eScholarship

International Journal of Comparative Psychology

Title
Two Approaches to the Distinction between Cognition and ‘Mere Association’

Permalink
https://escholarship.org/uc/item/1x28x459

Journal
International Journal of Comparative Psychology, 24(4)

ISSN
0889-3675

Author
Buckner, Cameron

Publication Date
2011

DOI
10.46867/ijcp.2011.24.04.06

Copyright Information
Copyright 2011 by the author(s).This work is made available under the terms of a Creative
Commons Attribution License, available at https://creativecommons.org/licenses/by/4.0/

Peer reviewed

eScholarship.org Powered by the California Digital Library

University of California
International Journal of Comparative Psychology, 2011, 24, 314-348.
Copyright 2011 by the International Society for Comparative Psychology

Two Approaches to the Distinction between Cognition and

‘Mere Association’
Cameron Buckner
Indiana University, Bloomington, U.S.A.
The standard methodology of comparative psychology has long relied upon a distinction between
cognition and ‘mere association’; cognitive explanations of nonhuman animals behaviors are only
regarded as legitimate if associative explanations for these behaviors have been painstakingly ruled
out. Over the last ten years, however, a crisis has broken out over the distinction, with researchers
increasingly unsure how to apply it in practice. In particular, a recent generation of psychological
models appear to satisfy existing criteria for both cognition and association. Salvaging the standard
methodology of comparative psychology will thus require significant conceptual redeployment.
In this article, I trace the historical development of the distinction in comparative psychology,
distinguishing two styles of approach. The first style tries to make out the distinction in terms of the
properties of psychological models, for example by focusing on criteria like the presence of rules &
propositions vs. links & nodes. The second style of approach attempts to operationalize the
distinction by use of specific experimental tests for cognition performed on actual animals. I argue
that neither style of criteria is self-sufficient, and both must cooperate in an iterative empirical
investigation into the nature of animal minds if the distinction is to be reformed.

It is now commonplace in comparative psychology to find nonhuman

animal behaviors explained in terms of “cognitive” capacities such as episodic
memories, concepts, transitive orderings, and cognitive maps. When a comparative
psychologist attempts to justify a cognitive explanation of animal behavior,
however, it is widely-accepted that she must do more than merely show that the
behavior is consistent with the exercise of a cognitive capacity. Rather, both
proponents and skeptics of cognitive explanations agree that they can only be
regarded as legitimate if “simpler” associative explanations of the behavior have
been carefully considered.1 As such, rigorous experimental controls ruling out
“merely associative” explanations are routine and required component of any
experiment purporting to show evidence of animal cognition. I will call this
approach to comparative cognition research “Standard Practice.”
Standard Practice has recently been challenged by some of its most
prominent practitioners, who now argue that the distinction between cognition and
mere association on which it relies is “oversimplified,” “antiquated,” and even
1
Really, Standard Practice involves default concern for at least a “trichotomy,” with even “simpler”
processes like innate-releasing mechanisms placing even “lower” on the scale than mere association.
However, for the sake of simplicity in this paper I will speak as others have of a dichotomy, as the
distinction between such “innate” behaviors and learned ones merits its own extensive discussion.

Thanks go to Stanley Weiss and Aaron Blaisdell for organizing the focus session of the Winter
Conference on Animal Learning and Behavior where an earlier version of this paper was first
delivered. The paper also benefitted from feedback received with conference participants; special
thanks are due to Andy Baker for extensive comments on an earlier version of this draft. This
research was funded in part by the Center for the Integrative Study of Animal Behavior at Indiana
University. Author is now a Visiting Assistant Professor, Department of Philosophy, University of
Houston in Houston, TX, and Postdoctoral Researcher at Ruhr-Universität Bochum in Bochum,
Germany. Correspondence should be addressed to Cameron Buckner, University of Houston,
Department of Philosophy, 513 Agnes Arnold Hall, Houston, TX, 77204-3004, U.S.A.
([email protected]).
“specious and unproductive” (Allen, 2006; Papineau & Heyes, 2006; Penn &
Povinelli, 2007). Rather than presenting a systematic defense of Standard Practice
against these critics — a task I have attempted elsewhere (Buckner, 2011) — I will
here attempt to put this practice in historical context and provide a partial etiology
of the current crisis. In particular, I suggest that amongst comparative
psychologists there are actually two broad classes of approach to the distinction,
and that improper coordination between these two perspectives has led to
unnecessary disagreements and talking-past. On the one hand, there is a “model-
based” approach, which tries to make out the distinction in terms of two different
kinds of psychological model — for example, models constructed from rules and
propositions vs. those built from links and nodes. On the other hand, there is a
more behavior-based approach, which relies on tests of behavioral flexibility to
operationalize the distinction in experimental practice — for example, the “short-
cutting” test for cognitive mapping (O’Keefe & Nadel, 1978) or the 5-element
series test for transitive inference (Bryant & Trabasso, 1971).
In this paper, I will discuss the role each of these perspectives has played
in the history of the distinction, along the way drawing some general morals about
the proper relationship between the two. While both perspectives must play an
important role in the reform of Standard Practice, to do so each must be carefully
distinguished and put in its proper place.

Standard Practice: From the Ancient to the Current Crisis

As the empirical and philosophical contributions to the recent volume

Rational Animals? (Hurley & Nudds, 2006) demonstrate, a cloud of confusion now
obscures the distinction between “cognition” and “mere association,” with
researchers from different backgrounds bringing to the table very different
assumptions woven together from diverse historical threads. Over the centuries,
associationism has been variously conflated with empiricism, behaviorism, and
connectionism; and cognitivism has been conflated with rationalism, vitalism,
introspectionism, the computational theory of mind, and, when applied to
nonhuman animals, anthropomorphism. Any justification of Standard Practice will
thus require some redeployment, as no account of the distinction could — or
should — answer to all of these disparate uses.
The distinction we have today is the product of millennia of debates about
the nature of animal psychological capacities and their similarity to those found in
humans. These debates occurred in very different contexts, and as a result, our
modern concept of cognition straddles many disparate elements: the ability to use
logical rules, employ concepts or propositions, provide verbal justifications for
actions, infer causes and effects, select the optimal course of action, and many
others. Sorabji (1993) opines that this collection can seem “higgledy-piggledy” in
the absence of the historical milieu from which it emerged.
Current debates regarding the nature of animal cognition often center on
the question of whether animal behavior is best explained as a product of mere
perceptual association or some form of “cognitive” reasoning. Sorabji (1993)
argues persuasively that the basic shape of this dialectic coalesced long ago, when
Aristotle denied that animals could possess reason (logos) and intellect (nous).
- 315 -
Aristotelian psychology focused on a hierarchy of three faculties — nutrition,
perception (aisthesis), and intellect, with animals possessing only nutrition and
perception. According to Sorabji, Aristotle’s denial of intellect to animals2 made it
difficult for the Ancients to explain the intelligent flexibility apparent in animal
behavior — known to the ancients primarily through a series of famous anecdotes,
such as the tale of Chryssipus’ dog sniffing two roads at a three-way fork and then
straightaway taking the third. As a result, the content and inferential capabilities of
perception had to be correspondingly expanded.3 In Aristotle we find the roots of
the modern conceptual challenges, for we see perhaps the first concerted attempt to
make out a cluster of “middle ground” mental capacities which are powerful
enough to explain the flexibility of animal behavior without thereby granting
animals full rationality or reason.
Later commentators were greatly puzzled by the details of Aristotle’s
distinction between perception and intellect. On the one hand, Aristotle granted
animals sophisticated perceptual content which could be propositional in nature
(meaning, roughly, that their perceptual states can have the content that something
is the case and that these contents could be assessed in terms of truth or falsity). In
Nicomachean Ethics, for example, a lion is described as perceiving that the oxen is
near and rejoicing that he will get a meal (Aristotle, trans. 1925, 3.10 1118a20-23);
in De Anima, animals are credited as able to perceive that certain stimuli are
pleasant or unpleasant, as well as to make judgments based on assessment of the
“common” properties such as likeness and difference (Aristotle, trans. 1931, 3.2
426b12-427a14). On the other hand, he insisted that animals, lacking logos, could
not possess beliefs or reasoning (Aristotle, trans. 1931, 3.3, 428a19-24, 3.10,
433a12). Yet since in Aristotelian thought memory is derived from perception and
experience (empeiria) from memory (Aristotle, trans. 1928, 100a3-6), animals
must also possess sophisticated capacities to perceive sensory impressions over
time, compare those impressions to one another, remember and learn from those
experiences, and act accordingly. Notably, these faculties may provide animals
with all of the raw materials they need to form concepts (as specified in Aristotle’s
famous “battle metaphor,” a locus classicus of empiricist concept acquisition).4

2
For evidence of the denial, see Aristotle’s De Anima (trans. 1931): 1.2, 404b4-6; 2.3, 414b18-19
with 23-3 and 415a7-8; 3.3, 428a19-24; 3.10, 433a12; 3.11, 434a5-11.
3
For examples of the expanded content of perception, Aristotle allows sensory perception to report
the “common sensibles” such as “movement, rest, number, figure, and magnitude” in De Anima 2.6;
at 3.2 426b12-427a14, he suggests that assessments of sameness and difference are also assigned to
pure perception (trans. 1931). For a more complete case, see Sorabji (1993, 17-20).
4
“But when perception is present, in some animals the sense image (aisthima) comes to remain, and
in some not. Where it does not at all or does not for certain things, there and for those things there is
no awareness outside perception. But there is [such awareness] in those animals which can keep it in
their minds after perception. And when many such [sense images] have come [to remain], there
comes then to be a certain distinction, so that reason (logos) develops out of their remaining in the
case of some animals, and in the case of others not. Thus out of perception arises memory, so we say,
and out of memory often repeated of the same thing, experience (empeiria). For memories which are
many in number are a single case of experience. And from experience, or from the whole universal
come to rest in the mind, the one beside the many which is one and the same in all of them […] Thus
dispositions are neither present as determinate entities nor developed out of cognitively superior
dispositions, but out of perception. As when a rout occurs in a battle, if one man makes a stand, so
does another and then another, until it has got back to the start of the rout.” An. Post. 2.19 99b32-
100a6. Quoted with editorial insertions from Sorabji (1993, p. 33).
- 316 -
Having apparently given animals so much, these commentators wondered, how
could Aristotle consistently deny them full-fledged thought and reason?
The debate continued between those supporting and criticizing the
Aristotelian position; Epicureans and Stoics tended to side with Aristotle, while
Neoplatonists (most notably Porphyry) argued that at least some animals could
possess genuine reason. These debates touched on a wide variety of criteria, and
the conflation between “model-based” and “behavior-based” approaches to the
distinction can, I suggest, already be found there. On the more model-based side,
we find Aristotelian and Stoic suggestions that animals could not engage in reason
due to an inability to represent the propositional content of their mental states
linguistically (with the Stoics taking literally the Platonic suggestion that thought is
“the soul conversing with itself”). This was significant because it left animals
unable to place their reasoning in the form of a syllogism — at the time, the sole
device available for formalizing inferences. On the more behavior-based side,
proponents of animal reason retorted with anecdotes suggesting that animals can
respond successfully to commands, themselves use verbal commands or signs, and
exhibit evidence of deliberation and the ability to plan for future needs (based on
criteria like hesitating and dithering when making choices and preparing
environmental circumstances, such as lairs and refuges, for future needs).5 The
relationship between the two kinds of criteria was, then as now, hotly debated.
In the early modern period, the distinction reappeared in debates between
rationalists like Descartes and Leibniz and empiricists like Locke and Hume. In the
early modern discussion, however, there was less need for subtle criteria to
distinguish reason from association for two reasons. First, subtle empirical criteria
applicable to nonhuman animals were not needed because early modern empiricists
and rationalists largely agreed that animals lacked any form of reason. Second,
positions had hardened in the early modern debate; the issue became less a matter
of how to distinguish two capacities which both humans (and perhaps animals)
possessed, but rather a question of whether all thought was best conceived of as
associative or propositional in nature.6 This hardening illustrates a long-term
pattern of oscillation between “monistic” and “dichotomous” approaches to the
mind — with the more extreme swings often being driven by enthusiasm about
developments in other areas, such as the syllogisms, Pavlovian nerve theory, or the
advent of the digital computer.
Nevertheless, Hume’s (1748/1993) associationism is worth mention
because his famous “three principles of association” — similarity, contiguity, and
cause and effect (the latter based on nothing more than “habit” together with
constant conjunction and priority of cause to effect) — offered perhaps the first
concerted attempt to codify the mechanisms of associative thought (Section III).
Presaging behaviorism in both mechanism and method, Hume (1748/1993)
steadfastly defended the position that animals possessed no inferential capacities
beyond those provided by his three principles, as well as the optimistic appraisal

5
See Porphyry’s fascinating On Abstinence from Killing Animals (trans. 2000) for a summary of these
ancient debates.
6
Or, at least, nearly all thought — even Hume granted a limited role to reasoning in humans, to
puzzle out pure relations between ideas (as in the fields of geometry, algebra, and arithmetic) or
engage in practical means-end reasoning.
- 317 -
that these mechanisms were so powerful that appeal to more advanced forms of
learning or reasoning was not required to explain the full intelligent flexibility of
animal behavior (Section IX).7
Descartes (1637/1985) also denied any intellective faculties to animals,
infamously regarding them as mindless automata (Part 5 57-59); this position was,
in the history of Standard Practice, a double-edged sword. While Descartes’ early
accounts of the nervous system must be credited as foundational in the history of
neuroscience, leading the way towards mechanistic explanations of animal
behaviors, his identification of reason and mentality with an immaterial soul
stunted materialist psychology for centuries. This identification drove many of the
debates between vitalists/introspectionists and materialists in the 19th century, and
the influence of this doctrine still surfaces in the odd place even today.8
Reacting to these introspectionists and vitalists, C. Lloyd Morgan sought
to place comparative psychology on firmer empirical footing by basing hypotheses
about animal abilities less on anecdote, introspection, and anthropomorphism and
more on objective empirical observation. As a check against these influences,
Morgan (1894) etched the modus operandi of Standard Practice into the
foundations of comparative psychology with his famous “Canon”: “In no case
may we interpret an action as the outcome of the exercise of a higher psychical
faculty if it can be interpreted as the outcome of the exercise of one which stands
lower on the psychological scale” (pp. 53). Since its publication, nearly every word
in the Canon has been the subject of disputed interpretation (Radick, 2000); the
most pressing questions surround the default preference for explanations appealing
to “lower” processes, with associative processes commonly taken to be lower on
Morgan’s scale than cognitive ones. The justification for this bias has often been
interpreted as a matter of parsimony (sometimes taking the Canon to be a special
case of Occam’s Razor), implying that explanations appealing to association are
simpler than those appealing to cognition. However, the relevant metric of
simplicity is rarely specified,9 and Sober (2005) has convincingly argued that the
standard, parsimony-based justifications for the Canon do not withstand critical
scrutiny.
I suggest that the Canon is just an extension of the standard scientific
practice of ruling out alternative explanations of data, specialized for the context of
psychology. A default concern for associative mechanisms is justified by the fact
that they, unlike cognitive mechanisms, are supposed to be ubiquitous amongst
animals. Basic forms of conditioning have been investigated in Aplysia and
Melanogaster, and have even been demonstrated to occur in the spinal cords of rats
after the connection to the brain has been severed (Allen, Grau, & Meagher,
2009).10 The same is not true of cognition, which is presupposed to be a
comparatively less common cause of behavior amongst animals which, when

7
Notably, Hume also grants animals instinct; cf. footnote 1.
8
E.g., Allen (2006) detects such a “Cartesian residue” in the apparent opposition between “associative
mechanisms” and “intentional processes” set up by Clayton, Emery, and Dickinson (2006).
9
Zentall (2001) does suggest that explanations positing unobservables are more complex than those
that do not, though it is not clear how this form of simplicity should be weighed against others.
10
One may retort that the effects observed in spinally-transected rats are not due to associative
learning. Allen et al. (2009) argue extensively against this response.
- 318 -
present, depends upon specialized neural mechanisms. In short, the comparative
ubiquity of associative mechanisms explains the ubiquity of the controls.11
We may still wonder, however, why we should not simply consider the
presence of two competing explanations a draw. Sober (2005) suggests an
interpretation of “higher” and “lower” that will work for present purposes: “One
internal mechanism is higher than another if and only if the behavioral capacities
entailed by the former properly include the behavioral capacities entailed by the
latter” (pp. 236). If cognition is “higher” than mere association in this sense, then if
the additional capacities of the cognitive mechanism are not observed when the
organism is put in appropriate circumstances, then our failure to observe them is
evidence against the cognitive explanation. The elaborate controls of comparative
psychology can thus be understood as the attempt to ensure that the animal has
been put in a situation where the “higher” (cognitive), but not “lower” (merely
associative), capacities should manifest.
While Morgan’s psychology was significantly more empirical than what
came before, he reserved a place for introspection in theorizing about the mental
capacities of nonhuman animals — recommending what he called the “double
inductive” method, which combined empirical observation of animal behavior with
human introspection when performing similar tasks. Radical behaviorists such as
Watson (1928) and Skinner (1953) thus felt that Morgan’s reforms had not gone
nearly far enough. In his zeal to turn psychology into an objective, empirical
science, Watson declared all mentalist concepts off-limits. Watson painted
psychology as the study of behavior rather than the mind, rejecting mental states
like beliefs and desires as subjectivist fictions. While the basic principles of
Watson’s radical behaviorism eventually found wide acceptance in the United
States, his inheritors fiercely debated how to proceed. Some sought to reform
mentalist constructs under a broadly behaviorist methodology; Hull (1943) is well-
known for his work in quantifying various aspects of motivation, and “cognitive
behaviorists” such as Tolman (1948) argued that some mental constructs such as
goals and curiosity could be experimentally operationalized. Skinner, however,
furthered Watson’s anti-mentalism by explicitly banning reference to any
intermediary processing states between stimulus and response. By the late 1950’s,
Skinner’s camp had become the dominant — or at least most vocal — faction in
American psychology.
After the broader “cognitive revolution” repudiated Skinner’s (1953)
strictures, Standard Practice gradually again became mainstream. As cognitivism
about nonhuman animal behavior sought legitimacy, however, the behaviorist’s
mechanisms of instrumental and operant conditioning became the associative “null
hypothesis” which must be ruled out for cognitive explanations to be deemed
legitimate (Dennett, 1983; Wasserman & Zentall, 2006). In other words, the
mechanisms studied by various strains of behaviorism were all conflated as “mere
association” (which by now included at least all of the basic principles of classical
Pavlovian and operant Skinnerian conditioning) in Standard Practice — despite the
fact that many radical behaviorists had actively resisted the label of
“associationism,” which smacked to them of internal mental relations.
11
A phyletic explanation of these distributions may redeem Morgan’s contention that evolution
legitimizes the Canon.
- 319 -
 Completing the final step to the present crisis, I will eschew the standard
bedtime story about the cognitive revolution which is, I suspect, familiar to most
readers; but one brief qualification of this familiar fable is worth mentioning. It is
commonly overlooked that the more advanced associationists were just as stymied
by behaviorist epistemology as the cognitivists (Smith, 2000). While we emerged
from the cognitive revolution with a fairly firm understanding of the most basic
forms of conditioning, we had almost no precise understanding of the more
advanced forms of learning involving internal interactions between stimulus
elements. Many members of the latest generation of “associative” models do
involve sophisticated between-stimulus interactions. The distinction’s most
prominent critics are explicitly driven by the worry that, using currently popular
criteria, these models cannot be neatly classified as either cognitive or associative
(e.g., Penn & Povinelli, 2007). The solution to the current crisis, I will suggest
below, will require anchoring both our nascent model-based approaches to
association and behavior-based approaches to cognition in something more stable.

The Modern Crisis

Many erstwhile proponents of Standard Practice have begun to call for the
wholesale rejection of the distinction (Allen, 2006; Papineau & Heyes, 2006; Penn
& Povinelli, 2007). The problem is that in the interim since the cognitive
revolution, there has been a dramatic increase in the power and diversity of
“associative” models of learning, and as a result, there seem to be few behaviors
which cannot be modeled using some combination of associative tools. This poses
an existential threat to the field of comparative cognition research, a threat which
can be summarized by the following deductive argument:

1. Cognitive and associative explanations of behavior are exclusive

alternatives.
2. It is highly likely that associative models will eventually explain all
behavior.
3. Often implicit: If cognition does not explain behavior, it doesn’t exist;
there is no other work for it to do.
4. Therefore, it is highly likely that cognition does not exist.

Premise 1 is the contemporary interpretation of Morgan’s Canon

(Shettleworth, 2010). Premise 2 is taken from the “optimistic appraisal” inherited
from Hume and seemingly borne out by the most recent generation of sophisticated
“associative” models.12 Premise 3 is a standard “realist” interpretation of the
distinction (often left implicit), and premise 4 follows deductively from the others.

12
The “optimistic” appraisal can be found in many places; for example, Mitchell, De Houwer, and
Lovibond (2009) attest that “from experience we have learned that it is difficult to produce a pattern
of data that cannot be explained by one or the other variant of these associative models” (pp. 194).
Plausibility for this premise, however, can be derived from very general observations, such as that
connectionist models trained by backpropagation are Turing-complete, or that the brain itself is a
vastly more sophisticated kind of associative network.
- 320 -
Responses to the current crisis can be categorized in terms of how they respond to
this argument.
Some associationists and neo-behaviorists may welcome the conclusion,
arguing either that there is no role for cognitive theorizing in psychology generally
or only in the study of non-human animals (Wynne, 2004). A more moderate
version of this position rejects instead premise 3, holding that while cognition
strictly speaking does not exist, cognitive theorizing may yet have some heuristic
utility (Blaisdell, 2009; Zentall, 2001); I will not argue against these heuristic
positions here, supposing merely that a realist account of the distinction is to be
preferred if viable. Mitchell, De Houwer, & Lovibond (2009) reject premise 2,
courting controversy by arguing that there is little evidence for the “pure link-
forming” mechanism they use to characterize mere association and that even the
simplest forms of conditioning are better interpreted as propositional in nature.
The most promising response to this troubling argument holds that its
conclusion only follows if we commit a problematic equivocation. In particular,
the tension is resolved by avoiding an equivocation on the word “associative”
between premises 1 and 2. In other words, the sense in which the more
sophisticated models are “associative” is not the sense of “associative” which
Standard Practice takes to be mutually exclusive with cognition. The most popular
form of this response suggests that cognitive capacities are “implemented by” or
“emerge from” associative ones of a certain level of complexity. Crucially, these
solutions are not available unless we avoid this equivocation, for they cannot be
coherently described if the relevant senses of ‘cognition’ and ‘association’ are
mutually exclusive.
To clarify this solution, we should here distinguish two senses of the
words “cognitive” and “associative” (Buckner, 2011). On the one hand, there is the
“exclusionary” use of the terms found in Standard Practice (and premise 1), which
I will hereafter denote as “cognitiveex” and “associativeex.” These terms range over
psychological processes, and diagnoses of cognitionex and mere associationex are
mutually exclusive. A psychological process here is a series of events (which may
or may not be causally responsible for some observed behavior) in an organism’s
mind, presumably a temporally-extended causal sequence taking place in the
organism’s nervous system. On the other hand, there is another use of these terms
ranging instead over models of psychological processes, meaning roughly that
those models are “constructed according to cognitive/associative principles”; this
model-based sense of the distinction is the only sense in which premise 2 is
plausible. In this latter sense, the distinction ranges not over psychological
processes, but rather over our representations of those processes. This second use
of these terms will hereafter be denoted by “cognitivemod” and “associativemod”, and
apply to models based on whether they are built out of things like rules and
symbolic propositions on the one hand or nodes representing stimuli and links
representing the associations between them on the other.13 (Note that my goal in
this paper is not to provide a new theory of cognition or association with specific

13
Note that the “mere” modifier in this article’s title is not accidental; in the vernacular of
comparative psychology, it provides fairly reliable indication that the author intends to invoke the
exclusionary interpretation of the distinction.
- 321 -
criteria for applying these terms, but rather to discuss general tensions found in the
literature across a variety of diverse accounts.)
Separating these two distinctions helps us avoid fallacious inferences
which can arise from confusing the natural phenomena studied by science with our
representations of those phenomena. Just as a photograph of person has different
properties than the subject of that photograph (e.g., being black-and-white, two-
dimensional, or out-of-focus), psychological models are human-made artifacts that
possess properties beyond those of the psychological processes they depict. For
example, models can consist of diagrams, verbal descriptions, or a series of
mathematical equations, whereas psychological processes are none of these things.
Furthermore, models can omit aspects of the processes they are about for the sake
of simplicity or in order to focus attention on particular aspects of those processes.
While models can be evaluated in terms of the degree of correspondence between
their properties and the properties of their target phenomena, abstractions,
simplifications, and even assumptions which are outright false (such as frictionless
surfaces in a physical model) may be beneficial depending on the purposes to
which the model is put (Parker, 2011). Furthermore, the mapping from models to
natural phenomena can be many-to-one, and different representations of the same
process may be suitable for different purposes.
These general morals carry important implications for the present context.
For example, if, as many have suggested, the same psychological process can be
aptly described at one “level of analysis” by an associativemod model and at another
“level of analysis” by a cognitivemod model, then the verdicts of associativemod and
cognitivemod are not necessarily mutually exlusive. If, as some have suggested
(e.g., Fodor & Pylyshyn, 1988), a cognitiveex process is one which admits of
precise description with a cognitivemod model, then this entails that associativemod
and cognitiveex are also not mutually exclusive. This possibility succinctly
describes a solution to the “modern crisis”: Many of the most recent associativemod
models actually depict processes which are cognitiveex, rather than associativeex, in
nature.
This solution has, in broad strokes, long been familiar in certain areas of
cognitive science — in particular, in the debate about cognitive architecture
between connectionists and classical cognitive scientists which raged from the ‘80s
to mid-‘90s (e.g., Fodor & Pylyshyn, 1988; Smolensky, 1988). However, the
pressing pragmatic question still concerns how and where to draw the line between
the associativemod models depicting merely associativeex processes and those
depicting cognitiveex ones. For this kind of strategy to be of any practical help to
comparative psychologists, we must characterize cognitionex and mere
associationex with enough precision such that the two can be experimentally
distinguished. This is a difficult challenge, for there are many model-based and
behavior-based candidates to choose from (many of which are mutually-
inconsistent), and no clear winner is favored by the current state of psychological
data.
The key way that model-based and behavior-based approaches must work
together, I suggest, is in helping us choose which of the candidate criteria will be
the most conducive to future productive science. I here only suggest that this role
is pragmatic or methodological; I remain agnostic on the troubled metaphysical
- 322 -
question of the intrinsic nature of psychological states, e.g., as to whether their
“essences” are behavioral, functional, or neural. In the following two sections, I
will review elements of the contemporary debate which focus on more model-
based and behavior-based approaches to the distinction, concluding with some
suggestions as to how the two might be usefully integrated in future research.

General Comments about the Two Approaches

The difference between the two approaches that I will discuss can be
expressed simply enough: “Model-based” criteria attempt to constrain the class of
models that could be used to aptly describe cognitiveex (and/or merely
associativeex) processes, and “behavior-based” criteria attempt to specify the
behavioral capacities required of processes which are cognitiveex (and/or merely
associativeex). However, a few general caveats must be added to this simple picture
before reviewing specific model- and behavior-based approaches to the distinction.
First, unlike the distinction between cognitionex and mere associationex
which they are designed to illuminate, the two approaches are not intended to be
mutually exclusive. Quantifiable psychological models will entail that particular
behaviors will be produced in response to particular stimuli; and particular
behavioral tests for cognitionex may only be satisfiable by systems describable
using a limited range of models. The degree of “distance” between a model-based
and a behavior-based approach to the distinction depends in part on the degree of
abstraction present in the model or behavioral criteria (“abstraction” here
signifying “lack of detail”). There may be a tight relationship between the
specification of the behavioral capacities and the model; one might argue that to
specify behavioral criteria with enough precision just is to specify a model. Some
models, such as the purely input-output models of the behaviorist, provide only
very weak modeling constraints, placing few restrictions on what happens “inside”
the system; yet these models also provide specific behavioral constraints as to how
the system will respond in a wide variety of situations. On my scheme, these input-
output models would thus fall in-between pure model-based and pure behavior-
based accounts. These examples illustrate that the distinction I wish to draw
between the two styles of approach is a continuum or smear, with the most abstract
specifications of families of models at one pole and most abstract behavioral
criteria on the other.
Second, though the distinction between cognition and mere association in
comparative psychology is often tied to “dual systems” theories from human
psychology, this identification should be regarded with suspicion. A variety of
such theories draw the boundaries of the two systems differently; here, I shall only
focus on a few of the most prominent and oft-cited. An obvious problem is that
many dual-system theorists suppose their “cognitive” system (often referred to as
“System II”) to be uniquely human, and in such cases it is almost guaranteed that
the distinction drawn there and that of Standard Practice do not coincide. Rather,
despite the fact that these systems are often labeled as “associative” and
“cognitive”, both systems satisfy standard criteria for cognitionex in Standard
Practice (Mitchell et al., 2009; Smolensky, 1988).

- 323 -
 Finally, while I have defined ‘associationex’ and ‘cognitionex’ as properly
ranging over processes and not models, I will at times speak of model-based
approaches to the distinction between these two (ex) terms. This way of speaking
should be understood as shorthand for “a way of making out the distinction
between cognitiveex and associativeex processes based on whether those processes
are aptly describable using model(s) picked out using certain model-based
criteria.” This must be distinguished from criteria which are merely designed to
make out a distinction between associativemod and cognitivemod models, lest we fall
back into the problematic equivocation pointed out above.

Figure 1. Example of some standard model- and behavior-based criteria arranged on a continuum
from most abstract model-based to most abstract behavior-based. Criteria typically considered
cognitive are above the line, those typically considered associative are below. More specific models
and behavioral criteria (which imply both behavioral and modeling constraints) are located in the
middle. Arrangement of criteria on the continuum is meant to be monotonic, but distances from each
other and from continuum line are not significant.

Model-Based Approaches

Examples

In almost all standard accounts of the distinction between cognition and

mere association, one can find appeals to some model-based criteria. Where
comparative psychologists offer accounts of the distinction, they are often parasitic
upon “dual systems” accounts in human psychology just mentioned. In these
accounts, model-based criteria proliferate, as the accounts emerged from the
modeling debate which raged in the ‘80s and ‘90s between the classical
computationalists -- typified by Chomsky (1965), Newell and Simon (1976), and
Fodor and Pylyshyn (1988) and connectionists -- led by McClelland and
Rumelhart (1986), Smolensky (1988), and Elman (1996). Like the empiricists and
rationalists in the 17th and 18th centuries, positions in these debates tended towards
monistic extremes and thus away from subtle criteria which could help disentangle
the “middle ground” capacities so troubling to Standard Practice. Connectionists
- 324 -
often claimed (and some still do) that all thought was better understood on the
model of connectionist networks, and classical computationalists responded that
any mental activity deserving the label “cognition” must instead be modeled on the
rule-based perspective of Newell and Simon’s Physical Symbol System hypothesis
(1976).
As these two factions fought more or less to a stalemate and productive
research continued to emerge from both camps, interest in the early ‘90s began to
grow in ecumenical compromises. As a result, “dual system” views on cognition
began to proliferate, a perspective which perhaps remains the dominant view in
human psychology today. The early roots of the two-systems views can perhaps be
found in Smolensky’s (1988) “connectionist proposal” that the brain contains both
a conscious rule-interpreter which is properly described at one “level of analysis”
as a physical symbol system, and a subsymbolic connectionist system which
admits of no accurate classical description. Notably, Smolensky took both systems
to be implemented on fundamentally connectionist architecture, so the two systems
could not be distinguished using this associativemod model-based criterion. Rather,
his “reflective” system differed from his “intuitive” system in that only the former
admitted of precise description at the “symbolic” level.
In later accounts of the two systems views, more abstract model-based
criteria proliferate. Sloman’s (1996) influential account of the two systems view
consists almost entirely of the abstract model-based variety; the associative and
rule-based systems are distinguished along the dimensions of “principles of
operation,” “sources of knowledge,” “nature of representations,” “relations,”
“nature of processing,” and “illustrative cognitive functions.” The associative
system is characterized as operating on similarity and contiguity, as manipulating
representations consisting of concrete images, stereotypes, and feature sets, and as
relating representations with associations and “soft constraints.” Some more
behavior-based features are included in the list as “illustrative cognitive functions;”
such functions for System 1 including a capacity to engage in fantasy, creativity,
imagination, and associative memory tasks, for System II deliberation, analysis,
ascription of function, and strategic memory. Notably, Gigerenzer and Regier
(1996) worry that these many criteria cross-cut one another and thus are unsuitable
for drawing a coherent dichotomy.

Figure 2. Proposed mapping of terminologies from literatures on Standard Practice, dual-systems

theory, and Morgan’s Canon.

- 325 -
 While, again, it is unlikely that this distinction corresponds to the one
found in Standard Practice, these debates show that there are, broadly, two kinds of
model-based approach one could adopt. The first kind of approach focuses on one
or two specific models as exemplars of cognition or mere association. This kind of
approach might draw the line between cognition and mere association by taking,
for example, a classical model of associative learning as such as Estes (1950)
model of elemental conditioning or the Rescorla-Wagner model (1972) of cue
competition as the paradigm associative model which is supposed to define the
limits of associativeex learning. The utility of this kind of approach is limited,
however, in that there are a number of different associative models and it is not
clear from behavioral evidence alone which (if any) is the right one. Models of
associative learning often have non-overlapping strengths and weaknesses; models
can differ in their perceived “simplicity,” the number of free parameters which
must be fit, the breadth or type of phenomena they explain, and their degree of
biological plausibility. It is clear from the current literature that nearly all
associative theorists agree that it would be premature, given the current state of
evidence, to nominate one of these models as “the model” of associativeex learning.
What is much less clear, unfortunately, is what future evidence should be gathered
that would help us make such a selection.
To avoid these complications, the second style of model-based approach
instead defines classes of models by focusing on more abstract properties of those
models, such as whether they are constructed from things like nodes representing
stimuli and links representing the associative strength between them, or rules and
propositions. Discussion of individual models can still be relevant to this approach,
but only insofar as they illuminate principles uniting a whole class of models. For
example, Wallace and Fountain (2002) consider their Sequential Pairwise
Associative Memory (SPAM) model as a paradigm stand-in for mere associationex
(i.e., they presume the behavioral capacities of SPAM and mere associationex to be
equivalent), but argue that SPAM is equivalent to a broad class of models that
share common features (see also Fountain and Doyle, this issue). Popular
contemporary candidates for such abstract criteria for cognitivemod models include
a tendency to use rules and propositions (Mitchell et al., 2009; Smolensky, 1988)
or deploy representations vehicled in a “language of thought” (Fodor & Pylyshyn,
1988); for associativemod models, criteria include being based on “contiguities
between cues” (Penn & Povinelli, 2007) and/or similarity (Gigerenzer & Regier,
1996), or resemblance to a simple link-forming mechanism (Mitchell et al., 2009).
(As I emphasized in Section 3, we should here exercise extreme caution in
reasoning from cognitivemod/associativemod to cognitiveex/associativeex.)

Problems with Purely Model-Based Approaches

Model-based approaches are saddled by a number of challenges which

make it very unlikely that a purely model-based approach could provide a solution
to the current crisis. Many of the problems issue from a kind of trade-off faced by
model-based approaches: they must be both general enough to capture a wide
enough class of phenomena to leave cognitionex and mere associationex with
- 326 -
enough empirical significance to merit their default roles in Standard Practice, and
yet specific enough to issue precise, quantifiable predictions which could be
verified through a wide variety of experiments. Let us call these desiderata
“Generality” and “Empirical Reference” (“generality” here signifying “applies to
more phenomena”). Let us thus summarize these three desiderata as follows:

Desideratum #1: Generality. The criteria nominated to characterize

associationex (and cognitionex) should cover the full range of associativeex (and
cognitiveex) processes.
Desideratum #2: Empirical Reference. The criteria nominated to
characterize associationex (and cognitionex) should issue quantifiable predictions
which could allow a hypothesis that a behavior is the result of “associationex” (and
“cognitionex”) to be empirically confirmed or falsified.

To the two desiderata in this trade-off, we should also add the constraint
that all cognitiveex phenomena should have much in common with one another,
and all associativeex phenomena should have much in common with one another.
This desideratum is justified by the appraisal that if cognitiveex or associativeex
phenomena were instead radically dissimilar (and merely shared a label for, say,
historical reasons), then it would not be useful to group them into unitary classes in
Standard Practice. It could be, for example, that the processes enabling the
different “cognitive” capacities such as episodic memory, cognitive mapping,
transitive inference, conceptual abilities, metacognition, mindreading, and so on
share nothing in common beyond the fact that behaviorists have difficulty
explaining them. In this case, it would then be more useful to concede to the
distinction’s critics that it cannot answer to the needs of comparative psychology
and thus should be rejected.

Desideratum #3: Inductive Unity. The criteria nominated to characterize

associationex (and cognitionex) should allow a significant number of inductive
generalizations to be true of all associativeex processes (and a significant number to
be true of all cognitiveex processes).14

Problem 1: Picking the Winning Horse

Model-based approaches which focus on particular models as exemplars of

cognitionex or mere associationex tend to do a decent job on satisfying the
Empirical Reference desideratum but have a comparatively tenuous grip on
Generality. As mentioned above, model-based approaches which focus on
particular models as exemplars of cognitionex or associationex overreach. We
simply do not know enough at present about the distribution of behavioral
properties amongst animals to say whether specific models carve nature at its
joints. Many models seem clearly too specific because they make predictions only
about a limited range of phenomena in particular experimental settings. For
example, the input and output nodes of De Lillo, Floreano, and Antinucci’s (2001)

14
Note that there is work to do to ensure that these properties are “projectible” and not disjunctive,
but I will not draw these specifications here.
- 327 -
connectionist model of transitive inference are specialized for a dyadic choice
situation (the only output nodes being “left response” and “right response”),
leaving it unclear what the model would say, if anything, about more complex
decision environments. Other models with a more general reach, such as Rescorla-
Wagner, might be too general, crossing the line from mere associationex to claim
some territory better described as cognitiveex — a charge pressed against this
model by Mitchell et al. (2009), for example.

Problem 2: Runaway Abstraction

Model-based approaches which try to delineate general classes of models

might easily include all the properties we attribute to cognitionex and mere
associationex, but only at the cost of being too abstract to satisfy the Empirical
Reference criterion. Very abstract model-based criteria such as “links & nodes”
and “rules & representations” fit into this category. As is often observed, many
abstract model-based criteria possess too much representational power to
effectively draw a distinction. Penn and Povinelli (2007) press this charge against
Blaisdell et al.’s (2006) Causal Bayes Net approach to causal cognition, calling
the framework a “lingua franca that can give posthoc explanations for nearly any
nonpathological causal inference” (pp. 105). Smolensky (1988) emphasizes that
both subsymbolic computational processes and theorem provers can, in principle,
be implemented in connectionist models. At the very least, such abstract criteria
must be constrained in further ways if they are to be taken as model-based criteria
for cognitionex. Obvious additions would include implementation constraints, such
as that we should not consider “associativeex” anything which a connectionist
network could in principle be trained to do if given an infinite number of nodes, an
optimal training set, and an infinite amount of time to master it. Rather, only those
tasks which biologically-relevant architectures could perform when trained in
biologically-relevant learning conditions should be considered relevant. A similar
modification to the Causal Bayes Net approach might precisely specify how causal
models themselves are acquired (a process which may or may not itself be
Bayesian), in order to constrain the class of causal models which could be induced
from a given training regime.

Problem 3: Contingent Notational Properties

The greatest danger posed by model-based criteria is that they encourage

us to focus on contingent notational properties of models. As I use the phrase, a
“contingent notational property” is a property of a model which is irrelevant to the
predictions that model makes about the phenomena it purports to be about. In other
words, the same model could be expressed without any reference to those
properties; those properties could be changed or excised, and the model would
make the same behavioral predictions. Without further constraints or elaboration,
“rules & propositions” vs. “links & nodes” are, I suggest, just such contingent
criteria. The Rescorla-Wagner model is often taken as an exemplary associative
model; but it has been expressed using rules and symbols in a PROLOG program
(Shanks & Pearson, 1987). In fact, the representational power of both rules &
- 328 -
propositions and of links & nodes ensure that, without additional constraints,
almost any model could be expressed using either. When appealing to model-based
criteria in an evaluation of the distinction in Standard Practice, we must carefully
base our classification not on contingent properties of models but rather on deeper
properties of the processes those models describe.

Behavior-Based Approaches

Due to the difficulties in outlining purely model-based approaches to the

distinction, one might suppose our attention to be better-directed towards the
phenomena of associationex and cognitionex themselves, rather than our
(preliminary, partial, diverse) models of them. This approach would suppose that
we could identify some rough behavioral criteria which could be used to
experimentally distinguish cognitionex and mere associationex, and then gradually
home in on the specifics and sharpen the boundaries as our level of knowledge
improved. While modeling could be seen to provide important steps on the road to
the ultimate account of the distinction, all modeling would be regarded with some
suspicion as tentative, and researchers would place more trust in experimental
assessments of the properties of cognitionex and mere associationex themselves. The
mantra of this approach to the distinction might be that “the animal itself is its own
best model.”
Admittedly, this “pure” way of presenting the approach is an
oversimplification; behavioral criteria are often initially proposed to distinguish
between two kinds of model, though they can quickly thereafter take on a life of
their own. Whatever their origin, behavior-based approaches to the distinction rely
on a series of operational criteria which can be assessed in experimental settings to
determine whether some behavior is driven by cognitionex or mere associationex.
Behavior-based approaches to mere associationex often stray little from the
behaviorist’s conditioning methodologies, so I will comment little on them here.
By contrast, the last four decades have seen a great surge of interest in behavioral
tests for cognitionex, especially those which, by requiring particular forms of
behavioral flexibility, rule out associativeex explanations of the data.
These criteria can be formulated simultaneously at two levels. In a
pessimistic appraisal of the utility of grand dichotomies, Newell (1973) described
the nature of psychological research as simultaneously investigating the level of
“phenomena” and the level of “conception.” The “phenomenal” level pertains
directly to observed effects that can be reliably produced in experimental settings;
Newell produced a list of 59 such effects, including things like categorical
conceptualization, subitizing, chunking, hierarchical effects in long-term memory,
serial position effects in recall, backward associations, reversal learning, and so on.
At the same time, however, psychologists attempt to tie their “low-level”
experimental or modeling research into conceptualizations “at the high level of
grand theory” — which Newell supposed to unfold “mostly…by the construction
of [binary] oppositions” (of which cognition vs. association was one of the most
prominent).
While the list of phenomena and our models of them might have evolved
during the intervening decades, this basic two-tiered approach is still alive and well
- 329 -
in comparative cognition research. The important point for present purposes is that
we can direct our attention towards formulating very specific behavioral criteria
for individual cognitiveex or associativeex capacities, or towards very abstract
criteria at the level of the grand dichotomy. As with the model-based criteria, we
find a trade-off between the desiderata of Generality and Empirical Reference. In
order to also satisfy Inductive Unity, the criteria at the two levels should not be
unrelated; for example, the criteria for assessing whether a behavior was driven by
a particular cognitiveex capacity should be seen as specific expressions of the
characteristics attributed to cognitionex more generally.

Figure 3. A conceptual map of current research into Newell’s “levels,” with examples of some
common operational criteria.

Most of the abstract behavioral criteria for cognitionex can be seen to

follow from the intuition that cognitionex permits a kind of behavioral flexibility
which is unavailable to mere associationex. This general intuition is expressed by
Smolensky’s (1988) criteria for cognition (which includes both his “intuitive” and
“reflective” systems):

A cognitive system can, under a wide variety of environmental conditions,

maintain a large number of goal conditions. The greater the repertoire of
goals and variety of tolerable conditions, the greater the cognitive capacity
of the system. (p. 15)

Mitchell et al. (2009) rely on indicators of “effort” and “attention” as behavioral

criteria for “propositional” processes. Penn and Povinelli (2007) regard the ability
“to understand the ‘web of possibilities’ that connects causes and effects” as a
sufficient criterion for cognition, as this is critical to their argument that the
extended comparator hypothesis is both cognitive and associative.
These discussions of behavioral flexibility are often too abstract to be
useful, however; such criteria must be cashed out in specific experimental tests to

- 330 -
be empirically meaningful.15 More interesting and illuminating criteria arise from
the behavioral tests for specific cognitive capacities that have been worked out by
comparative cognition researchers. A case study of these criteria can reveal the
specific forms of behavioral flexibility attributed to cognitionex by comparative
psychologists.

Examples of Behavior-Based Approaches

Case Studies: Cognitive Maps. Cognitive (or mental) mapping is a

paradigm cognitiveex phenomena. Like real maps, they are characterized as
representing the environment which they model. Furthermore, work on cognitive
maps was historically influential in shifting attention in animal psychology from
the “behavioral stereotypy” emphasized by radical behaviorists to the “behavioral
flexibility” that has been the focus of comparative cognition research (Olton,
1979). While Chomsky’s research in human linguistics is often cited as the most
important work driving the cognitive turn in human psychology, Tolman’s (1948)
work on cognitive maps was perhaps more influential amongst researchers on
nonhuman animals.
In his classic work Cognitive maps in rats and men (1948), Tolman
insisted that the “stimulus-response” school of learning could not account for the
navigation behaviors of his rats. Instead, he suggested the existence of more
sophisticated learning mechanisms:

The stimuli, which are allowed in, are not connected by just simple one-to-
one switches to the outgoing responses. Rather, the incoming impulses are
usually worked over and elaborated in the central control room into a
tentative, cognitive-like map of the environment. (p. 193)

The primary distinguishing factor of this latter form of learning is its “selectivity”
and the way in which stimuli are affected by each other and by what is already
known (especially about other cues). On Tolman’s more sophisticated form of
learning, stimuli do not simply act individually and independently on the rat’s
nervous system. Instead, only contingencies between stimuli deemed relevant and
informative are attended to and integrated into a global, map-like representation of
the environment — the selection and integration processes being based in large
part on what is already known.
Tolman cited five categories of behavioral criteria which he used to
operationalize his “cognitive control room” (leading to his classification as a
“cognitive behaviorist”): “latent learning,” “vicarious trial and error,” “searching
for the stimulus,” “hypothesis,” and “spatial orientation,” each of which
investigated the rat’s abilities to make use of information in ways which were more

15
To their credit, some proponents of abstract criteria have often attempted to discharge this burden.
Mitchell et al. (2009) dismiss purported evidence for “link-forming mechanism” explanations of
learning data by using diminishment under cognitive load as an operational criterion for propositional
learning; Penn and Povinelli (2007) suggest that the cognitive understanding of causes and effects
can be operationalized in terms of retrospective reevaluation effects, supporting their contention that
the Extended Comparator Hypothesis satisfies cognitive criteria.
- 331 -
flexible than would be predicted by the S-R behaviorism of the day. In his famous
experiments on latent learning, for example, Tolman demonstrated that rats could
acquire navigational knowledge about a maze through idle exploration, in the
absence of explicit reinforcement.
Further experiments on cognitive mapping since Tolman have continued to
emphasize the importance of being able to flexibly and adaptively use information
gained on one task in different contexts or to achieve different goals. For example,
Walker and Olton (1979) conducted a series of experiments showing that rats
could spontaneously deploy spatial information gained from exploring one part of
a maze to navigate to the goal from a novel starting position at which they had
never before been placed. Menzel (1973, 1978) showed that chimpanzees could
also use spatial information flexibly in a variety of tasks in a way which could not
be explained by simple conditioning. For example, in one case experimenters
carried chimpanzees around a room in crisscrossing and convoluted paths while
hiding food in 18 locations; they then released the chimps to see if they would
replicate the tortured paths or recover the food in a more efficient manner. The
chimps tended to utilize novel, efficient routes around the perimeter of the
enclosures to recover the food, rarely crossing their paths.
The concept of a cognitive map was perhaps developed into its most
sophisticated form by O’Keefe and Nadel in their influential interdisciplinary
work, The Hippocampus as a Cognitive Map (1979). There, the cognitive map
concept is described in detail and contrasted with more basic systems, such as
“instinctual,” single-cue, and route-based strategies for spatial navigation
(cognitive maps have also since been distinguished from path integration and dead
reckoning (see Mackintosh, 2002) though it is not clear that all authors would
endorse a neat division). Again, O’Keefe and Nadel (1979) claim that the
distinguishing feature of cognitive maps is the behavioral flexibility they support:

The first striking feature of a map is its flexibility. Whereas a route

specifies a starting point, a goal, and a particular direction of movement
from the former to the latter, a map specifies none of these, either in its
construction or its usage. It can be used with equal facility to get from any
particular place to any other. Additional flexibility derives from the
freedom from specific objects and behaviours. If one path is blocked
another can be easily found and followed. (p. 87)

O’Keefe and Nadel unpack the flexibility of cognitive maps into five
characteristics (all but #4 of which are straightforwardly behavior-based — and
operationalizations of #4 have been proposed) which they hold can be used to
distinguish cognitive maps from other navigation strategies:

1. Learning is driven by “curiosity” and “novelty-seeking,” rather than by

mere exposure to stimulus co-occurrences or repetition. Redundant or
uninteresting cues may be ignored entirely, whereas novel or surprising
cues are sought out.

- 332 -
 2. The representation is suitable to reach a range of goals and from a range of
starting points, many of which may have never been explored. When using
a cognitive map, the creature can get from many points on the map to
many other points, whereas routes lead to only one goal.

3. Maps are robust in the face of environmental change and navigational

error. They do not depend on any particular subset of landmarks or turns.
Routes and instinctual strategies are brittle in the face of change,
depending upon a small subset of cues or a particular series of turns or
distances, which, if missed through mistake, environmental change, or
lapse in attention, render the goal inaccessible until the creature can re-
locate a specific position on that route.

4. Maps are extremely efficient and non-redundant. Many routes may

terminate in the same sequence of turns, for example, but these sequences
will be redundantly represented separately in each route. Maps, on the
other hand, represent an enormous number of possible routes with a single,
integrated representation.

5. Maps are at least potentially multi-modal; they do not depend on any one
sensory modality. Involving a true representation of space, navigation is
possible by sight, touch, and (for many species) smell, sound, and taste.

Case Studies: Episodic Memory and Future Planning. Episodic

memory, another paradigm cognitiveex capacity, is typically characterized by the
ability to flexibly deploy a mental record of a specific experienced event in terms
of its “what,” “when,” and “where.” Furthermore, episodic memory is almost
universally distinguished from other forms of associativeex learning such as
procedural memory — and thus serves as an illustrative case study for present
purposes.
Tulving’s (1983, 1985) influential account of episodic memory has been
the launching point for many studies on the faculty. According to Tulving,
episodic memory was built on earlier memory systems, the procedural (stimulus-
response) and the semantic (which involves the construction of “response-neutral,”
“third-person” mental models of the world). Episodic memory is to be
distinguished from these two earlier systems by virtue of its ability to enable what
Tulving (1985) called “mental time travel”:

Episodic memory affords the additional capability of acquisition and

retention of knowledge about personally experienced events and their
temporal relations in subjective time and the ability to mentally "travel
back" in time. (p. 387)

In short, an organism with episodic memory is able to “re-experience” remembered

events and use information acquired from that re-experiencing to make decisions.
The ability to engage in mental time travel is purportedly enabled by a difference

- 333 -
in the kind of representations involved in episodic memory, which records
particular experienced events in their autobiographical context.
According to these theorists, procedural memory allows only rather
inflexible behavior; responding according to a relatively rigid format to stimuli
determined at the time of learning is obligatory (Hirsh, 1974; Mishkin & Petri,
1984). On the other hand, knowledge acquired in both semantic and episodic
memory can be expressed flexibly, in a variety of distinct behavioral and
perceptual circumstances. In influential studies on episodic memory, Clayton and
Dickinson (2009) have proposed to investigate these forms of flexibility in caching
bird species like scrub jays. Clayton and Dickinson capitalize on the idea that one
of the proposed functions of an ability to relive past experiences is to utilize that
information in anticipating and planning for similar events occurring in the future.
As they put it,

The function of episodic memory lies primarily in its constructive rather

than reconstructive ability: its purpose is to mentally simulate multiple
future scenarios by flexibly recombining details from past events without
having to physically engage in the actual behavior. Inevitably there is a
trade-off between flexibility and stability, and therefore the cost is that the
episodic memory system is much more vulnerable than the semantic
knowledge system to memory errors such as misattribution and false
recognition. (p. 62, emphasis added).

As such, they have investigated the ability of caching birds to recall details of past
caching activities. In these tests, experiments are designed to determine whether
birds are capable of recording this information rapidly and deploying it flexibly to
plan for the future.
For example, one of the key experiments that Clayton and Dickinson
(2009) performed was designed to demonstrate their ability to plan for the future
by differentially caching a food item that would be preferred in anticipated future
conditions, but not preferred at the time of caching. In one experiment (Raby,
Alexis, Dickinson, & Clayton, 2007), scrub jays were confined at breakfast time in
one of two enclosures, and given the opportunity to learn that they would receive
either no food or a particular type of food in one enclosure, and another type of
food in the other. One evening, the birds were then given the opportunity to cache
a particular type of food; if they were operating with “episodic-like” memory, we
would predict that they would preferentially cache the type of food in the enclosure
where they had not previously received it, remembering their previous deprivation
and anticipating their future breakfast needs. This prediction was confirmed; across
several different variations of the experiment, the birds preferentially cached foods
in the enclosures where they had not previously received it before.
Why should we suppose that the explanation of this behavior merits appeal
to episodic memory? Again, the contrast class for Clayton and Dickinson (2009)
were various forms of “instinctive” behaviors and associativeex learning. First, the
experimenters noted that the abilities must not be found to depend invariantly upon
a handful of cues and be extremely limited in their means of expression; as they
point out, relatively “hard-wired” behaviors like migration, nest-building, and
- 334 -
hibernation, despite being apparently “future-oriented,” would not count as
cognitiveex unless we can “rule out simpler accounts in terms of behavior triggered
by seasonal cues or previous reinforcement of the anticipatory act” (p. 65).
Previous experiments had found significant flexibility in scrub jay caching
abilities, such as a sensitivity to the type of item cached and an ability to learn and
utilize information about the rates at which these cached items degraded (reviewed
in Clayton & Dickinson, 2009), ruling out the simplest “instinctual” explanations.
It remained to be proven, however, that the bird’s apparent abilities to plan for the
future were not based in past conditioning or in current motivational states. Their
experiments thus included elaborate controls to rule out such “simpler”
explanations, which included that:

1. The birds were confined the same number of times in each compartment in
an attempt to rule out an associative preference for one over the other.

2. The birds were fed to satiety prior to the caching event, to rule out an
explanation in terms of their current motivational state.

3. The birds had never before been given the opportunity to cache in either
enclosure, and thus caching in those locations was unexpected and novel.

These controls have convinced some who were skeptical of earlier studies
— for example, Shettleworth (2007) has suggested two requirements for evidence
of future planning in nonhuman animals, that “the behavior involved should be a
novel action or combination of actions” and “it should be appropriate to a
motivational state other than the one the animal is in at that moment.” On
surveying the controls in the Raby et al. (2007) experiment, Shettleworth (2007)
conceded that it comprised “the first observations [on nonhuman animals] that
unambiguously fulfill both requirements’’ (p. 825).16
This is not to say that all skeptics have been convinced. Suddendorf and
Corballis (2007) have challenged the conclusions of Raby et al. (2007) on the basis
of — again — a lack of the appropriate kind of flexibility for cognitiveex
ascriptions. One complaint stems from the fact that a “genuine” manifestation of
episodic memory and future planning should be able to exhibit itself domain-
generally, and there is not currently any evidence that jays can exhibit episodic
memory in any domain other than caching. While the “flexibility criteria” of
Clayton and her colleagues (for a review, see Clayton, Russel, & Dickinson 2009)
have focused on the rapidity of learning from caching episodes and the ability to
generalize the learned information to novel circumstances, Suddendorf and
Corballis instead emphasize a need to manifest memory and future-oriented
abilities across different. Suddendorf and Corballis (2007) suggest that absent
further information, a simple associative mechanism may explain the data and that
the “ability in linking their caching and retrieval they have demonstrated may be
more akin to Garcia-type learning than to human mental time travel” (p. 2). Rapid
one-trial learning recording what-when-where data, they suggest, may not be

16
Shettleworth has since expressed some additional skepticism; see Shettleworth, 2010.
- 335 -
sufficiently flexible to count as even episodic-like memory if the context of
application is so limited.
Whatever the fate of the Raby et al. (2007) conclusions, from this brief
review of the dispute on episodic memory and future planning, it is supposed that
behaviors driven by episodic memory:

1. Should deviate from what would be predicted from simple conditioning

and other invariant S-R mechanisms, and should deviate in ways which are
flexible and adaptive. (How intelligent? Adaptive in what circumstances?)

2. Should manifest themselves across a diverse range of tasks and domains

(How many?)

3. Should manifest in novel perceptual circumstances on tasks never before

attempted. (How novel?)

4. Should cease to manifest when no longer appropriate/needed (always, or

only when there is some cost — and if so, how quickly and how costly?).

Each criterion here is faced with a problem which limits its objective applicability:
they require an evaluative judgment for which precise criteria are unspecified,
indicated in the parentheticals above. Different researchers often disagree on the
evaluative yardstick to be applied; this difficulty will be commented on again in at
the end of this section.

Case Studies: Transitive Inference. The final cognitiveex capacity which

I will discuss here is transitive inference. The ability to pick out the easiest prey,
most nourishing food, or dominant troop mate is of obvious adaptive significance.
Two kinds of solution to this problem are possible. First, animals can learn
rankings by rote, encountering each possible set of options and learning their
relative rank(s) by trial and error. Secondly, and more efficiently, animals can
draw inferences about novel sets of options (that is, between sets never before
encountered) by flexibly integrating information learned in previous encounters. In
other words, animals can use encounters with two or more distinct choice sets
which involve common elements to rank items on a common monotonic dimension
(tastiness, defensive potential, dominance, etc.). For example, if the R relation is
transitive and I know that aRb and bRc, I can conclude that aRc without needing to
evaluate an ac pair. Recognizing this would allow an organism to display correct
behavior on the perceptually novel pair the first time it is encountered, without
needing to go through the potentially costly evaluation process. This latter solution
to the ranking problem has been considered another paradigm cognitiveex capacity.
The “gold standard” behavioral test for transitive inference has, for many
years, been the 5-element series (Bryant & Trabasso, 1971; McGonigle &
Chalmers, 1977). In the 5-element series, subjects are trained on four successive
discrimination dyads — A+B-, B+C-, C+D-, and D+E-. Subjects are then tested on
the novel BD dyad, and above-chance responding to “B” is taken as evidence that
animals are capable of transitive inference. The key to this test is that B and D
- 336 -
have been rewarded an equal number of times, and thus should have a similar
elemental reinforcement value. A correct choice on a novel AE dyad, for example,
would provide no evidence of transitive inference, for A was rewarded 100% of
the time and E 0%. On the other hand, the organism responding on the basis of
mere elemental conditioning should show no preference for B over D (this is why
all authors agree that we must, at minimum, go to a five-element series to find
genuine evidence of transitive inference).
At the time of writing, a number of species have “passed” the 5-element
test, including several species of monkey, rats, and pigeons. This is not to say,
however, that skeptics have been satisfied that all these animals are implementing
the “cognitive” solution to the problem. von Fersen, Wynne, Delius, and Staddon
(1991) suggested an “associative” mechanism of “value transfer” to account for
animal performance on the 5-element task. On the value transfer account, stimuli
can get a reward “boost” from co-occurring with more frequently rewarded stimuli.
So while A has been rewarded 100% of the time and B 50% of the time, when A
and B co-occur, B gets a boost in its own reinforcement value for having co-
occurred with a more highly-rewarded A. Value transfer is here construed as a
merely associativeex mechanism, given that it is based only on the spatiotemporal
contiguity of cues, and thus to provide a “simpler,” deflationary explanation for
observed animal data on the 5-element task.
Models built on the value transfer principle (De Lillo et al., 2001; Frank,
Rudy, & O’Reilly, 2003) have since purportedly done a better job of accounting
for animal data than some of the more traditional, rule-based approaches to
transitive behaviors (e.g., Dusek & Eichenbaum, 1997). Van Elzakker, O’Reilly,
and Rudy (2003), for example, collected additional data on a six-element task
(A+B-, B+C-, C+D-, D+E-, E+F-), arguing that the traditional, rule-based
conclusion should predict no difference in responding between BD and BE pairs
(since each correct choice receives equal reward), whereas they observed much
stronger responding to B when paired with E than when paired with D. However,
more sophisticated versions of the cognitivemod models of transitive inference have
predicted a Symbolic Distance Effect (McGonigle & Chalmers, 2006), where
animals find it “easier” to make discriminations amongst items spaced far apart on
the monotonic scale than those ranked nearby (and thus can make them more
quickly). Since B and E would be farther apart on the integrated preference
ranking, the thought goes, we would predict quicker and more successful
discriminations on BE than on BD. Van Elzakker et al. (2003) retort, however, that
even if we allow the cognitivist the normalized symbolic distance scales that allow
them to make the right predictions about BD and BE on the 6-element task, the
account seems unable to explain why the observed BE responding in their 6-
element task was essentially identical to the BD responding in the 5-element task.
We must accept that a 0.1 value difference between the normalized distances
between B and D on the 5- and 6-element scales (0.5 vs. 0.4) translates into a 30%
difference in response time.
Should we thus conclude the discussion with a clear associationist win?
Nowhere are the conceptual tensions surrounding the distinction between
cognitionex and mere associationex more apparent than in the debate surrounding
transitive inference, for we have come farther with the experiments, models, and
- 337 -
underlying neuroscience here than with any other capacity. While Frank, Rudy,
and O’Reilly (2003) present a sophisticated neural (and apparently associativemod)
model of transitive inference in a companion paper to Van Elzakker et al. (2003),
how are we certain that they have offered a deflationary associativeex alternative
for the cognitiveex accounts of transitive inference, as opposed to a neural
implementation of them? While Van Elzakker et al.’s value transfer account makes
slightly different predictions about the shape of learning curves than the Symbolic
Distance account of McGonigle and Chalmers (1992), they concede that BD
responding would improve on the 6-element task with additional training. Their
model also posits a variety of mechanisms for flexible representation which go
well beyond what one might expect of simpler associativeex approaches to learning
— for example, they interpret the hippocampal system as “construct[ing]
conjunctive representation[s] of the choice stimuli that resolves the ambiguity of
their associative values.” Furthermore, their model takes the hippocampal system
to record and constantly update excitatory strengths for each stimulus by tracking
the reward and co-occurrence history in these conjunctive representations,
effectively creating a ranking which will roughly reflect the position of those
stimuli as ordered along a monotonic dimension. Their account thus requires a
variety of processes which would seem cognitiveex according to many criteria. The
question remains as to whether their model counts as implementing an integrative,
cognitiveex solution to transitive inference, or rather specifies a sophisticated
associativeex alternative to it.

Problems with Purely Behavior-Based Criteria

Problem 1: It Matters How You Do It

The first observation, mirroring those found in the case of purely model-
based criteria, is that purely behavior-based criteria cannot be taken as self-
standing indicators of cognitionex. To illustrate this point, consider the debate over
John Searle’s (1980) (in)famous “Chinese Room” thought experiment. Searle
invites us to imagine a man who speaks no Chinese in a room with an input slot,
and output slot, drawers full of Chinese symbols, and a rulebook. The rulebook
contains a series of instructions about which Chinese symbols to put in the output
slot in response to certain symbols being presented at the input slot. We are also
invited to imagine that this rulebook is so comprehensive that native Chinese
speakers cannot distinguish the responses to queries released in the output slot by
the man in the room from those of a native Chinese speaker.
The intuition Searle (1980) expects us to have is that the man in the room
does not understand Chinese, because the procedure he uses to generate his
responses in fact involve no real intelligence. While Searle intended to generalize
this intuition to support a pessimistic appraisal of the prospects for true “artificial
intelligence,” others have derived from it a broader and more moderate lesson.
Block (1981), for example, has argued that the moral of the story is not that no
artificial system could be intelligent, but rather that it matters for the purported
intelligence of the system how it implements the behavioral capacities it appears to
display. If that system displayed a degree of internal functional sophistication that
- 338 -
went beyond the series of simple input-output rules suggested by Searle’s
rulebook, then it becomes increasingly more plausible to say that, in fact, the
whole system of the room — that is, the man together with the rulebook — does
understand Chinese.
The same moral carries over to behavioral criteria for cognitionex in the
present context, as behavior-based methods can be seen as input-output tests. The
problem arises from the Inductive Unity desiderata, for simple behavioral tests
could be passed by different processes which internally had little in common with
one another. Really, the idea of a merely associativeex “stimulus-response
automaton” that produced behavior truly indistinguishable, in all circumstances,
from that of a cognitiveex organism is not a live empirical possibility worth
worrying about. However, if we are to engage in an iterative investigation where
our behavioral criteria begin as fairly thinly-specified, this kind of concern can
remind us to regard these initial criteria with humility. It may be possible for a
“stimulus-response automata” to produce behavior which passes some of the
behavioral tests nominated above if we are not careful to implicitly constrain these
tests by appeal to certain forms of internal processing. In short, one must go further
than specify what an organism must do to count as cognitiveex; we must also link
these tests to some constraints on how the organism can do it.

Problem 2: The Motley Crew

A further problem for the behavior-based criteria for cognitionex and mere
associationex is that these criteria can appear quite diverse. While many
philosophers and psychologists have tended to focus on one or two criteria for
cognitionex such as systematicity or effortfulness in isolation, there seems little a
priori reason why these properties should truck together. Why would systematic
thought necessarily be effortful? The situation seems even more puzzling when we
make closer contact with the experimental work surrounding the distinction in
Standard Practice, as we can come up with scores of additional experimental tests
from a systematic review of the literature. Why, for example, should “fast
learning” and “ability to detect abstractions” both be part of cognitionex? Why
should an ability to detect abstractions permit context-sensitivity? Why should any
of these properties imply the use of multiple modalities? The questions become
even more baffling if we move to specific behavioral criteria at Newell’s (1973)
“phenomenal” level discussed in the case studies of the previous section; why
should an ability to pass the 5-element series imply an ability to find novel routes
in mazes or dissociate current from future motivations when planning for
breakfast?17
Again, it matters for the Inductive Unity desiderata whether they do. While
it may be possible to answer this question through brute force, by systematically

17
It is a live possibility — perhaps even a likely one — that these properties only co-occur due to
historical reasons. For example, they may all be due to the fact that evolutionary homologies cause
certain functionally-distinct brain structures to regularly develop together across a wide swathe of the
animal kingdom. That the properties attributed to cognitionex may only co-occur for historical
reasons, however, is no bar to its legitimacy as a respectable scientific posit — for example, see
Millikan (1999) and Boyd (1999).
- 339 -
investigating whether these properties co-occur across the full range of behaviors
and species of relevance to comparative psychology, we should surely avoid such
an inefficient method if possible. It is not clear that there are more efficient
methods available at the purely behavioral level, however. Rather, obtaining
answers to these questions more efficiently may require appeal to processing
stories inspired by reflection on the internal workings of particular models, which
can make predictions about the behavioral capacities which would be exhibited
under a wide variety of circumstances.

Problem 3: How much? Vague and Relational Criteria

Another persistent threat to the independence of behavior-based criteria is

that they are often specified in vague or relational terms. “Fast learning” is often
attributed to cognitionex in tests of learning set or reversal learning, for example;
but how fast must the learning be to count as cognitiveex? These vague, relational
criteria were found to be present in each of the criteria which could be derived
from the debates over the Raby et al. (2007) studies, for example. If we had
particular models which we could appeal to in order to provide anchor points for
the comparatives, we might be able to answer this challenge. We could specify, for
example, how many trials an organism will require to master a learning set, or the
range of different contexts in which an organism should be able to satisfy a goal,
for the processes behind those behaviors to count as cognitiveex.
Admittedly, simple answers which would apply invariantly across all
species or classes is unlikely to be in the cards; sloths might simply take longer to
cognizeex than hummingbirds. What is needed is not only the ability to reference a
particular model of cognitionex to garner an estimate of, for example, the number
of trials required to master a learning set, but rather also some indication of how to
fit the model to the particular abilities possessed by disparate species. As I will
suggest in the conclusion, having some idea as to how to tie the cognitiveex
abilities to underlying neuroanatomy and then comparing the similarities and
differences of the neuroanatomical structures possessed by the different species
might provide some guidance in this respect — showing both what different
organisms have in common and making quantitative predictions as to how they
might differ.

Problem 4: Mixed, “Unmoored” Criteria — Where Things Go Really Awry

Here is a pattern that has been repeated several times in the debate
surrounding the distinction which led directly to the current crisis: A comparative
cognition researcher begins investigating some behavioral capacity at the level of
Newell’s (1973) “phenomena.” At least two possible explanations for a
phenomena are proposed, one cognitive and the other associative. A behavior-
based criterion is proposed which would discriminate between those explanations,
and a successful experimental paradigm designed around that behavioral criteria
leads to a cottage industry in determining which animals can “pass the test.” Later,
however, that behavior-based test — initially justified in virtue of its ability to
distinguish between two specific model-based explanations — becomes
- 340 -
“unmoored” from the explanations from which it originally derived its
justification. For example, one or both of the original models might have fallen out
of favor. In other cases, a new associative(mod) model is proposed which is a slight
modification or tweak of the prior associative(mod) model for which the behavioral
criterion was devised, but which can now “pass” the test for cognitionex that the
prior associativemod model could not (and, often, do little extra besides). It is then
proclaimed that the original capacity was non-cognitiveex, or, worse, that the
distinction itself is untenable and should be rejected.
Examples of this pattern are not difficult to find: traditional associative
models were augmented with “value transfer” (von Fersen et al., 1991) to pass the
5-element transitive inference test; the comparator hypothesis was “extended” with
higher-order comparisons to allow it to show retrospective revaluation effects
(Denniston, Svastano, Blaisdell, & Miller, 2001); and multi-stimuli “elements”
were “added-“ or “replaced-“ to elemental models of learning to allow them to
solve certain configural and non-linear classification tasks (Wagner, 2008). While
it is always good to try to devise models that explain more data, all other things
being equal, turning around and then using such developments to criticize the
distinction itself, based on the fact that the augmented model can now pass a
behavioral test nominated only for its (now obsolete) ability to distinguish between
two specific explanations, is surely a mistake.
Certainly, it is possible that an increasing number of tweaks defeating
traditional tests for cognitionex, if derived from some unified, independent
justification, could indicate that a “dichotomous” approach to the psychology is
destined for the dustbin. The fact that there seems to be little structurally in
common across the tweaks mentioned in the previous paragraph, however, does
not offer the kind of “concerted accumulation of evidence” that would support this
narrative. Rather, it looks like a number of models can be extended with a number
of (possibly artificial) tweaks which help those models pass specific behavioral
tests. While again one should not fault psychologists for attempting to produce
models which can accommodate more data, I think we should be wary when
“unmoored” behavioral tests are deployed in criticisms of the distinction without
further argument justifying those behavioral tests themselves. For many of these
tests, I suggest that no further argument can be given. Divorced from specific
cognitiveex and associativeex models between which it can distinguish, there is little
reason to suppose that “passing” the five-element series test establishes anything of
significance.
No defensible justification is typically given, in these dialectics, to support
the conclusion that the tweaked associativemod model which passes the (now
unmoored) behavioral test for cognitionex in fact still describes capacities which
are associativeex. A significant possibility is that the tweaked model in fact
describes a process which is cognitiveex, or — much worse — is a gerrymandered
representation which does not really correspond to any actual process. If this is
right, then such models should not be offered as evidence for the triumphant
conclusion that all thought is really associativeex in nature.
What is needed are criteria to distinguish between “virtuous” and “vicious”
tweaks to models which would allow us to place the “genuine advances” to one
side and the “ad hoc tweaks” to the other. Standard scientific principles already
- 341 -
supply us with much to draw upon here — we should investigate whether the
modified models match other features of the “performance” data such as response
times, make novel predictions about behavioral capacities which are unrelated to
the behavioral tests they are explicitly designed to defeat, involve fewer free
parameters, are more biologically plausible, match patterns of deficits found in real
organisms or lesion models, and so on. How well the recent generation of
“tweaked” models will fare on these broader criteria for model selection remains to
be seen. At present, I simply mean to temper the enthusiasm encouraged by the
mere existence of an associativemod model which can defeat a behavioral test for
cognitionex.

Conclusion: How Comparative Psychology Might get by with

a Little Help from its Friends

From behaviorism, we inherited a fairly firm understanding of the most

basic forms of associationex. From classical computationalism (e.g., Newell &
Simon 1976), we also possess a fairly firm understanding of the most advanced
forms of cognitionex. The problem — the basic shape of which has changed little
from the time of Aristotle — is that we still don’t know how to divide up the great
middle ground between them, which surely constitutes the majority of the mental
capacities of humans and animals, and neither model- nor behavior-based criteria
are alone up to the task. Yet we have already conceded that both behavior- and
model-based criteria must be continuously revised as our continual accumulation
of experimental evidence allows us to improve our tenuous grasp on the
distinction, resulting in better models and behavioral predictions. How can we
engage in this iterative inquiry, constantly revising both model- and behavior-
based criteria, without, as Newell (1973) feared, having the distinction constantly
shift under our feet, preventing the accumulation of conceptual progress?
The rough solution is to carve up the middle ground such that both
associationex and cognitionex simultaneously fare as well as possible on the
aforementioned desiderata of Generalization, Empirical Reference, and Inductive
Unity. The challenge, of course, is that we do not yet know the best way to do this
given the current state of empirical knowledge available to us in psychology. What
is needed is thus a way to structure an iterative inquiry which makes optimal use of
what little we do know to gradually sharpen the distinction while minimizing the
chance that Newell’s (1973) fears will be realized.
My suggestion is that the best way to do this is to be guided by other
sciences which have often been supposed “irrelevant” to the ontology of
psychology. Much ink has been spilt about the purported “autonomy” of
psychology (Fodor, 1974, 1997; Jones, 2004). It is important to note, however, that
these autonomy arguments, even if successful, establish only the freedom of
psychology from metaphysical reduction to lower-level sciences, and even then
only on relatively strong conceptions of reduction. All that is needed in the present
context are methods to estimate the suitability along the aforementioned desiderata
of various ways of grouping together behavioral properties and psychological
models under the labels “cognitiveex” and “associativeex.” We already know a great
deal about comparative neuroanatomy which generates predictions about the
- 342 -
presence or absence of behavioral capacities. The same goes, to perhaps a lesser
degree, with evolutionary biology, under the assumption that species with shared
phyletic backgrounds and ecological pressures will tend to have similar
psychological capacities and behavioral abilities. Attempting to assess the
legitimacy of the distinction in Standard Practice without making use of this
clearly-relevant data seems to me like trying to run a marathon with one leg
hobbled.
Being guided by neuroscience and evolutionary biology is not a panacea,
of course. In many ways, the interdisciplinary move simply introduces additional
moving parts into the debate, and issues of Generality, Empirical Reference, and
Inductive Unity arise again in other sciences. Neuroanatomical regions can
themselves be described at varying levels of abstraction, and it is not always clear
which level is appropriate when linking a behavioral capacity to its neural
substrates (Craver, 2009). Though I have elsewhere defended the importance of
medial temporal lobe structures as the locus of representational flexibility
characteristic of cognitionex (Buckner, 2011), there have been significant
difficulties in cleanly delineating this region from nearby brain structures, and
there continue to be heated debates as to the region’s function — for example
whether it is best construed as primarily spatial (O’Keefe & Nadel, 1978), as
supporting all forms of relational memory (Dusek & Eichenbaum, 1997), or as an
all-purpose declarative memory system (Squire, 1992).
Progress in neuroscience, however, is arguably more stable than progress
in psychology. While our divisions at the level of gross neuroanatomy have
changed dramatically over the last hundred years, some accumulation of
information has clearly been achieved. The exaggerated rhetorical differences
between competing research groups working on medial temporal lobe function
obscures the vast amount of ground now shared between them. There is now little
doubt that the medial temporal lobes play a critical role in a variety of cognitiveex
capacities such as cognitive mapping, episodic memory, and transitive inference.
Similarly, we have made some significant progress in recent years in establishing
homologies in brain structures between disparate classes, such as the verdict of
homology between the mammalian and avian hippocampus (Columbo &
Broadbent, 2000; Reiner et al., 2004). These hard-won verdicts summarize vast
amounts of anatomical and behavioral data, have been endorsed by a consensus of
the neuroscientific community, and carry significant implications about the
cognitiveex capacities that disparate species are likely to possess. These
neuroscientific findings can thus ground an iterative inquiry with a reasonable
hope of avoiding stagnation in Newell’s (1973) quagmire.
The best way to make gradual progress on our quarry is to begin with
model-based criteria which are focused on large-scale differences in gross
neuroanatomy — at a balanced level of abstraction which is specific enough to
illustrate compelling differences in the functional capacities of different areas of
the brain while still abstracting away from details likely to be unique to particular
species and models which will be revised as our future neuroscience evolves. We
should then develop behavior-based criteria which distinguish between these
models, and then use the evidence we accumulate by testing animals on these
criteria to elaborate those models. I have suggested (Buckner, 2011) that the
- 343 -
distinction in the cortico-hippocampal model of Gluck and Myers (2001) provides
an example of how such a balance might be struck, with learning processes
mediated by the medial temporal lobes (and functional homologues in other
classes) uniting the cognitiveex side of the distinction and less flexible learning
systems such as the cortices and spinal cord uniting the associativeex side. Other
neurally-based schemes should of course be explored — though I know of no story
with as wide and relevant a scope as that based on the medial temporal lobes.
Like any model, Gluck and Myers’ (1994) involves a number of
idealizations which may be discarded as our knowledge develops. Nevertheless,
their model is based in gross differences in underlying neuroanatomy — in
particular on the well-established flow of highly-processed sensory information
through the entorhinal cortex, to the CA3 region, and back out to the same
stimulus input stream — which are common across many species and likely to be
conserved in future discoveries. However, their model is still specific enough to
issue a variety of predictions which appear to be borne out by currently-available
experimental data for diverse species and learning paradigms. Thus, a distinction
supported by this or similar neuroanatomically-guided models would seem to fare
well on our three desiderata.
That all being said, monistic or pluralistic perspectives on animal
psychology may still turn out to be correct — and Standard Practice might turn out
to be misguided. The comments in this concluding section were meant as a defense
of how we should conduct our investigations to obtain a fair assessment of the
hypothesis that the dichotomous perspective presumed by Standard Practice is
cogent. If we had the “optimally predictive global model of the mind” which
outperformed all other models on all tasks, we could simply read off the legitimacy
of the distinction from that model by seeing whether it posits one, two, or many
basic kinds of psychological process. In the absence of this optimal model,
however, we should continue to simultaneously explore all plausible perspectives
— and to the winner go the spoils. What we should not continue to do, I have
argued above, is haphazardly mix model- and behavior-based approaches in our
assessments of the cogency of Standard Practice.

References

Allen, C. (2006). Transitive inference in animals: Reasoning or conditioned associations?

In S. Hurley & M. Nudds (Eds.), Rational animals? (pp.175-185). Oxford, UK:
Oxford University Press.
Allen, C., Grau, J., & Meagher, M. (2009). The lower bounds of cognition: What do spinal
cords reveal? In J. Bickle (Ed.), The Oxford handbook of philosophy of
neuroscience (pp. 129-142). Oxford, UK: Oxford University Press.
Aristotle. (1925). Nicomachean ethics. Trans. W. D. Ross. In R. McKeon (Ed.),
Introduction to Aristotle (pp.151-249). New York: Random House.
Aristotle. (1928). Posterior analytics. Trans. G. R. G. Mure. In R. McKeon (Ed.),
Introduction to Aristotle (pp.151-249). New York: Random House.
Aristotle. (1931). De anima. Trans. J. A. Smith. In R. McKeon (Ed.), Introduction to
Aristotle (pp.151-249). New York: Random House.
Blaisdell, A. (2009). The role of associative processes in spatial, temporal, and causal
cognition. In S. Watanabe, L. Huber, A. P. Blaisdell, & A. Young (Eds.), Rational
animal, irrational human. Tokyo, Japan: Keio University Press.
- 344 -
Blaisdell, A., Sawa, K., Leising, K., Waldmann, M. (2006). Causal reasoning in rats.
Science, 311, 1020-22.
Block, N. (1981). Psychologism and behaviorism. Philosophical Review, 90, 5 - 43.
Boyd, R. (1999). Kinds, complexity and multiple realization. Philosophical Studies, 95, 67-
98.
Bryant, P., & Trabasso, T. (1971). Transitive inferences and memory in young children.
Nature, 232, 456-458.
Buckner, C. (2011). Learning from mistakes: Error-correction and the nature of cognition.
(Doctoral dissertation). ProQuest/UMI. (Publication No. 3456443)
Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press.
Clayton, N., Emery, N., & Dickinson, A. (2006). The rationality of animal memory: The
cognition of caching. In S. Hurley & M. Nudds (Eds.), Rational animals? (pp.
197–216). Oxford, UK: Oxford University Press.
Clayton, N., & Dickinson, R. (2009). Are animals stuck in time or are they chronesthetic
creatures? Topics in Cognitive Sciences, 1, 59-71.
Columbo, M., & Broadbent, N. (2000). Is the avian hippocampus a functional homologue
of the mammalian hippocampus? Neuroscience and Biobehavioral Reviews, 24,
465-484.
Craver, C. (2009). Mechanisms and natural kinds. Philosophical Psychology, 22, 575-594.
De Lillo, C., Floreano, D., & Antinucci, F. (2001). Transitive choices by a simple, fully
connected, backpropagation neural network: Implications for the comparative
study of transitive inference. Animal Cognition, 4, 61-68.
Dennett, D. (1983). Intentional systems and cognitive ethology — the ‘Panglossian
paradigm’ defended. Behavioral and Brain Sciences, 6, 343-390.
Denniston, J., Svastano, H., Blaisdell, A., & Miller, R. (2001). The extended comparator
hypothesis: Learning by contiguity, responding by relative strength. In R. Mowrer
& S. Klein (Eds.), Handbook of contemporary learning theories (pp. 65-117).
Mahway, NJ: Earlbaum.
Dusek, J., & Eichenbaum, H. (1997). The hippocampus and memory for orderly stimulus
relations. Proceedings of the National Academy of the Sciences, 94, 7109-7114.
Elman, J. (1996). Rethinking innateness: A connectionist perspective on development.
Cambridge, MA: MIT Press.
Estes, W. K. (1950). Toward a statistical theory of learning. Psychological Review, 57, 94-
104.
Fodor, J. (1974). Special sciences, or the disunity of science as a working hypothesis.
Synthese, 28, 97-115.
Fodor, J. (1997). Special sciences: Still autonomous after all these years. Philosophical
Perspectives, 11, 149-163.
Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A
critical analysis. Cognition, 28, 3-71.
Fountain, S., & Doyle, K. (this issue). Association and abstraction in sequential learning:
‘What is learned?’ revisited. International Journal of Comparative Psychology.
Frank, M., Rudy, J., & O'Reilly, R. (2003). Transitivity, flexibility, conjunctive
representations and the hippocampus: II. A computational analysis. Hippocampus,
13, 341-354.
Gardner, H. (1985). The mind’s new science: A history of the cognitive revolution. New
York: Basic Books.
Gigerenzer, G., & Regier, T. (1996). How do we tell an association from a rule? Comment
on Sloman (1996). Psychological Bulletin, 119, 23-26.
Gluck, M., & Myers, C. (1994). A neurocomputational theory of hippocampal function in
stimulus representation and learning. In S. Zournetzer, J. Davis, T. Mckenna, & C.

- 345 -
 Lau (Eds.), An introduction to neural and electronic networks, 2nd ed. (pp. 77-90).
San Diego, CA: Academic Press.
Gluck, M., & Myers, C. (2001). Gateway to memory: An introduction to neural network
modeling of the hippocampus and learning. Cambridge, MA: MIT Press.
Hirsh, R. (1974). The hippocampus and contextual retrieval of information from memory:
A theory. Behavioural Biology, 12, 421-444.
Hull, C. (1943). Principles of behavior. New York: Appleton-Century.
Hume, D. (1748/1993). An enquiry concerning human understanding. Trans. E. Steinberg.
Indianapolis, IN: Hackett Publishing.
Hurley, S. & Nudds, M. (2006). Rational animals? Oxford, UK: Oxford University Press.
Jones, T. (2004). Special sciences: Still a flawed argument after all these years. Cognitive
Science, 28, 409-432.
Mackintosh, N. (2002). Do not ask whether they have a cognitive map, but how they find
their way about. Psicológica, 23, 165-185.
McGonigle, B., & Chalmers, M. (1977). Are monkeys logical? Nature, 267, 694-696.
McGonigle, B., & Chalmers, M. (1992). Monkeys are rational. Quarterly Journal of
Experimental Psychology, 45, 189-228.
McGonigle, B., & Chalmers, M. (2006). Ordering and executive functioning as a window
on the evolution and development of cognitive systems. International Journal of
Comparative Psychology, 19, 241-267.
Menzel, E. W. (1973). Chimpanzee spatial memory organization. Science, 822, 943-945.
Menzel, E. W. (1978). Cognitive mapping in chimpanzees. In S. H. Hulse, H. F. Fowler, &
W. K. Honig (Eds.), Cognitive processes in animal behavior (pp. 375-422).
Hillsdale, NJ: Erlbaum.
Millikan, R. (1999). Historical kinds and the special sciences. Philosophical Studies, 95,
45-65.
Mishkin, M., & Petri, H. L. (1984). Memories and habits: Some implications for the
analyses of learning and retention. In L. R. Squire & N. Butters, (Eds.),
Neuropsychology of memory (pp. 287-296). New York: The Guilford Press.
Mitchell, C. J., De Houwer, J., & Lovibond, P. F. (2009). The propositional nature of
human associative learning. Behavioral and Brain Sciences, 32, 183-246.
Morgan, C. L. (1894). An introduction to comparative psychology. London: Walter Scott.
Newell, A. (1973). You can’t play 20 questions with nature and win: Projective comments
on the papers of this symposium. In W. G. Chase (Ed.), Visual information
processing (pp. 283-308). New York: Academic Press.
Newell, A & Simon, H. A. (1976). Computer science as empirical inquiry: Symbols and
search. Communications of the Association for Computing Machinery, 19(3), 113–
126.
O’Keefe, J., & Nadel, L. (1978). The hippocampus as cognitive map. Oxford, UK:
Clarendon Press.
Olton, D. (1979). Mazes, maps, and memory. American Psychologist, 34, 583-596.
Papineau, D., & Heyes, C. (2006). Rational or associative? Imitation in Japanese quail. In
S. Hurley & M. Nudds, (Eds.), Rational animals? (pp. 198-216). Oxford, UK:
Oxford University Press.
Parker, W. (in press).Scientific models and adequacy for purpose. The Modern Schoolman,
87.
Penn, D., & Povinelli, D. (2007). Causal cognition in human and nonhuman animals: A
comparative, critical review. Annual Review of Psychology, 58, 97-118.
Porphyry. (2000). On abstinence from killing animals. Trans. Gillian Clark. London:
Duckworth Press.
Raby, C., Alexis, D., Dickinson, A., & Clayton, N. (2007). Empirical evaluation of mental
time travel. Behavioral Brain Sciences. 30, 330-331.
- 346 -
Radick, G. (2000). Morgan’s canon, Garner’s phonograph, and the evolutionary origins of
language and reason. The British Journal for the History of Science, 33, 3-23.
Reiner, A., Perkel, D., Bruce, L., Butler, A., Csillag, A., Kuenzel, W. . . .Jarvis, E. D.
(2004). Revised nomenclature for avian telencephalon and some related brainstem
nuclei. The Journal of Comparative Neurology, 473, 377-414.
Rescorla, R. & Wagner, A. (1972). A theory of Pavlovian conditioning: Variations in the
effectiveness of reinforcement and nonreinforcement. In A. Black & W. Prokasy
(Eds.), Classical Conditioning II (pp. 64-99). New York: Appleton-Century-
Crofts.
Rumelhart, D., McClelland, J., & the PDP research group. (1986). Parallel distributed
processing: Explorations in the microstructure of cognition. Volume I.
Cambridge, MA: MIT Press.
Searle, J. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3, 417-457.
Shanks, D., & Pearson, S. (1987). A production system model of causality judgment.
Proceedings of the Ninth Annual Conference of the Cognitive Science Society (pp.
210-220). Hillsdale, NJ: Lawrence Erlbaum Associates, Inc.
Shettleworth, S. (2007). Cognition, evolution, and behavior, 1st ed. Oxford, UK: Oxford
University Press.
Shettleworth, S. (2010). Cognition, evolution, and behavior, 2nd ed. Oxford, UK: Oxford
University Press.
Skinner, B. F. (1953). Science and human behavior. New York: Macmillan.
Smith, L. (2000). Avoiding association when it’s behaviorism you really hate. In R.
Golinkoff & K. Hirsh-Pasek (Eds.), Breaking the word learning barrier (pp. 169-
174). Oxford, UK: Oxford University Press.
Sloman, P. (1996). The empirical case for two systems of reasoning. Psychological
Bulletin, 119, 3-22.
Smolensky, P. (1988). On the proper treatment of connectionism. Behavioral and Brain
Sciences, 11, 1-23.
Sober, E. (2005). Comparative psychology meets evolutionary biology – Morgan’s Canon
and cladistic parsimony. In L. Daston & G. Mitman (Eds.), Thinking with animals
– New perspective on anthropomorphism (pp. 85-99). New York: Columbia
University Press.
Sorabji, R. (1993). Animal minds and human morals. New York: Cornell Press.
Squire, L. (1992). Memory and the hippocampus: A synthesis of findings with rats,
monkeys, and humans. Psychological Review, 99, 195-231.
Suddendorf, T., & Corballis, M. C. (2007). New evidence of animal foresight. Animal
Behaviour, 75, e1-e3.
Tolman, E. (1948). Cognitive maps in rats and men. The Psychological Review, 55, 189-
208.
Tulving, E. (1983). Elements of episodic memory. Oxford, UK: Clarendon Press.
Tulving, E. (1985). Memory and consciousness. Canadian Psychology, 26, 1-12.
Van Elzakker, M., O’Reilly, R., & Rudy, J. (2003). Transitivity, flexibility, conjunctive
representations, and the hippocampus: I. An empirical analysis. Hippocampus,
13(3), 334-340.
von Fersen, L., Wynne, C., Delius, J., & Staddon, J. (1991). Transitive inference formation
in pigeons. Journal of Experimental Psychology: Animal Behavior Processes, 17,
334-341.
Wagner, A. (2008). Evolution of an elemental theory of Pavlovian conditioning. Learning
and Behavior, 36, 253-265.
Wallace, D., & Fountain, S. (2002). What is learned in sequential learning? An associative
model of reward magnitude serial-pattern learning. Journal of Experimental
Psychology: Animal Behavior Processes, 28, 43-63.
- 347 -
Walker, J. & Olton, D. (1979). The role of response and reward in spatial memory.
Learning and Motivation, 10, 73-84.
Wasserman, E., & Zentall, T. (2006). Comparative cognition: A natural science approach
to the study of animal intelligence. In E. Wasserman & T. Zentall (Eds.),
Comparative cognition (pp. 619-636). Oxford, UK: Oxford University Press.
Watson, J. B. (1928). The ways of behaviorism. New York: Harper.
Wynne, C. (2004). The perils of anthropomorphism. Nature, 428, 606.
Zentall, T. (2001). The case for a cognitive approach to animal learning and behavior.
Behavioural Processes, 54, 65-78.

- 348 -