Next Section:
Conclusions
IV. Successive Same-Different Discrimination
Thus far, we have discussed the pigeon's classification of variability for simultaneously presented array of icons. It was possible that pigeons may be successful at extracting variability within an array of icons, but be incapable of extracting the variability present within a list of icons. Perceiving array variability makes no memory demands and leverages the perceptual system's ability to determine whether two or more items are the same or different from one another. Perceiving list variability makes significant memory demands (only one of the items is present at any moment) and relies on the cognitive system to compare a viewed item with one or more items that are stored in memory. Thus, the classification of list variability may be far more difficult than the classification of array variability.
We initially trained pigeons to discriminate lists of 16 identical icons from lists of 16 different icons, where Same lists involve minimal variability and Different lists involve maximal variability (Young, Wasserman, & Dalrymple, 1997). Pigeons were subsequently tested with novel lists consisting of all new icons (a transfer test) and novel lists consisting of (a) mixtures of same and different icons in various temporal locations within the list or (b) different numbers of same and different items (Young et al., 1999). Through the systematic exploration of the pigeon's responding to these novel lists, we sought to determine the effective stimulus that was controlling the pigeons' same-different report responses.
The successive version of the same-different task required that the pigeon peck once at each of the 16 icons in the list before the next icon was displayed. The simultaneous version of the task required that the pigeon peck 16 times anywhere on the 7 cm x 7 cm display. Forty-eight highly distinguishable Macintosh icons were chosen as the total item pool; these icons were randomly sorted into three sets of 16 icons each (Set 1, Set 2, and Set 3), and from these three 16-icon sets, 16-icon lists were constructed. For any given Same list, a single icon from the appropriate set was randomly chosen and used to make up a list of 16 identical icons. For the Different arrays, all 16 of the icons of a set were used with no repetitions. The 16 same or 16 different icons were randomly distributed over 25 locations in a 5 x 5 grid.
During training, pecks to the green button on Same trials or to the red button on Different trials were correct and were reinforced with food; pecks to the red button on Same trials or to the green button on Different trials were incorrect and were punished by repetition of the trial until the correct response was made. Button color was reversed for half of the subjects.
A Java applet is available on-line that demonstrates the task (with you playing the part of the pigeon). There are some differences between the on-line task and the actual task (e.g., the Java script alternates Same and Different trials for demonstration purposes, whereas the original program used block randomization); consult Young, Wasserman, and Dalrymple (1997) for procedural details.
Four pigeons were initially trained to discriminate 16-icon lists constructed from either Set A or Set B (reaching an 85% performance criterion) and were later tested with lists constructed from a second set (Set B or Set A, respectively). In our group of four pigeons, initial acquisition of same-different responding was rapid (averaging 35 days), but we did not observe strong transfer to novel 16-item lists; accuracy on training lists averaged 92% correct, but only 57% correct on new testing lists that were created from 16 untrained icons. The pigeons were given further training with lists comprising icons from either Set A or Set B in an attempt to produce a more generalizable concept and were then tested with lists constructed from a third set (Set C). Acquisition of the same-different discrimination for the new set (Set A or Set B) was very rapid (averaging 11 days), and we now observed strong transfer to the novel 16-item lists constructed from Set C; accuracy averaged 94% correct on training lists and 73% correct on the new testing lists. This finding closely corresponds with other research on natural and abstract conceptualization by both humans and nonhumans showing that larger sets of training stimuli promote stronger generalization performance (see Wasserman, 1993 for a review).
This experiment therefore contributed an unprecedented finding: namely, memory-based conceptualization by a nonhuman animal. Unlike all previously reported results, successive same-different discrimination had to be based on the bird's remembering 1 or more of the 16 icons that had been presented in a list. Although this result was remarkable, we were interested in whether the addition of memory demands had produced a fundamental change in how the pigeon solved this successive same-different discrimination: would entropy explain these results or was the pigeon relying on one ore more discriminative processes?
In Experiment 1 of Young et al. (1999), we tested 16-item lists comprising either 1, 2, 4, 8, or 16 different types of icons. Furthermore, these icon types were temporally organized in different ways to produce differences in the number of transitions between icons. For example, a list comprising 4 icon types could be one of the following: aaaabbbbccccdddd, aabbccddaabbccdd, or abcdabcdabcdabcd with 3, 7, or 15 transitions, respectively (each letter in the list represents a randomly chosen icon type). The results revealed that increasing the number of icon types led to more "different" responses and that temporally distributing those different types of items (i.e., increasing the number of transitions) increased the likelihood of a
"different" response (Figure 10). Because entropy predicts that variability is a function of the number and distribution of icon types, but that the temporal distribution of the icons should have no effect on performance, these results suggest that entropy (as previously applied to simultaneous arrays) is not a complete account of successive same-different discrimination performance.
In Experiment 2A, we more fully investigated the role of time on our pigeons' same-different discrimination behavior. We anticipated that items at the end of the list would have a greater influence on pigeons' behavior than items at the beginning of the list (a recency effect). To test this idea, we used lists containing a mixture of same and different items wherein the same items all occurred at the beginning of the list (with the different items occurring at the end of the list) or the same items all occurred at the end of the list (with the different items occurring at the beginning of the list). For example, a "same-first" mixture list comprising a mixture of 8
same items and 8 different items would be presented as aaaaaaaabcdefghi. A pigeon more influenced by later items than by earlier items would be expected to respond different more often to that list (where the different items are all at the end of the list) than to abcdefghiiiiiiii (where the same items are all at the end of the list). This is precisely what we observed (Figure 11).
Experiment 2A thus documented a strong recency effect on our pigeons' behavior. It was possible that a primacy effect could also have been in operation, but that this effect was much weaker than the observed recency effect. We examined this possibility in Experiment 4 by using tests in which a set of 4 different items appeared either at the beginning, in the middle, or at the end of a list otherwise containing same items (we also used lists in which a set of same items occurred at various positions within a list of different items, but the effect of 4 same items had no significant effect on choice behavior in any of the temporal positions). Our analyses revealed a strong recency effect, but no primacy effect in our task.
It is apparent that memory was affecting our pigeons' same-different list discrimination. In Experiment 3, we manipulated the amount of time between each icon in the list. We observed a decrease in performance accuracy for Same and Different lists as the interstimulus interval increased. This result is consistent with pigeons' failing to remember all of the list items.
If the pigeons were not remembering all of the list items, then the entropy of the entire list should not (and did not) account for the birds' discriminative behavior. Extending our behavioral account by including the presence of memory processes thus had the potential of explaining the data; perhaps the pigeons were still using entropy as the discriminative stimulus, but entropy was computed only on those items that were recalled at choice time. Further tests revealed that this simple extension was insufficient.
In Experiments 2A, we tested the birds with same and different lists involving either 2, 4, 8, 12, or 14 items. In Experiments 2B, we tested the birds with same and different lists involving either 2, 4, 8, 12, 20, or 24 items. We found that increasing the number of list items raised discrimination accuracy on both same and
different trials (Figure 12). This result has an important implication. The systematic effect of list length on same lists indicates that entropy is an insufficient account of our pigeons' same-different list discrimination. Any recalled subset of a same list will always have an entropy of 0.0. If choice behavior were solely a function of the entropy of recalled items, then the birds should have pecked at the "same" key for all of the same lists regardless of their length; this result would parallel that observed in pigeons trained with the simultaneous same-different discrimination. We did not, however, observe an asymmetry in the effects of list length on same and different trials; accuracy approached chance for both same and different lists when list length was decreased. We believe that these results implicate an evidence accumulation process.
Our analysis of the pigeons' discrimination of variability in simultaneously presented arrays of items provided strong evidence that the variability in list items had a strong effect on the choice behavior of our subjects. Similar examinations of the pigeons' discrimination of variability in successively presented lists of items provided strong evidence that the entropy of the list's items was an insufficient explanation of the choice behavior of our subjects. An account of the successive same-different discrimination appears to require two additional processes: a memory mechanism (wherein the likelihood of an item being recalled is a direct function of its recency) and an evidence accumulation mechanism (wherein more list items produce more definitive choice behavior). A computational account involving all three mechanisms was successful in accounting for a much higher proportion of the variance in our pigeons' behavior (89% variance accounted for) than an account that included only entropy (68% variance accounted for).
We next considered whether the account offered here for lists of successively presented icons could be applied to the previous experiments involving arrays of simultaneously presented icons; such an integration would provide a parsimonious account of the pigeon's behavior in two very different tasks.
Pigeons' behavior in the simultaneous same-different discrimination involved a series of pecks at the presented icons (pigeons in the simultaneous task are routinely required to peck at the array between 20 and 30 times). This pecking behavior could be functionally equivalent to viewing a list of icons, with each peck at an icon being equivalent to viewing a single icon in a list. The sampling of icons in the simultaneous array is under the control of the pigeon, however, not the experimenter; thus, nothing prevents the pigeon from pecking at 1 icon 30 times rather than at 15 icons 2 times. The pigeons' successful discrimination of simultaneous same arrays from simultaneous different arrays (Young & Wasserman, 1997) suggests, however, that the pigeons are sampling more than a single icon in the display.
Under the assumption that each peck samples a single icon, each "list" of icons that is produced by sequential pecking at a simultaneous array would constitute 30 items (under an FR 30 schedule) and thus be equated for accumulated evidence. In addition, item memory would be factored out because there would be no control of when or how long a pigeon viewed each item (thus making item memory a random factor). These methodological differences leave list variability as the only factor of our three-factor model that could account for discriminative performance in the simultaneous same-different task. And, indeed, a metric of list variability, entropy, was documented to be an excellent predictor of such performance.
This integration of the two discrimination tasks assumes that pigeons sequentially process the icons in simultaneous displays; but, this assumption may not be completely correct. It is possible that pigeons view and process collections of array items in a simultaneous display. To best determine the extent to which our account of successive same-different discrimination can be extended to simultaneous same-different discrimination, it would be necessary to conduct studies involving the manipulation of evidence (e.g., by decreasing the FR requirement for some test trials) and the manipulation of item memory (e.g., by removing items from the display as they are pecked) in a simultaneous same-different task. The integration of pigeons' performance on these two discriminations is an intriguing possibility that we will likely pursue.
Next section: Conclusions
Figures 10, 11, and 12 are adapted from: Young, M. E., Wasserman, E. A., Hilfers, M. A., & Dalrymple, R. M. (1999). An examination of the pigeon's variability discrimination using lists of successively presented stimuli. Journal of Experimental Psychology: Animal Behavior Processes, 25, 475-490.