer of 66% correct to novel stimuli, which was below their baseline performance of 78% correct. The intermediate transfer to novel stimuli indicates that they had partially learned the task relationally. Individual subjects in the FR10 group of Experiment B showed this result of partial concept learning (60%, 66%, 68%, and 70% correct). Thus, the group average result was not the result of some pigeons learning the concept completely and others not learning it at all. The superior baseline performance (but not superior ‘untrained’ display performance) indicates configural-pattern learning as well as relational learning.

Another piece of evidence that pigeons in the FR20 groups learned the MTS task both configurally and relationally comes from the slow learning by the trial-unique group of pigeons trained with 152 stimuli, scrambled daily (Wright et al., 1988). These pigeons had no possibility of learning the task configurally because individual displays were seldom, if at all, repeated. Thus, they had to learn the task relationally. They learned the abstract concept, but required about 10 times as much training as pigeons in the FR20 groups. Configural learning by pigeons in the FR20 groups may have enhanced their relational learning.

The possibility of configural learning enhancing relational learning argues for testing these two types of learning during acquisition. If configural learning enhances relational learning, then configural learning might develop somewhat earlier than relational learning and provide evidence for synergism between these two types of learning. Such a result might provide insights into how concepts are learned, open new ways for exploring mechanisms of concept learning, and perhaps lead to better ways to train concepts.

 

    V. References

       Berryman, R., Cumming, W. W., Cohen, L. R., & Johnson, D.  F. (1965). Acquisition and transfer of simultaneous oddity. Psychological Reports, 17, 767-775.

           Carter, D. E., & Eckerman, D. A. (1975). Symbolic matching by pigeons: Rate of learning complex discriminations predicted from simple discriminations. Science, 187, 662-664.  

            Carter, D.  E.,  & Werner, T. J. (1978).  Complex learning and information processing by pigeons: A critical analysisJournal of the Experimental Analysis of Behavior, 29, 565-601.

             Chase, S. & Heinemann, E. G. (2001). Exemplar memory and discrimination. In R. G. Cook,  (Ed.), Avian visual cognition [On-line]. Available: pigeon.psy.tufts.edu/avc/chase/

            Cumming, W. W., Berryman, R., & Cohen, L. R. (1965).  Acquisition and transfer of zero delay matching.  Psychological Reports, 17, 435-445.

            Daehler, M. W., & Bukatko, D. (1985). Cognitive development. New York, NY: Knopf.

            D'Amato, M. R., Salmon, D. P., & Colombo, M. (1985). Extent and limits of the matching concept in monkeys  (Cebus apella). Journal of Experimental Psychology: Animal Behavior Processes, 11, 35-51.

            Farthing, G. W., & Opuda, M. J. (1974). Transfer of matching-to- sample in pigeons. Journal of the Experimental Analysis of Behavior, 21, 199-213.

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

            Holmes, P. W. (1979).  Transfer of matching performance in pigeons. Journal of the Experimental Analysis of Behavior, 31, 103- 114.

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

             Lombardi, C.  M., Fachinelli, C. C., & Delius, J.  D.  (1984). Oddity of visual patterns conceptualized by pigeons. Animal Learning & Behavior, 12, 2-6.

            Nisbett, R.  E.  (1993). Rules for reasoning.  Hillsdale, NJ: Lawrence Erlbaum.

            Nisbett, R. E., Fong. G. T., Lehman, D. R., & Cheng, P. W. (1987). Teaching reasoning. Science, 238, 625-631.

            Piaget, J., & Inhelder, B. (1969). The psychology of the child (H.  Weaver, Trans.). New York, NY: Basic Books.  (Original work published 1966)

            Pisacreta, R., Redwood, E., & Witt, K.  (1984).  Transfer of matching-to-figure samples in the pigeonJournal of the Experimental Analysis of Behavior, 42, 223-237.

            Premack, D. (1978). On the abstractness of human concepts: Why it would be difficult to talk to a pigeon. In S. H.  Hulse, H. Fowler, & W. K. Honig (Eds.), Cognitive processes in animal behavior (pp. 423-451).  Hillsdale, NJ: Lawrence Erlbaum.

             Premack, D.  (1983).  Animal cognition.  Annual Review of Psychology, 34, 351-362.

            Sacks R.A., Kamil, A.C. & Mack, R. (1972).  The effects of fixed-ratio sample requirements on matching-to-sample in the pigeon.  Psychonomic Science, 26, 291-293.

            Santi, A. (1978). The role of physical identity of the sample and correct comparison stimulus in matching-to-sample paradigms. Journal of the Experimental Analysis of Behavior, 29, 511-516.

            Santi, A.  (1982). Hue matching and hue oddity in pigeons: Is explicit training not to peck incorrect hue combinations a sufficient condition for transfer? Psychological Record, 32, 61-73.

            Skinner, B.  F.  (1950). Are theories of learning necessary? Psychological Review, 57, 193-216.

            Urcuioli, P. J., & Nevin, J. A. (1975). Transfer of hue matching in pigeonsJournal of the Experimental Analysis of Behavior, 24, 149-155.

            Vaughan, W.,  Jr., & Greene, S.  L.  (1983).  Acquisition of absolute discriminations in pigeons. In M. L. Commons &  R. J.  Herrnstein (Eds.), Quantitative Analysis of Behavior: Vol. IV  (pp. 231-238).  New York, NY: Harper & Row.

           Wilson, B., Mackintosh, N. J., & Boakes, R. A. (1985).  Matching and oddity learning in the pigeon: Transfer effects and the absence of relational learning.  Quarterly Journal of Experimental Psychology, 37, 295-311.

           Wright, A. A. (1997). Concept learning and learning strategies. Psychological Science, 8, 119-123.

            Wright, A. A., Cook, R. G., & Kendrick D. F. (1989). Relational and absolute stimulus learning by monkeys in a memory task. Journal of the Experimental Analysis of Behavior, 52, 237-248.

           Wright, A. A., Cook, R. G., Rivera, J. J., Sands, S. F, & Delius, J. D. (1988). Concept learning by pigeons: Matching to sample with trial-unique video picture stimuli. Animal Learning and Behavior, 16, 436-444.

           Wright, A. A., Santiago, H. C., Sands, S. F., & Urcuioli, P.  J. (1984).   Pigeon and monkey serial probe recognition: Acquisition, strategies, and serial position effects.  In H. L. Roitblat, T. G. Bever, & H. S. Terrace  (Eds.), Animal cognition (pp. 353-374).  Hillsdale, NJ: Lawrence Erlbaum.

            Zentall, T. R., Edwards, C. A., Moore, B. S., & Hogan, D.  E. (1981).  Identity:  The basis for both matching and oddity learning in pigeons.  Journal of Experimental Psychology: Animal Behavior Processes, 7, 70-86.

            Zentall, T. R., & Hogan, D. E. (1978). Same/different concept learning in the pigeon: The effect of negative instance and prior adaptation to transfer stimuli. Journal of the Experimental Analysis of Behavior, 30, 177-186.  

    Acknowledgement

    This research was supported by grants R01 MH-35202 and R01 DA-10715 to the author. The author thanks Jacquelyne J. Rivera for her assistance in conducting the experimental sessions. Inquiries should be addressed to the author: University of Texas Medical School, Department of Neurobiology and Anatomy, P.O. Box 20708, Houston, Texas 77225 or by email: aawright@nba19.med.uth.tmc.edu.