Professional Documents
Culture Documents
Spontaneous Learning Crows
Spontaneous Learning Crows
063
Report
Crows Spontaneously Exhibit
Analogical Reasoning
Anna Smirnova,1 Zoya Zorina,1 Tanya Obozova,1 (Table S2) as well as to stimuli from an altogether different
and Edward Wasserman2,* category (Figure S2)—stimuli varying in size—suggesting
1Department of Biology, Lomonosov Moscow State that the birds had acquired a general rule that was based on
University, Moscow 119899, Russia physical identity [14].
2Department of Psychology, The University of Iowa, Iowa City, The key behavioral assessments followed, in which novel
IA 52242, USA pairs of items served as the sample and test stimuli [7]. The
visual dimensions in these assessments were size, shape,
and color (Figures 1, 2, and 3 show several illustrative, not
Summary actual, trial sequences). On identity trials, we arranged differ-
ential reinforcement (food was given only after correct
Analogical reasoning is vital to advanced cognition and choices) to promote continued discriminative responding;
behavioral adaptation. Many theorists deem analogical accurate choice responding here could be based on either
thinking to be uniquely human and to be foundational to physical or relational matches between the sample pair and
categorization, creative problem solving, and scientific dis- the correct test pair. On the critical relational trials, we
covery [1]. Comparative psychologists have long been inter- arranged nondifferential reinforcement (food was given after
ested in the species generality of analogical reasoning, but all choices) because we did not want to teach the crows the
they initially found it difficult to obtain empirical support very behavior that we were assessing; accurate choice re-
for such thinking in nonhuman animals (for pioneering sponding here could only be based on relational matches be-
efforts, see [2, 3]). Researchers have since mustered consid- tween the sample pair and the correct test pair. Assessment
erable evidence and argument that relational matching-to- sessions were conducted 5 days a week and contained 48
sample (RMTS) effectively captures the essence of analogy, trials: 36 identity trials and 12 relational trials. The sessions
in which the relevant logical arguments are presented visu- contained six blocks (each block containing six identity trials
ally [4]. In RMTS, choice of test pair BB would be correct if and two relational trials); the trial order was randomized within
the sample pair were AA, whereas choice of test pair EF each of the six blocks. On all trials, we scored as ‘‘correct’’
would be correct if the sample pair were CD. Critically, no those choices that accorded with relational matching. Each
items in the correct test pair physically match items in the assessment phase lasted eight sessions.
sample pair, thus demanding that only relational sameness The first assessment phase involved size. Figure 1 shows
or differentness is available to support accurate choice re- that on half of the trials the sample pair involved shapes of
sponding. Initial evidence suggested that only humans and the same size, whereas on the other half of the trials the sam-
apes can successfully learn RMTS with pairs of sample ple pair involved shapes of different sizes. With the different-
and test items [4–7]; however, monkeys have subsequently sized sample and choice pairs, the large shape always
done so [8–12]. Here, we report that crows too exhibit rela- appeared on the left and the small shape always appeared
tional matching behavior. Even more importantly, crows on the right in order to keep the number of sample-compari-
spontaneously display relational responding without ever son combinations within manageable limits. The palette of
having been trained on RMTS; they had only been trained simple black shapes from which the sample and test pairs
on identity matching-to-sample (IMTS). Such robust and could be selected contained a square, a rectangle, a circle,
uninstructed relational matching behavior represents the an oval, a wide triangle, and a narrow triangle. These shapes
most convincing evidence yet of analogical reasoning in a could be large or small; the overall area of the two test pairs
nonprimate species, as apes alone [7] have spontaneously was equated, so that this stimulus property could not control
exhibited RMTS behavior after only IMTS training. the crows’ behavior. Identity matching trials were arranged in
which one test pair presented the same shapes in the same
Results and Discussion sizes as the sample pair; each of the sample and test pairs
involved the same shape, and only correct choices were
Our findings come from three separate behavioral assess- reinforced on these trials. Relational matching trials were
ments that followed pretraining and pretesting on identity arranged in which neither of the test pairs matched the sample
matching-to-sample (IMTS; see the Supplemental Experi- pair in shape, thereby eliminating control by physical identity;
mental Procedures available online), deploying behavioral on these trials, either correct or incorrect choices were
methods that were earlier used by Smirnova, Lazareva, and reinforced.
Zorina [13]. In that prior IMTS period, two hooded crows The second assessment phase involved shape. Figure 2
were shown several different kinds of visual stimuli: single shows that on half of the trials the sample pair involved iden-
stimuli of varying colors or shapes and clusters of stimuli tical shapes, whereas on the other half of the trials the sample
containing varying numbers of items. The birds were taught pair involved nonidentical shapes. In this assessment phase
to match stimuli by color, shape, and number of items, and the next, the items in each pair were arranged along the
but not size. After learning these various IMTS tasks, the positive diagonal in order for the spatial arrangement to differ
crows reliably transferred discriminative responding to new from the horizontal arrangement in the prior assessment phase
stimuli from the same categories that were used in training and to keep the number of sample-comparison combinations
within manageable limits. The palette of shapes contained
equal-sized circles, squares, triangles, and crosses that could
*Correspondence: ed-wasserman@uiowa.edu be colored red, yellow, blue, or green. Identity matching trials
257
Trial Left Test Sample Right Test Trial Left Test Sample Right Test
1 1
2 2
3 3
4 4
5 5
6 6
7 7
8 8
Figure 1. Examples of Identity and Relational Trials across Eight Exemplary Figure 2. Examples of Identity and Relational Trials across Eight Exemplary
Trials for Size Stimuli Trials for Shape Stimuli
On three-fourths of the trials (1–3 and 5–7), the correct test stimulus was an On three-fourths of the trials (1–3 and 5–7), the correct test stimulus was
identity match to the sample. On one-fourth of the trials (4 and 8, shaded an identity match to the sample in shape. On one-fourth of the trials
gray), the correct test stimulus was a relational match to the sample. (4 and 8, shaded gray), the correct test stimulus was a relational match to
the sample.
8
matching (averaging 77.78% correct) that was just as robust
as their identity matching (averaging 72.69%) in session 1
Figure 3. Examples of Identity and Relational Trials across Eight Exemplary across all three assessment phases. Although physical iden-
Trials for Color Stimuli tity could have guided the crows’ choice behavior on IMTS tri-
On three-fourths of the trials (1–3 and 5–7), the correct test stimulus was als, physical identity could not have done so on RMTS trials, as
an identity match to the sample in color. On one-fourth of the trials (4 and no physical matches were possible between the sample pairs
8, shaded gray), the correct test stimulus was a relational match to the
and the correct test pairs. These results perhaps surprisingly
sample.
suggest that physical identity contributed little or nothing to
our crows’ testing performance; relational processing seems
also analyzed choice behavior in session 1 of each assess- to have been of prime importance to controlling the birds’
ment phase (third column in Table 1) using a similar logistic choice behavior.
regression as in the prior analyses. In session 1, crow 1 re- Our results thus constitute unprecedented behavioral evi-
sponded at significantly above chance levels of accuracy dence of analogical reasoning by a nonprimate animal. They
(B = 1.08, SE = 0.20, Z = 5.39, p < 0.0001), but there were no sig- therefore add to growing research undermining the influential
nificant effects of trial type or stimulus dimension; summed claims of such famous philosophers as René Descartes and
across all three dimensions, accuracy on identity trials aver- John Locke that only humans are capable of abstract thought.
aged 73.15% correct, and accuracy on relational trials aver- Relational reasoning—particularly appreciating the relation
aged 75.00% correct. Crow 2 also responded at significantly between relations, as in analogies—can no longer be deemed
above chance accuracy levels in session 1 (B = 1.11, SE = to be the unique pinnacle of human cognition.
0.20, Z = 5.53, p < 0.0001), but there were no significant effects It may be no accident that crows performed so impressively
of trial type or stimulus dimension; summed across all three in our study; they stand out among birds in their highly devel-
dimensions, accuracy on identity trials averaged 72.22% oped neuroanatomy [16, 17]. More generally, mounting evi-
correct, and accuracy on relational trials averaged 80.56% dence indicates that although birds do not have a brain
correct. structure that is homologous to the mammalian prefrontal
We thus found that, when tested during the size-assess- cortex, the avian nidopallium caudolaterale may effectively
ment phase immediately after initial IMTS training, our crows mediate complex cognitive functions, perhaps representing
not only responded discriminatively on identity matching a case of convergent evolution [18].
trials, but also did so on relational matching trials, all of It should nevertheless be appreciated that we are not claim-
these trials for the first time involving two-item sample and ing that our crows’ spontaneous relational matching behavior
test stimuli. These results represent striking behavioral evi- arose entirely de novo. Indeed, we believe that their earlier
dence that the crows spontaneously perceived the relation IMTS training is likely to have enabled them to grasp a broadly
between relations without ever having been explicitly trained applicable concept of sameness that could apply to novel two-
to do so. item sample and test stimuli involving only relational sameness
That initial size assessment entailed sample and compari- [19, 20]. Just how that remarkable transfer is accomplished
son stimuli that were horizontally ordered large (left) to small represents an intriguing matter for future study. Nor are we
(right) with different-sized pairs. Could that incidental spatial claiming that crows will prove to be the only nonprimate ani-
cue have affected the accuracy of the crows’ later choice mals that are capable of exhibiting such spontaneous rela-
behavior? Evidently not. The crows continued to respond at tional matching behavior. Future research must be undertaken
similarly high levels of accuracy on identity and relational trials in which different species are given comparable pretraining
during the shape and color assessment phases, during which experience to our crows. Until such systematic comparative
the items in the sample and comparison pairs were diagonally research is conducted, it would be premature to offer specula-
arranged. tive evolutionary accounts as to why crows appear to have
This documentation of RMTS behavior is particularly note- excelled in solving this challenging cognitive task. It would
worthy because our crows exhibited discriminative relational also be premature to suggest that simply because crows
259
successfully master RMTS tasks, they process the stimuli and repeated that trial. If the bird again did not choose either card, then the
relations in the same way as do humans and other nonhuman experimenter ended the session.
animals.
Finally, we should underscore the importance of research on Supplemental Information
animal behavior to the role of language or symbol systems in
Supplemental Information includes Supplemental Experimental Proce-
abstract conceptualization. The first evidence of analogical dures, two figures, two tables, and one movie and can be found with this
reasoning in animals came from symbol-trained chimpanzees article online at http://dx.doi.org/10.1016/j.cub.2014.11.063.
[2, 5]. Accordingly, the hypothesis was advanced that only if an
organism had acquired an elaborate symbol system might it Author Contributions
encode and process abstract analogical relations [5]. The re-
sults of more recent primate studies [6, 7, 12], as well as the A.S. and Z.Z. designed and conducted the studies and interpreted the data.
findings from our present experiment with crows, suggest T.O. and E.W. helped conduct the statistical analyses and interpret the data
and composed the paper in collaboration with the other coauthors.
that extensive prior experience with the abstract concept of
‘‘sameness’’ per se—rather than a linguistic symbol for such
Acknowledgments
a concept—is sufficient for animals to succeed in solving
analogical reasoning tasks. This research was supported by Russian Foundation for Basic Research
(RFBR) grant 13-04-00747. We thank Leyre Castro, Bob McMurray, and
Experimental Procedures Mark Blumberg for their help in preparing this report.
15. Håstad, O., Victorsson, J., and Ödeen, A. (2005). Differences in color
vision make passerines less conspicuous in the eyes of their predators.
Proc. Natl. Acad. Sci. USA 102, 6391–6394.
16. Emery, N.J. (2006). Cognitive ornithology: the evolution of avian intelli-
gence. Philos. Trans. R. Soc. Lond. B Biol. Sci. 361, 23–43.
17. Portmann, A. (1947). Études sur la cérébralisation chez les oiseaux. II.
Les indices intracérébraux. Alauda 15, 1–15.
18. Güntürkün, O. (2012). The convergent evolution of neural substrates for
cognition. Psychol. Res. 76, 212–219.
19. Truppa, V., Piano Mortari, E., Garofoli, D., Privitera, S., and Visalberghi,
E. (2011). Same/different concept learning by capuchin monkeys in
matching-to-sample tasks. PLoS ONE 6, e23809.
20. Wasserman, E.A. (2008). Development and evolution of cognition: one
doth not fly into flying!. Behav. Brain Sci. 31, 400–401.
Current Biology, Volume 25
Supplemental Information
Figure S2. Examples of sample and comparison stimuli across 8 exemplary IMTS testing trials
outside the training categories. On three-fourths of the trials (1-3 and 5-7), only correct choices
were reinforced; these were trials where the sample stimulus was identical to one of the
comparison stimuli. On one-fourth of the trials (4 and 8, shaded grey), both correct and incorrect
choices were reinforced; these were trials where the sample stimulus was not identical in shape
to either of the comparison stimuli, but matched the comparison stimuli on the basis of the novel
category of size.
Bird Dimension Trial Type Session 1 Session 2 Session 3 Session 4 Session 5 Session 6 Session 7 Session 8 Sessions 1 - 8
Identity 77.78 75.00 86.11 86.11 72.22 80.56 77.78 80.56 79.51
Size
Relational 75.00 83.33 83.33 66.67 66.67 50.00 75.00 83.33 72.92
Identity 75.00 80.56 61.11 69.44 80.56 69.44 63.88 66.67 70.83
Crow 1 Shape
Relational 91.67 75.00 83.33 66.67 75.00 66.67 66.67 66.67 73.96
Identity 66.67 80.56 86.11 75.00 69.44 75.00 63.88 91.67 76.04
Color
Relational 58.33 75.00 50.00 66.67 75.00 50.00 50.00 66.67 61.46
Identity 75.00 72.22 72.22 66.67 86.11 72.22 86.11 80.56 76.39
Size
Relational 91.67 75.00 83.33 75.00 58.33 80.56 66.67 75.00 76.04
Identity 63.89 63.88 63.88 75.00 83.33 66.67 75.00 77.78 71.18
Crow 2 Shape
Relational 75.00 75.00 91.67 58.33 100.00 83.00 66.67 66.67 77.08
Identity 77.78 72.22 72.22 80.56 86.11 83.33 77.78 80.56 78.82
Color
Relational 75.00 66.67 66.67 75.00 75.00 83.33 58.33 75.00 71.88
Table S1. Mean percentage of correct choices by Crows 1 and 2. Accuracy scores are from Identity (IMTS) and Relational (RMTS) trials on
Table S2. Examples of sample and comparison stimuli across 8 exemplary trials of IMTS testing
within the training categories. On three-fourths of the trials (1-3 and 5-7) only correct choices
were reinforced; these were trials with the familiar training stimuli. On one-fourth of the trials (4
and 8, shaded grey) both correct and incorrect choices were reinforced; these were trials with the
The crows’ pretraining and pretesting was on IMTS tasks in the apparatus shown in
Figure S1. Here, the crows’ choice responses were reinforced for selecting the comparison
stimulus that was identical to the sample. It is important to stress that, in order to conclude that
crows had acquired a “matching” concept, one must provide clear evidence of successful
learning as well as successful transfer of matching behavior to novel stimuli. If the birds
accurately apply the matching rule to completely novel stimuli, then it may be concluded that the
We thus trained the crows to acquire the concept of sameness based on color, Arabic
numerals, and the number of items composing the stimuli. We later tested their ability to match
novel stimuli from the same categories and from an entirely new category—size.
IMTS Training
IMTS training continued until the acquisition criterion (80% correct choices over 96
consecutive trials; binomial test, p < .001) had been reached for each of four successively given
stimulus sets: (a) black and white, (b) Arabic numerals 1 and 2, (c) arrays of 1 and 2 elements,
and (d) intermixed pairs (black and white, light grey and dark grey, Arabic numerals from 1 to 4,
To reach the criteria for each of four stimulus sets, Crow 1 needed 2,096, 1,056, 128, and
96 trials, respectively; Crow 2 needed 3,136, 128, 104, and 96 trials, respectively.
After training, we explored whether the crows could transfer their IMTS behavior to
novel stimuli within the training categories. On three-fourth of the trials, all of the stimuli came
from the training set. On another one-fourth of the trials, none of the stimuli came from the
training set.
Each crow completed 288 trials over 6 sessions. On each of the trials, the correct
comparison stimulus was an exact match to the sample. On 216 trials, all of the stimuli had
previously been seen during training. On another 72 trials, none of the stimuli had previously
been seen during training. The testing stimuli came from the same categories: four different
Binomial tests for choice accuracy on both familiar stimulus training trials and novel
stimulus testing trials revealed that each crow performed at significantly above chance levels
(50%), p < 0.001, thereby attesting to the ability of the birds to transfer a matching rule to new
stimuli from the same training category. The mean percentage of correct choices for Crow 1 was
81.48% on training trials and 83.33% on testing trials. The mean percentage of correct choices
for Crow 2 was 83.33% on training trials and 79.16% on testing trials.
This final step explored whether the crows could transfer their IMTS behavior to stimuli
from a novel category—size. On three-fourth of the trials, the correct comparison stimulus was
an exact match to the sample. On another one-fourth of the trials, the correct comparison
stimulus matched the sample in size, but depicted a different shape. Each testing stimulus
involved six different black shapes on a white background. Each shape appeared in one of two
Both crows completed 480 trials over 10 sessions. On 360 trials, the correct comparison
stimulus was an exact match to the sample (exact match). On another 120 trials, the correct
comparison stimulus matched the size, but not the shape of the sample (size match). [A series of
8 exemplary trials is in Figure S2.] Because of the importance of assessing the immediacy of
transfer outside of the training categories, only the scores from Session 1 are reported here.
Binomial tests for choice accuracy in Session 1 of testing revealed that each bird
consistently performed at above chance levels (50%), all p < 0.001, thereby documenting the
ability of crows to transfer IMTS to novel stimuli outside the training categories. The percentage
of correct choices for Crow 1 was 83.33% on exact match trials and 83.33% on size match trials.
The percentage of correct choices for Crow 2 was 88.89% on exact match trials and 91.67% on
size match trials. These scores indicate that the crows had indeed acquired a general matching
rule based on the abstract concept of sameness before the final step of RMTS testing had begun.
To control for any inadvertent visual cues, the stimulus tray was prepared for each trial
out of the field of the bird’s vision. An opaque plastic screen (70 × 40 cm) was placed between
the experimental cage and the experimenter, minimizing the chance of unintentional cues from
the experimenter. The bird could not see the experimenter at the moment of choice nor could the
experimenter see the bird. This methodological measure precluded the possibility of a “Clever
Hans” error. Note that because the experimenter could not see the bird, she could only judge the
outcome of each trial by the sounds made by the bird. If she heard the sound of a card falling
from the cup and the later sound of the bird pecking the cup containing the worms, then the
choice was correct. If no pecking sound occurred, then it meant that the choice was incorrect and
the experimenter quickly removed the tray to prevent the bird from uncovering the second cup.
To control for any inadvertent auditory cues, the tray was always prepared in the same
sequence for each trial: the left cup was placed on the tray, then right cup was placed on the tray.
One of the cups contained two mealworms as reinforcement. Subsequently, the left cup was
covered with a comparison stimulus, then right cup was covered with a comparison stimulus.
The mealworms were alive and may have emitted sounds, but the importance of such sounds
could be eliminated on RMTS trials because mealworms were placed into each cup.
Statistical Analysis Details During Identity and Relational Assessment Phases
Our primary data were analyzed with logistic regressions run separately for each crow.
Each analysis examined accuracy as a function of three independent factors. Session (1-8) was a
linear predictor and centered. Trial Type (Identity vs. Relational) was dummy coded (Relational
= 1) and then centered. Stimulus Dimension was coded as two dummy variables: one was set to
1 for Size and 0 otherwise; the other was set to 1 for Color and 0 otherwise. Both were then
centered. Because the main effect of Stimulus Dimension (and its various interactions) was
spread across two variables, the significance of fixed effects was established with the χ2 of
model comparison. Main effects were evaluated by comparing a model with all three main
effects to one without the effect in question. Two-way interactions were evaluated by comparing
a model with all two-way interactions to one without the interaction in question. The three-way
interaction was evaluated by adding the three-way interaction terms to a model with only two-
way interactions. The intercept of the model (and its significance) tests the hypothesis that
accuracy is above (or below) chance; we report the Wald Z statistic for this purpose. Models
were run using the GLM command (family = “binomial”) in R (Version 3.1.0).