Anim Cogn (2008) 11:587–597 DOI 10.1007/s10071-008-0149-0
ORIGINAL PAPER
Inferential reasoning by exclusion in pigeons, dogs, and humans Ulrike Aust · Friederike Range · Michael Steurer · Ludwig Huber
Received: 7 September 2007 / Revised: 18 February 2008 / Accepted: 18 February 2008 / Published online: 29 February 2008 © Springer-Verlag 2008
Abstract The ability to reason by exclusion (which is deWned as the selection of the correct alternative by logically excluding other potential alternatives; Call in Anim Cogn 9:393–403 2006) is well established in humans. Several studies have found it to be present in some nonhuman species as well, whereas it seems to be somewhat limited or even absent in others. As inconsistent methodology might have contributed to the revealed inter-species diVerences, we examined reasoning by exclusion in pigeons (n = 6), dogs (n = 6), students (n = 6), and children (n = 8) under almost equal experimental conditions. After being trained in a computer-controlled two-choice procedure to discriminate between four positive (S+) and four negative (S¡) photographs, the subjects were tested with displays consisting of one S¡ and one of four novel stimuli (S⬘). One pigeon, half of the dogs and almost all humans preferred S⬘ over S¡, thereby choosing either by novelty, or by avoiding S¡ without acquiring any knowledge about S⬘, or by inferring positive class membership of S⬘ by excluding S¡. To decide among these strategies the subjects that showed a preference for S⬘ were then tested with displays consisting of one of the S⬘ and one of four novel stimuli (S⬘⬘). Although the pigeon preferentially chose the S⬘⬘ and by novelty, dogs and humans maintained their preference for S⬘, thereby showing evidence of reasoning by exclusion. Taken together, the results of the present study suggest that
Electronic supplementary material The online version of this article (doi:10.1007/s10071-008-0149-0) contains supplementary material, which is available to authorized users. U. Aust (&) · F. Range · M. Steurer · L. Huber Department for Neurobiology and Cognition Research, University of Vienna, Althanstrasse 14, 1090 Vienna, Austria e-mail:
[email protected]
none of the pigeons, but half of the dogs and almost all humans inferred positive class membership of S⬘ by logically excluding S¡. Keywords Reasoning by exclusion · Categorization · Pigeons · Dogs · Humans
Introduction In their environment, animals often encounter inconsistent or incomplete information. One possible way to deal with this problem is to use inferential reasoning, the ability to associate a visible and an imagined event (Premack 1995). The type of inference investigated in the present study is inference by exclusion. Under laboratory conditions, the Wrst step to determine whether a subject is able to draw such inferences usually consists of investigating whether it will choose an undeWned stimulus (i.e., a stimulus that does not already have a learned association with a category or a verbal label) over a deWned one (i.e., a stimulus that is already associated). However, diVerent mechanisms may underlie such choice behavior. First, the subject may simply have a preference for novel stimuli (neophilia). Second, it may reject the deWned alternatives, but without making any inferences about the undeWned stimulus (avoidance). Third, by excluding the deWned alternatives because of their prior associations, the subject may infer class membership (or the label) of the undeWned stimulus (inferential reasoning by exclusion). For example, if a novel stimulus is presented with a stimulus already deWned as a member of the negative class, the subject may (by logically excluding the latter) conclude that the novel stimulus must be positive. Deciding among these three mechanisms requires a more stringent test than just determining whether or not a subject
123
588
displays a preference for undeWned stimuli. Only if stability of the novel association in the presence of unfamiliar rather than familiar alternatives is maintained, reasoning by exclusion can be inferred (see Kastak and Schusterman 2002). Humans are known to learn by exclusion, which is particularly evident in vocabulary learning by children (Dixon 1977; Ferrari et al. 1993; Markman and Wachtel 1988; McIlvane et al. 1987; Meehan 1995; Stromer 1989; Wilkinson et al. 1998), and the ability to map a newly heard word to an object that does not have a known lexical entry is already present at a very early age (Behrend et al. 2001; Halberda 2003; Jaswal and Markman 2001; Mervis and Bertrand 1994; Vincent-Smith et al. 1974). Several studies with nonhuman species have suggested that at least some animals may also be able to solve inferences by exclusion. In others, however, such an ability was not revealed or was found to be rather limited. Sea lions and bottlenosed dolphins have been exhibited to reason by exclusion (Herman et al. 1984; Kastak and Schusterman 2002; Schusterman and Krieger 1984), whereas data on chimpanzees are inconclusive (Beran and Washburn 2002; Cerutti and Rumbaugh 1993; Hashiya and Kojima 2001; Savage-Rumbaugh 1986; Tomonaga 1993; Tomonaga et al. 1991). Kaminski et al. (2004) demonstrated that the Border Collie Rico was able to infer the referent of a new word by exclusion learning and retain this knowledge over time. Evidence from pigeons suggests that smaller brained animals may also be capable of exclusion performances. In contrast to earlier claims (e.g., Cumming and Berryman 1961) it has more recently been shown that they can be encouraged to choose by exclusion (i.e., to select undeWned stimuli in preference of deWned ones) when tested with appropriate experimental designs (Clement and Zentall 2000, 2003; Zentall et al. 1981). However, evidence of learning by exclusion is, to date, still missing in pigeons. While there is evidence of exclusion performances in various species, due to diVerences in the methods used it is usually not possible to directly compare the Wndings obtained in diVerent studies. One of the most frequently used paradigms for testing inference by exclusion is matching-to-sample. An animal may for instance be trained on a conditional discrimination whereupon a novel sample is introduced, with a choice between a novel and a familiar comparison. Pigeons, for example, have been tested for exclusion abilities mainly in matching-to-sample paradigms (e.g., Clement and Zentall 2000, 2003; Cumming and Berryman 1961; Zentall et al. 1981). In the domain of sequential learning paradigms, researchers have used wildcard tests, a variant of substitution tests (see, e.g., D’Amato and Colombo 1989; Terrace et al. 1995; Tomonaga 1999; Tomonaga and Matsuzawa, 2000) that also give insight into a subject’s exclusion abilities. In an experiment with chimpanzees, for example, the subjects were trained to sequen-
123
Anim Cogn (2008) 11:587–597
tially respond to stimuli of two-item lists by touching them (Tomonaga 1999). They were then transferred to a test in which one of the members of a list was replaced with a novel stimulus (the wild-card) to examine whether the subjects would respond to the tested pairs in the order deWned by the trained functional classes (which had been established previously through repeated reversals of the order in which the items had to be chosen). Other paradigms have refrained from the touchscreen methodology and used more natural setups. For example, in one method pioneered by Premack and Premack (1994) and mainly used with primates, food is hidden in two containers. The animal is allowed to choose one of them after witnessing the experimenter removing the food from one of the containers (e.g., Call 2006). Similarly, Erdöhegyi et al. (2007) have shown that dogs can Wnd a hidden toy under one of two containers if they have seen where the toy was missing. Dogs have also been tested in fetching paradigms where the experimenter asks them to fetch a novel object from a selection of familiar ones (by using the respective novel verbal label), and, in a second step, to pick that object from a selection of familiar and novel items to see if it has been associated with the novel verbal label (Kaminski et al. 2004). In addition to using diVerent paradigms, studies on exclusion abilities vary with respect to apparatus, training history, stimulus material, reinforcement conditions, and many other aspects of experimental design. As inconsistent methodology might have contributed to the revealed interspecies diVerences, it seems necessary to test various species under almost equal experimental conditions. To this end, we developed a computer-controlled two-choice procedure suitable for testing diVerent species. In principle, it was an abstract analogue of the traditional fetching paradigm as was used, for example, in the study with Rico (Kaminski et al. 2004; see S1 for a comparison of the two procedures). As a starting point, we considered it reasonable to test an avian and a mammal (nonhuman) species and compare their performance with that of humans. The abilities of the latter to reason by exclusion are well established. However, the conditions under which humans naturally acquire new words are considerably diVerent from the most common laboratory word-learning situations (see, e.g., Jaswal and Markman 2001), and are probably even more diVerent from the automated computer-controlled procedure used in the present experiment. Therefore, we believed humans as a good control for the general appropriateness of our procedure. Further, we wanted to compare performance of children to that of adults for possible diVerences in their learning by exclusion abilities in an abstract task. As for the avian species, pigeons are a good choice as they have already been shown to be capable of choosing by exclusion, and it would thus be interesting to see if they might also learn/reason by exclusion. Also, they have so far been
Anim Cogn (2008) 11:587–597
tested for exclusion mainly with matching-to-sample paradigms. From previous experiments on unrelated topics we know that pigeons learn fast and easily in two-choice procedures and therefore suspected that such an approach may also be appropriate for testing exclusion abilities in this species. As regards the nonhuman mammal species, we decided to use dogs because previous studies (Erdöhegyi et al. 2007; Kaminski et al. 2004) have yielded evidence of this species being able to draw inferences by exclusion in paradigms that involve experimenter–subject interactions (such as verbal commands). It is, however, not entirely clear whether and in what way human interference may aVect a dog’s performance in such a task. On the one hand, dogs are known to be extremely responsive to social attachment and sensitive to inadvertent cues given by humans, which are transmitted, e.g., through eye contact, body language, or verbally (Agnetta et al. 2000; Miklosi et al. 1998; Riedel et al. 2006; Szetei et al. 2003). Therefore, social cues may contribute to the display of a “desired” response behavior (i.e., choice according to an exclusion-based logic). On the other hand, Erdöhegyi et al. (2007) have shown that dogs may actually be prevented from showing reasoning abilities when given social-communicative cues (directional gesture and gaze cues). Therefore, we considered it particularly interesting to investigate if the ability to draw inferences by exclusion would be displayed by dogs in an abstract task and in an automated experimental setup without any human interference.
589
monitor. Data acquisition and device control were handled with a microcomputer interfaced through a digital input– output board. The apparatus was the same for the dogs, except for some slight adjustments that had to be carried out to account for diVerences between the two species regarding physique and behavior. Their box had no rear wall, which allowed the subjects to reach the touchscreen while at the same time their vision was shielded to avoid distractions from the side and above. Reinforcement was administered in the form of small commercial dog food pellets, which were dispensed by an automated feeding device through a small hole beneath the touchscreen. The experimenter had to be present during the tests to reduce nervousness in the dogs, but was standing behind the apparatus and was therefore ignorant of the presented stimuli. The apparatus for the human subjects consisted only of a monitor that could be operated with a computer mouse. Separated from the testing monitor by an opaque wall, the experimenter could specify the testing parameters and supervise the experiment on a control monitor (Fig. 1a).
Methods Subjects The subjects were six adult pigeons (P1-6), six adult dogs (D1-6), six students (S1-6), and eight children (C1-8). Four pigeons were homing pigeons, two were of a local Austrian strain, called “Strasser”. The dogs were four Border Collies, one Australian Shepherd, and one mongrel. The adult human participants were under- and postgraduate students of the University of Vienna, between 19 and 38 years. The children were between 7½ and 10 years old. Apparatus The pigeons were trained in a closed experimental indoor chamber (“Skinner-box”). A 15⬘⬘ TFT computer screen mounted behind an infrared touchframe (Carroll Touch, Round Rock, TX, USA; 32 vertical £ 42 horizontal resolution) constituted the frontal wall. A receptacle into which grain was delivered by an automated feeding device was situated directly below the touchscreen panel. The birds’ activities inside the box could be observed on a control
Fig. 1 a A pigeon, a dog, and a student working with their respective versions of the computer-controlled two-choice procedure. Pigeons and dogs were trained in modiWed Skinner-boxes to prevent distraction and social cueing. The human participants were visually separated from the experimenter by an opaque wall. b The stimuli used to investigate reasoning by exclusion: Top panel Four training stimuli belonged to the positive class (S+) for half of the subjects of each group (Group A or Group B, respectively) and to the negative class (S-) for the other subjects. Bottom panel The four test stimuli in the top row (S⬘) were shown together with S- in Test 1, and together with novel stimuli (S⬘⬘; bottom row) in Test 2
123
590
Stimuli Pictures were color photographs obtained from the Internet. They were resized to 4.5 cm£ 4.5 cm by using photo software and were presented at a resolution of 72 dpi. The eight training stimuli showed everyday objects, four of which were deWned as positive (S+), the other four as negative (S-) with assignment being arbitrary. However, we made sure that—to the best of our knowledge—no spurious perceptual cue or semantic principle reliably separated the two classes. For example, we were careful about not assigning all objects that contained a particular color, bore some similarities regarding shape, or shared some functional properties (e.g., objects used in the kitchen) to the same class. Similar precautions were taken for the test stimuli (four in each of the two tests) by avoiding close similarities with any of the training stimuli regarding color, form, and function. The stimuli are on display in Fig. 1b. Procedure Each trial involved the simultaneous presentation of two stimuli. They were projected onto a white background in Wxed positions, with one stimulus appearing somewhat left of the middle of the screen and the other stimulus appearing somewhat right. The positions (left/right) of the individual stimuli varied randomly from trial to trial. The subjects had to decide on one of them, whereupon they were given some feedback by the computer in case of training trials. Choice of S+ resulted in reinforcement, whereas choice of S- was followed by a correction trial, i.e., stimulus presentation was terminated and the color of the screen turned red for 3 s. Then the stimuli of the previous trial were shown again in identical positions as before. Another wrong choice led to another correction trial, whereas a correct choice terminated the trial and led, in the case of the animals, to food access. To enhance learning, correct and incorrect choices on training trials were furthermore accompanied by diVerent acoustic signals. (See S2-4 for videos of a pigeon, a dog, and a child being trained with this procedure.) The children received (independent of performance) some sweets after each phase of the experiment and were given a little present at the very end. Test trials were nonreinforced, i.e., the Wrst choice of any of the two stimuli immediately terminated the trial without any acoustic or visual feedback, correction trial, or reward. For the animals, each trial (except correction trials) was preceded by an intertrial interval (ITI) during which an empty (white) background was shown. The ITI was 4 s for the pigeons. For the dogs it was reduced to 2 s because longer ITIs were found to negatively aVect attention. For both species, touching the screen during the ITI led to its prolongation by another 2 s. In the experiments with
123
Anim Cogn (2008) 11:587–597
students and children there were no ITIs at all. As they did the entire experiment in one go, it seemed advisable to keep the individual sessions as short as possible to avoid decreases in concentration and motivation. The pigeons indicated their choices by pecking at a stimulus, the dogs touched it with their nose, and the humans navigated the mouse cursor onto one of the stimuli and clicked at it. The animals were familiar with the basic procedure and the reinforcement conditions (including experience with nonreinforced trials) at the onset of training, as they had already participated in unrelated visual discrimination experiments carried out with the same forced two-choice procedure as used in the present study. For example, the pigeons had participated in an experiment where they had to discriminate between color photographs with and without human Wgures, and most of the dogs had been employed as subjects in a dog/landscape discrimination task (Range et al. 2007). The human participants received a short instruction prior to the experiment (see S1 for details). All subjects were arbitrarily assigned to one of two groups (A and B) which diVered in the contingencies of the individual training pictures. The pigeons were trained 5 days a week, the dogs one to two times a week. The human subjects did the entire experiment within 30–45 min. Training. The subjects were Wrst trained to discriminate between the four S+ and the four S- stimuli. Each session consisted of 32 trials, with one S+ and one S- being shown on each trial. Each of the 16 possible S+/S- pairings was presented twice per session. The criterion of mastery was set at ¸28 correct Wrst choices (which equals 87.5%) in Wve of seven consecutive sessions and at least 22 correct Wrst choices (which equals 68.75%) in the two others (if needed at all). The minimum number of sessions to reach criterion was thus Wve. When the subjects had acquired the training task, they were transferred to two consecutive tests. Test 1. Test sessions involved 28 training trials and four interspersed test trials. The latter consisted of one of the four S- and one of four novel stimuli S⬘ (Fig. 1b). The S⬘ therefore replaced the S+ shown during training, but were themselves not yet associated with either class (i.e., they were undeWned). Each of the 16 possible test combinations was shown twice, i.e., the subjects ran two “cycles”, each of which consisted of four sessions. The rationale of the test was the following. Subjects choosing by exclusion should reject the S- because of their prior associations with the negative class and instead choose the hitherto undeWned S⬘ to avoid categorical inconsistencies. By contrast, subjects with a propensity to choose a familiar stimulus in the presence of a novel alternative should prefer S-. Test 2. A preference for S⬘ in Test 1 alone would, however, be insuYcient to infer reasoning by exclusion as the
Anim Cogn (2008) 11:587–597
underlying mechanism. To exclude possible alternative strategies such as neophilia and avoidance of S- without any untrained relation between the S⬘ and the positive class emerging, the subjects that had displayed a preference for S⬘ in Test 1 were tested for its persistence when the S⬘ was presented together with a novel rather than with a familiar alternative. Again, test sessions involved 28 training trials and four interspersed test trials. The latter consisted of one of the S⬘ and one of four novel and thus undeWned alternatives (S⬘⬘). As in Test 1, each subject ran two cycles of four sessions each. Inference of positive class membership of S⬘ by exclusion in Test 1 would make Test 2 a choice task with a (just recently) deWned positive stimulus (S⬘) and an undeWned (novel) stimulus (S⬘⬘) as alternatives on test trials. Consequently, reasoning by exclusion in Test 1 should manifest in maintenance of the preference for S⬘ in Test 2. By contrast, subjects whose preference for S⬘ in Test 1 was due to neophilia should now prefer the (even “more novel”) S⬘⬘ over S⬘. Subjects that had simply avoided S- (and therefore chosen S⬘) in Test 1 without making any inferences about class membership of S⬘ should not show a clear preference for either S⬘ or S⬘⬘ in Test 2. To them, both stimuli should be undeWned and should therefore be chosen arbitrarily. After the experiment the human subjects were interviewed about their choice strategies (see S1 for details). Data analysis Performance was assessed by means of a binomial test with the hypothesis of equal choice-probability for either S⬘ or S- in Test 1 and S⬘ or S⬘⬘ in Test 2, respectively. Since we used a discrete statistic (count of choices of one stimulus type) no value existed that exactly corresponded to the signiWcance criterion of P = 0.05 which is commonly used with continuous statistics. If the data of two cycles are pooled (n = 32 trials) the probability for ¸21 choices of a speciWc stimulus type equals 0.055. For ¸22 choices the probability equals 0.025, which we considered a reasonable level of reliability. Thus, the preference criterion was set at ¸22 choices (which equals ¸68.75%) in analyses based on pooled cycles. In order to assess preferences in individual cycles as well, data were also inspected unpooled (n = 16 trials). The probability for ¸11 choices of one stimulus type equals 0.105 and 0.038 for ¸12 choices. We thus decided on ¸12 out of 16 scores (which equals ¸75%) as indicating preference with suYcient reliability. Generally, “preference” for a particular stimulus type in a test (S⬘ or S- in Test 1; S⬘ or S⬘⬘ in Test 2) was inferred, when a subject chose the respective stimuli in ¸12 trials out of 16 in at least one of the two cycles and in ¸22 trials out of 32 when the data of the two cycles were pooled, i.e., a “near miss” in one of the cycles was tolerated.
591
Results The training results are shown in Fig. 2. Pigeons and humans acquired the initial discrimination task very quickly with pigeons needing an average (§SD) of 7.5 (§1.05) sessions to reach criterion (range: 6–9), students 5.2 (§0.41) sessions (range: 5–6), and children 5.5 (§0.93) sessions (range: 5–7). The dogs, by contrast, required an average of 32.2 (§21.89) sessions, and learning performances within the dog group were characterized by large inter-individual variations (range: 12–70). Test 1 yielded two major results which are illustrated in Fig. 3 and summarized in Table 1. (1) There was an increase in the ratio of subjects that met the criterion for preference for S⬘ from pigeons (one out of six: P4), to dogs (three out of six: D2, D4, D6), to humans (Wve out of six students: S1–4, S6, and all eight children: C1–8). Only one student (S5) did not show a preference for either stimulus type. In the post-experimental interview S5 reported having searched for perceptual response rules (mainly based on color and form features) as an alternative to reasoning by exclusion (which appeared to him “too simple a solution to be the ‘correct’ one”). Although he realized that the rules he inferred did not work reliably, they were not dismissed but persistently interfered with and overshadowed reasoning by exclusion, which was also present. The other students (S1–S4, S6) reported inferential reasoning as their only choice strategy, i.e., they deduced positive contingencies of S⬘ from rejecting the negative alternative. As for the children, one (C7) reported a shift from an initial tendency to choose S⬘ (because of its novelty) to inferential reasoning. All other children (C1–6, C8) reported inferential reasoning as the only strategy underlying their exclusively choosing S⬘ on test trials. (2) The preference for S⬘ was stronger in humans than in the two nonhuman species. Ten human subjects (three students and seven children) even chose S⬘ in 100% of the test trials in both cycles. The highest score reached by one of our animals (D6 in Cycle 1) was 15 (which equals 93.75%). The results of Test 2 are illustrated in Fig. 4 and summarized in Table 1. The pigeon’s (P4) preference for S⬘ in Test 1 was lost in Test 2. Instead, it showed a strong preference for the novel S⬘⬘. By contrast, the three dogs (D2, D4, D6) maintained their preference for S⬘. Also the human subjects strongly preferred the S⬘ stimuli, with three out of Wve students and six out of eight children even choosing them in 100% of the trials. Only two children preferred S⬘⬘ over S⬘. In summary, Test 2 failed to yield evidence for the emergence of novel untrained associations through exclusion and thus inferential reasoning in the pigeon (P4), whereas the results of the three dogs (D2, D4, D6) suggest the presence of such an ability. Also the majority of the human subjects were able to infer class membership of S⬘ on the basis
123
592
Fig. 2 Acquisition performance of pigeons, dogs, students, and children when trained to discriminate between four S+ and four S-. The results are shown separately for each species and subject as % correct Wrst choices per session (i.e., correction trials were excluded from anal-
Anim Cogn (2008) 11:587–597
ysis). The dashed line indicates the level of performance that had to be obtained Wve times within seven sessions (87.5%) to meet the criterion of mastery. _A = subject of Group A; _B = subject of Group B
Fig. 3 Percentage of test trials in which the undeWned stimuli (S⬘) were chosen in preference to the negative ones (S-) in Test 1. The results are shown separately for each species and subject and with the data obtained in cycles 1 and 2 being pooled (§SD between cycles). The upper dashed line indicates the levels of performance beyond which preference for S⬘ was inferred (¸68.75%; choice by novelty, avoidance, or reasoning by exclusion). The lower dashed line indicates the level of performance below which preference for S- was inferred (·31.25%; choice by familiarity). _A = subject of Group A; _B = subject of Group B; no novelty; av avoidance; ex exclusion; fa familiarity
of exclusion. This was conWrmed by the post-experimental interviews where most candidates reported reasoning by exclusion as their dominant strategy also in Test 2. Alternative solutions such as choice by novelty were rather the exception than the rule (C1, C3). Occasionally, subjects reported having used a mix of diVerent strategies or a shift from alternative strategies towards reasoning by exclusion in the course of testing.
123
Most human participants had no problems recalling their choice behavior and describing the underlying cognitive mechanisms in the individual phases of the experiment. Some children, however, had diYculties remembering individual stimuli, stimulus pairings that had occurred, and their choice preferences on test trials. Furthermore, some were unable to spontaneously explain the reasons for their choice preferences, but did so only on repeated request (if
Anim Cogn (2008) 11:587–597
593
Table 1 Choices of S⬘: scores out of 16 in cycles 1 and 2; scores out of 32 in both cycles (pooled) Group
Pigeons
Dogs
Students
Subject
Test 2
cy 1
cy 2
Pooled
cy 1
cy 2
Pooled
P1_A
12 (0.038)
05 (0.105)
17 (0.430)
P3_A
09 (0.402)
05 (0.105)
14 (0.298)
P5_A
11 (0.105)
08 (0.598)
19 (0.189)
P2_B
05 (0.105)
09 (0.402)
14 (0.298)
P4_B
13 (0.011)
12 (0.038)
25 (0.001)
03 (0.011)
02 (0.002)
05 (·0.0001)
P6_B
10 (0.227)
04 (0.038)
14 (0.298)
D1_A
11 (0.105)
08 (0.598)
19 (0.189)
D3_A
08 (0.598)
07 (0.402)
15 (0.430)
D5_A
11 (0.105)
09 (0.402)
20 (0.108)
D2_B
12 (0.038)
12 (0.038)
D4_B
12 (0.038)
11 (0.105)
24 (0.004)
11 (0.105)
13 (0.011)
24 (0.004)
23 (0.010)
11 (0.105)
12 (0.038)
D6_B S1_A
15 (0.0003) 16 (·0.0001)
23 (0.010)
11 (0.105) 16 (·0.0001)
26 (0.0003) 32 (·0.0001)
12 (0.038)
26 (0.0003) 30 (·0.0001)
11 (0.105)
16 (· 0.0001)
27 (·0.0001)
14 (0.002) 16 (·0.0001)
14 (0.002) 16 (·0.0001)
S3_A
16 (·0.0001)
32 (·0.0001)
S5_A
12 (0.038) 16 (·0.0001)
06 (0.227) 16 (·0.0001)
18 (0.298) 32 (·0.0001)
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
S4_B
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
S6_B
15 (0.0003) 16 (·0.0001)
27 (·0.0001)
13 (0.011)
11 (0.105)
C1_A
12 (0.038) 16 (·0.0001)
32 (·0.0001)
02 (0.002)
01 (0.0003)
24 (0.004) 03 (· 0.0001)
C3_A
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
C5_A
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
02 (0.002) 16 (·0.0001)
07 (0.402) 16 (·0.0001)
09 (0.010) 32 (·0.0001)
C7_A
15 (0.0003) 16 (·0.0001)
16 (·0.0001)
31 (·0.0001)
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
16 (·0.0001)
32 (·0.0001)
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
S2_B
Children
Test 1
C2_B C4_B
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
C6_B
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
C8_B
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
16 (·0.0001)
16 (·0.0001)
32 (·0.0001)
P-values (binomial test) in parenthesis. Probabilities refer to the respective preferences. To contrast biases in favor of S⬘ with those in favor of the alternative stimuli S- or S⬘⬘, respectively, the former are given in bold typeface. Scores ·4 and ¸12 (for cy 1 and cy 2) and ·10 and ¸22 (for pooled data) are given in italics Cy cycle
at all), sometimes in a vague and confusing manner. Further, children reported a mix of strategies more often than the students, and sometimes the strategies they described were inconsistent with their actual choice behavior.
Discussion The computer-controlled forced two-choice procedure ruled out the risk of social cueing and provided maximum objectivity and equal conditions across trials, sessions, subjects, and species, thereby allowing direct comparisons of the cognitive strategies employed to solve the present task. Nevertheless, the fact that all three species mastered the initial discrimination does not necessarily mean that they did so with equal ease (see, e.g., Macphail 1987). For example, our pigeons as well as the human participants
were well experienced in using a computer monitor, whereas the procedure was relatively new for the dogs. Also, the latter did not practice on a daily basis as the pigeons did. In addition, physiological particularities of the three species’ visual systems may have aVected performance. In contrast to primates and birds, which strongly depend on vision, this sense is not so dominant in dogs, as reXected, e.g., by their restricted color perception. Thus, the dogs may have found the stimuli harder to distinguish and to memorize than the pigeons and humans. Also, variations in visual acuity and spatial resolution may have played a role (Aguirre 1978; Coile et al. 1989; Murphy et al. 1992; Neitz et al. 1989; Odom et al. 1983; Peichl 1991; Range et al. 2007). Despite their excellent perceptual abilities, extensive experience with the procedure and their independence from social attachment, the pigeons were the only species that
123
594
Anim Cogn (2008) 11:587–597
Fig. 4 Percentage of test trials in which the S⬘ were chosen in preference to novel undeWned stimuli (S⬘⬘) in Test 2 by those subjects that had displayed a preference for S⬘ in Test 1. The results are shown separately for each species and subject and with the data obtained in cycles 1 and 2 being pooled (§SD between cycles). The upper dashed line indicates the level of performance beyond which preference for S⬘ was inferred (¸68.75%; reasoning by exclusion). The lower dashed line indicates the level of performance below which preference for S⬘⬘ was inferred (choice by novelty). _A = subject of Group A; _B = subject of Group B; no novelty; ex exclusion
did not show any sign of inferential reasoning. The only pigeon that preferred S⬘ in Test 1 failed to do so in Test 2. This would suggest choice by novelty as the underlying strategy, which is interesting insofar as pigeons are reputed to be neophobic rather than neophilic (Clement and Zentall 2003). One possible explanation may be in terms of the nature of the stimuli we used. Actually, it has repeatedly been shown that pigeons choose diVerent ways of approaching a task, depending on the appearance of the presented stimuli. Maybe the relative “naturalness” of the stimuli we used (i.e., color photographs) slightly biased the pigeons (in particular P4) towards neophilia. Alternatively, one may think of an explanation for the results of P4 other than in terms of choice by novelty. Namely, responding could have been on the basis of avoiding unrewarded stimuli, which the S⬘ would have become by the end of Test 1. Consequently, P4 would have preferred S⬘⬘ in Test 2. It should, however, be considered that P4 neither chose S⬘ in Test 1, nor S⬘⬘ in Test 2 in 100% of the cases, and the fact that the respective choice alternatives were not rewarded either, should have somewhat devalued avoidance of nonreward as a basis for responding. By contrast, in half of the dogs and almost all humans, we found evidence that they had deduced positive class membership of S⬘ by logically excluding the negative alternatives. This shows that the ability to reason by exclusion is, in principle, dependent not on “natural” contexts (such as fetching paradigms), on social encouragement, or on verbal commands. Thereby, our results stand in contrast to the claim that such emergent processes require verbal behavior (e.g., Horne and Lowe 1996). At the same time, they are in line with the results of Erdöhegyi et al. (2007) who found
123
that pre-existing biases for social cues may prevent dogs from reasoning by exclusion. Two children (C1, C3) displayed a clear preference for S⬘⬘ over S⬘ in Test 2, which was probably based on neophilia, as indicated in the post-experimental interviews. For most other human participants, the interviews conWrmed reasoning by exclusion as the dominant strategy. Only the student that failed to show a preference for S⬘ (S5) in Test 1 reported the additional use of a strategy based on spurious and imperfect correlations between perceptual features of the stimuli and class membership. The only diVerence found between the students and at least some of the children concerned short-term memory (e.g., of the presented stimuli and the combinations in which they appeared) and the ability to verbalize and discuss the employed strategies. Namely, deWcits regarding both aspects were occasionally found in children during the interviews, which may possibly be due to their ability of active reXection and mental manipulation of more abstract cognitive contents not yet being fully developed. However, the small number of subjects and the absence of a strictly standardized protocol for the interviews make such considerations speculative. Although the Wndings of the present study are suggestive of an ability to reason by exclusion in half of the dogs and almost all humans, the sceptic might develop an alternative account for the results, in which the eVects of a mild neophobic tendency would have led to the observed preferences for S⬘. Namely, the subjects might have displayed a small amount of initial neophobia towards the S⬘-stimuli in Test 1, which would thus have acquired aversive properties. However, the S- were even more aversive because they had been associated with nonreward during training. So, the
Anim Cogn (2008) 11:587–597
subjects should nevertheless have preferred S⬘. Due to repeated experience with S⬘ during Test 1, neophobia towards these stimuli should have gradually been alleviated. This would account for them being preferred over the S⬘⬘-stimuli in Test 2, which were completely novel and thus more aversive than were the S⬘ at that point. However, the following should be considered. First, the S- were even more familiar than the S⬘ in Test 1, which should have “devalued” familiarity as an option. Second, a decrease in neophobia towards the S⬘ through repeated presentation should have resulted in an increase in the subjects’ proneness to choose them (over S-) in Cycle 2 of Test 1 as compared to Cycle 1. However, the opposite was the case in all but one dog (see Table 1). This means that the eVects of relative familiarity and neophobia, if present at all, must have been weak and easy to overrule by other factors. As such they were probably not suYcient for explaining the degree to which S⬘ was preferred by the dogs. Because of a ceiling eVect (10 out of 14 subjects chose S⬘ in 100% of the trials in both cycles of Test 1) the above argument cannot be veriWed (or falsiWed) for the human participants. However, from the post-experimental interviews it was very evident that neophobia was not the mechanism underlying choice of S⬘ in Test 2. Instead, subjects that actually preferred S⬘ (almost all) indicated having done so because they “knew from Test 1 that S⬘ was positive”. Not one of them reported having been prevented from choosing S⬘⬘ because of its unfamiliarity. Third, the three dogs that preferred S⬘ over S- in Test 1 showed almost equal scores in Test 2 (i.e., a similar preference for S⬘ over S⬘⬘). But would one not expect extensive experience with the negatively deWned S- causing a stronger preference for S⬘ over S- in Test 1 than for S⬘ over the undeWned S⬘⬘ in Test 2? In other words: Would the small amount of “familiarity” attributed to S⬘ that is lacking in S⬘⬘ really be suYcient for causing an equally strong preference for S⬘ in Test 2 as in Test 1, where the S⬘ were pitted against the S- with their strong negative contingencies? (Again, this argument cannot be applied to the human subjects because of ceiling eVects). So, while such an alternative account cannot be ruled out completely, the pattern of results we obtained rather seems to argue against a prominent role of relative familiarity and neophobia. What remains to be answered is how the inter- and intraspeciWc diVerences may possibly be captured under a single theoretical umbrella. Perhaps, we may best explain the present Wndings in terms of two diVerent learning strategies with prevalence of one or the other being a matter of individual and/or species-speciWc preferences. On the one hand, a subject may easily be conditioned to respond correctly to positive and negative stimuli through reward or timeout, respectively (“rote learning”), which would allow for rapid acquisition of perceptual discrimination problems
595
but would lead to failure on more abstract tasks. Such a strategy may well explain the results of the pigeons, which quickly acquired the initial discrimination but failed to reason by exclusion in the subsequent tests. There is in fact abundant data to show that pigeons readily solve problems that have a perceptual basis but usually fail on tasks that require abstraction and the application of logic (for reviews, see Aust et al. 2005; Huber 2001; Huber and Aust 2006; Mackintosh 2000; but see Delius et al. 2000 for a contrasting view). Therefore, it is possible (although absence of evidence is not evidence of absence) that also the ability to make inferences by exclusion—which requires logical reasoning independent of perceptual features—may be out of a pigeon’s reach. Similarly, the three dogs that failed to show a preference for S⬘ in Test 1 may have relied on rote learning, which enabled them to acquire the initial discrimination between S+ and S- stimuli during training by being conditioned to respond correctly to each stimulus, but prevented them from reasoning by exclusion in the subsequent tests. On the other hand, a subject may prefer a logic-based strategy as required for solving the test tasks in terms of inferential reasoning over a rote learning strategy as required for mastering the training task. The fact that the three dogs that reasoned by exclusion in the tests needed comparatively more training to acquire the initial discrimination may be indicative of a propensity to search for abstract, logic-based solutions, rather than to rely on rote learning. Equal eYciency in both training and test phase required the subjects to quickly and Xexibly switch from rote learning to logic. Possible reasons for the diVerences in the subjects’ readiness to switch strategy are manifold. Failure to do so does not necessarily imply absence of the ability to reason by exclusion but may instead be due to insuYcient cognitive Xexibility. Pigeons, in particular, tend to become blind to alternative solutions once they have learned a task by means of simple, perceptually based associations (Aust and Huber 2006). A similar lack in Xexibility may have contributed to the results of the three dogs and the student (S5) that did not show a preference for S⬘. A factor that could have aVected the results of all three species was the absence of reward on test trials. It is, for example, conceivable that initial response biases towards S⬘ in Test 1 were weakened, extinguished, or even reversed as a consequence of the lack of reinforcement (see Table 1)1. 1 Of course we could have counteracted such eVects by rewarding choices of S⬘. However, what we wanted to investigate were possible preferences for S⬘ as a result of reasoning by exclusion, not of reinforcement. Also, we suspected that providing partial reinforcement or rewarding the one or the other stimulus type in half of the trials may have led the subjects into searching for some spurious rule according to which reinforcement may be given, and would thus have distracted their focus from more logic-based solutions.
123
596
The fact that choices of S⬘ were (although sometimes only slightly) lower in Cycle 1 than in Cycle 2 of Test 1 in Wve pigeons, Wve dogs, and even one student may have been due to such an eVect. Namely, the observed decreases could have been indicative of the repeated experience of not getting positive feedback for choosing S⬘ on S⬘/S- trials tempting the subjects into increasingly trying the S- stimuli, despite their hitherto negative contingencies. The possible eVect of nonreward on performance of P4 has already been discussed earlier. In Test 2, it may be argued that the slight increase in choices of S⬘ from Cycle 1 to Cycle 2, which was found in the three dogs and one student, could have resulted from the subjects trying out the novel S⬘⬘ Wrst and, when they were found to be unrewarded, too, turning back again to some extent towards S⬘. In most cases, however, diVerences in choice numbers both between cycles and tests were rather small so that we are reluctant to attribute a major role to the eVects of nonreward in explaining the present results. Other factors that might possibly have inXuenced the subjects’ performances include sex, age, and the amount of training. For example, male dogs were found to be slower learners (42.3 § 23.97) than female dogs (16.3 § 5.86). Furthermore, the three dogs that showed evidence of inferential reasoning (the two slowest male dogs and the slowest female learner), needed comparatively more training to acquire the initial discrimination than the fastest dogs within their sex class. It is thus possible that sex and the quantity of training may have interacted in inXuencing the subjects’ preference for either rote learning or a logic-based strategy. Similarly, the two children that failed to reason by exclusion were female, but, at the same time, were also the youngest participants. Thus, one might equally infer an inXuence of sex, age, or both. However, considerably more data from tests that systematically investigate the possible contributions of age, sex, and training history would be needed to decide whether and how these factors inXuence response behavior in inferential reasoning tasks, or whether the observed variations were merely due to individual diVerences. Similarly, eVects of group aYliation cannot be completely excluded—the only animal subjects that showed a preference for S⬘ (P4 in Test 1, and D2, D4, and D6 in both tests) were all members of Group B—but regarding the relatively small number of subjects in each group the observed diVerence may also merely be due to chance. In a nutshell, there are a variety of factors that may have impeded the display of reasoning by exclusion in the present task. All the more, the Wnding that three dogs and most humans not only showed a spontaneous preference for the undeWned stimuli but also maintained that preference in the presence of novel rather than familiar alternatives gains in value and weight. For these subjects, the present results are in fact indicative of an ability to draw inferences by exclusion.
123
Anim Cogn (2008) 11:587–597 Acknowledgments The research was supported by the Austrian Science Foundation through Grant V3-B03 to Ulrike Aust and by the European Community’s Sixth Framework Programme under contract number: NEST 012929. Thanks are due to Wilfried Apfalter, Johanna Kramer, Katharina Kramer, and Michael Pollirer for their assistance in the pigeon laboratory, and to Karin Bayer and Stefanie Riemer for their help in carrying out the experiments with the dogs. We would also like to thank Christian Palmers for providing the facility for the dog experiments. Also, we wish to thank Julia Fischer for valuable comments and discussion. We declare that the experiments comply with the current Austrian laws.
References Agnetta B, Hare B, Tomasello M (2000) Cues to food locations that domestic dogs (Canis familiaris) of diVerent ages do and do not use. Anim Cogn 3:107–112 Aguirre G (1978) Retinal degeneration in the dog: rod dysplasia. Exp Eye Res 26:233–253 Aust U, Huber L (2006) Does the use of natural stimuli facilitate amodal completion in pigeons? Perception 35:333–349 Aust U, Apfalter W, Huber L (2005) Pigeon categorization: classiWcation strategies in a non-linguistic species. In: Grialou IP, Longo G, Okada M (eds) Images and reasoning. interdisciplinary conference series on reasoning studies. Keio University Press, Tokyo, pp 183–204 Behrend DA, ScoWeld J, Kleinknecht EE (2001) Beyond fast mapping: young children’s extensions of novel words and novel facts. Dev Psychol 37:698–705 Beran MJ, Washburn DA (2002) Chimpanzee responding during conditional matching-to-sample: control by exclusion. J Exp Anal Behav 78:497–508 Call J (2006) Inferences by exclusion in the great apes: the eVect of age and species. Anim Cogn 9:393–403 Cerutti DT, Rumbaugh DM (1993) Stimulus relations in comparative primate perspective. Psychol Rec 43:811–821 Clement TS, Zentall TR (2000) Stimulus relations in comparative primate perspective. Psychol Sci 11:261–264 Clement TS, Zentall TR (2003) Choice based on exclusion in pigeons. Psychon Bull Rev 10:959–964 Coile DC, Pollitz CH, Smith JC (1989) Behavioral determination of critical Xicker fusion in dogs. Physiol Behav 45:1087–1092 Cumming WW, Berryman R (1961) Some data on matching behavior in the pigeon. J Exp Anal Behav 4:281–284 D’Amato MR, Colombo M (1989) Serial learning with wild card items by monkeys (Cebus apella): implications for knowledge of ordinal position. J Comp Psychol 103:252–261 Delius JD, Jitsumori M, Siemann M (2000) Stimulus equivalencies through discrimination reversals. In: Heyes C, Huber L (eds) The evolution of cognition. MIT Press, Cambridge, pp 103–122 Dixon LS (1977) The nature of control by spoken words over visual stimulus selection. J Exp Anal Behav 27:433–442 Erdöhegyi A, Topal J, Virányi Z, Miklosi à (2007) Dog-logic: inferential reasoning in a two-way choice task and its restricted use. Anim Behav 74:725–737 Ferrari C, de Rose JC, McIlvane WJ (1993) Exclusion vs. selection training of auditory-visual conditional relations. J Exp Child Psychol 56:49–63 Halberda J (2003) The development of a word-learning strategy. Cognition 87:B23–B34 Hashiya K, Kojima S (2001) Hearing and auditory–visual intermodal recognition in the chimpanzee. In: Matsuzawa T (ed) Primate origins of human cognition and behavior. Springer, Tokyo, pp 155–189
Anim Cogn (2008) 11:587–597 Herman LM, Richards DG, Wolz JP (1984) Comprehension of sentences by bottlenosed dolphins. Cognition 16:129–219 Horne PJ, Lowe CF (1996) On the origins of naming and other symbolic behavior. J Exp Anal Behav 65:185–241 Huber L (2001) Visual categorization in pigeons. In: Cook RG (ed) Avian visual cognition [On-line] Huber L, Aust U (2006) A modiWed feature theory as an account to pigeon visual categorization. In: Zentall TR, Wasserman E (eds) Comparative cognition: experimental explorations of animal intelligence. Oxford University Press, New York, pp 325–342 Jaswal VK, Markman EM (2001) Learning proper and common names in inferential versus ostensive contexts. Child Dev 72:768–786 Kaminski J, Call J, Fischer J (2004) Word learning in a domestic dog: evidence for “fast mapping”. Science 304:1682–1683 Kastak CR, Schusterman RJ (2002) Sea lions and equivalence: expanding classes by exclusion. J Exp Anal Behav 78:449–465 Mackintosh NJ (2000) Abstraction and discrimination. In: Heyes C, Huber L (eds) The evolution of cognition. MIT Press, Cambridge, pp 123–141 Macphail EM (1987) The comparative psychology of intelligence. Behav Brain Sci 10:645–695 Markman EM, Wachtel GF (1988) Children’s use of mutual exclusivity to constrain the meanings of words. Cogn Psychol 20:121–157 McIlvane WJ, Kledaras JB, Munson LC, King KAJ, de Rose JC, Stoddard LT (1987) Controlling relations in conditional discrimination and matching by exclusion. J Exp Anal Behav 48:187–208 Meehan EF (1995) Emergence by exclusion. Psychol Rec 45:133–154 Mervis CB, Bertrand J (1994) Acquisition of the novel name—nameless category (n3c) principle. Child Dev 65:1646–1662 Miklosi Ã, Polgardi R, Topal J, Csanyi V (1998) Use of experimentergiven cues in dogs. Anim Cogn 1:113–121 Murphy CJ, Zadnik K, Mannis MJ (1992) Myopia and refractive error in dogs. Invest Ophtmol Vis Sci 33:2459–2463 Neitz J, Geist T, Jacobs G (1989) Color vision in the dog. Visual Neurosci 3:119–125 Odom JV, Bromberg NM, Dawson WW (1983) Canine visual acuity: retinal and cortical Weld potentials evoked by pattern stimulation. Am J Physiol 245:R637–R641 Peichl L (1991) Catecholaminergic amacrine cells in the dog and wolf retina. Vis Neurosci 7:575–587 Premack D (1995) Cause/induced motion; intention/spontaneous motion. In: Changeux JP, Chavaillon J (eds) Origins of the human brain. Oxford University Press, Oxford, pp 286–308
597 Premack D, Premack AJ (1994) Levels of causal understanding in chimpanzees and children. Cognition 50:347–362 Range F, Aust U, Steurer M, Huber L (2007) Visual categorization of natural stimuli by domestic dogs (Canis familiaris). Anim Cogn. doi:10.1007/s10071-007-0123-2 Riedel J, Buttelmann D Call J, Tomasello M (2006) Domestic dogs (Canis familiaris) use a physical marker to locate hidden food. Anim Cogn 9:27–35 Savage-Rumbaugh ES (1986) Ape language: from conditioned response to symbol. Columbia University Press, New York Schusterman RJ, Krieger K (1984) California sea lions are capable of semantic comprehension. Psychol Rec 34:3–23 Stromer R (1989) Symmetry of control by exclusion in humans’ arbitrary matching to sample. Psychol Rep 64:915–922 Szetei V, Miklosi A, Topal J, Csanyi V (2003) When dogs seem to lose their nose: an investigation on the use of visual and olfactory cues in communicative context between dog and owner. Appl Anim Behav Sci 83:141–152 Terrace HS, Chen S, Newman AB (1995) Serial learning with a wild card by pigeons (Columba livia): eVect of list length. J Comp Psychol 109:162–172 Tomonaga M (1993) Tests for control by exclusion and negative stimulus relations of arbitrary matching to sample in a “symmetryemergent” chimpanzee. J Exp Anal Behav 59:215–229 Tomonaga M (1999) Establishing functional equivalence in a chimpanzee (Pan troglodytes) with a two-item sequential-responding procedure. J Exp Anal Behav 72:57–79 Tomonaga M, Matsuzawa T (2000) Sequential responding to arabic numerals with wild cards by the chimpanzee (Pan troglodytes). Anim Cogn 3:1–11 Tomonaga M, Matsuzawa T, Fujita K, Yamamoto J (1991) Emergence of symmetry in a visual conditional discrimination by chimpanzees (Pan troglodytes). Psychol Rep 68:51–60 Vincent-Smith L, Bricker DD, Bricker W (1974) Acquisition of receptive vocabulary in the toddlerage child. Child Dev 45:189–193 Wilkinson KM, Dube WV, McIlvane WJ (1998) Fast mapping and exclusion (emergent matching) in developmental language behavior analysis and animal cognition research. Psychol Rec 48:407–422 Zentall TR, Edwards CA, Moore BS, Hogan DE (1981) Identity: the basis for both matching and oddity learning. J Exp Psychol: Anim Behav Process 7:70–86
123