rate of responding during test trials with ABC was faster than during a compound composed of 3 ... tered the discrimination, test trials with the prototype re-.
Journal of Experimental Psychology: Animal Behavior Processes 1994, Vol. 20, No. 3, 264-277
Copyright 1994 by the American Psychological Association, Inc. 0097-7403»4/$3.00
Prototype Effects in Categorization by Pigeons Ay dan Ay din and John M. Pearce In 4 experiments pigeons received a categorization task involving 6 simultaneous compounds in which the elements A, B, and C were more frequently paired with food than were the elements D, E, and F. Food was delivered after compounds ABF, AEC, and DEC but not after DEC, DBF, and AEF. Subsequent testing revealed a higher rate of responding during ABC than during any of the compounds that had signaled food and a lower rate of responding during DBF than during any of the compounds that had not signaled food. Experiments 2, 3, and 4 further demonstrated that the rate of responding during test trials with ABC was faster than during a compound composed of 3 elements that had individually been paired with food. The results are more consistent with a configural than an elemental analysis of discrimination and categorization.
A frequent finding from studies of categorization with humans is that exemplars that can be said to represent the central tendency of a category are the easiest to classify correctly. Reed (1972) described one example of such a prototype effect; Reed showed students cartoon drawings of faces that belonged to two categories. The faces for one category had, on average, smaller foreheads, more closely spaced eyes, and shorter noses than those in the other category. Once the subjects were able to sort the faces correctly, test trials with novel patterns were given. The test patterns that were most accurately classified were those that corresponded to the averages of the faces belonging to the two categories, that is, the prototypes. In a rather different study, Posner and Keele (1968) generated four prototypical patterns from a number of dots. They then distorted these patterns to some extent by moving the dots, and the resultant figures were used for a categorization study. Once subjects were able to sort the distorted patterns into groups on the basis of their prototypes, they were required to sort a number of novel patterns, including the prototypes. The more similar a novel pattern was to a category prototype, the easier it was to classify. For the above-mentioned experiments, the prototypes were based on the mean values of the various dimensions from which the exemplars of the respective categories had been constructed. Prototype effects can also be observed when the test stimulus is based on a rather different measure of central tendency. Goldman and Homa (1977) conducted a study based on cartoon drawings of faces, in which the prototypical faces were composed of the features that most frequently appeared in the training faces. Such modal faces were classified more efficiently than any other faces, inAydan Aydin and John M. Pearce, School of Psychology, University of Wales College of Cardiff, Cardiff, United Kingdom. This research was supported by a research studentship awarded by the Turkish government to Aydan Aydin, and by a grant from the United Kingdom Science and Engineering Research Council. Correspondence concerning this article should be addressed to Aydan Aydin, School of Psychology, University of Wales College of Cardiff, Cardiff CF1 3YG United Kingdom.
264
eluding those presented during training. Examples of related findings have been reported by Hayes-Roth and Hayes-Roth (1977) and Neumann (1977). There is now abundant evidence to show that animals can solve categorization problems. To cite two examples, pigeons are able to categorize photographs of natural objects such as trees (Herrnstein, Loveland, & Cable, 1976), and they can also categorize drawings of faces similar to those Reed (1972) used (Huber & Lenz, 1993). There remains, however, some uncertainty as to whether it is possible to obtain prototype effects with animals. Only a few experiments have looked for such effects, and most of these have been unsuccessful. One experiment, conducted by Watanabe (1988), was based on the design used by Posner and Keele (1968). Pigeons were shown a number of different patterns that were generated from a prototype that consisted of a triangle formed from dots. Pecking on a response key during patterns that were based on this prototype occasionally resulted in the delivery of food, whereas responses during other patterns never resulted in food. Although the subjects mastered the discrimination, test trials with the prototype revealed a level of responding that was by no means superior to that during other test patterns. Another study, conducted by Pearce (1989), used patterns that were composed of three vertical bars that each varied in height on a scale from 1 to 7. Patterns belonging to the tall category were composed of bars with a total height of 15 units, whereas the combined height of the bars belonging to the short category was 9 units. Using an autoshaping design, the delivery of food to pigeons was signaled by the short patterns but not by the tall patterns, and eventually a higher rate of responding was recorded during presentation of members of the short category rather than the tall category. Test trials with novel patterns then revealed that responding during presentation of the short prototype (a pattern composed of three bars, each 3 units high) was slower than during a pattern composed of 3 bars that were each 1 unit high. Conversely, responding during presentation of the tall prototype (a pattern composed of three bars that were 5 units high) was faster than during a pattern composed of 3
265
PROTOTYPE EFFECTS
bars that were each 7 units high. In other words, subjects showed a better discrimination between patterns that represented the extremes of the short-tall continuum than between patterns that represented the average of the exemplars used for training. The remaining experiments that have sought evidence of prototype effects were conducted by Lea and Harrison (1978), Jitsumori (1993), and Fersen and Lea (1990). In Lea and Harrison's study, pigeons were trained in a discrimination task in which patterns constructed from three features signaled whether a reward was available for pecking a response key. The patterns consisted of two shapes (Al or A2) that varied in brightness (B1 or B2) and were presented against backgrounds of different colors (Cl or C2). All six features were paired with food, but the positive features (Al, Bl, and Cl) were paired with food more frequently than were the negative features (A2, B2, and C2). Thus, food was signaled by the three patterns that could be constructed from two positive features and one negative feature but not by the three patterns that could be constructed from two negative features and one positive feature. In a test phase, subjects were presented with patterns A1-B1-C1 and A2-B2-C2. These patterns can be regarded as the prototypes for the two categories because they contained, respectively, the features that occurred more frequently than any other on the reinforced and nonreinforced trials (see Goldman & Homa, 1977). There was no indication that the latency to respond to these stimuli was any different to that recorded during other members of the category to which the stimuli belonged. Moreover, when subjects were presented simultaneously with a prototype and another member of the same category, there was no systematic preference for one stimulus over the other. This was true for both categories. Jitsumori's (1993) experiment was based on the design that Lea and Harrison (1978) developed, and, despite the use of more complex stimuli, the results from this study were essentially the same as those of its predecessor. One can find more promising evidence of a prototype effect in animals in Fersen and Lea's (1990) study. This had the same type of design as the one used by Lea and Harrison (1978), except that there were five binary-valued dimensions from which the stimuli were created. Features Al, Bl, Cl, Dl, and El can be referred to as positive features and A2, B2, C2, D2, and E2 can be referred to as negative features. Any pattern that contained three or more positive features signaled food, whereas food was never presented during patterns composed of three or more negative features. In this experiment, the prototypes were presented throughout training, and eventually every bird showed a higher rate of responding during the positive prototypes (Al, Bl, Cl, Dl, and El) than during any other type of stimulus. There was no evidence, however, that the negative prototype resulted in a consistently slower rate of responding than did the remaining stimuli. Fersen and Lea's (1990) study thus suggests that obtaining prototype effects with animals is possible, but overall the evidence for this conclusion is by no means compelling. Accordingly, the initial concern of the experiments described in this article is to confirm that prototype effects can
be reliably demonstrated with animals. Having achieved this aim, our further concern in these experiments was to evaluate various explanations for our findings.
Experiment 1 We based the design of the first experiment, which used pigeons, on the one used by Lea and Harrison (1978) but with a different method for evaluating the properties of the prototypes. The prototype for the positive category, which signaled food, was composed of the positive features A, B, and C; the prototype for the negative category, which signaled the omission of food, was composed of the negative features D, E, and F. Members of the positive category were composed of two positive features and one negative feature, ABF+ AEC+ DBC+, and members of the negative category were composed of two negative features and one positive feature, DECo DBFo AEFo. The stimuli were presented on a television screen behind a transparent response key. We anticipated that autoshaped responding would be more vigorous during presentation of members of the positive category than of the negative category. Once this pattern of responding emerged, test trials assessed the level of responding during each prototype. If it is possible to obtain prototype effects with animals, then the rate of responding during the positive and negative prototypes will be, respectively, faster and slower than during any other pattern. We used a variety of methods to assess the level of responding to the test stimuli. First, once the pigeons had mastered the discrimination task, occasional probe trials with the prototypes were presented during the training sessions. The indication from these tests was that performance during the prototypes was marginally superior to some, but not all, of the members of their categories. The modest success of these tests may be attributed to the fact that the training stimuli elicited a high rate of responding in the positive category and a low rate in the negative category. Any special properties of the prototypes may thus have been masked by an insensitivity in our test procedure. To circumvent this obstacle, we conducted further tests with the positive prototype once the rate of responding during the positive category had been reduced by extinction. We also conducted further tests with the negative prototype once responding during the negative category had been enhanced by conditioning. Method Subjects. The subjects were 20 experimentally naive adult pigeons (Columbia livia; Abbott Brothers, Norfolk, UK). They were housed in pairs and had free access to water and grit in their home cages. Before the start of the experiment, they were gradually reduced to 80% of their free-feeding weights and were maintained at this level throughout the experiment by being fed a restricted amount of food following each experimental session. The pigeons were maintained in a light-proof room in which the lights were on for 14.5 hr each day. They were tested on successive days, at the
266
AYDAN AYDIN AND JOHN M. PEARCE
same time, during periods when the lights were on in their holding room. Apparatus. The experimental apparatus consisted of eight pigeon test chambers (30 cm X 33 cm X 35 cm); each originally contained a three-key pigeon panel (Camden Instruments Ltd., London, Model CI 4171). The center key was replaced by a 4.5 cm X 5.0 cm clear Perspex panel that was hinged at the top. Pecks on the panel were detected by a reed relay that was operated whenever a magnet attached to its lower edge was displaced by a distance greater than 0.7 mm. The midpoint of the panel was 24 cm above the floor of the chamber. A Panasonic microcolor television with a 5.5 cm X 4.4 cm screen was located 4.0 cm behind the Perspex panel. Food was delivered through a grain feeder (Campden Instruments Ltd., Model CI 443), which was located immediately below the center panel. The feeder was illuminated whenever grain was made available. The chambers were permanently lit during all experimental sessions by a 2.8-W bulb, operated at 24 V, located 2.5 cm above the top of the Perspex panel. Two British Broadcasting Corporation microcomputers (Acorn Computers Ltd., Cambridge, England) were separately connected to four of the chambers. The microcomputers, which were programmed in SPIDER (Paul Fray Ltd., Cambridge, England), were used for the control of events, the recording of responses, and the generation of stimuli on the televisions. Stimuli. The stimuli were composed of rectangles that occupied the height of the television screen and were 1.5 cm wide. For half of the subjects Rectangle A was composed of alternating red and cyan squares the sides of which were equal to one-tenth of the height of the screen, Rectangle B consisted of eight alternating blue and white vertical stripes of equal width, Rectangle C was yellow, Rectangle D consisted of three vertical columns of alternating green and white circles (with a radius equal to 0.24 cm) on a black background, Rectangle E was magenta, and Rectangle F was composed of 40 alternating blue and cyan horizontal stripes of equal width. Rectangles A and D were always shown in the left third of the screen, B and E in the center third, and C and F in the right third. Whenever adjacent rectangles were presented, they were separated by a 0.5-cm wide black stripe, and whenever fewer than three rectangles were presented simultaneously the unfilled screen was black. For the remaining half of the subjects Rectangles A, B, C, D, E, and F were respectively the same as Rectangles D, E, F, A, B, and C for the subjects considered above. Procedure. The first three sessions consisted of magazine training. For the final session of this training food was presented for 5 s on 48 occasions according to a fixed time 25-s schedule. Discrimination training began in the following session. In each of 40 sessions, members of the positive category—ABF, AEC, and DEC—were followed by food for 4 s, and members of the negative category—DEC, DBF, and AEF—were not followed by food. The compounds were presented an equal number of times in a random sequence with the restriction that no more than three trials with members of the same category could occur in succession. There were 48 trials in a session, the stimuli were presented for 10 s, and the intertrial interval was 60 s. Sessions 22 and 27 each contained 4 test trials in addition to the 48 training trials. The stimuli presented for testing were the positive prototype, ABC, and the negative prototype, DBF. During these sessions, a test trial occurred after each block of 10 training trials, and the two test stimuli were each presented twice in a random sequence. Food was delivered after the first but not the second trial with each prototype in each test session. Further test trials were given with the prototypes in Sessions 41 and 45. The procedural details that are omitted for these sessions were the same as for the original discrimination training.
The manner of training for the first half of Session 41 was the same as for the first half of the discrimination training sessions. For the next 24 trials, the members of the positive category and the positive prototype, ABC, were presented in extinction. The stimuli were presented once in each of six blocks of 4 trials. The presentation sequence within a block was random. In Session 42, subjects were given 6 trials with ABC in addition to 36 trials with the members of the positive and the negative categories. During this session, all stimuli were presented once in every block of seven trials, the presentation sequence within a block was random, and all members of the positive category, including ABC, were followed by food. Sessions 43 and 44 were conducted in the same manner as the discrimination training sessions. In Session 45, subjects received the first 24 trials in the manner described for the first half of Session 41. For the last 24 trials of the session, members of the negative categories, DEC, DBF, and AEF, and the negative prototype, DBF, were followed by food. The stimuli were presented once in each of six blocks of 4 trials. The presentation order within a block was random.
Results and Discussion All statistical tests were evaluated with respect to an alpha level of .05. The discrimination training was successful and eventually resulted in the group pecking the Perspex key in front of the television screen more rapidly in the presence of stimuli belonging to the positive than to the negative category. The group mean rates of responding that were recorded during the six training patterns in the first two test sessions (Sessions 22 and 27) are shown in Figure 1. The rate of responding was considerably faster during patterns belonging to the positive than to the negative category. An analysis of combined individual mean response rates in the presence of the three training patterns from the positive category and the three training patterns from the negative category revealed that responding was significantly more rapid in the presence of the former, r(19) = 55.7. Figure 1 also shows the results from the test trials with the prototypes in Sessions 22 and 27. It is evident from the left panel that responding was slightly faster in the presence of the positive prototype, ABC, than for two of the three training stimuli. A one-way analysis of variance (ANOVA), based on the combined individual mean rates of responding from Sessions 22 and 27, for ABC and for each training stimulus for the positive category, revealed a significant effect of stimulus, F(3, 57) = 7.2. Subsequent comparisons, using the Newman-Keuls procedure, indicated that responding during ABC and DEC was significantly faster than during either ABF or AEC. However, none of the other comparisons was significant. The results in the right panel of Figure 1 suggest that responding during the negative prototype, DBF, was slower than during the three training stimuli for the negative category. This suggestion, however, was not supported by statistical analysis. Although a one-way ANOVA similar to that reported in the previous paragraph revealed a significant effect of stimulus, F(3, 57) = 6.2, subsequent Newman-Keuls comparisons revealed only that responding during DEC was higher than during DEF and AEF.
267
PROTOTYPE EFFECTS POSITIVE CATEGORY
NEGATIVE CATEGORY
• DEF 0 DEC • DBF E3 AEF
z i DC UJ
d. CO UJ
in o D. in
UJ
DC
z UJ
Figure 1. Group mean response rates and standard errors, recorded in the presence of the various stimuli presented in Sessions 22 and 27 of Experiment 1. The results for the positive category are shown in the left panel, and the results for the negative category are shown in the right panel. Letters represent sequences of experimental stimuli.
The results from these tests indicate that responding during the prototypes differed from some, but not all, of the training members of the two categories. It would thus be unreasonable to infer that these tests have demonstrated a reliable prototype effect. The results of the remaining test sessions provide a more convincing demonstration of such an effect. The left panel of Figure 2 shows the response rates that were recorded in the second half of Session 41, in which extinction trials were given with the positive prototype and the training patterns from the positive category. As the rate of responding during the training stimuli declined, a higher rate of responding during the prototype than during any other stimulus became apparent. A two-way ANOVA, using individual rates of responding for each of the stimuli during each of the six trials, revealed a significant effect of stimulus, F(3, 57) = 9.2; and of trial, F(5, 95) = 24.0. The interaction between these variables was also significant, F(15, 285) = 3.0. Subsequent tests of simple main effects indicated a significant difference among the response rates recorded during the various stimuli on Trials 5 and 6, Fs(3, 342) > 9.7. Further comparisons using the Newman-Keuls test revealed that responding during ABC was significantly faster than during any other stimulus for the final two extinction trials. None of the other comparisons was significant. The right panel of Figure 2 shows the results from the second half of Session 45, in which the negative prototype and the members of negative category were paired with food. Responding in the presence of all four stimuli increased as conditioning progressed, but responding was consistently slower for the negative prototype, DEF, than
for any other stimulus. A two-way ANOVA, using the individual rates of responding during each of the four stimuli for the six trials, revealed a significant effect of stimulus, F(3, 57) = 4.8; a significant effect of trial, F(5, 95) = 12.6; and a nonsignificant Stimulus X Trial interaction (F < 1). Subsequent Newman-Keuls comparisons confirmed that responding was significantly slower during the negative prototype than during any other stimulus. None of the other comparisons was significant. The present experiment examined the levels of responding during the compounds ABC and DEF, after subjects had received an ABF+ AEC+ DBC+ DECo DBFo AEFo categorization task. When steps were taken to weaken responding during the compounds that signaled food, responding during ABC was more vigorous than during any of the training patterns from the positive category. Similarly, when responding during the compounds that originally signaled the absence of food was enhanced by pairing them with food, responding during DEF was weaker than during any of the negative training patterns. The test compounds were composed of the elements that occurred most frequently in their respective categories, and thus they can be regarded as the modal prototypes for these categories (Goldman & Homa, 1977). Accordingly, the present results provide a clear demonstration of a prototype effect in a categorization task by pigeons. The design of the experiment was essentially the same as the one used by Lea and Harrison (1978) and Jitsumori (1993), yet these earlier experiments were unable to reveal the effects that we report. There are a number of procedural factors that may account for the unique outcome of the present study. For instance, Lea and Harrison did not con-
268
AYDAN AYDIN AND JOHN M. PEARCE 140
120-
100-
oc 111 a. in ui
80-
CO
O &
CO UJ
60-
60-
40-
40-
DEC
20-
20-
AEF
oc
