Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/5371058
CITATIONS READS
499 149
3 authors:
Daniel J Povinelli
University of Louisiana at Lafayette
143 PUBLICATIONS 7,125 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Keith J Holyoak on 03 June 2014.
Keith J. Holyoak
Department of Psychology, University of California – Los Angeles, Los Angeles,
CA 90095
holyoak@lifesci.ucla.edu
http://reasoninglab.psych.ucla.edu/
Daniel J. Povinelli
Cognitive Evolution Group, University of Louisiana, Lafayette, LA 70504
ceg@louisiana.edu
http://www.cognitiveevolutiongroup.org/
Abstract: Over the last quarter century, the dominant tendency in comparative cognitive psychology has been to emphasize the
similarities between human and nonhuman minds and to downplay the differences as “one of degree and not of kind” (Darwin
1871). In the present target article, we argue that Darwin was mistaken: the profound biological continuity between human and
nonhuman animals masks an equally profound discontinuity between human and nonhuman minds. To wit, there is a significant
discontinuity in the degree to which human and nonhuman animals are able to approximate the higher-order, systematic,
relational capabilities of a physical symbol system (PSS) (Newell 1980). We show that this symbolic-relational discontinuity
pervades nearly every domain of cognition and runs much deeper than even the spectacular scaffolding provided by language
or culture alone can explain. We propose a representational-level specification as to where human and nonhuman animals’
abilities to approximate a PSS are similar and where they differ. We conclude by suggesting that recent symbolic-
connectionist models of cognition shed new light on the mechanisms that underlie the gap between human and nonhuman
minds.
Keywords: analogy; animal cognition; causal learning; connectionism; Darwin; discontinuity; evolution; human mind; language;
language of thought; physical symbol system; reasoning; same-different; theory of mind
forces and that our ability to do so relies on a unique rep- numerous researchers have shown, the propensity to
resentational system that has been grafted onto the cogni- evaluate the similarity between states of affairs based on
tive architecture we inherited from our nonhuman the causal-logical and structural characteristics of the
ancestors (Povinelli 2000; 2004; Povinelli & Giambrone underlying relations rather than on their shared percep-
2001; Povinelli & Preuss 1995; Povinelli & Vonk 2003; tual features appears quite early and spontaneously in
2004; Vonk & Povinelli 2006). Independently, Holyoak, all normal humans – as early as 2– 5 years of age, depend-
Hummel, and colleagues have argued that the ability to ing on the domain and complexity of the task (Gentner
reason about higher-order relations in a structurally sys- 1977; Goswami 2001; Halford 1993; Holyoak et al. 1984;
tematic and inferentially productive fashion is a defining Namy & Gentner 2002; Rattermann & Gentner 1998a;
feature of the human mind and requires the distinctive Richland et al. 2006).
representational capabilities of a “biological symbol In short, there appear to be at least two kinds of simi-
system” (Holyoak & Hummel 2000; 2001; Hummel & larity judgments at work in human thought: judgments
Holyoak 1997; 2001; 2003; Kroger et al. 2004; Robin & of perceptual similarity based on the relation between
Holyoak 1995). Herein we combine, revise, and substan- observed features of stimuli; and judgments of non-
tially expand on the hypotheses proposed by these two perceptual relational similarity based on logical, func-
research groups. tional, and/or structural similarities between relations
We argue that most of the salient functional discontinu- and systematic correspondences between the abstract
ities between human and nonhuman minds – including roles that elements play in those relations (Gentner
our species’ unique linguistic, mentalistic, cultural, 1983; Gick & Holyoak 1980; 1983; Goswami 2001;
logical, and causal reasoning abilities – result in part Markman & Gentner 2000). The question we are
from the difference in degree to which human and nonhu- interested in here is whether or not there is any evidence
man cognitive architectures are able to approximate the for non-perceptual relational similarity judgments in
higher-order, systematic, relational capabilities of a phys- nonhuman animals as well.
ical symbol system (Newell 1980; Newell & Simon
1976). Although human and nonhuman animals share
many similar cognitive mechanisms, our relational reinter-
2.2. Same-different relations
pretation hypothesis (RR) is that only human animals
possess the representational processes necessary for sys- Among comparative researchers, the most widely repli-
tematically reinterpreting first-order perceptual relations cated test of relational concept learning over the last
in terms of higher-order, role-governed relational struc- quarter century has been the simultaneous same-different
tures akin to those found in a physical symbol system (S/D) task, in which the subject is trained to respond one
(PSS). We conclude by suggesting that recent advances way if two simultaneously presented stimuli are the same
in symbolic-connectionist models of cognition provide and to respond a different way if the two stimuli are differ-
one possible explanation for how our species’ unique ent. In the purportedly more challenging relational match-
ability to approximate the higher-order relational capabili- to-sample (RMTS) task, the subject must select the choice
ties of a physical symbol system might have been grafted display in which the perceptual similarity among elements
onto the proto-symbolic cognitive architecture we inher- in the display is the same as the perceptual similarity
ited from our nonhuman ancestors in a biologically plaus- among elements in the sample stimulus. For example, pre-
ible manner. sented with a pair of identical objects, AA, as a sample
stimulus, the subject should select BB rather than CD;
presented with a pair of dissimilar objects, EF, as the
2. Similarity sample stimulus, the subject should select GH rather
than JJ (see Thompson & Oden 2000 for a seminal
2.1. Perceptual versus relational similarity
discussion).
We begin our review of the similarities and differences Although Premack (1983a; 1983b) initially reported that
between human and nonhuman cognition with what only language-trained chimpanzees passed S/D and
William James (1890/1950) called “the very keel and back- RMTS tasks, success on two-item S/D tasks has since
bone of our thinking”: sameness. The ability to evaluate the been demonstrated in parrots (Pepperberg 1987), dol-
perceptual similarity between stimuli is clearly the sine phins (Herman et al. 1993b; Mercado et al. 2000),
qua non of biological cognition, subserving nearly every baboons (Bovet & Vauclair 2001), and pigeons (Blaisdell
cognitive process from stimulus generalization and Pavlo- & Cook 2005; Katz & Wright 2006), among others.
vian conditioning to object recognition, categorization, Thompson et al. (1997) showed that language-naive chim-
and inductive reasoning. Humans, however, are not panzees with some exposure to token-based symbol
limited to evaluating the similarity between objects systems are able to pass a two-item RMTS task
based on perceptual regularities alone. Humans not only (cf. Premack 1988). Vonk (2003) has reported that three
recognize when two physical stimuli are perceptually orangutans and one gorilla were able to pass a complex
similar, they can also recognize that two ideas, two two-item RMTS task without any explicit symbol or
mental states, two grammatical constructions, or two language training at all. Fagot et al. (2001) have shown
causal-logical relations are similar as well. Even pre- that language-naı̈ve baboons can pass an RMTS task invol-
school-age children understand that the relation between ving arrays of elements (see discussion below); and Cook
a bird and its nest is similar to the relation between a and Wasserman (in press) have reported successful
dog and its doghouse despite the fact that there is little results on an array-based RMTS task with pigeons. So
“surface” or “object” similarity between the relations’ con- passing S/D and RMTS tasks does not appear to be
stituents (Goswami & Brown 1989; 1990). Indeed, as limited to language-trained apes or even primates.
their responses were correct 80% of the time on trials perceptually disparate relations, allowing the cognizer to
involving novel pairs of objects. draw novel inferences about the target domain indepen-
The available evidence therefore suggests that the for- dently from the perceptual similarity between the
mative discontinuity in same-different reasoning lies not relations’ constituents (Gentner 1983; Gentner &
between monkeys and apes, as Thompson and Oden Markman 1997; Holyoak & Thagard 1995). Accordingly,
(2000) proposed, but between nonhumans and humans. analogical relations sensu stricto cannot be reduced via
Chimpanzees and other nonhuman apes can pass chunking and segmentation, but require the cognizer to
RMTS tasks with only 2 items in the sample display evaluate the abstract, higher-order relations at stake in a
(e.g., Thompson et al. 1997; Vonk 2003). Baboons can structurally systematic and inferentially productive
pass RMTS tasks with as few as 3 – 4 items in each fashion.
sample (Fagot et al. 2001); and pigeons can pass RMTS Analogical reasoning is a fundamental and ubiquitous
tasks with 16 items in each sample (Cook & Wasserman, aspect of human thought. It is at the core of creative
in press). The difference between the performance of problem solving, scientific heuristics, causal reasoning,
language-naive pigeons and language-trained chimps on and poetic metaphor (Gentner 2003; Gentner et al.
these tasks often comes down to a question of the 2001; Holyoak & Thagard 1995; 1997; Lien & Cheng
number of items in each set and the number of trials 2000). And it is also central to the more prosaic ways
necessary to reach criterion. As Katz and colleagues that typical human children learn about the world and
point out (see Katz & Wright 2006; Katz et al. 2002), each other (Goswami 1992; 2001; Halford 1993; Holyoak
this strongly suggests that there is a difference in degree et al. 1984). To date, however, the only evidence that
between various nonhuman species’ sensitivity to simi- any nonhuman animal is capable of analogical reasoning
larity discriminations (influenced by training regimen), sensu stricto comes from the unreplicated feats of a
not a difference in kind between their conceptual abilities single chimpanzee, Sarah, reported more than 25 years
to predicate same-different relations. ago by Gillan et al. (1981). Sarah reportedly constructed
The performance of human subjects, on the other hand, and completed two distinct kinds of analogies. The first
contrasts sharply with the performance of all other animal was based on judging whether or not two geometric
species. Humans manifest an abrupt, categorical distinc- relationships were the same or different (e.g., large blue
tion between displays in which there is no variability and triangle is to small blue triangle as large yellow crescent
displays in which there is any variability at all (Cook & is to small yellow crescent). The second was based on
Wasserman 2006; Wasserman et al. 2004). More impor- judging the similarity between two “functional” relation-
tantly, contra Castro et al. (2007), we believe that human ships (e.g., padlock is to key as tin can is to can opener).
subjects possess a qualitatively distinct system for reinter- Gillan et al. (1981) reported that Sarah was successful on
preting sameness and difference in a logical and abstract both tests.
fashion that generalizes beyond any particular source of Savage-Rumbaugh was the first to point out that Sarah’s
stimulus control. In short, even with respect to the most performance on the geometric version of the original tests
basic and ubiquitous of all cognitive phenomena – judg- could have been the result of a simple, feature-matching
ments of similarity – there is already a distinctive seam heuristic (cited by Oden et al. 2001). In response, Oden
between human and nonhuman minds. et al. (2001) followed up Gillan et al.’s original experiment
on geometric analogies with a series of more carefully con-
structed tests designed to flesh out Sarah’s actual cognitive
2.3. Analogical relations
strategy. These new experiments used geometric forms
Premack (1983a, p. 357) suggested that the RMTS task is that varied along one or more featural dimensions (e.g.,
an implicit form of analogy and claimed that “animals that size, color, shape, and/or fill). After extensive testing,
can make same/different judgments should be able to do Oden et al. showed that Sarah was actually tracking the
analogies.” Indeed, it is still widely accepted that the number of within-pair featural differences rather than
ability to pass an RMTS task is the “cognitive primitive” the kind of relation between pairs of figures. For
for analogical reasoning (see, e.g., Thompson & Oden example, whereas a human would see a color plus a
2000, p. 378). We disagree. While recognizing perceptual shape change as differing from a size plus a fill change,
similarities is certainly a necessary condition for making Sarah saw these two transformations as equivalent
analogical inferences (inter alia), there is a qualitative because they both entailed two featural changes.
difference between the kind of cognitive processes necess- Oden et al. (2001) argued that this strategy still demon-
ary to pass an S/D or RMTS task and the kind of cognitive strates Sarah’s ability to reason about the “relation
processes necessary to reason in an analogical fashion. The between relations.” But there is a profound difference
relations at issue in S/D and RMTS tasks are based solely between the feature-based heuristic Sarah apparently
on the perceptual features of the constituents; and the adopted and the role-based structural operations that are
constituents play undifferentiated and symmetrical roles the basis of analogical inference sensu stricto. To be
in those relations (e.g., two objects are symmetrically sure, keeping track of the number of within-pair featural
either the same or different). changes certainly requires quite sophisticated represen-
Most true analogies, on the other hand, are based on tational processes. But the fact that Sarah apparently
relations in which the constituents play asymmetrical, ignored the structure of the relation between pairs of
causal-logical roles (e.g., the role that John plays in figures suggests that she represented any featural change
forming the relation, John loves Mary, is not equivalent as an undifferentiated chunk for the purposes of this
to the role that Mary plays, perhaps to John’s dismay). task. Therefore, her strategy on this task appears to be
Furthermore, genuine analogical inferences are made by computationally equivalent to the kind of chunking and
finding systematic structural similarities between segmentation strategies other nonhuman primates use
relation between C and A? Is A dominant to C to a greater ages created five-cup sets less consistently than the nonhu-
or lesser extent than B is dominant to C? (Goodwin & man subjects did, and they were rarely able to place a sixth
Johnson-Laird 2005; Halford et al. 1998a). cup into a seriated set. Bizarrely, at least for a purely rela-
In short, whereas at least some nonhuman animals tional interpretation of the results, monkeys were more
clearly are able to make transitive inferences about their successful than either apes or human children on the
own relation to potential rivals to a degree that rules out more challenging six-cup trials, yet were also the most inef-
purely associative learning mechanisms, the comparative ficient (in terms of number of moves) of the three
evidence accumulated to date is nevertheless consistent populations.
with the hypothesis that nonhuman animals’ under- Fragaszy et al.’s (2002) explanation for these anomalous
standing of transitive relations is punctate, egocentric, results is quite sensible (see also Fragaszy & Cummins-
non-logical, and context-specific. Sebree 2005): They hypothesize that the seriation task
does not, in fact, require the subject to reason about com-
binatorial, hierarchical relations per se, but depends more
6. Hierarchical relations simply on situated, embodied sensory-motor skills that are
experientially, rather than conceptually, driven. Apes and
Being able to process recursive operations over hierarchi- monkeys do better than children because they are more
cal relations is unarguably a key prerequisite for using a physically adept than 11- to 21-month-old children
human language (Hauser et al. 2002a). And most normal are – not because they have a more sophisticated rep-
human children are capable of reasoning about hierarchi- resentation of the combinatorial and hierarchical relations
cal class relations in a systematic and combinatorial fashion involved. Although subassembly may be a more physically
by the age of five (Andrews & Halford 2002; cf. Inhelder & “complex” strategy than other methods of seriation, it does
Piaget 1964). Given the ubiquity and importance of hier- not necessarily require the subject to cognize the spatial-
archical relations in human thought, the lack of any physical relations involved as hierarchical; and therefore
similar ability in nonhuman animals would therefore con- there is no reason to claim an isomorphism between the
stitute a marked discontinuity between human and nonhu- embodied manipulation of nested cups and the cognitive
man minds. manipulation of symbolic-relational representations
(cf. Greenfield 1991; Matsuzawa 1996).
6.1. Seriated cups and hierarchical reasoning
6.2. Hierarchical relations in the wild
A number of comparative researchers have reinterpreted
the behavior of nonhuman animals in hierarchical terms The strongest evidence to date in support of the claim that
(e.g., Byrne & Russon 1998; Greenfield 1991; Matsuzawa nonhuman animals can reason about hierarchically struc-
1996). In each of these cases, however, there is no evi- tured relations in the social domain comes from
dence that the nonhuman animals themselves cognized Bergman et al.’s (2003) study of free-ranging baboons.
the task in hierarchical terms or employed hierarchically Bergman et al. designed an elegant playback experiment
structured mental representations to do so. The most in which female baboons heard a sequence of recorded
widely cited case of hierarchical reasoning among nonhu- calls mimicking a fight between two other females. Mock
man animals, for example, has come from experiments agonistic confrontations were created by playing the
involving seriated cups. It has been claimed that “subas- “threat-grunt” of one individual followed by the subordi-
sembly” (i.e., combining two or more cups as a subunit nate screams of another. On separate days, the same
with one or more other cups) requires the subject to rep- subject heard one of three different call sequences: (1)
resent these nested relations in a combinatorial and an anomalous sequence mimicking a rank reversal
“reversible” fashion (Greenfield 1991; Westergaard & between members of the same matrilineal family (i.e.,
Suomi 1994). Indeed, Greenfield (1991) argued that chil- sisters, mothers, daughters, or nieces); (2) an anomalous
dren’s ability to nest cups develops in parallel with their sequence mimicking a between-family rank reversal (i.e.,
ability to employ hierarchical phonological and grammati- between members of two different matrilineal families in
cal constructions, and therefore, that the ability of nonhu- which one of the families is dominant to the other); or
man primates to seriate cups is the precursor to (3) a control sequence replicating an existing dominant-
comprehending hierarchical grammars (see Matsuzawa subordinate relationship (i.e., no rank reversal) using
1996 for claims of a similar “isomorphism” between tool between-family or within-family dyads. As predicted,
and symbol use). there was a significant difference in the focal subjects’
But is it actually necessary to cognize hierarchically responses to the three different kinds of call sequences.
structured relations in order to assemble nested cups? Subjects looked longest at between-family rank reversals.
To date, Johnson-Pynn, Fragaszy, and colleagues have There was no significant difference between within-
provided the most convincing evidence that a nonhuman family reversals and no-reversal control sequences.
animal can use subassembly to assemble seriated cups According to Bergman et al., the reason the baboons
(Fragaszy et al. 2002; Johnson-Pynn & Fragaszy 2001; responded more strongly to between-family rank reversals
Johnson-Pynn et al. 1999). Yet, Johnson-Pynn and Fra- than within-family sequences is because the baboons
gaszy themselves dispute the claim that this behavior recognized that the former imply a superordinate reorgan-
requires hierarchical relational operations of the kind ization of matrilineal subgroups. Bergman et al. (2003,
suggested by Greenfield (1991). p. 1236) conclude: “Our results suggest that baboons
Fragaszy et al. (2002), for example, presented seriated organize their companions into a hierarchical, rule-
cups to adult capuchin monkeys, chimpanzees, and 11-, governed structure based simultaneously on kinship and
16-, and 21-month-old children. Children of all three rank” (see also Seyfarth et al. 2005).
solve the task within 100 trials. However, none of the something more specific than being a socially savvy
chimps showed any evidence of distinguishing between animal: it means being able to impute unobservable, con-
the trap-up and trap-down versions of the task. By way tentful mental states to other agents and then to reason in
of comparison, it should be noted that children as young a theory-like fashion about the causal relation between
as 3 years of age successfully solve the trap-tube task these unobservable mental states and the agents’ sub-
after only a few trials (see Limongelli et al. 1995). sequent behavior (see Penn & Povinelli 2007b for a
Recently, Mulcahy and Call (2006b) tested ten great more extensive discussion of this point). Of course,
apes on a version of the trap-tube task that allowed sub- theory-like inferences are not the only way in which a cog-
jects to choose whether to pull or push the reward nizer might reason about other agents’ mental states (see
through the tube. Three out of the ten subjects learned Carruthers & Smith 1996 for a review of the possibilities).
to avoid the trap when pulling rather than pushing. Mentalistic simulation, for example, provides an alterna-
However, the majority of subjects still failed the task. tive and popular explanation. However, all but the most
Indeed, even the three successful subjects took an radical simulation-oriented theories do not deny that
average of 44 trials to achieve above-chance performance, humans represent causal relations involving other agents’
and then continued to fail Visalberghi and Limongelli’s unobservable mental states. They simply propose an
(1994) push-only version of the task. Therefore, these alternative, analogical mechanism for how humans do so.
latest results seem to confirm two earlier hypotheses: (1) Whiten (1996; 2000) has proposed another, influential
nonhuman apes are more adept at pulling than pushing hypothesis about how nonhuman apes (and young chil-
in tool-use tasks such as these (see, e.g., Povinelli 2000, dren) might represent the mental states of their conspeci-
Ch. 5); and (2) nonhuman primates’ causal knowledge is fics without relying on theory-like metarepresentations.
tightly coupled to specific task parameters and bodily Whiten proposed that nonhuman apes use “intervening
movements: in particular, they do not appear to grasp variables” to stand in for generalizations about the causal
the abstract, analogical similarity between perceptually role played by a given mental state in a set of disparate
disparate but functionally equivalent tasks (Penn & Povi- behavioral patterns. For example, a chimpanzee that
nelli 2007a; Povinelli 2000; Visalberghi & Tomasello encodes the observable patterns “X saw Y put food in
1998). bin A,” “X hid food in bin A,” and “X sees Y glancing at
Nonhuman primates are not the only animals that seem bin A” as members of the same abstract equivalence
to be incapable of cognizing the general causal principles class could be said, on Whiten’s account, to recognize
at issue in the trap-tube task. Seed et al. (2006) recently that “X knows food is in bin A” and, therefore, be
presented eight rooks with a clever modification to Visal- capable of “explicit mindreading” (Whiten 1996).
berghi’s trap-tube task in which each tube contained two Notice that Whiten’s example of “explicit mindreading”
traps, one which was functional and one which was not. is a textbook example of analogical reasoning: Whiten’s
Seven out of eight rooks rapidly learned to pull the food hypothetical chimpanzee must infer a systematic higher-
away from the functional trap and successfully transferred order relation among disparate behavioral patterns that
this solution to a novel but perceptually similar version of have nothing in common other than a shared but unobser-
the task. Nevertheless, when presented with transfer tasks vable causal mechanism: that is, what X “knows.” If this is
in which the visual cues that were associated with success an “intervening variable,” it is an intervening variable that
in the initial tasks were absent or confounded, only one of requires reasoning about the higher-order, role-governed
the seven subjects passed. In a follow-up experiment relational similarity between perceptually disparate
(Tebbich et al. 2007), none of the rooks passed the transfer causal relations in order to be produced.
tasks. We believe Whiten is right in this sense: If a nonhuman
Seed et al.’s (2006) results add to the growing evidence animal were capable of inferring that these disparate beha-
that corvids are quite adept at using stick-like tools (see, vioral patterns were actually instances of the same super-
e.g., Weir & Kacelnik 2007). But as Seed et al. (2006) ordinate causal relation, then the animal would surely
point out, these results also suggest that rooks share a have demonstrated that it possessed a ToM and the
common cognitive limitation with nonhuman primates: ability to reason analogically, as well. There is, however,
they do not understand “unobservable causal properties” no such evidence on offer. Indeed, until recently, there
such as gravity and support; nor do they reason about has been a fragile consensus that nonhuman animals lack
the higher-order relation between causal relations in an anything even remotely resembling a ToM (Cheney &
analogical or theory-like fashion. Instead, rooks, like Seyfarth 1998; Heyes 1998; Tomasello & Call 1997; Visal-
other nonhuman animals, appear to solve tool-use pro- berghi & Tomasello 1998).
blems based on evolved, domain-specific expectations A few years ago, however, Hare et al. (2000; 2001)
about what perceptual features are likely to be most reported “breakthrough” evidence that chimpanzees do,
salient in a given context and a general ability to reason in fact, reason about certain psychological states in their
about the causal relation between observable contingen- conspecifics (see, particularly, Tomasello et al. 2003a;
cies in a flexible, goal-directed but task-specific fashion 2003b). And since then, there have been a flurry of
(see also Penn & Povinelli 2007a). similar claims on behalf of corvids and monkeys based
on similar protocols (Bugnyar & Heinrich 2005; 2006;
Dally et al. 2006; Emery & Clayton 2001; in press; Flom-
8. Theory of mind baum & Santos 2005; Santos et al. 2006). Because Povi-
nelli and colleagues have provided detailed critiques of
Nonhuman animals certainly manifest many sophisticated Hare et al.’s (2000; 2001) protocol and results elsewhere
social-cognitive abilities. But having a theory of mind (see Penn & Povinelli 2007b; Povinelli 2004; Povinelli &
(ToM) sensu Premack and Woodruff (1978) means Vonk 2003; 2004), here we will focus on the best available
been proposed to date – ToM and language – do not do a Dennett 1996). Gentner and colleagues, for example,
good job of accounting for the comparative evidence. have shown that relational labels play an instrumental
role in facilitating young human learners’ sensitivity to
9.2. The ToM hypothesis relational similarities and potential analogies (Gentner &
Rattermann 1991; Loewenstein & Gentner 2005). Our
A number of comparative researchers believe that the dis-
ability to reason about large quantities of countable
continuity between human and nonhuman minds can be
objects in a generative and systematic fashion seems to
traced back to some limitation in nonhuman animals’
require the acquisition of numeric symbols and a linguistic
social-cognitive abilities (e.g., Cheney & Seyfarth 1998; counting system (Bloom & Wynn 1997). Numerous
Terrace 2005a; Tomasello et al. 2005). Although we cer-
studies have shown that subjects with language impair-
tainly agree that nonhuman animals do not appear to
ment exhibit a variety of cognitive deficits (e.g., Baldo
possess anything remotely resembling a ToM, the hypoth- et al. 2005) and that deaf children from hearing families
esis that some aspect of our ToM alone is responsible for
(i.e., “late signers”) show persistent deficits in ToM tasks
the disparity between human and nonhuman cognition
(see Siegal et al. 2001 for a review). Furthermore, there
seems difficult to sustain. For example, it is very hard to is good evidence that a child’s ability to pass certain
see how a discontinuity in social-cognitive abilities alone
kinds of ToM tests is intricately tied to the acquisition of
could explain the profound differences between human
specific sentential structures (de Villiers 2000). Normal
and nonhuman animals’ abilities to reason about causal
human cognition clearly depends on normal linguistic
relations in the physical world or nonhuman animals’ capabilities.
inability to reason about higher-order spatial relations.
But although natural language clearly subserves and cat-
Even Tomasello and his colleagues have admitted that
alyzes normal human cognition, there is compelling evi-
trying to explain all the differences between human and dence that the human mind is distinctively human even
nonhuman cognition in terms of a difference in ToM
in the absence of normal natural language sentences (see
skills is “highly speculative” at best (Tomasello & Call
Bloom 2000; Garfield et al. 2001; Siegal et al. 2001).
1997, p. 418). Indeed, in a different context, Tomasello Varley and Siegal (2000), for example, studied the
has himself argued (e.g., Tomasello 2000) that human
higher-order reasoning abilities of an agrammatic
language learners rely on cognitive capacities – such as ana-
aphasic man who was incapable of producing or compre-
logical reasoning and abstract rule learning – that are inde- hending sentences and whose vocabulary was essentially
pendent from ToM and absent in nonhuman animals. So
limited to perceptual nouns. In particular, he had lost all
while our ability to participate in collaborative activities
his vocabulary for mentalistic entities such as “beliefs”
and to take each others’ mental states into account may
and “wants.” Yet this patient continued to take care of
be a distinctive feature of the human lineage, it is clearly the family finances and passed a battery of causal reason-
not the only or even the most basic one.
ing and ToM tests (see also Varley et al. 2001; 2005).
Although late-signing deaf children’s cognitive abilities
9.3. The language-only hypothesis may not be “normal,” they nevertheless manifest gramma-
The oldest and still most popular explanation for the wide- tical, logical, and causal reasoning abilities far beyond
ranging disparity between human and nonhuman animals’ those of any nonhuman subject (Peterson & Siegal
cognitive abilities is language (for recent examples of this 2000). And the many remarkable cases of congenitally deaf
venerable argument, see Bermudez 2003; Carruthers children spontaneously “inventing” gestural languages
2002; Clark 2006). Dennett (1996, p. 17) described the with hierarchical and compositional structure provide
extreme version of this hypothesis in characteristically further confirmation that the human mind is indomitably
pithy terms: “Perhaps the kind of mind you get when human even in the absence of normal linguistic encultura-
you add language to it is so different from the kind of tion (see, e.g., Goldin-Meadow 2003; Sandler et al. 2005;
mind you can have without language that calling them Senghas et al. 2004).
both minds is a mistake.” Of course, the process of learning a language may
To be sure, language clearly plays an enormous and “rewire” the human brain in ways that make certain
crucial role in subserving the differences between kinds of cognition possible that would not be possible
human and nonhuman cognition. But we believe that otherwise, even if the subject subsequently loses the
language alone is not sufficient to account for the discon- ability to use language later in life. But this ontogenetic
tinuity between human and nonhuman minds. In order version of the “rewiring hypothesis” (Bermudez 2005)
to make our case, we need to distinguish between three begs the question of what allows language to so profoundly
distinct versions of the language-only hypothesis: (1) that rewire the human mind, but no other.
verbalized (or imaged) natural language sentences are Over the last 35 years, comparative researchers have
responsible for the disparity between human and nonhu- invested considerable effort in teaching nonhuman
man cognition; (2) that some aspect of our internal animals of a variety of taxa to use and/or comprehend
“language faculty” is responsible for the disparity; and language-like symbol systems. Many of these animals
(3) that the communicative and/or cognitive function of have experienced protracted periods of enculturation
language served as the prime mover in the evolution of that rival those of modern (coddled) human children.
the uniquely human features of the human mind. The stars of these animal language projects have indeed
been able to approximate certain superficial aspects of
9.3.1. Are natural language sentences what makes the human language, including the ability to associate arbi-
human mind human? Natural language tokens clearly trary sounds, tokens, and gestures with external objects,
play an enormous role in “extending” and even in “rewir- properties, and actions and a rudimentary sensitivity to
ing” the human mind (Bermudez 2005; Clark 2006; the order in which these “symbols” appear when
Finally, while LF representations may very well be of the language faculty narrowly construed. Our coevolu-
necessary to reason about certain kinds of higher-order tionary story does not make language an exaptation
relations, particularly those involving linguistically (cf. Hauser et al. 2002a); nor does it make our prelinguistic
mediated representations (Bermudez 2003; 2005), there relational capabilities a “pre-adaptation” for language
is little reason to believe that LF representations are (cf. Christiansen & Kirby 2003); nor does it deny the enor-
necessary in order to reason about any sort of higher- mous evolutionary importance that language has had in
order relations at all. Indeed, the available evidence “rewiring” the human mind (cf. Bermudez 2005). We
suggests otherwise. We noted above that humans are simply hypothesizing that the communicative function
without any appreciable grammatical or linguistic ability of language may have been just one among a number of
are nevertheless often able to reason quasi-normally factors that pushed the cognitive architecture of our
about higher-order causal relations and mental states species in a relational direction.
(e.g., Siegal et al. 2001). Conversely, subjects who suffer In any case, regardless of which factors most strongly
from frontal forms of frontotemporal dementia show contributed to the unique evolution of the human brain,
selective impairment in the ability to integrate higher- language alone is no longer directly and entirely respon-
order visuospatial relations (Waltz et al. 1999) and to sible for the functional discontinuity between extant
pass ToM tests (Gregory et al. 2002) even when their lin- human and nonhuman minds.
guistic abilities are still largely normal (see also Blair et al.
2007). This double dissociation suggests that the ability to
reason about higher-order relations is not entirely encap- 10. On the proper treatment of symbols in a
sulated within our language faculty narrowly construed nonhuman cognitive architecture
(i.e., FLN).
The crux of the matter, then, is to identify the specific
9.3.3. Did the adaptive functions of language drive the changes to the hominoid cognitive architecture that
evolution of the human mind? The third – and, we enabled Homo sapiens sapiens to reason about higher-
believe, most plausible – version of the language-only order relations in a structurally systematic and inferentially
hypothesis is that the communicative and/or cognitive productive fashion, and ultimately resulted in the evol-
functions of language played an instrumental role in the ution of our unique linguistic, mentalistic, logical, and
evolution of the human brain. Learning a language causal reasoning abilities. Behavioral evidence from
seems to require the ability to cognize higher-order extant animal species alone cannot tell us what changed
abstract relations in a systematic, generative, and struc- in the neural architecture of the human brain since the
tural fashion. And it seems indisputable – at least to split from our nonhuman ancestors. But when that evi-
us – that the language faculty, broadly construed, is the dence is combined with recent advances in computational
product of extensive evolutionary tinkering (Pinker & models of biological cognition, it becomes possible to
Bloom 1990; Pinker & Jackendoff 2005). So it is possible sketch a fairly detailed representational-level specification
that our ability to reason about higher-order relations of the kind of changes we should be looking for.
evolved first in order to accommodate the requirements
of language, and then was co-opted, exported, and/or
10.1. The PSS hypothesis
duplicated for other purposes in nonlinguistic domains.
But there are good reasons, we would argue, to favor a The classical school of thought in cognitive psychology
more complex, coevolutionary relationship between human has insisted for more than three decades that both
thought and human language (see also Bloom 1994; 2000; human and nonhuman minds are the product of a phys-
Bloom & Keil 2001). While the advantages of symbolic com- ical symbol system (Fodor 1975; 1997; Fodor &
munication are enormous, the adaptive advantages of being McLaughlin 1990; Fodor & Pylyshyn 1988; Newell
able to reason in a relational fashion have a certain primacy 1980; 1990; Newell & Simon 1976; Pinker & Prince
over the communicative function of language. It is quite dif- 1988). According to the now familiar tenets of the phys-
ficult to imagine how communicating in hierarchically struc- ical symbol system (PSS) hypothesis, mental represen-
tured sentences would be of any use without the ability to tations are composed of discrete, symbolic tokens,
entertain hierarchically structured thoughts. But it is quite which can be combined into complex representations
easy to imagine how the ability to reason about higher- by forming syntactically structured relations of various
order relations – particularly causal and mentalistic rela- types. Cognitive processes, according to the classical
tions – might be highly adaptive without the ability to com- view, are rule-governed algorithms that operate over
municate those thoughts to anyone else. If one is a tool-using the formal structure of these mental representations in
bipedal ape in a rapidly-changing environment surrounded a truth-preserving fashion. The classic defense of the
by ambitious and conniving conspecifics, the evolutionary PSS hypothesis is that it provides a computational
advantages of reasoning about higher-order relations go far account for several of the most spectacular aspects of
beyond the ability to communicate hierarchical thoughts to human thought, including our abilities to generalize
those conspecifics. rule-like relations over abstract categorical variables, to
Over the course of this target article, we have argued reason in an inferentially coherent fashion, and to use
that our ability to reason about higher-order relations sub- the artificial symbols of a natural language in a systematic,
serves a wide variety of distinctively human capabilities. It recursive, and generative manner (Fodor & Pylyshyn
seems possible that the adaptive advantages of one or 1988; Marcus 2001; Newell 1980; Pinker & Prince 1988).
more of these capabilities might have played a critical The PSS hypothesis is certainly not the “only game in
role in pushing the human brain in a relational direction town” (Fodor 1975). Indeed, the PSS hypothesis has
either in conjunction with, or even prior to, the evolution been roundly criticized for a variety of reasons, ranging
simpler representations can be used to represent different functionally discrete, particular-involving, and syntacti-
states of affairs in a combinatorial fashion. Few dispute the cally structured without being concatenatively compo-
fact that human thought can approximate the functional sitional in the sense postulated by the PSS hypothesis
effects of compositionality (cf. Prinz & Clark 2004). The (e.g., Plate 1991; Pollack 1990; Smolensky 1990; van
purported compositionality of nonhuman animals’ Gelder 1990; Wilson et al. 2001). Many of these proposals
mental representations, on the other hand, has been the can account for the kind of compositionality manifested by
object of innumerable, hard-fought battles, particularly nonhuman animals. Therefore, none of the comparative
between the “associationist” and “symbolic” theoretical evidence available to date warrants the widespread
camps that have dominated comparative debate for assumption among comparative cognitive researchers
many decades. In our view, the comparative evidence (e.g., Gallistel 2006) that nonhuman animals necessarily
accumulated over the past quarter-century comes down form compositional representations in the concatenative
firmly in favor of neither of these venerable theoretical fashion proposed by the PSS hypothesis.
alternatives. Instead, the available comparative evidence Indeed, as we will see below, nonhuman animals do not
suggests that compositionality is a ubiquitous feature of even come close to approximating any of the other, more
animal cognition, albeit not necessarily the kind of compo- distinctive, higher-order features of a PSS. Therefore, at
sitionality posited by the PSS hypothesis or “symbolic” least in biological organisms, the various representational
accounts of nonhuman cognition. capabilities putatively associated with a PSS are not a
The PSS hypothesis argues not only that mental rep- package deal as a matter of nomological necessity (see
resentations are compositional but also that they are com- Hadley 1997).
positional in a specific fashion: Complex compositional
representations in a PSS are formed by concatenation,
10.4. Types and tokens
thereby retaining the identity of the original constituents,
rather than by some other conjunctive mechanism that The distinction between types (e.g., kinds, classes, roles,
sacrifices the integrity of the original constituents (see dis- variables) and tokens (e.g., individuals, instances, fillers,
cussions by Aydede 1997; Horgan & Tienson 1996; van values) is one of the essential characteristics of a genuine
Gelder 1990). Van Gelder (1990) has suggested that any PSS. A PSS maintains explicit information about the syntac-
cognitive system should be considered “functionally com- tic type identity of each structural relation it employs and
positional” if it possesses generally reliable and effective the type identity of its allowable constituents as distinct
mechanisms for (1) producing a complex representation from the constituents involved in any particular relational
given its constituents, and (2) decomposing a complex rep- instance. For example, in a PSS, the abstract characteristics
resentation back into those constituents – regardless of of the loves relation is explicitly represented and invariant to
whether these complex representations are formed by whether John loves Mary or Mary loves John.
concatenation or by some other means. The ability to reason about the relation between types
As we see it, the comparative evidence leaves no doubt and tokens pervades many aspects of human thought.
that the nonhuman mind employs enduring, functionally Role-governed rules appear to be a formative feature of
discrete, particular-involving mental representations that all human languages; and the universal ability of humans
are at least functionally compositional in van Gelder’s to learn novel role-governed rules is evident not only in
agnostic sense. As Horgan and Tienson (1996) point out, their mastery of natural human languages but also in
the complexity of social relationships among nonhuman their ability to extract the abstract rules of artificial gram-
animals would be literally unthinkable without the ability mars in AGL experiments. “Role-governed categories”
to represent novel dyadic relations by combining discrete (Markman & Stilwell 2001) also play a central role in
representations associated with each individual in a combi- human concept formation far beyond the abstract gram-
natorial fashion. More generally, the well-documented matical structures of language. A human subject is per-
ability of nonhuman animals to keep track of means-ends fectly capable of reasoning about a role-based category
contingencies and predicate-argument relationships in a such as “lovers” or “mothers” or “tools” without there
combinatorial fashion implies that they possess some gen- being any set of perceptual features that all lovers,
erally reliable and productive mechanism for encoding the mothers, or tools have in common. Moreover, the ubiqui-
relation between particular constituents. Such a mechan- tous human capacity to find analogical correspondences
ism is necessary in order to ensure that when multiple between perceptually disparate relations appears to
relations predicate the same property, the fact that it is require an ability to find systematic correspondences
the same property in each case is somehow manifest in between the roles defined for those relations as distinct
the structural similarity between the representations. from the perceptual similarity between the fillers of
Horgan and Tienson (1996) argue that this is all it these roles (Gentner 1983; Gentner & Markman 1997;
should take in order for a representational system to Markman & Gentner 1993; 2000). Therefore, analogical
qualify as “syntactically structured”; and therefore one inferences – one of the hallmarks of the human mind
must conclude, they argue, that nonhuman animals and a prominent feature of abstract causal reasoning and
employ syntactically structured mental representations, ToM – seem to require the ability to distinguish
albeit not necessarily in the concatenative sense posited between roles and their fillers and to dynamically “bind”
by the PSS hypothesis. one with the other without corrupting the independence
We agree. And this conclusion rules out most traditional of either. Notably, the ability to maintain role-filler inde-
associative and distributed connectionist models as plaus- pendence while dynamically binding roles and fillers to
ible accounts of the nonhuman mind (see again Marcus particular relations seems to require the kind of concate-
2001). However, any number of researchers have pro- native compositionality posited by the PSS hypothesis
posed nonclassical connectionist architectures that are (for more extensive discussions of this point, see Doumas
animal capable of thinking the thought dominates(A, B) is van Gelder 1995; 1998). But the comparative evidence
likely to be able to think the thought dominates(B, A) for definitively rules out any nonrepresentational, purely
any arbitrary pair of conspecifics of the appropriate age embodied, or traditional associative account of animal cog-
and gender. And this kind of systematicity is often cited nition; and it strongly suggests that nonhuman minds, like
by advocates of the classical school of thought to support human ones, are highly structured, information-proces-
extending the PSS hypothesis to nonhuman animals sing devices in a way that stomachs and Watt governors
(e.g., Carruthers 2004; Fodor & Pylyshyn 1988). But are not (cf. Clark 2001). Indeed, nonhuman minds
there is a fundamental difference between the kind of sys- approximate certain features of a PSS that are extremely
tematicity manifested by nonhuman animals and the kind problematic for the kind of traditional distributed connec-
of systematicity posited by the PSS hypothesis. tionist systems that have been the principal antagonist to
The kind of systematicity manifested by nonhuman the PSS hypothesis for more than a quarter century
animals is limited to perceptually based relations in (Elman 1996; Hinton et al. 1986; Rumelhart & McClel-
which the values that each argument can take on in the land 1986). Whatever kind of architecture the nonhuman
relation are constrained only by observable features of mind employs, it is certainly not based solely on traditional
the constituents in question (e.g., the gender and age of distributed connectionist networks or associative learning.
the conspecific). Feature-based systematicity such as this The comparative evidence therefore leads us to propose
does not require the cognizer to posit relational roles dis- a hybrid alternative to the orthodox debate between clas-
tinct from the relations’ constituents, nor to cognize the sical and nonclassical theories of cognition: what we call
fact that certain relations logically imply certain other the relational reinterpretation (RR) hypothesis. Povinelli’s
relations. Not coincidentally, this is also the kind of sys- “reinterpretation hypothesis” previously suggested that
tematicity that happens to be easily implemented by humans alone are able to “reinterpret” the world in
many nonclassical connectionist models (e.g., Niklasson & terms of unobservable causal forces and mental states
van Gelder 1994). But as Fodor and colleagues have repeat- (e.g., Povinelli 2000; Povinelli et al. 2000). According to
edly argued (e.g., Fodor 1997; Fodor & McLaughlin 1990), our relational reinterpretation hypothesis, the discontinu-
the kind of systematicity posited by the PSS hypothesis is ity between human and nonhuman minds extends much
not statistical or accidental or a by-product of domain- farther: to any cognitive capability that requires reinter-
specific adaptations; rather, it arises as a matter of nomolo- preting perceptual relations in terms of higher-order,
gical necessity from the fact that a PSS defines relations structural, role-governed relations.
structurally. Classical systematicity entails cognizing the According to the RR hypothesis, animals of many taxa
fact that certain relations necessarily imply other relations employ functionally compositional, particular-involving,
independently of any particular domain or learning syntactically structured mental representations about
context: for example, for all R, the relations R(a,b) and observable features, entities, and relations in the world
R(b,c) necessarily imply the relation R(a,c) provided that around them. Furthermore, they form abstract represen-
R is a transitive relation. tations about statistical regularities they perceive in the
There is no evidence for this kind of classical, inferential, behavior of certain classes of physical objects (e.g., obser-
role-governed, domain-independent systematicity among vable causal relations) and other animate agents (e.g.,
nonhuman animals. affiliative interactions) and are capable of using these rep-
resentations off-line to make decisions in a flexible,
reliable, and ecologically rational (i.e., adaptive) fashion.
11. The relational reinterpretation hypothesis Human animals alone, however, possess the additional
capability of reinterpreting these perceptually grounded
Here’s the pickle. On the one hand, despite its many flaws, representations in terms of higher-order, role-governed,
the PSS hypothesis lays out a package of representational inferentially systematic, explicitly structural relations – or,
capabilities that appear to be well – though imperfec- to be more precise, of approximating these higher-order
tly – approximated by normal human minds. On the features of a PSS, subject to the evolved, content-specific
other hand, whereas nonhuman minds approximate biases and processing capacity limitations of the human
some of these same capabilities to some degree, they do brain. Ex hypothesi, the discontinuity between the cogni-
so to a significantly lesser degree than human minds do, tive abilities of human and nonhuman animals – including
and in some cases, not at all. our unique linguistic, logical, mentalistic, cultural and
The comparative evidence therefore poses a serious causal reasoning abilities – largely results from the sub-
challenge to the classical version of the PSS hypothesis. stantial difference in degree to which human and nonhu-
All of the strongest empirical arguments for the PSS man minds are able to approximate the relational
hypothesis rest on representational capabilities that capabilities of a PSS.
appear to be largely absent from nonhuman species – for Our RR hypothesis bears more than a nominal resem-
example, inferential systematicity, types and tokens, con- blance to Karmiloff-Smith’s (1992) “representational
catenative compositionality, and explicitly hierarchical redescription” hypothesis and to the growing family of
relations. In short, the evidence for a classical PSS “dual-process” accounts of reasoning (Evans 2003). The
among infraverbal organisms is a lot less “secure” than case for two broad “systems” of reasoning within the
Fodor and Pylyshyn (1988) assumed. human mind is already well-founded on the basis of the
The comparative evidence poses an equally serious evidence from human cognitive behavior alone. Our
challenge for many prominent nonclassical theories of cog- review of the comparative evidence suggests that a dual-
nition. The most extreme critics of the classical school have process account is well-founded from a comparative
argued that one can do without the notion of “represen- perspective as well. And our version of the RR hypothesis
tation” and “computation” altogether (e.g., Brooks 1991; is indebted to Karmiloff-Smith’s (1992) earlier and
If LISA is broadly correct, the substantive difference we applaud the efforts of these researchers and believe that
between human and nonhuman brains will be found in they are already shedding new light on our species’ unique
the prefrontal cortices, and specifically in synchronized relational capacities, relatively little effort has been invested
activity among prefrontal neural populations that support in modeling the relational abilities of other cognitively soph-
working memory, as well as among neural populations in isticated animals. In our view, the entire field of cognitive
the frontal and posterior cortical areas (see Lu et al. science – not just our particular hypothesis – would
2006; Morrison et al. 2004; Robin & Holyoak 1995; benefit if more effort were focused on constructing biologi-
Waltz et al. 1999; 2004). Of course, we are not suggesting cally plausible, behaviorally accurate, computationally feas-
that temporal synchrony among prefrontal neural popu- ible models of the cognitive abilities of honeybees, corvids,
lations is the only possible neural-level explanation for and chimpanzees, in addition to the cognitive abilities of
the functional differences between human and nonhuman enculturated, language-wielding humans.5
relational cognition, nor that it provides a full explanation Fortunately, the fate of our RR hypothesis does not ride
(see, e.g., Jung & Haier 2007). We are simply suggesting on the success or failure of any particular computational
that computational models of biological cognition such as proposal. Our most important claim in this target article
LISA provide an important tool for comparative research- is simply that whatever “good trick” (Dennett 1996) was
ers wishing to formulate biologically plausible, represen- responsible for the advent of human beings’ ability to rein-
tational-level hypotheses concerning the similarities and terpret the world in a symbolic-relational fashion, it
differences between human and nonhuman minds. evolved in only one lineage – ours. Nonhuman animals
didn’t (and still don’t) get it.
11.2. Moving forward ACKNOWLEDGMENTS
Admittedly, our RR hypothesis has a number of substan- The authors would like to thank José Luis Bermúdez, Paul Bloom, Dedre
tial holes. With respect to the empirical evidence, we Gentner, Arthur Markman, and seven other anonymous reviewers for
their invaluable comments on earlier drafts. Thanks also to Aaron
have not directly addressed a number of important cogni-
Blaisdell, Patricia Cheng, Nicola Clayton, Leonidas A. A. Doumas,
tive domains. In some cases – for example, numeracy, Graeme Halford, John Hummel, Laurie Santos, and Michael
cooperation, and mental time travel – others have Waldmann for helpful suggestions and discussions along the way. This
already proposed analyses of the functional discontinuity work was supported in part by NIH Grant MH072613 to KJH and a
between human and nonhuman animals’ capabilities that Centennial Fellowship to DJP by the James S. McDonnell Foundation.
are largely consistent with the hypothesis defended in the
present article (see, e.g., Dehaene 1997; McElreath et al. NOTES
2003; Suddendorf & Corballis 2007a). In other cases – for 1. Let us be clear: All similarities and differences in biology are
example, empathy and metacognition – the discontinuity ultimately a matter of degree. Any apparent discontinuities
between human and nonhuman minds continues to be between living species belie the underlying continuity of the evol-
challenged (cf. Preston & de Waal 2002; Smith et al. utionary process and largely result from the fact that many, and
2003). We believe our analysis and hypothesis can (and often all, of the intermediate steps are no longer extant. In the
should) be extended to these latter domains as well. present article, our claim that there is a “discontinuity” between
Indeed, we believe that our RR hypothesis offers a power- human and nonhuman cognition is based on our claim that there
ful framework for explaining what all these disparate is a significant gap between the functional capabilities of the
human mind and those of all other extant species on the planet.
cases – from cooperation and mental time travel to
Our point, to cut to the chase, is that the functional discontinuity
numeracy and metacognition – have in common. But we between human and nonhuman minds is at least as great as the
acknowledge that we have not had the space to extend much more widely acknowledged discontinuity between human
our analysis to these other domains herein. and nonhuman forms of communication. But we do not doubt
With respect to our representational-level claims, we have that both evolved through standard evolutionary mechanisms.
not specified how our proposed symbolic-relational super- 2. It is important to distinguish between causes that are in
module combines inputs from such a motley collection of principle unobservable (such as gravity and mental states) and
perceptual and conceptual modules in a computationally causes that are temporarily absent or hidden in a particular
feasible fashion. Fodor (2000) has argued that this context. We do not doubt that nonhuman animals can
problem is unsolvable, and therefore that the human mind learn about the latter (e.g., Blaisdell et al. 2006; Call 2004;
Wasserman & Castro 2005). What we doubt is that nonhuman
cannot, in the end, be entirely computational. We do not animals can learn or reason about the former.
have a complete solution to Fodor’s challenge; but, like 3. Importantly, we are not claiming (cf. Emery & Clayton, in
many others, we do not believe it is in principle unsolvable press; Tomasello et al. 2003b) that corvids – or other nonhuman
(Barrett 2005; Carruthers 2005b; Pinker 2005). Hybrid sym- animals – are limited to reasoning about concrete, occurrent
bolic-connectionist architectures such as LISA provide one cues in the immediate environment. To the contrary, we believe
possible solution that Fodor has not considered. it is obvious that nonhuman animals are perfectly capable of
The most glaring weakness in our hypothesis is that we keeping track of past events, as well as forming general abstractions
have no complete, biologically plausible model of nonhu- about observed behavioral regularities, and that they can use these
man cognition to propose. Many who have adopted some multifarious representations in a flexible and adaptive fashion (see
form of the Symbolic Approximation hypothesis have again Penn & Povinelli 2007b; Povinelli & Vonk 2004). Our claim
is simply that nonhuman animals’ representations do not extend to
taken on the ambitious goal of trying to determine how higher-order relations involving mental states.
the unique symbolic capabilities of the human mind 4. It is important to note that LF representations are not the
might be implemented in a neurally plausible architecture same thing as a “language of thought” (LoT). LF representations,
(e.g., Holyoak 1991; Hummel & Holyoak 2003; Marcus Carruthers (2002) explains, contain lexical items drawn from the
2001; Plate 1991; Pollack 1990; Shastri & Ajjanagadde specific natural language spoken by the cognizer, whereas a LoT
1993; Smolensky 1990; 1999; Wilson et al. 2001). Although is purportedly independent of any particular natural language.
An amicus for the defense: Relational More importantly, perhaps, they also allow extension of the rep-
reasoning magnifies the behavioral resentation of one concept by virtue of its similarity to another.
Information about one domain that is connected to the relational
differences between humans and nonhumans match can be carried over to the other domain as an analogical
doi: 10.1017/S0140525X08003671 inference (Falkenhainer et al. 1989; Markman 1997; Spellman
& Holyoak 1996). Thus, analogies provide a powerful mechanism
Arthur B. Markman and C. Hunt Stilwell of representational change. The importance of analogy in scien-
Department of Psychology, University of Texas, Austin, TX 78712.
tific innovation, for example, has been well demonstrated, both
markman@psy.utexas.edu.
in historical and naturalistic settings. Gentner et al. (1997)
chstilwell@mail.utexas.edu
provide a detailed analysis of the importance of analogy in
http://www.psy.utexas.edu/psy/FACULTY/Markman/index.html
Kepler’s discovery of the elliptical motion of the planets,
showing it was his knowledge of light and magnetism that
Abstract: Relational representation abilities are a crucial cognitive suggested solutions to the problems that planetary motion pro-
difference between human and nonhuman animals. We argue that duced for seventeenth-century scientists. Similarly, Dunbar
relational reasoning and representation supports the development of (1997) found that molecular biologists frequently rely on analogi-
culture that increases in complexity. Thus, these abilities are a force cal mappings to problems with familiar solutions to produce
that magnifies the apparent difference in cognitive abilities between novel solutions. Finally, Christensen and Schunn (2007) find
humans and nonhumans.
that analogies are crucial for the development of innovative
Penn et al. demonstrate the many ways in which humans and non- ideas in engineering design teams.
human animals are behaviorally similar. As they point out, despite As argued by the target article, role-governed categories and
these similarities, there are significant differences between analogical reasoning are a result of straightforward differences
human and nonhuman animals outside the lab. Penn et al. argue in representational capacity between human and nonhuman
convincingly that nonhuman animals rely largely on situational animals. We suggest that these abilities serve to magnify the
cues (e.g., perceptual features and associations), whereas humans apparent cognitive differences between human and nonhuman
are able to reason using complex relational representations. We animals, because they are crucial for the development of cultural
expand on this point, arguing that the ability to understand and systems that increase in complexity across generations. Animals
reason with relational information magnifies the cognitive differ- who have only feature-based concepts have no way of escaping
ences between humans and nonhuman animals. In particular, our the attributes of existing objects to suggest similarities across
ability to form role-governed categories expands our conceptual items that are based on relational similarities or on the possibility
and linguistic repertoire, allowing us to transcend mere situational that two objects or individuals play the same relational role. Once
and perceptual cues and represent concepts as related to functions animals develop a representational capacity that allows them to
and goals (Markman & Stilwell 2001). In addition, our ability to con- represent that two items play the same relational role within a
struct analogies based on relational mappings between domains that relational system, it becomes possible to envision additional
are dissimilar on the surface makes representational change and objects that might also fulfill that relational role. Thus, if a rock
conceptual innovation possible. Finally, these differences may is used to break a nut, this rock becomes just one kind of
help to explain the uniquely human phenomenon of cumulative breaker. Finding other objects that could fill this same role
culture (Tomasello et al. 2005). Role-governed categories allow (perhaps more effectively) is crucial to the development of a
humans to posit the existence of objects that fill a particular rela- technology. Role-governed technologies are the ones that are
tional role regardless of the perceptual properties of that object. central to cultures that increase in their complexity (Tomasello
Human concepts can be loosely divided into three types: feature- 1999). Thus, role-governed categories are crucial prerequisites
based categories, which are represented as collections of features; to the development of human-like tool cultures.
relational categories, which represent a particular relational struc- This view helps to explain how the cognitive abilities of human
ture; and role-governed categories, which refer to items that play a and nonhuman animals could simultaneously appear to be very
particular role within a relational structure (Markman & Stilwell similar and very different. Small differences in representation
2001). Many nonhuman animals have feature-based categories. ability support large differences in the available knowledge
Feature-based categories require only representations of co-occur- base that humans and nonhuman animals have to reason with.
rences among features. Humans, however, are able to construct What this work does not explain is how the leap from feature-
relational categories. Some relational categories refer to particular based representations to relational representations is made.
important relationships in the world (e.g., kinship terms, which Future work in cognitive science must examine the important
specify relationships between people). In addition, verbs specify influence of language on the development of relational represen-
relationships among a set of items that are part of a sentence tations for insight into the development of relational systems
(Gentner & Kurtz 2005; McRae et al. 1997). Verbs are particularly (Gentner 2003).
interesting, because any given verb must be completed with a set of
objects that play different roles within the relational structure ACKNOWLEDGMENTS
named by the verb (Ferretti et al. 2001). This work was supported by AFOSR grant FA9550-06-1-0204 and NIMH
For example, in the sentence “The EMT treated the accident grant R01 MH0778, and a fellowship in the IC2 institute to the first
victim,” the EMT and accident victim play particular roles within author.
these scenes (Ferretti et al. 2001). These roles themselves can in
turn be named by role-governed categories, of which the typical
agents associated with a relation are members. So a (medical)
patient is someone who is the object of the relation X treats Putting Descartes before the horse (again!)
Y. Having relational and role-based concepts allows humans to
categorize entities based on goals and functions, and to associate doi: 10.1017/S0140525X08003683
entities that, based on surface features alone, would be con-
sidered very different. The role-governed category patient Brendan McGonigle1 and Margaret Chalmers
allows humans and hamsters to be part of the same category, pro- Psychology Department, School of Philosophy, Psychology and Language
vided they are both undergoing medical treatment. Sciences, University of Edinburgh, Edinburgh, EH8 9JZ, Scotland, United
Analogical mappings – relational mappings between two Kingdom.
domains – are ubiquitous in human reasoning. They allow the M.McGonigle@ed.ac.uk
detection of subtle relational similarities between domains. www.psy.ed.ac.uk
or even homogeneous. The RMTS task may be more com- 1965), a formal analogy is defined as an isomorphism
putationally complex than the S/D task for primates. And between systems of relations (e.g., the analogy between
the two-item RMTS task may require a greater sensitivity groups in algebra and topological manifolds in geometry).
to variability than the multi-item RMTS task does. These Sarah’s strategy for solving geometric “analogies” – equat-
are all plausible hypotheses that deserve further exper- ing number of featural changes – does not establish an
imental scrutiny. But the fact that only certain species isomorphism, and hence does not exemplify an analogy
possess the evolved heuristics and/or processing capacity under any established definition. If the term formal
necessary to solve two-item RMTS tasks does not imply analogy is now to be used to refer to relational tasks that
that the RMTS task requires subjects to reason about can be solved by comparing analog measures of variation,
higher-order structural relations. To make this latter then indeed Sarah is capable of solving “formal analo-
claim, Thompson & Flemming would need to show gies” – but so are many other species, including pigeons
that the RMTS task requires subjects to reason about (Cook & Wasserman, in press). Thompson & Flemming’s
higher-order relations in a structurally sensitive fashion. change in terminology simply shifts the semantics, not the
And this, they have not done. substance, of the debate.
The substantive debate is not about how to define the
term analogy but about whether or not there is a disconti-
R2.4. Analogical relations
nuity in the cognitive mechanisms that human and nonhu-
We defined “analogical reasoning” as the ability to draw man animals employ to make relational inferences.
inferences about a target domain based on systematic, struc- Thompson & Flemming propose that there is a disconti-
tural similarities between a source domain and the target nuity between the symbolic-relational abilities of apes and
domain. Importantly, the relevant similarities in analogical all other species (see also Thompson & Oden 2000). We
inferences are based on the roles various entities play in believe this “analogical ape” hypothesis fails twice: It
their respective relations, and on structural similarities severely underestimates the symbolic-relational abilities of
between the relations, rather than (and distinct from) per- other non-primate species (see, e.g., the commentaries by
ceptual similarities between the entities involved in the Herman et al. and Pepperberg). And it glosses over the
relations (Gentner 1983; Gentner & Markman 1997; fundamental, qualitative difference between the feature-
Holyoak & Thagard 1995). Like any other form of relational based strategy employed by Sarah and the non-domain-
reasoning, analogical inferences vary in their degree of specific, role-based analogies made universally by modern
abstraction, structural sophistication, and domain specificity. humans. Even Thompson & Flemming admit that the
Even 4-year-old children understand simple analogies invol- sole evidence of a nonhuman animal having solved a
ving familiar visuospatial relations – for example, “If a tree “material analogy” is Sarah’s unreplicated performance on
had a knee, where would it be?” (Gentner 1977). But it Experiment 3 reported by Gillan et al. (1981). As we
takes quite a bit of linguistic scaffolding, inter-domain pointed out in our target article, Sarah’s remarkable and
mapping, and content-specific enculturation to make sense unreplicated success in this experiment constitutes exceed-
out of Donald Rumsfeld’s assertion that installing democ- ingly thin support for the “analogical ape” hypothesis. (See
racy in Iraq is like teaching a child to “ride a bike” (Silver- our Appendix [sect. R7] for examples of experimental pro-
stein 2007). According to our RR hypothesis, reasoning tocols that could provide evidence for various kinds of ana-
about even the simplest, most modality-specific analogies logical reasoning in nonverbal subjects.)
is a human cognitive autapomorphy.
Many commentators agree with us that analogical
R2.5. Rules
reasoning is a distinctively human capability (e.g., Bermú-
dez, Gentner & Christie, Hallinan & Kuhlmeier, Tetzlaff & Carruthers are right to emphasize the fact
Markman & Stilwell). Thompson & Flemming, that rule learning (or, at least, rule-like learning) can be
however, argue that at least one chimpanzee, Sarah, is found among minds as distantly related to humans as
also capable of comprehending some analogies. We have those of honeybees and desert ants. We made the same
considerable sympathy with their point of view, as one of point in our target article. But we hypothesized that only
us (Holyoak) reached similar conclusions at one time humans possess the ability to learn rules that involve
(see Holyoak & Thagard 1995). However, more recent non-perceptual, structural relations among role-based
findings (i.e., Oden et al. 2001) have shown that Sarah’s variables. Tetzlaff & Carruthers provide no reason to
performance does not merit this conclusion. doubt this hypothesis.
Thompson & Flemming admit that Sarah’s perform- For example, the location of a particular object with
ance on Oden et al.’s (2001) replication does not qualify respect to specific landmarks is the epitome of a perceptual
as a “material analogy” and acknowledge that Sarah’s per- (i.e., spatial) relation between observable stimuli. There-
formance was functionally equivalent to the performance fore, the fact that honeybees’ path integration mechanisms
of other primates and birds on S/D tasks. Thompson & use the distance and angle between arbitrary landmarks is
Flemming nonetheless claim that Sarah’s performance evidence that they can represent these spatial relations in
counts as a “formal analogy,” and they find our own defi- a rule-like fashion; but it hardly counts as evidence that hon-
nition of analogy “overly exclusive.” eybees are able to reason in terms of “non-perceptually-
Researchers clearly use the term analogy to refer to a based information” as Tetzlaff & Carruthers claim.
wide variety of relational inferences (see, e.g., Gentner
& Christie, Halford et al., Herman et al., Lupyan,
R2.6. Higher-order spatial relations
Markman & Stilwell). But Thompson & Flemming’s
definition of a “formal analogy” is exceptionally idiosyn- Hallinan & Kuhlmeier do not challenge our claim that
cratic. In the philosophical literature (e.g., Hempel reasoning about higher-order spatial relations is a uniquely
McGonigle et al. (2003) provide no evidence that capuchin term as meaning “tool-specific.” We meant the term to
monkeys are able to recombine hierarchically organized refer to the domain of physical causal reasoning in
sequences in a systematic or generative fashion. For general, not tools in particular. Many cognitive psycholo-
example, there is no evidence that the monkeys would gists believe that human subjects reason about the physical
be able to switch from sorting on the basis of shape and world using formal and substantive assumptions such as
then size, to sorting on the basis of size and then shape, temporal priority, causal directionality, and Michottean
without learning the entire sequence over from scratch. perceptual causal principles that are specific to the
If it took human language learners thousands of trials to domain of physical causality, but not specific to tool use
acquire a single invariant sentence, human language per se (see Gopnik et al. 2004; Lagnado et al. 2005).
would be of little interest. There is abundant evidence that nonhuman animals
make many of the same causal assumptions as humans
(see Penn & Povinelli 2007a for a review). In our view,
R2.9. Causal relations the fact that a non-tool-using species such as rooks was
A key claim in our target article is that the ability to reason able to quickly master the initial version of Seed et al.’s
(2006) task is compelling evidence that at least some non-
about unobservable causal mechanisms is a uniquely
human capability (see also Penn & Povinelli 2007a; Povi- human animals are able to reason about novel tool-use
nelli 2000; Vonk & Povinelli 2006). We interpreted Seed tasks using knowledge and expectations that are specific
to physical causal relations but not to tool use per se (see
et al.’s (2006) results as further evidence for this hypoth-
also Santos et al. 2006).
esis. Emery & Clayton claim that we were guilty of “mis-
interpretations, absences, and misrepresentations” in our
portrayal of Seed et al.’s (2006) experiment. What is “at R2.10. Theory of mind
issue,” Emery & Clayton write, is the performance of a
In our target article, we criticized Dally et al.’s (2006)
single rook, Guillem, which passed the crucial transfer
experiment with scrub-jays as providing no new positive
test.
evidence for theory of mind (ToM) abilities. Emery &
We fail to see where the alleged “misrepresentations”
Clayton did not challenge our interpretation of Dally
are to be found. Indeed, our interpretation of Guillem’s
et al. (2006). Instead, they reasserted that scrub-jays are
singular behavior is the same as that proposed by the
capable of “experience projection” based on evidence
authors of the original paper:
reported by Emery and Clayton (2001).
Given that six of the seven rooks failed to transfer to Tubes C In these experiments, Emery & Clayton investigated
and D which had no visual features in common with the first the propensity of scrub-jays to re-cache food that they
task, it seems unlikely that they had an understanding of the had previously cached in front of a conspecific, and
unobservable causal properties of the task at their disposal. . . . found that scrub-jays only re-cached food when they had
The surprising performance of Guillem, who solved all four had prior experience stealing another bird’s caches.
tasks despite the lack of a constant arbitrary visual cue, deserves “This result raises the exciting possibility,” Emery (2004,
further attention. . . but the result of one bird among seven must p. 21) wrote, “that birds with pilfering experience can
be interpreted with caution. (Seed et al. 2006, p. 700)
project their own experience of being a thief onto the
We certainly agree that Guillem’s behavior deserves observing bird, and so counter what they would predict
“further attention,” but Tebbich et al. (2007) subsequently a thief would do in relation to their hidden food” (see
replicated the same task on seven new rooks and found also Emery & Clayton 2004b).
that only three out of seven passed the perceptual transfer As noted by Penn and Povinelli (2007b), this may be an
task, and none of them passed the crucial nonperceptual “exciting possibility”; but it is certainly not the only, or
transfer task. In the Abstract to this paper, Tebbich et al. even the most cogent, explanation. Unfortunately, the
(2007) write: existing evidence sheds almost no light on the internal
mental representations or cognitive processes being
We found no evidence compatible with the formation of a employed by the birds in question. For example, all of
mental representation of physical problems given that none the birds involved in this experiment had had previous
of these 3 birds passed the transfer tasks. This is not surprising experience being pilfered (see discussion in Emery &
given that there is no evidence to date that any tool-using
Clayton, in press). But Emery & Clayton do not
animal has a causal understanding of the trap-tube problem.
explain how scrub-jays could have the cognitive prowess
If anything, our interpretation of Seed et al.’s (2006) necessary to reason by analogy to their own subjective
results seems to be more generous than that of the original experience as pilferers but not have the cognitive where-
authors: Contra Tebbich et al. (2007), we posit that rooks withal to realize they should start caching once they have
as well as other nonhuman animals do, indeed, have a been victims of pilferage themselves. Indeed, Emery &
“mental representation of physical problems” and a Clayton do not explain why a species with the ability to
“causal understanding of the trap-tube problem” – albeit reason by analogy cannot understand, prior to pilfering
not one that involves unobservable causal mechanisms another’s cache, that caching food from potential compe-
(see again Penn & Povinelli 2007a; Povinelli 2000). titors might be a good idea.
Emery & Clayton also argue that the rooks’ perform- To make matters worse, the existing evidence has not
ance on the two-tube task could not be due to “domain- ruled out the obvious possibility that pilfering changes
specific expectations” because rooks do not use tools in the subjects’ motivation to cache their own food rather
the wild. Here, we suspect that both we and Emery & than change their cognitive understanding of the func-
Clayton were tripped up by the protean term “domain- tional value of caching per se. This is the point of our
specific.” Emery & Clayton interpreted our use of the analogy to redirected aggression in primates (see Penn &
test the capacity to reason about higher-order relations or nonhuman primates perform as well as 3-year-old children
relational roles in the fashion that we have posited is on some (but not other) tasks has little bearing on our
unique to modern humans. In their past research, claim that there is a fundamental discontinuity between
Halford and colleagues have employed a wide variety of human and nonhuman minds. The monumental fact of
protocols to test higher-order, role-based relational capa- the matter – a fact which Hallinan & Kuhlmeier do not
bilities in human subjects (e.g., Andrews & Halford deny – is that the ontogenetic trajectory of one particular
2002; Andrews et al. 2003; Halford 1984; Halford et al. primate species’ relational abilities distinguishes itself
2005; Halford & Busby 2007). Many of these protocols from that of all other extant species on the planet. As
can (and should) be adapted to probe the similarities Shatz points out, by the second year of life, the cognitive
and differences between human and nonhuman LoTs differences between humans and other primates are
(see our Appendix for examples). unmistakable. And by age 5, the functional discontinuity
is so enormous that even the most generous comparative
psychologist cannot deny the disparity.
R3. Who gets to become human? Hallinan & Kuhlmeier end up proposing a theoretical
account for the disparity between human and nonhuman
As Shatz points out, the cognitive abilities of even the relational cognition that appears to be the same as our
most highly encephalized and enculturated nonhuman own. Citing Povinelli (2001), they postulate that humans
pale in comparison with the typical human child. Some possess “an additional system that sits side by side with
of our commentators, however, tried to use the ontogen- evolutionarily older systems” and that this additional
etic evidence against us (e.g., Hallinan & Kuhlmeier, system allows for analogical reasoning that is “not con-
McGonigle & Chalmers, Siegal & Varley, Wasser- strained by superficial or context-specific correspon-
man): If human infants start out with cognitive abilities dences.” To our ears, that sounds a lot like our
less sophisticated than that of some adult nonhuman hypothesis (see also Povinelli 2000; Povinelli & Bering
animals, they argued, how can we claim that there is an 2002; Povinelli et al. 2000). Like Hallinan & Kuhlmeier,
innate, genetically-prespecified “discontinuity” between we believe that our uniquely human system for higher-
human and nonhuman animals? Darwin, of course, order, role-based relational reasoning continues to interact
relied on a similar argument to bolster his case for the with cognitive systems that are evolutionarily more ancient
mental continuity between humans and other animals and that come on-line earlier in normal human ontogeny
(Darwin 1871, p. 84; cited approvingly by Wasserman). (the second ‘R’ in our RR hypothesis stands for “reinter-
Let’s take apart this venerable argument piece by piece. pretation,” not “replacement”). But Hallinan & Kuhlmeier
are mistaken, in our opinion, to believe that our “reinter-
R3.1. Nature and nurture (and more nature) pretation” hypothesis is inconsistent with the claim that
there is a fundamental discontinuity between human and
Many of our commentators (e.g., Hallinan & Kuhlmeier, nonhuman minds. Both we and Hallinan & Kuhlmeier
Lupyan, McGonigle & Chalmers, Siegal & Varley, postulate that there is an “additional system” responsible
Wasserman) assumed that because we postulated an for subserving our uniquely human ability to reason
“innate” or “genetic” basis for the discontinuity between about higher-order relations and that the emergence of
human and nonhuman cognition, we were necessarily this additional system is unique to the ontogeny of
denying the importance of ontogeny, environment, members of our species. Unless Hallinan & Kuhlmeier
language, enculturation, and everything else. For want to argue that the profound disparity between the cog-
example, Wasserman acknowledges that the neural nitive ontogenies of human and nonhuman primates is
systems of humans may differ from those of nonhumans solely the result of environmental factors, there must be
but asks, “Do these systems merely mature as the child something unique about the potential of the human
approaches adulthood? Or must these systems be carefully mental architecture from day one.
cultivated by enriching experiences to fully flower?”
Wasserman’s rhetorical question poses a false
dilemma. There is no either/or when it comes to nature R3.3. Constructing the human mind
and nurture. No biological system, least of all a neural Lupyan seems to believe that we consider the human
one, “merely” matures on its own. The ontogeny of any mind to be “innately symbolic and propositional.” Our
biological system is substantially modulated by its environ- RR hypothesis could not be farther from this strawman.
ment. But this does not mean that genetic factors play no We explicitly denounced the classical view of the mind
role in shaping an organism’s ontogeny. There is a as biologically implausible and functionally impoverished.
complex, nonlinear, epigenetic relationship between To argue that humans and nonhumans differ in their
genes and the environment that plays out over the entire potential for symbolic-relational cognition from con-
lifespan of an organism – even an enculturated organism. ception forward does not entail – or even suggest – that
the human cognitive architecture is born with its adult-
state symbolic-relational abilities all wired up and ready
R3.2. Does primate phylogeny recapitulate human
to go. Our claim is that the human genotype has the
ontogeny?
unique potential to produce a neural architecture
Hallinan & Kuhlmeier argue that there would be a true capable of higher-order relational reasoning. Without the
cognitive discontinuity between human and nonhuman appropriate internal and external inputs, however, this
minds only if the behavior evident in the first stages of genetic potential is sure to be thwarted.
human development looked strikingly different from the Lupyan goes on to point out that some of the authors
capacities we see in other species. But the fact that we cited on the subject of human language learning do
computational architecture do they have? There is a con- language “rewired” the human mind (see also Bickerton)
spicuous dearth of biologically plausible, computationally leaves most of the interesting representational-level ques-
feasible, behaviorally adequate answers to this question. tions unanswered as well. It is one thing to identify the
Indeed, there are so few researchers willing to even ask functional characteristics of the discontinuity between
this daunting question that we happily accord two kudos human and nonhuman cognition. It is quite another to
to Tomlinson & Love just for showing up and making explain how the functional abilities specific to human cog-
the effort. nition are implemented in the neural matter of the human
This said, the BRIDGES (Building Relations through brain.
Instance Driven Gradient Error Shifting) model touted Representational-level computational models such as
by Tomlinson & Love begs the question at issue in our LISA (see also Tomlinson & Love) have an invaluable
target article. The BRIDGES model solves S/D and but undervalued role to play in cognitive science. It is all
RMTS tasks by combining exemplar-based category learn- too common for psychologists and philosophers to create
ing (what we call perceptual relational learning) with high-level models of a given cognitive behavior without
structured relational mapping (which we claim is unique giving due consideration to whether such models are com-
to humans). No one doubts, of course, that S/D and putationally feasible or biologically plausible. Although
RMTS tasks can be solved by structured relational map- there are clearly multiple distinct “levels” of explanation
ping – human subjects may very well solve RMTS tasks in cognitive science, even Marr (1982) did not counte-
in this manner under certain conditions. But the issue at nance ignoring all but the highest level of analysis.
stake in our target article is whether or not there is any evi- Working implementations of a cognitive capability have
dence that other extant species employ this particular the potential to challenge or support the plausibility and
mechanism as well. coherence of higher-level specifications, to provide new
Tomlinson & Love point out that an explanation based insights into the operational characteristics of that cogni-
on sensitivity to categorical entropy alone does not explain tive capability, and to serve as models bridging the
the degree to which pigeons are influenced by the featural (often quite large) gap between functional-level and
similarity between the test array and previous arrays the neural-level descriptions.
animal has been trained with. We agree. Categorical The bulk of our target article focused on identifying the
entropy is certainly inadequate to account for all of the functional characteristics of the discontinuity between
patterns of relational responding manifested by pigeons human and nonhuman cognition. Developers of
or any other animal (as Cook and Wasserman [2006] them- symbolic-connectionist computational models such as
selves point out). But all of the additional influences on LISA are trying to understand what kind of rewiring
pigeons’ relational responses, including those documented changes are necessary in order to subserve the higher-
by Gibson and Wasserman (2004), are further examples of order relational capabilities that both we and Bermúdez
feature-based relations, not the higher-order structural believe are unique to the human mind – including those
relations we have argued are unique to humans. And Tom- that are necessary for language itself. LISA, in particular,
linson & Love give no reason to believe that pigeons or any provides one possible example of how the higher-order
other nonhuman animals employ higher-order mappings relational capabilities of the human mind might be
between structured relations in order to solve S/D or implemented on top of the lower-order, perceptually
RMTS tasks. grounded capabilities of the nonhuman mind. At the
very least, then, LISA provides some confirmation that
our RR hypothesis is neither computationally infeasible
R5.3. Does LISA earn its explanatory keep?
nor neurally implausible.
We discussed the LISA (Learning and Inference with But LISA’s explanatory neck is stuck out a good deal
Schemas and Analogies) model of analogical reasoning farther. If LISA is correct, the substantive discontinuity
proposed by Hummel and Holyoak (1997; 2003) as one between human and nonhuman cognition came about
possible representational-level model for how higher- because only the hominid lineage evolved the ability to
order relational reasoning might be implemented in a neu- use synchronized activity among prefrontal neural popu-
rally plausible architecture. But Tetzlaff & Carruthers lations to support dynamic-binding among roles, fillers,
are not unjustified to point out LISA’s numerous limit- and structured relations. Although neural synchrony is
ations. To put it bluntly, LISA is a rudimentary, highly sty- used by many species for coding contextual associations
lized model of analogical reasoning that accounts for only a of various sorts (see Fries et al. 2007), LISA suggests
small part of what makes human cognition human that co-opting this mechanism for role-based relational
(although it is getting better; see Doumas et al. 2008). In coding was responsible for the “Great Move” (Newell
our view, LISA is the worst model of higher-order reason- 1990) in human cognition. Certainly, neural synchrony is
ing currently on offer, except for all the others. If Tetzlaff not the only possible mechanism by which the human
& Carruthers have a better model to suggest, we are all brain might approximate the higher-order properties of a
ears. PSS (for other possibilities, see Smolensky 1999; Wilson
Bermúdez largely concurs with our analysis of the dis- et al. 2001a). And the hypothesis that some form of
continuity between human and nonhuman minds but neural synchrony is the critical innovation subserving
argues that the LISA model simply “appears to recapitu- higher-order human cognition requires much further
late” our functional-level description and does not empirical support before it can be deemed anything
“explain” how this discontinuity evolved in the first more than a plausible possibility (but see Uhlhaas &
place. It is true that we did not provide an “explanation” Singer 2006 for a start). Nevertheless, Bermúdez is
for how higher-order relational reasoning evolved in the surely mistaken to argue that LISA is merely a “redescrip-
human brain; but simply invoking a story about how tion” of the functional-level facts.
anthropomorphic than Darwin’s. Burghardt’s commen- recognize relational mappings – that is, subjects begin
tary confirms those suspicions: “Research over the last by recognizing highly similar examples of a relation and
40 years,” Burghardt writes, “has shown that Darwin actu- are progressively encouraged to compare more and more
ally underestimated the mentality of apes, for example, in disparate instances. Then, on each test trial, role-based
tool making, numerosity, and communication.” Indeed, in and perceptual similarity are pitted against one another.
many respects, the hypothesis we proposed in our target Train a dolphin to mimic the relational action demon-
article lies closer to Darwin’s views on the matter than to strated by a trainer as closely as possible given the set of
those of many of our contemporaries. Darwin (1871) at objects available to the dolphin in the pool. In the begin-
least acknowledged the “immense” and “enormous” differ- ning, the objects in the pool should allow the dolphin to
ence between “the lowest savages” and even the “most mimic the trainer’s actions quite closely: for example, the
highly organised ape” – whereas many comparative trainer touches a stick to a Frisbee, and the dolphin has
researchers today believe that a human child magically available a different but identical stick and a different
kept alive alone on a desert island would “not differ very but identical Frisbee. Then, the dolphin can be trained
much” from other great apes (Tomasello & Rakoczy on more and more challenging problems by using
2003; see also Gardner, Lupyan, Wasserman). And objects that can no longer serve, literally, to imitate the
unlike some researchers (e.g., Bickerton) who believe trainer’s actions. For example, when the trainer touches
that language alone can explain what is distinctive about a stick to a Frisbee, there is only a stick and a ball in the
the human mind, Darwin argued – as we do – that “the pool; when the trainer puts a ball in a box, there is only
mental powers of some early progenitor of man must a Frisbee and a basket in the pool.
have been more highly developed than in any existing Once the dolphin has learned to mimic the relevant
ape, before even the most imperfect form of speech relations using perceptually disparate objects, the
could have come into use” (Darwin 1871, p. 57). dolphin can be tested on tasks in which perceptually
In short, any differences we may have with Darwin con- similar objects play conflicting roles. For example, the
cerning the cognitive limitations of nonhuman animals trainer could put a ball in a basket; and the dolphin
pale in comparison to the differences we have with most could be given a small basket, a larger box, and an even
of our contemporaries in comparative psychology. And larger ball (where the ball is too large to go in the box
any mistakes Darwin may have made over the course of and the box is too large to go in the basket, so that the “ana-
his career seem trivial when weighed against the monu- logous” solution is for the dolphin to put the basket in the
mental insights he provided into the evolution of life in box). Or the trainer might blow a ping-pong ball through a
general and the origins of the human mind in particular. cylinder with her mouth; and the dolphin could be given
In hindsight, we erred: “Darwin’s Triumph” makes a an identical cylinder, a hoop, and a ball (where the ball
better title. is too large to pass through either the hoop or the cylinder,
so that the analogous solution is for the dolphin to push or
blow the cylinder through the hoop with its mouth).
R7. Appendix: Falsifying the relational Of course, success on any one test trial is of little signifi-
reinterpretation hypothesis cance. Any given trial could be passed using some feature-
based heuristic, or indeed, simply by chance. In order to
Hereinbelow, we sketch examples of experimental proto- provide convincing evidence of analogical reasoning, sub-
cols capable of falsifying our functional-level claims for jects must demonstrate their ability to systematically
each of the distinctively human relational capabilities dis- mimic the trainer’s actions across a variety of visuospatial
cussed in our target article. relations.
salient perceptual relations between elements in a given protocol described by Bergman et al. (2003), present sub-
sequence. The same test can be readily adapted to nonver- jects not in the A, B, or C matrilines with unexpected
bal animals. rank reversals between the dominant members of the A
and B matrilines and the B and C matrilines (i.e., C . B
R7.4.2. Experiment 9: Structural rule learning using a and B . A but not C . A). Subjects that are capable of
contingency learning task. The following experiment is reasoning about social dominance relations in a hierarchical
adapted from a contingency learning task first employed fashion should be more “surprised” at hearing cx , ay than
by Shanks and Darby (1998) with human subjects. at hearing cx . ay for any arbitrary member of the two non-
During the initial training, subjects are presented with adjacent matrilines. Furthermore (contrary to the results
cues (e.g., lights, tones) and outcomes in the following reported by Bergman et al. [2003], the subjects should be
combinations: ABþ, A-, B-, CD-, Cþ, Dþ, Eþ, Fþ and more “surprised” by rank reversals between subordinates
GH- (where Xþ means that presentation of the cue X is in more distant matrilines (e.g., ax . cy) than by rank rever-
paired with a reward, and X- means that presentation of sals in more closely ranked matrilines (e.g., bx . cy).
the cue X is not paired with any reinforcer). At test, the
subjects are presented with the novel cue-outcome combi-
nations EF, G, and H. R7.6. Causal relations
In Shanks and Darby’s (1998) original experiment, R7.6.1. Experiment 12: Nonhuman primates’ under-
human subjects who learned the patterns shown during standing of weight. One of us (Povinelli) has already pro-
training anticipated that EF would not be paired with a posed numerous experiments capable of testing a
reward, whereas cues G and H presented separately nonhuman primate’s ability to reason about unobservable
would be paired with a reward. In other words, these sub- causal mechanisms (see Povinelli 2000). Here we suggest
jects learned the “rule” that the likelihood of the outcome one further experiment that would be probative.
after a compound of two cues is the inverse of the likeli- Nonhuman primates are quite familiar with the effort
hood of the outcome when those same cues are presented required to lift objects. Chimpanzees in particular have
separately. Shanks and Darby (1998) concluded, and we been shown to use the weight of an object instrumentally.
concur, that success on this protocol demonstrates that For example, free-ranging chimpanzees learn to use the
the subject can learn structural rules about contingencies weight of heavy stones to crack open hard nuts and
that are distinct from (and even contrary to) the outcomes lighter stones to crack open softer nuts, in a population-
predicted by associative conditioning. specific fashion. The following experiment tests whether
nonhuman primates actually reinterpret the effort
R7.5. Hierarchical relations
required to lift objects and the accompanying sensorimo-
tor cues into a causal notion of “weight” (for more
R7.5.1. Experiment 10: Hierarchically structured sen- details, see Povinelli, in press). First, present subjects
tences. Herman et al. (1984) report that bottlenosed with two balls that are visually identical but of radically dif-
dolphins comprehend sentential constructions such as fering weights (one is heavy, the other is very light). Train
“LEFT FRISBEE FETCH RIGHT HOOP,” which can the subjects to sort the balls into one of two containers
be glossed as “take the Frisbee to your left to the hoop based on the “weight” of the balls (e.g., “heavy” ball goes
located to your right.” To test whether or not dolphins in the container to the left, “light” ball goes in the con-
can understand hierarchically structured, recursive con- tainer on the right). Separately, allow the subjects to
structions, one could test their ability to comprehend sen- freely play with the ramp apparatus shown in Figure R1,
tential constructions of the following form: NP þ V þ NP, allowing them to launch balls of various weights down
where V is an action such as FETCH, and NP is a noun the ramp without the reward apple being present.
phrase that can be either an object (e.g., FRISBEE), or
a location modifier and an object (e.g., LEFT
FRISBEE), or another noun phrase followed by a location
modifier and an object (e.g., RIGHT FRISBEE LEFT
BALL). For example, the construction “RIGHT
FRISBEE LEFT BALL FETCH SPIGOT” instructs the
dolphin to take the ball that is to the left of the Frisbee
that is to the right of the subject over to the water
spigot. If dolphins – or any other nonhuman animal –
could comprehend and act on constructions such as
these in a systematic fashion, this would constitute defini-
tive evidence that they are able to comprehend hierarchi-
cally structured grammatical relations.