COGNITION

Cognition 65 (1998) 197–230

Similarity and rules: distinct? exhaustive?

empirically distinguishable?
Ulrike Hahn*, Nick Chater
Department of Psychology, University of Warwick, Coventry CV4 7AL, UK

Abstract

The distinction between rule-based and similarity-based processes in cognition is of funda-

mental importance for cognitive science, and has been the focus of a large body of empirical
research. However, intuitive uses of the distinction are subject to theoretical difficulties and
their relation to empirical evidence is not clear. We propose a ‘core’ distinction between rule-
and similarity-based processes, in terms of the way representations of stored information are
‘matched’ with the representation of a novel item. This explication captures the intuitively
clear-cut cases of processes of each type, and resolves apparent problems with the rule/
similarity distinction. Moreover, it provides a clear target for assessing the psychological
and AI literatures. We show that many lines of psychological evidence are less conclusive
than sometimes assumed, but suggest that converging lines of evidence may be persuasive.
We then argue that the AI literature suggests that approaches which combine rules and
similarity are an important new focus for empirical work.  1998 Elsevier Science B.V.

Keywords: Similarity-based process; Rule-based process

1. Introduction

The contrast between rule- and similarity-based accounts of cognition is central to

cognitive science. The two approaches correspond to different research traditions,
and the contrast between them is the focus of vigorous empirical and theoretical
debate across a wide range of cognitive domains.
The idea that cognition involves following mental rules lies at the heart of the
classical picture of the cognitive system (Newell and Simon, 1990). Mental rules
encode general facts about the world and these facts are applied to specific instances
* Corresponding author. e-mail: [email protected]

0010-0277/98/$19.00  1998 Elsevier Science B.V. All rights reserved

PII S00 10-0277(97)000 44-9
198 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

in cognitive activity. A paradigm example is language: linguistics aims to specify

rules which explicate the structure of language, and it is assumed that these rules are
mentally represented and applied in language processing. The same pattern is
assumed to hold in knowledge of naive physics (Hayes, 1979), arithmetic (Young
and O’Shea, 1981), social conventions (Cheng and Holyoak, 1985), and so on. In all
these cases, knowledge is stored in collections of rules, which are organized into
theories. These are assumed to have the same structure as explicitly-described
theories in science: collections of general statements from which predictions and
explanations for specific aspects of the everyday world can be constructed. The
emphasis on rules is embodied in many formalisms used in psychological modeling,
most directly in systems based on production rules (Newell and Simon, 1972;
Anderson, 1983; Newell, 1991) or on logical inference (Inhelder and Piaget,
1958; Braine, 1978; Rips, 1994). It is also embodied in much practical artificial
intelligence research (within what Haugeland calls GOFAI – good old fashioned AI
(Haugeland, 1985)) ranging from early game-playing programs (Newell, 1963) to
expert systems (Dayal et al., 1993).
Similarity, in conjunction with sets of stored instances, suggests an alternative
model of cognition. Instead of deriving general rules concerning the structure of the
world, past situations (‘instances’ in psychology; ‘cases’ in AI) are stored in a
relatively unprocessed form. Reasoning concerning a new situation depends on its
similarity to one or more past situations. Here, we shall call such methods similarity-
based reasoning, to emphasise the centrality of similarity. Such theories have been
proposed in many contexts: in exemplar theories of concepts (Medin and Schaffer,
1978; Nosofsky, 1984), in instance-based models of implicit learning (Berry and
Broadbent, 1984, 1988; Vokey and Brooks, 1992; Redington and Chater, 1996), in
theories of reasoning (Ross, 1984, 1987; Ross and Kennedy, 1990), in ‘case-based
reasoning’ in AI (Kolodner, 1991; Aamodt and Plaza, 1994), and ‘lazy learning’
(Aha, 1997) in machine learning. Moreover, similarity-based approaches to cogni-
tion are strongly rooted in behaviorist theories of human learning. The fundamental
behaviorist claim is that behavior is mediated by a set of stimulus-response associa-
tions. As no stimulus is exactly the same as any previously-encountered stimulus,
behavior must depend on some form of generalization, depending on the similarity
between the new stimulus and previous stimuli. Thus, similarity has been stressed by
behaviorists in psychology (Pavlov, 1927) and philosophy (Quine, 1960), as well as
being important in cognitive science.
The difference between rule- and similarity-based accounts is clearly of central
theoretical importance to cognitive science, but can they be distinguished empiri-
cally? There have been many attempts to do so in areas of cognitive psychology
as diverse as categorization (Komatsu, 1992), implicit learning (Reber, 1989;
Shanks and John, 1994), problem-solving (Gentner, 1989) and the development of
reading skills (Goswami and Bryant, 1990). Our own research has concentrated
on empirically distinguishing specific rule-based accounts from similarity-based
(and other) accounts in the context of language and implicit learning (Redington
and Chater, 1994, 1996; Nakisa and Hahn, 1996; Hahn et al., 1997; Nakisa et al.,
1998).
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 199

However, the interpretation of the empirical evidence is difficult, because both

classes of account are very heterogeneous, and hence each can capture a wide
range of data. Possibly, these two classes are too broad to really allow an overall
empirical assessment. Perhaps, the best empirical research can do is to test particular
models of each kind, not ‘rules’ or ‘similarity’ generally. Furthermore, the empirical
literature contains many confusingly correlated distinctions such as symbolic versus
subsymbolic, abstract versus specific or deductive versus inductive. The resulting
problems of interpretation which we have encountered in our own work have
acted as a personal motivation for developing the ideas in this paper: specifically,
to provide a core conceptual distinction between rule- and similarity-based
models as a framework for interpreting empirical and computational considera-
tions.
One reaction to the difficulties encountered in distinguishing between such broad
classes of account is simply to abandon the attempt. However, we believe that this is
at best a last resort: there are strong reasons to attempt to maintain the rule- versus
similarity-based distinction. If viable, it allows general theoretical statements to be
made about broad classes of account, which are not tied to specific models or
implementations. This is important for relating different theoretical proposals, and
for unifying what must otherwise remain fractionated literatures on different cog-
nitive domains. Moreover, the viability of a general distinction has been routinely
presupposed by the empirical literature, wherever the terms ‘rule-’ or ‘similarity-
based’ are used without further clarification, such as in the many experimental
studies which seek to distinguish rule- and similarity-based reasoning on the basis
of their putative effects, without further commitment to more specific models (e.g.
Reber, 1989; Shanks, 1995). The issue of whether a general distinction can be
maintained is therefore clearly a pressing one. This paper has three main sections.
In Section 2, we show how both classes of account appear so general that they seem
to collapse into each other. Section 3 provides a core account which successfully
separates rule- and similarity-based processes. We show that this core account
correctly decides intuitively clear cases of each type more adequately than a
range of alternative criteria that might be suggested. In Section 4, we re-assess
how empirical and computational evidence can be brought to bear on distinguishing
between the two classes of model and, finally, consider implications for future
research.

2. Rules and similarity: the problem

The intuitive notions of both rule- and similarity-based processes seem alar-
mingly general1. Almost any aspect of thought may be viewed as determined by

1
Moreover, the very notions of ‘rule’ and ‘similarity’ have been attacked in the philosophical literature
(Goodman, 1972; Kripke, 1982). However, these problems are so general that they threaten the entire
program of cognitive science, rather than providing specific difficulties for the present debate (Hahn,
1996; Hahn and Chater, 1997).
200 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

rules, at least in the sense that laws of nature are rules of a kind; and similarity seems
an essential ingredient of an extremely wide range of paradigms and phenomena –
connectionism, case-based reasoning, exemplar- and prototype-theories, and possi-
bly even metaphor and analogy.
The threat that follows from the generality of both ‘rule’ and ‘similarity’ can be
illustrated by the apparent possibility of each account ‘mimicking’ the other.
First, as suggested by Nosofsky et al. (1989), ‘rule’ can be used to include
procedures for computing similarity as special cases. Indeed, specific theories of
similarity, such as geometric models (Shepard, 1980) or the contrast model
(Tversky, 1977) appear to provide suggestions about what this rule might be.
Second, ‘similarity’ appears so general that it can include any rule. Suppose we
view a rule, R, as a function from inputs to outputs. Define a dissimilarity measure,
D, such that
D(x, y) = 0 iff R(x) = R(y)

D(x, y) = 1 otherwise
That is, two inputs are similar when the rule gives the same output for both and
dissimilar otherwise.
Therefore, similarity-based reasoning might be viewed as involving a kind of
rule; and rule-based reasoning might be viewed as involving a kind of similarity.
The notions seem so general that they collapse into each other.
The artificiality of this ‘mimicry argument’ may lead one to underestimate the
extent of the problem. However, more realistic variants abound. Allen and Brooks
(1991) discuss ‘additive rules of thumb’ of the form ‘At least two of (long legs,
angular body, spots) then builder.’ These rules, however, are equivalent to a special
case of a psychological prototype model (Smith and Medin, 1981) – where the
prototype is defined by n features, of which m must be present – which seemingly
involves similarity comparison of the new item with the prototype. Moreover, the
same behavior can be obtained from a single-layer connectionist network with a
linear threshold unit. Therefore, identical behavior appears consistent with rules and
similarity, as well as with connectionist networks.
Connectionist networks themselves further illustrate the problem, in that they
might be seen to fall in both camps. Back-propagation networks are often described
as depending on similarity (Rumelhart and Todd, 1993). However, they are also
often described as using ‘implicit rules’ which can be extracted using appropriate
analysis (Bates and Elman, 1993; Hadley, 1993; Andrews et al., 1995; Davies,
1995). Therefore, back-propagation networks appear rule- and similarity-based.
These concerns suggest that the intuitively sharp distinction between rule- and
similarity-based processing may be illusory. If this conclusion is accepted, then the
empirical debate aimed at testing between the two is futile. We will argue that this
pessimistic conclusion is not justified, that a core distinction can be made, and that
empirical evidence, both from experimental and computational sources, can be
brought to bear on whether specific cognitive processes are similarity-based, rule-
based, or neither.
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 201

3. Rules versus similarity: an explication

An explication of the core distinction between rules and similarity must balance
two forces. It must be sufficiently specific that it solves the problems of generality
that we have outlined. However, it must also be sufficiently general to take in the
great diversity within each type of account. Thus, rule-based processes may invoke
symbolic statements with logical connectives, with or without explicit variables (as
in classical AI, or some parts of the psychology literature (Nosofsky et al., 1989;
Sloman, 1996)); they may operate over banks of connectionist units (Touretzky and
Hinton, 1988) or have the form of the additive rules of thumb (Allen and Brooks,
1991) mentioned above. Equally, similarity-based models range from case-based
reasoning (CBR) systems in AI, where similarity is assessed between graph struc-
tures (Branting, 1991), to spatial and set-theroretic models in psychology where
similarity is defined in terms of spatial distance or feature overlap, respectively
(Shepard, 1957; Tversky, 1977).
One approach to constructing a core distinction proposes that the two classes can
be distinguished because they use different types of representation. Perhaps rules
contain variables but things entering into similarity comparisons do not; or rules are
general whereas similarity-based reasoning applies to specific claims (e.g. describ-
ing specific instances)2; or rules are rigid, whereas representations used in similarity
comparison are in some sense fuzzy. Whether explicitly or implicitly, such criteria
underlie many definitions of rule-following in cognitive science (e.g. Sloman,
1996).
This focus on different types of representation is undermined by the fact that the
very same representation can be used both in rule- and similarity-based processing.
Consider a representation of the information that monkeys like bananas. This can be
used as a rule, on encountering a particular monkey, and classifying it as liking
bananas. However, it can also be used in similarity-based reasoning in proposing the
generalization that gorillas also like bananas. The core distinction cannot simply be
based on different types of representation; rather, it must involve the way in which
representations are used.
To clarify, let us consider a specific scenario. Suppose that we are presented with
a new item, which is represented by the features {large, barks, brown, furry, has-
teeth... }. To classify this item, we must somehow relate its representation to our
existing knowledge. Rule- and similarity-based processes differ regarding the way
the representation of the new item is integrated with existing knowledge.
A paradigmatic case of rule-based processing runs as follows. Existing knowledge
is stored in conditional rules (e.g. ‘if something barks and is furry, then it is a dog’).
If the antecedent of a rule is satisfied (it barks and is furry), then the category in the
consequent applies (it is a dog). A paradigm case of similarity-based processing is as
2
From a logical point of view, a natural fromulation of this type of claim is that rules involve universal
quantification, whereas similarity is defined over instances which are represented by existential quanti-
fication. Aside from the difficulty pointed out below, this approach collapses because of the purely logical
result that any representation involving existential quantification can be converted into a sentence invol-
ving universal quantification, and vice versa, by applying negation.
202 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

follows. Knowledge is stored as a set of past instances, with associated category

labels. The new item is classified as a dog if the past instance to which it is most
similar was classified as a dog. In both paradigms, there is a ‘match’ between the
representation of the new item and a representation of stored knowledge. Crucially,
however, the nature of the matching process differs in two ways.
First, the antecedent of the rule must be strictly matched, whereas in the similarity
comparison matching may be partial. In strict matching, the condition of the rule is
either satisfied or not – no intermediate value is allowed. Partial matching, in
contrast, is a matter of degree – correspondence between representations of novel
and stored items can be greater or less. Notice that there is no restriction on the
nature of the representation that is matched, whether strictly or partially. Our exam-
ple is implicitly a conjunction (the item must be furry and a dog for the rule to be
satisfied), but the condition of a rule could equally well be disjunction (furry or a
dog), or have any form whatever.
Second, the rule matches a representation of an instance (large, barks, brown,
furry, has-teeth...) with a more abstract representation of the antecedent of the rule
(barks, furry), whereas the similarity paradigm matches equally specific representa-
tions of new and past items. The antecedent ‘abstracts away’ from the details of the
particular instance, focusing on a few key properties.
Note, crucially, that abstraction here is relative, not absolute. Thus, a similarity-
based process could operate over highly abstract representations, such as logical
forms of sentences. Rule-based processes can apply to arbitrarily specific represen-
tations (e.g. specifying minute detail about the perceptual properties of the objects it
applies to) if the representations of new objects are even more detailed. Thus evi-
dence for highly abstract mental representations is not thereby evidence for rule-
based processing; and evidence for highly specific mental representations is not
evidence for similarity-based processing. This point will be important in our reeva-
luation of empirical criteria for distinguishing between rule- and similarity-based
processes below.
The need for relative abstraction stems from its link with generalization. If the
antecedent of a rule were as specific as the representation of the instance, then it
would apply to at most this single instance and thus provide no basis for generalizing
our knowledge about one case to another. Indeed. a system containing rules of this
form is simply a ‘memory bank’ of instances and their classifications.
Thus our paradigm cases differ along two dimensions, defining a space of possi-
bilities illustrated in Fig. 1.
This space provides a useful framework for differentiating ‘rules’ and ‘similarity’
because it allows us to think generally about the different ways in which stored
knowledge can be applied to enable the processing of novel items. The paradigm
case of processing corresponds to the top right corners ‘strict matching/abstraction.’
The paradigm case of similarity-based reasoning corresponds to the bottom left
corner, ‘partial matching/no-abstraction.’ Note that these locations are sufficiently
general to be compatible with the diverse array of specific instantiations of ‘rule’ and
‘similarity’ mentioned at the beginning of this section: strict matching to an abstrac-
tion is not a notion which refers to particular rule formats, nor does partial matching
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 203

Fig. 1. The space of possibilities for representation matching.

to an instance distinguish between the matching of graph structures, feature overlap

or occupying a nearby point in similarity space.
The bottom right corner, ‘strict-matching/no-abstraction’ corresponds to the
‘memory bank.’ What about the final corner, at the top left, combining partiality
and matching to an abstraction? A candidate for this class are versions of prototype
theory, which relax the classic definitional account: i.e. prototypes are construed as
lists of core features (hence abstraction) of which only sufficiently many, but not all,
must be matched by a new instances (see Komatsu, 1992), thus replacing strict with
partial matching. This raises the possibility, which we discuss below, that similarity
may encroach into this corner of the space.
Outside the space entirely are processes which do not involve matching the novel
instance to stored representations of any kind and, hence, are alternatives to both
rule- and similarity-based accounts. For example, generalization might be based on
simple failure to discriminate different perceptual stimuli, rather than on stored
knowledge. More interestingly, an input-output mapping might be performed with-
out consulting any stored representations.

3.1. Why representations matter

We have explained rule- and similarity-based processing in terms of matching

between representations: of the new item and of stored knowledge. Crucially, it is
not sufficient for a process to behave ‘as if’ it were matching representations. We
now illustrate why this is so, considering rules and similarity in turn.
Regarding rules, the issue is the vital distinction between rule following
and merely rule describable behavior (Chomsky, 1980, 1986; Searle, 1980;
Dreyfus, 1992; Smith et al., 1992; Marcus et al., 1995). Rule-based reasoning
implies rule-following: that a representation of a rule causally affects the behavior
204 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

of the system, and is not merely an apt summary description3. Thus, only claims
about rule-following are claims about cognitive architecture. To illustrate with the
classic example, the planets exhibit rule-describable behavior, concisely predicted
by Newton’s laws, but the planets do not rely on mental representations of Newton’s
laws to determine their orbits; thus, they are not rule-following. By contrast, explain-
ing why a motorist obeys traffic regulations makes reference to mental states, i.e.
knowledge of these regulations. Hence, the behavior in question exhibits rule-fol-
lowing. As any regularity can be stated in a format which fulfils our intuitions about
‘rule’, (e.g. as a universally quantified statement) any regular behavior would be
‘rule-based’ if the distinction between rule-following and merely rule-describable
were not maintained; the notion of rule-based processing would collapse into trivi-
ality.
Analogous considerations apply to similarity. Unless similarity-based models
involve comparison between representations, then any generalization (rule-based,
similarity-based or even non-representational) can be viewed as similarity-based in
the sense that items to which generalization applies can, by virtue of this fact, be
viewed as ‘similar.’ However, such post hoc measures of similarity have no expla-
natory value. If the constraint of representation-matching is relaxed, the splashes of
similar rocks thrown into water could be viewed as similarity-based processing on
part of the water, given that they cause similar splashes.
For the rule versus similarity debate to be meaningful, matching must apply to
actual representations of rules and instances. Consequently, non-representational
approaches to cognition, such as situated robotics (Brooks, 1991), stand outside
this debate altogether. Furthermore, mere procedures cannot constitute rule-based
reasoning. Some confusion over this exists with respect to inference rules in cogni-
tion, such as modus ponens. Smith et al. (1992), for instance, distinguish rule-
following and -describable behavior (they call the latter ‘conforming’ to a rule)
and state that they are only concerned with the former (p. 3). When it comes to
inference rules, however, they credit a system with rule following, albeit of ‘implicit
rules’, even if a rule is ‘only implemented in the hardware and is essentially a
description of how some built in processor works’ (p. 34). However, for modus
ponens to be followed, it is not sufficient for modus ponens to be ‘built in’ to the
procedures by which the system operates. Such a proceduralized notion of modus
ponens is found in production rule systems. Production rules ‘fire’ when their ante-
cedent is satisfied and produce a consequent; however, there is no representation of
modus ponens. In the rule-following sense, modus ponens itself is not a rule in a
production rule system, any more than the planets implement Newtonian mechanics.
If a proceduralized notion of rule and rule-following is allowed, then the distinction
between rule-following and rule-describable behavior is lost again, with the con-
sequences outlined above. As elsewhere, the central question of whether human
thought can be described by logical rules or norms must be carefully distinguished
from the issue of how such inference is realized in the cognitive system.

3
This also means that the philosophical debate on rule-following is of direct relevance here (Kripke,
1982; McDowell, 1984; Collins, 1992; Ginet, 1992).
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 205

Finally, these considerations also allow us to clarify the nature of standard back-
propagation networks, which, we noted above, are claimed both as rule- and simi-
larity-based. On our analysis these networks neither compute similarity nor apply
rules, because they do not involve matching to a stored representation of any kind.
What representations could be held to be ‘matched’ with the input pattern? Past
inputs are not stored, so that instance-based comparison seems ruled out. The only
candidate appears to be weight vectors, but these are not matched, i.e. brought into
correspondence with, the input at all. Instead, activation flows through the network
as a complex non-linear function of inputs and weights4.
That the network’s behavior can be described with rules and that the regularities
it uses may be ‘extracted’ (Andrews et al., 1995) is not to say that the network itself
is following rules. Likewise. it is true that networks to some extent depend
on similarity (Rumelhart and Todd, 1993); similar inputs will tend to produce
similar outputs. This, however, is a causal story, due to similarity between inputs
in the sense of ‘overlap of input representations’ and, thus, similar activation flow
through the network. It is not due to the fact that similarity is being computed, any
more than similar rocks producing similar splashes results from computation of
similarity.
In summary, for the debate between rule- and similarity-based accounts to be
meaningful, matching must apply to representations of rules or instances. Thus
important classes of cognitive architecture in which no matching to representations
takes place stand outside the rule- versus similarity-based processing debate
entirely.

3.2. Exploring representation matching: are rules and similarity exhaustive?

We have outlined a core account of the distinction between rule- and similarity-
based processing. We now consider some ways in which these notions may be made
more specific, and also whether there are other styles of processing, distinct from
rule- or similarity-based processing within the representation matching framework.
Leaving aside the ‘memory bank’ which does not generalize to novel items at all, we
consider each of the three non-trivial regions of our space – indicated in Fig. 2 – in
turn probing the exhaustiveness of ‘rule’ and ‘similarity’. This analysis also shows
why our core distinction does not succumb to the mimicry arguments, which appear
to collapse rule- and similarity-based processing.

4
The inner product between the input- and the first layer of weights can viewed as a measure of
similarity (Jordan, 1986), but only if input vectors are of standard length. If not, our basic intuitions on
similarity (Section 3.2.1.) are violated: (1) similarity is not maximal in the case of identity, (2) input
vectors – viewed as points in a multi-dimensional input ‘feature’-space – which are more distant, i.e. have
fewer properties overlapping with the weight vector, can have larger inner products than nearby input
vectors due to the effects of length. While normalization is used in some connectionist architectu es such
as self-organizing networks (Rumelhart and Zipser, 1985) – and here it may be useful to think of the
weight vector is representing a prototypical instance in input space – it is generally not true for back-
propagation networks. Even less can we see weight vectors representing rules, and mere procedures, on
our account. do not suffice.
206 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

Fig. 2. How far do rules and similarity exhaust representation?

3.2.1. Partial matching and no abstraction: relation to similarity

We have argued that paradigmatic similarity-based processes involve partial
matching with no abstraction. However, does similarity matching have additional
crucial features and hence are there other forms of matching of partial, non-abstract
matching? There are many positive examples of measures of similarity, including
the geometric (Shepard, 1980) and contrast (Tversky, 1977) models prominent in
psychology and a host of measures in the computational literature on instance-based
approaches such as case-based reasoning (CBR) (Herbig and Wess, 1992) and
nearest neighbor algorithms (Cost and Salzberg, 1993) in machine learning. How-
ever, when it comes to delimiting what exactly counts as ‘similarity’, our underlying
intuitions seem remarkably vague. Indeed, they appear to be exhausted by the
following criteria:
1. similarity is some function of common properties5
2. similarity is graded
3. similarity is maximal for identity
Thus, any function from common properties to a value on a multi-value scale,
which is maximal for identity, will fit the bill.
This is vague, but specific enough to defeat one half of the mimicry argument,

5
Where ‘property’ covers binary attributes, continuous valued dimensions and relations.
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 207

because our function D is ruled out: ‘common properties’ degenerate to one; the
function is not graded; nor is it really ‘maximal for identity’ because, although it
returns the maximal value for identical instances, it also returns this value for any
other instance which is an instance of the rule.
Are there other types of partial matching with no abstraction, which are not
based on similarity? Possible alternatives arise from considering other information
which might be included in the matching process in addition to ‘common
properties’. Information such as frequency or recency might usefully be used in a
function performing partial matching, but these do not seem to be aspects of
similarity. Hence, it seem that a whole range of ‘hybrid’ functions is conceivable.
Furthermore, to the extent that our vague intuitions (Goodman, 1972; Goldstone,
1994a) about similarity are made more specific, going beyond ‘partial matching
without abstraction’ this automatically means that matching functions in the bottom
left corner which do not meet these criteria will not be similarity functions.
So it seems that, whatever theory of similarity is ultimately adopted, there will be
other types of ‘partial matching without abstraction’ not classified as involving
similarity.

3.2.2. Strict matching with abstraction: relation to rules

Rule-application, we have argued, is a matter of strict matching to knowledge
which is more abstract than the new item to be matched. By definition, this means
that the set of objects consistent with the representation of the new instance is a
proper subset of the set of objects consistent with the antecedent of a rule.
How do we know that the stored information is more abstractly represented than
the new item? In some representation languages, different ‘types’ of representation
reflect different levels of abstraction. Many representation schemes make no such
overt distinctions, however. Indeed, natural language makes no surface syntactic
distinction between the statements: ‘the dinosaur is extinct’ from ‘the dinosaur is in
the museum.’ Yet the former can be used as a general rule when classifying newly-
encountered animals, whereas the latter describes a specific event and cannot be
applied to new instances of any kind.
In considering different representational schemes, from natural language state-
ments to templates with slots and fillers, semantic networks and production rules,
two ways emerge in which the general idea of abstraction is manifest. First, a
representation can be more abstract than another by ‘underspecification’: the
more abstract representation simply specifies fewer constraints. This is exemplified,
for instance, by the use of variables in production system rules (Anderson, 1983), or
unification of feature structures in computational linguistics (Shieber, 1986). It is
also present in our ‘dog-rule’, above, where only ‘furriness’ and ‘barking’ matter,
and other properties are irrelevant. The second way in which a representation can
display greater abstraction is through the use of ‘general terms.’ This implies a
hierarchy of terms (e.g. dog, mammal, animal). A description is more abstract if
it contains predicates of which the predicates of the less abstract description are
proper subsets. This relationship would hold between our ‘dog-rule’ and instance
descriptions phrased in terms of ‘short-fur’, ‘long-fur’ etc. The example also illus-
208 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

trates that underspecification and general terms can be, and frequently are, com-
bined6.
Therefore, there are many ways in which an internal representation might be more
abstract than the instance-representation with which it is matched, but what type of
internal representation counts as a rule? Artificial intelligence and cognitive psy-
chology offer a wide range of models for internal representation, from declarative
statements in Prolog, through semantic networks, property list, feature vectors to
symbolic systems implemented in connectionist hardware. Which of these constitute
‘rules’? Must these be propositional or expressed in a language, possibly encom-
passing symbols and logical connectives? Adopting any such further constraints on
the notion of rule restricts the scope of the notion within the top right corner of the
space (Fig. 2), making rule-based reasoning non-exhaustive even of strict matching
to an abstraction.
Again, however, even the core notion successfully deals with the second half of
the mimicry argument, i.e. that similarity comparison is rule-based because any
similarity metric can be specified as a rule. First, it is an empirical question whether
the similarity metric is in fact represented as a rule in specific cognitive processes.
although computational systems can contain an explicit representation of their simi-
larity metric, this metric can equally be proceduralized, just as modus ponens can
(see e.g. Kruschke’s (1992) implementation of Nosofsky’s (1988) generalized con-
text model). If the similarity metric is explicitly represented, the similarity compar-
ison, strictly speaking, does involve rule-application (of the metric). However, the
rule the system is applying is then so general that it is neither an interesting nor
useful claim to say that the system is ‘rule-based.’ In particular, this claim is not the
one that cognitive science is concerned with, because it concerns how the matching
process itself is implemented, not the crucial issue of what type of representation-
matching is used.
How does the core distinction deal with more realistic examples of mimicry? Let
us reconsider Allen and Brook’s ‘additive rule’ prototype models and the equivalent
connectionist network. Allen and Brook’s ‘additive rule’ is a rule, according to the
core distinction, because it requires strict matching of its antecedent (i.e. it applies
just when at least the specified number of criteria are fulfilled). However, compar-
ison with a prototype involves similarity (assuming that similarity is well-defined
between prototypes and exemplars – see below) and hence this model is similarity-
based. Although the two processes produce identical results, they involve different
kinds of matching to different representations (a declarative specification that m of n
features must be fulfilled vs. a prototype). Finally, the equivalent single-layer con-
nectionist network does not involve any kind of matching to stored knowledge and
hence it is neither rule- nor similarity-based. Thus, the core distinction preserves the
6
Our notion of abstraction requires some loss of information relative to a corresponding specific
representation. Hence, we reject the notion of ‘ideal abstraction’ whcih retains all information (Barsalou,
1990). In fact, information loss is present even in Barsalou’s examples where the abstractions contain all
the properties of the exemplars but ‘centralized’; the centralized, ‘abstract’, representation no longer
contains sufficient information to reconstruct the particular exemplars. As noted, an ‘absraction’ which
retained all information about a set of instances could not be used in generalization.
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 209

intuitive sense that the three models achieve the same result in very different
ways.
The fact that rule- and similarity-based processes can produce equivalent classi-
fications may seem to undermine the empirical testability and even the theoretical
importance of the distinction. It is important to stress, however, that these processes
do differ in a wide variety of cognitively important ‘secondary properties’ e.g. in the
learning procedures required for acquisition, ease with which modification can be
affected, or behavior under noise (see also Hahn, 1996) Thus, the rule/similarity
distinction is important for cognitive science, although not always for primary input-
output behavior.
In summary, the conclusions on the scope of rules parallel those on similarity. On
the one hand, new ways of achieving ‘strict matching to an abstraction’ may emerge;
on the other hand, the notion of rule might be tightened up by adopting further
constraints. Consequently, it seems unlikely that ‘rules’ will ultimately exhaust the
space of strict matching with abstraction.

3.2.3. Partial matching to an abstraction

This leaves the ‘top-left corner’: partial matching to an abstraction. Rules do not
seem to spill over into this corner; neither legal rules, physical laws, universally
quantified formulae in first order logic, probabilistic rules, nor defeasible rules allow
partial matching. However, similarity might extend into this corner, specifically in
those versions of prototype theory in which the prototype is an abstraction of typical
or core properties that, as neither necessary nor sufficient, need be matched only
partially.
Can such partial matching to an abstraction count as a similarity comparison?
Current theories of similarity differ on this issue. Tversky’s contrast model, for
example, allows similarity comparison between any two featural representations,
even an item and its category.
Geometric models of similarity do not allow comparison between an instance and
something that is an abstraction relative to this instance; instances are points in
similarity space, but abstractions over instances are regions therein, and the notion
of ‘distance’ (and thus, similarity) between a point and a region is not defined. Thus
it depends on the choice of similarity theory whether similarity-based processing
extends to partial matching with an abstraction. It is not the purpose of this paper to
decide such issues and the task of empirically distinguishing rules and similarity can
be investigated without legislating terminology on this point. In evaluating empiri-
cal evidence for rules or similarity, the existence of this part of the representation-
matching space must be born in mind. What it is called is of secondary importance,
last but not least, because, here too, it is unlikely that similarity would ultimately
exhaust partial matching to an abstraction, even where a theory of similarity allows
such matching.
Raised by these issues is the general relationship between similarity- and
instance-based reasoning. However, regardless of how the case of partial matching
to an abstraction is decided, ‘similarity-based’ is wider than ‘instance-based’ if (as is
generally the case in psychology (Medin and Schaffer, 1978; Nosofsky, 1988))
210 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

‘instance’ refers only to actually-encountered exemplars. Alternative notions of

prototype such as the central tendency or a modal exemplar (Posner and Keele,
1970; Rosch et al., 1976; Komatsu, 1992; Nosofsky, 1992) are straightforward
cases of partial matching between representations of the same level of abstraction
and thus constitute similarity-based processing for all similarity theories.

3.3. What the distinction is not

We now examine other criteria that might be viewed as relevant to the distinction
between rules and similarity. In contrast to our core distinction, these potential
alternatives, although relevant and important in their own right, turn out to cross-
classify rules and similarity or, at best, to partially correlate with one or the other.

3.3.1. Types of computational architecture

Similarity-based methods are sometimes associated with highly parallel, distrib-
uted computational architectures. Rule-based processes, by contrast, are sometimes
associated with serial, symbolic computation.

3.3.1.1. Serial versus parallel. The serial-parallel distinction does not distinguish
between similarity- and rule-based approaches. Production rule systems, which
are paradigm rule-based systems, have both serial and parallel implementations.
On the side of similarity, most CBR systems, paradigmatic similarity-based
approaches, have serial implementations, although (partially or completely)
parallel implementations are possible e.g. Myllymäki and Tirri, 1993; Brown and
Filer, 1995.

3.3.1.2. Symbolic versus connectionist. The border between rule- and similarity-
based processes also fails to coincide with the distinction between symbolic and
connectionist computation. First, ‘symbolic’ does not equate to ‘rule’: similarity-
based systems such as CBR systems (Aamodt and Plaza, 1994) and nearest neighbor
algorithms in machine learning (Aha et al., 1991; Cost and Salzberg, 1993) are
typically symbolic. Second, ‘connectionist’ does not equate to similarity –
indeed, we have seen that the most widely used connectionist networks, back-
propagation networks, are neither rule- nor similarity-based.

3.3.2. Structured versus non-structured representations

In psychology, similarity-based methods frequently apply to simple representa-
tions, such as vectors of binary feature-values (in Tversky’s contrast model) or
numerical values (in geometric models), whereas rule-based models typically use
structured representations, involving arbitrarily complex symbolic expressions. This
distinction, however, is also orthogonal to the division between rule- and similarity-
based reasoning. Similarity can be defined over structured representations (Gentner
and Markman, 1994; Goldstone, 1994b; Hahn and Chater, 1997) – frequently, this
similarity relation is called analogy, because structural similarities are of central
importance (Gentner, 1983; Gentner and Forbus, 1991). Conversely, rule-based
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 211

accounts may be sufficiently simple that structured representations are not required
(as in most statistical contexts).

3.3.3. Abstract versus concrete representations

In psychology, similarity-based models are often applied to very concrete, typi-
cally perceptual, representations of simple stimuli, such as schematic faces or geo-
metric shapes (Reed, 1972; Nosofsky, 1988). More abstract domains, involving
reasoning about social interactions (Cosmides, 1989), norms of behavior (Cheng
and Holyoak, 1985) or naive physics (Hayes, 1979) are often viewed as involving
general rules. We have seen, however, that the rule/similarity distinction does not
relate to absolute level of abstraction. Similarity can operate in highly abstract
domains (e.g. using analogy in mathematical or scientific reasoning (Gentner,
1989) or case-based reasoning in law (Ashley, 1990; Aamodt and Plaza, 1994));
and rules may be used to process highly specific representations (as, for example, in
‘blackboard’ models of perception (Selfridge, 1959)). Crucially, we saw that the
same representation in stored knowledge can serve either as a rule or an instance in a
similarity comparison, depending on the nature of the matching process with new
items.
Thus, psychological evidence concerning the level of abstraction of representa-
tions does not thereby provide evidence concerning whether cognitive processes are
rule- or similarity-based. We shall see the implications of this in discussing attempts
to empirically distinguish the two styles of processing below.7

3.3.4. Rigidity and gradedness of classification

The issue of rigid versus graded classification may appear to map on to the
distinction between strict versus partial matching. Strict versus partial matching,
however, concerns how a new input is related to existing knowledge; it does not
relate to the gradedness of an output such as a classification decision. These two
issues are distinct, although connected. Specifically, if we assume that the output is a
function of the input, then a strict match to an input implies a rigid output: thus rule-
based processes produce rigid, all-or-none classifications. However, a partial match
to the input is compatible with both a graded and a rigid output, depending on
whether information about degree of match is preserved. Degree of match could
be used, for instance, to establish graded category memberships. However, other
similarity-based classifiers produce yes/no decisions by using a threshold, or com-
petition between instances. For example, the strict nearest neighbor criterion, the
paradigm of similarity-based models, bases the decision on the single most similar
known item. Thus, the core distinction is related to, but different from, the distinc-
tion between processes producing rigid versus graded classifications.

3.3.5. Deductive versus non-deductive reasoning

The distinction between rule- versus similarity-based reasoning also does not map
onto the distinction between deductive (certain) and non-deductive inference. While
7
Evidence concerning abstraction can, of course, be crucially important in distinguishing between
specific theories – either kind, which postulate particular levels of representation.
212 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

it is true that similarity-based reasoning is never certain, and, hence, always non-
deductive, we find deductive reasoning which is not rule-based and rule-based
reasoning which is non-deductive.
Probabilistic rules, as well as the non-monotonic or defeasible inference (see e.g.
Ginsberg, 1987) necessary to capture how we actually reason with rules such as
‘birds fly,’ in the face of countless exceptions such as penguins, broken wings and so
on, are not deductive (at least in psychological parlance, see Johnson-Laird and
Byrne, 1991; Chater and Oaksford, 1996).
We can also find deductive reasoning which is not rule-based, however. ‘Or-
introduction,’ for instance, allows the inference from P(a) to P(a) or Q(a). Similarly,
we can infer from P(a) that Exists(x)P(x). In either case, such an inference consti-
tutes a case of rule-base reasoning only if the ‘inference rule’ (‘or-introduction’)
itself is explicitly represented and applied (see Section 3.1.), rather than implemen-
ted procedurally.

3.4. Summary

We have explicated the ‘core’ distinction between rule- and similarity-based

processing and argued that this explication is distinct from, and more appropriate
than, a range of alternatives. We have also shown that our explication is specific
enough to clarify unclear cases such as back-propagation networks, and to block the
mimicry argument. Having provided this support for our analysis, we now consider
its implications for the problem of distinguishing rule- and similarity-based pro-
cesses empirically.

4. Re-evaluating the empirical evidence

Empirically distinguishing similarity- and rule-based psychological theories has

proved to be extremely difficult. Moreover, there are at least two interpretations of
the question.
• Class distinguishability. Can empirical data distinguish between the classes
of rule-based and similarity-based theories? In other words, can we distin-
guish between rule- and similarity-based accounts of a task in general.
• Specific distinguishability. Given fully elaborated, specific similarity- and
rule-based theories, can these be distinguished empirically? This last ques-
tion is, of course, many questions, depending on what the specific theories
are.
In the literature, doubts about class distinguishability have variously been voiced
(e.g. Barsalou, 1990; Koh and Meyer, 1991), and we therefore give this issue
particular attention8. Note that if the classes of rule- and similarity-based theories
8
However, Barsalou’s (1990) argument rests crucially on his notion of ‘ideal abstraction’, which does
not involve any information loss – a notion not allowed in our framework, because information loss is
constitutive of what it means to be a.n abstraction, as noted above.
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 213

cannot be empirically distinguished, we may nonetheless be able to distinguish fully

elaborated, specific accounts, i.e. we may be restricted to comparative model fitting.
In terms of our analysis, class distinguishability requires finding empirical evi-
dence which locates a cognitive process in relevant regions of our space of possi-
bilities (Fig. 2) (including, recall, non-representational alternatives, as well as
different kinds of matching). Where ‘rule’ or ‘similarity’ are given more specific
elucidations, they correspond to yet smaller sub-spaces. Thus, empirically establish-
ing a more specific notion also requires eliminating the remaining, rival forms of
strict matching to an abstraction or partial matching to an instance. For example,
where the use of the term ‘rule’ is restricted to internal representations which are
symbolic and contain variables and logical connectives, an additional empirical case
for the symbolic nature of the internal representation has to be made. Likewise,
where ‘similarity’ is restricted to a metric of particular functional form, other forms
of partial matching must be ruled out. Hence, dealing only with the broad partitions
of the space entailed by our core distinction is the easiest version of the rule/simi-
larity class distiguishability problem and we will focus on how empirical evidence
relates here.
Again, the problem is that these classes of account are general, and distinguishing
general accounts is difficult in any scientific context. This is because relating general
accounts to empirical data requires additional auxiliary (and typically uncertain)
assumptions. Indeed, in principle, any theory can be made to fit any data set (Put-
nam, 1974), by postulating appropriate auxiliary assumptions, requiring additional
constraints on what assumptions count as plausible in order to rule out any general
account.
In psychology, where any specific mental processes lead to behavior only in
conjunction with many other processes which are highly complex but unknown,
additional hypotheses which save a theory are particularly easy to provide (Pyly-
shyn, 1984). Moreover, distinguishing between the classes of rule- and similarity-
based accounts of cognition is particularly difficult, because the auxiliary assump-
tions concern the very essence of the explanation – how instances are encoded, what
the actual rule is or what similarities are perceived. This is not a problem of having a
single free parameter to tie down: almost the entire predictive content of a rule-based
theory depends on what exactly the rules are; almost the entire predictive content of
a similarity-based theory depends on what the similarities are.
Perceived similarity depends on which properties of objects are selected as rele-
vant, how they are weighted and what metric is used to compare the resulting
representations. The number of possible rules that can be entertained is clearly
vast – in principle, there are infinitely many rules which would allow perfect
performance in any learning task9. Even allowing that this set is restricted by
representational and memory limitations, the set of possible rules that an agent
might entertain will in most cases still be large. This casts doubt on the possibility
of directly distinguishing the classes of rule- and similarity-based theories by a
single experimental measure.
9
For example, in a category learning task, there will be infinitely many rules compatible with the finite
set of data (labelled objects) – where the rules differ arbitrarily on the infinite number of unseen examples.
214 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

We now consider various suggestions concerning how the classes of theories can,
despite the apparent difficulties above, be distinguished, focusing first on experi-
mental criteria and then on computational criteria drawn from AI. Together these
will also suggest a different emphasis for future research concerning rule- and
similarity-based processing.

4.1. Experimental criteria

We are now in a position to survey empirical evidence which aims to distinguish

similarity- and rule-based accounts across various cognitive domains. The literature
contains a wealth of potential criteria; we have limited our discussion to those
experimental criteria which are widely applicable, i.e. not restricted to a particular
question or subject domain.10 We group this evidence under four headings: effects of
instances, effects of rules, generalization beyond the capabilities of instance-based
models and patterns of breakdown.

4.1.1. Effects of instances

Similarity-based theories typically assume that new items are compared with
representations of old items. If so, then some distinction between performance
with genuinely new items (where generalization is required) and old items (which
need only be ‘looked up’ in memory) may be expected.
By contrast, if all items are dealt with using a rule, there seems no reason to
suppose that new and old items will be classified differently. This is because the
general claim embodied in the rule must be applied to all specific instances; the rule
contains no information about which items have been seen before. The ensuing
criterion of ‘no observable difference between old and new items’ is used in
many studies (Nosofsky et al., 1989; Allen and Brooks, 1991; Smith et al., 1992)
and has been suggested as an ‘operational definition’ of rule-based performance
(Herrnstein, 1990; Shanks, 1995).

4.1.1.1. Old-new recognition. Instance-based accounts, in contrast to rule-based

accounts, also require that instances be remembered. Thus, they seem to predict
that people should be able to distinguish old from new items, and that their pattern of
old-new judgements should relate closely to their categorization performance.
Relating categorization and old-new judgements in detail requires a unified
psychological account of both. An example of this has been provided by
Nosofsky (1988), who obtained a good account of experimental data using simple
artificial stimuli. Such a unified account gives strong evidence for similarity-based
processing; but failure to find such an account is not equally strong evidence against
10
Thus we do not consider, e.g. arguments for rules based on linguistic analysis (Marcus et al., 1995), or
those of the criteria put forth by Smith et al. (1992) which are tailored specifically to identifying rules of
inference. We also omit Sloman’s S-criterion (Sloman, 1996), which – stemming from a somewhat
different interest – uses conflict to establish two cognitive sub-systems, because further, independent
evidence is required to identify these as ‘rule’ or ‘similarity’ systems. It is only this further evidence, not
the S-criterion, which directly applies to our question.
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 215

similarity-based processing: old-new discrimination might, for example, draw on a

memory store separate from that used in classification. More importantly, the
encoding of an instance may be sufficiently abstract that discrimination is not
possible (and, of course, some abstraction is inevitable, otherwise no item could
ever be re-recognized).

4.1.1.2. Manipulations of the instance-space. This approach is mainly used in

studies using stimulus sets which pit rule- and similarity-based classification
against each other. The interest lies in the classification of ‘critical instances’
which, although instances of the intended rule, are more similar to instances of
another rule or category. Rule-based reasoning seems to suggest that the ‘rule’
classification should prevail, whereas similarity-based reasoning seems to suggest
the opposite. However, sometimes, effects of both rules and similarity are found
(Nosofsky et al., 1989; Allen and Brooks, 1991; Vokey and Brooks, 1994).
Generally, a failure to find instance effects is evidence against instance-based
reasoning, at least on the experimenter’s assumption of instance encoding and
instance similarities, and, conversely, instance-effects can be viewed as refuting
reasoning with the rules intended by the experimenter. In both cases, however,
refutation of one provides only limited support for the other due to the non-exhaus-
tiveness of rule- and similarity-based reasoning. Where instance effects fail, proto-
type models (both with and without abstraction, see above) must also be ruled out;
they provide an important class of alternatives to rule-based accounts in this context,
because ‘critical instances’ by their very construction are peripheral, and, hence,
might be misclassified on prototype accounts as well. Where instance effects appear,
simple connectionist accounts of the type we have classified as similarity-dependent
but not similarity-based, in our usage, must be ruled out. Consequently, the domain
most intensely studied with tasks of this kind, artificial grammar learning (see
below), has seen a long series of claims and counterclaims (see Redington, 1996
for thorough discussion).

4.1.1.3. Summary. Instance-space manipulations are an effective tool, but the non-
exhaustiveness of rules and similarity means that empirical evidence is more
powerful in challenging than in supporting, either account. Also, specific
assumptions about rules, instances and instance-similarities must be made, so that
this criterion does not pertain to entire classes of account. Memory for instances
seems indicative only if a ‘unified account’ succeeds, making it a powerful but
demanding tool. Again, specific assumptions about instances and similarities are
required.

4.1.2. Effects of rules

4.1.2.4. Rule priming. Throughout cognition, repetition of the same, or a similar,

mental process (e.g. recognizing a word or a picture) speeds performance. Smith et
al. (1992) suggest that such priming effects might provide evidence concerning the
existence of internal rules. Specifically, they suggest that if priming were observed
216 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

between two cognitive tasks which share the same rule, but correspond to very
different instances, the rule-based view would be favored. Once we recognize that
similarity-based models may be defined over abstract representations, and not
merely superficial features of the stimulus, however, it is difficult to rule out the
class of similarity-based models.
Langston et al. (unpublished data), for example, use conditional sentences which
express either permission or obligation (see Cheng and Holyoak, 1985). They argue
that performance on Wason (1968) selection task using these two rules is primed if
the underlying rule-type is repeated, even though the surface form of the rules is
altered. They argue that this provides evidence for permission and obligation rules.
However, this important empirical result is equally consistent with the suggestion
that instances of conditional sentences have abstract codes, which distinguish per-
mission and obligation.
Another example is syntactic priming (H. Branigan et al., unpublished data),
where sentence production or comprehension is primed by previous sentences
which related syntactic structure. This is evidence for abstract representation of
syntactic information. However, again, this information may be embodied in rules
or as abstract information about the stored sentence-instances (e.g. sentences may be
stored not as strings of words, but as labelled tree structures)11. In both examples,
priming provides important evidence for a particular kind of abstract representation,
rather than evidence between rules and instances.

4.1.2.5. Rule complexity. If cognition is rule-based, then the number of rules

involved in a cognitive task may explain task difficulty (Smith et al., 1992). The
number of rules depends, of course, on the specific set of rules under consideration;
and difficulty (e.g. time) will also depend on how rules are implemented. In the same
way, task-difficulty predictions can be obtained from specific similarity-based
models (for example, reaction-time predictions from a specific model of
categorization, e.g. Lamberts, 1995). However, there do not appear to be task-
difficulty predictions associated with the classes of rule- and exemplar-based
models. Therefore, task-complexity considerations appear to be important in
testing specific rule-based accounts, but not suitable for distinguishing the rule-
and similarity-based classes of account.

4.1.2.6. Verbal protocols. If people use rules, it is possible that they may express
these rules, or aspects of them, in verbal protocols. Many production-rule theories
of problem solving and skill-learning are based on an interactive process of building
rule-based models and matching these models to verbal protocols and task
performance (Newell and Simon, 1972). Equally, protocols mentioning com-
parison with instances (e.g. analogical reasoning) might also provide evidence for
similarity-based processes. Protocol evidence is potentially very important.
Crucially, its strength depends on the degree to which protocols tie up with other
experimental measures, thus providing evidence that protocols are a reliable
11
Of course, a similarity-based approach to language processing may seem implausible for other
reasons.
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 217

indicator of the cognitive processes under study. Hence, production system

computer models of learning a computer programming language based on
protocols which provide good empirical data fits thereby provide evidence for
rules (Anderson, 1983). However, there are also circumstances where independent
evidence indicates that reported rules were not followed (see Nisbett and Wilson,
1977).

4.1.2.7. Summary. In short, putative effects of rules might provide useful, but not
decisive tests, between the classes of rule- and similarity-based accounts. Priming
effects may indicate that particular abstract information is represented, but not
whether that information is represented by rules or instances. Effects of rule
complexity depend on the specifics of the rule-based account and do not follow
from the class of rule-based accounts. Finally, verbal protocols may be suggestive,
but require evidence that protocols are reliable indicators of the underlying cognitive
mechanisms under study. All three criteria require specific rules, although for
protocols, these arise directly from criterion use.

4.1.3. Patterns of generalization

In favor of rules, it has been argued that several types of generalization seem
inexplicable by an instance-based approach.

4.1.3.8. Extrapolation. Extrapolation in so-called ‘function learning’ experiments

(Koh and Meyer, 1991) provides a potentially powerful source of evidence for rule-
based behavior (Shanks, 1995). For example, imagine subjects learning to press a
button with a duration proportional to the size of stimuli; larger stimuli requiring
longer durations. Correct performance outside the range of stimuli seen so far (e.g. a
very long button-press for a very large stimulus) is argued to be incompatible with
an instance-based model as the response is not that of the closest previously seen
instance, but a novel, i.e. longer, response. Behavior seems to depend on the
application of a rule specifying the function relating stimulus height and response
duration (Shanks, 1995).
Finding that people can generalize by extrapolation in this way (Koh and Meyer,
1991; Delosh, 1993) is an important empirical result and it constitutes strong evi-
dence against similarity-based approaches. Notice, however, that it does not show
that rules are being used. For example, a single-layer connectionist network, with
one input and linear output unit, could trivially learn to extrapolate from increasing
input to increasing output.

4.1.3.9. Transfer. A further approach considers the transfer of information learned

in one domain to another. Perhaps the paradigm example is transfer in artificial
grammar learning (AGL) (Reber, 1989). In AGL, subjects try to memorize a set
of letter strings, without being told that they are generated according to a set of rules
(typically a simple finite state grammar). They are then told about the existence of
the hidden rules, but not what the rules are, and asked to discriminate new test strings
which do or do not conform to these rules. Subjects perform significantly above
218 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

chance in this experiment. One hypothesis is that they have implicitly extracted
some of the underlying rules used to generate the items (‘implicitly’ because
subjects typically cannot verbally report any rules they have learned). Another is
that they are simply judging the similarity of new items to old, which can also lead to
above chance performance. The transfer condition seeks to rule out this possibility
by using a different vocabulary of letters in the memorization and discrimination
phases. The idea is that the new strings are not at all similar to the memorized
strings, and hence similarity cannot mediate generalization. Even here, subjects
do show (typically small) above chance transfer performance (Dulaney et al.,
1984; Redington and Chater, 1994).
One possible alternative to a rule-based account for this phenomenon is that
instance-encoding abstracts away from the specific alphabet used in training, so
that instances successfully classify transfer items. Again, we stress that evidence
for abstract representations in itself is equally consistent with rule- and similarity-
based processes. Abstraction in instance-encoding is perfectly possible, the question
is only how much abstraction is plausible.
A further alternative is that abstraction occurs only when the stored instances are
compared with the transfer stimuli, i.e. at transfer (e.g. Brooks and Vokey, 1991).
This is tantamount to analogy and models of this kind equal or surpass human
transfer performance without reference to rules (Redington and Chater, 1996).
The exact relationship between similarity-based reasoning and analogy is contro-
versial (Seifert, 1989). Thus, analogy presents either a version of similarity-based
reasoning or a ‘third account’.
Finally, attempts have been made to explain transfer with a connectionist network
(Altmann et al., 1995) which appears to involve no matching between input and
stored knowledge, and hence falls outside both rule- and similarity-based accounts.
Thus, it seems that transfer effects may be explained by rule- and by similarity-based
processes and by alternatives in neither framework.

4.1.3.10. Reversal. Another source of evidence comes from reversal of a learned

response (Shanks, 1995). In a typical experiment, people or animals are initially
trained to associate reliably two distinct responses to two sets of stimuli (Sidman and
Tailby, 1982; Vaughan, 1988). Then reversal occurs: it is now the other set which
demands the particular response. Subjects are trained to stable performance on the
‘reversed’ contingency, followed by a second reversal; and so on. After a number of
such reversals, both animal and human subjects display the ability to shift almost
immediately on new reversed trials, extending their behavior from these first
instances to the remaining members of the class. Thus, it is claimed, members of
each set are treated as an equivalence class.
This satisfies Shanks’ instrumental definition of rule as ‘no observable difference
between performance on trained items (here, of the reversal trial) and old items’
(here, the rest of the class) but it is insufficient on our account. This is because the
reversal could be happening solely through a switching of responses at the response
level. That is, subjects realize that they now have to respond the opposite way from
before, e.g. ‘yes’ now means ‘no’ and vice versa, but which response to choose (and
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 219

then reverse) is still determined solely through similarity comparison with past
instances. On this account, subjects need only realize that ‘responses have gone
funny again’; they need not treat the items as belonging to equivalence classes.

4.1.3.11. Summary. The above three criteria all seek to rule out the entire class of
instance-based models. The rule models they aim to support have particular rules in
mind – i.e. the underlying function, the rules of the underlying grammar, rules
describing the equivalence classes – but any rule which delivers the same
classification for the data seen will suffice; hence, these experimental criteria can
be seen as distinguishing classes of instances from classes of rules. Transfer appears
consistent with rules, similarity and connectionist alternatives. Reversal appears
consistent with both rules and similarity, because it can be explained by
switching at the response level. More positively, however, extrapolation provides
strong evidence against similarity-based models, although it is consistent with non-
matching models such as neural networks as well as with rules.

4.1.4. Error and patterns of breakdown

Patterns of breakdown and error provide another potentially valuable source of
evidence.

4.1.4.12. Memory failure. Suppose that people learn the rule NOT-RED OR
TRIANGLE in an artificial concept learning experiment. Later, they are the tested
on generalization to new instances. If their memory is incorrect, they might be
expected to classify according to: RED OR NOT-TRIANGLE, or NOT-RED
AND TRIANGLE. By contrast, errors on a similarity-based view (based on
instances) would not be expected to have this global character. Instead, individual
past instances might be misremembered, leading to local misclassifications of
nearby novel items. Global errors might, however, result if learning had yielded a
single prototype. We are not aware that anyone has aimed to make use of this
contrast, but it appears to be a potential direction for future research.

4.1.4.13. Neuropsychology. More dramatically, it is possible that neuropsycho-

logical patients may exhibit selective preservation of rules, but loss of exceptions.
This appears to occur in reading, with some patients appearing to lose exceptions
(surface dyslexia, McCarthy and Warrington, 1986) and others losing the ‘rules’ of
spelling to sound correspondence (phonological dyslexia, Funnell, 1983). This
‘double dissociation’ has been taken as evidence for rules in reading (e.g.
Shallice, 1988).
This may be over-interpreting the data. While connectionist modeling has made
us aware of the fact that a uniform connectionist architecture may not adequately
capture the neuropsychological data for reading, i.e. that we might need dual route
architectures, it has also alerted us to the fact that possibly neither route need
contain rules (Bullinaria and Chater, 1995). Nonetheless this source of evidence
may be difficult to account for by a similarity-based account, at least in the case of
reading (but for an attempt see Glushko, 1979).
220 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

4.1.4.14. Over-regularization. Rule-based accounts of partially regular domains

(such as the mapping between spelling and pronunciation or the English past-
tense) divide knowledge into two components: a set of rules and a list of
exceptions to those rules. Hence, over-extensions of rules might occur, where
they should have been blocked by an exception (e.g. the past tense of go is given
as goed). Such errors are observed in learning (Ervin, 1964) and have been taken as
evidence for rules. Caution is necessary, in that exemplar models overgeneralize
both irregular and regular past-tense forms in a manner which depends only on an
item’s location in phonological space (Hahn et al., 1997). Even where over-
regularization does not seem to depend on close similarity to other regular items,
connectionist models which are neither rule- nor similarity-based might behave
appropriately (e.g. Rumelhart and McClelland, 1986; Plunkett and Marchman,
1991 and Seidenberg and McClelland, 1989; Bullinaria, 1994) although
controversy on the (overall) adequacy of these models remains (Pinker and
Prince, 1988; Forrester and Plunkett, 1994; Westermann and Goebel, 1994;
Marcus et al., 1995; Nakisa and Hahn, 1996).

4.1.4.15. Summary. Data from neuropsychology and over-regularization may

present problems for similarity-based models in specific contexts (e.g. reading or
inflection). These criteria, however, do not appear to distinguish between rule-based
and non-matching connectionist accounts. Evidence from ‘memory failure’ may
provide a useful line of evidence, although this has currently not been explored
empirically. All of these criteria depend on specific rules and instance-
similarities, although in the case of reading and inflection these can be based on
large bodies of theoretical and empirical work.

4.1.5. The strength of experimental evidence

We have seen that most experimental criteria distinguish specific, fully elaborated
theories in particular domains more effectively than they decide between the entire
classes of rule- and similarity-based models. In distinguishing different classes,
although individually decisive empirical tests are difficult to provide, convergence
of several criteria may be persuasive (Smith et al., 1992). Thus, testing between the
classes of rule- and similarity-based processes seems possible, but may require
integration of a range of sources of data.

4.2. Computational criteria

We have so far focused on experimental evidence and ignored the computational

issues concerning the relative merits of rule- and similarity-based processing. Com-
putational considerations are crucial, however, because any viable cognitive theory
must be computationally viable.
Moreover, computational constraints provide a general perspective on rule and
similarity which contrasts usefully with that from the experimental literature. In
experiments such as those considered above, there is typically an inverse relation
between experimental precision and generality of the result. Therefore, the construc-
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 221

tion of ever more specific experimental contexts and tasks runs the risk of contribut-
ing relatively little to our understanding of rules and similarity in normal cognition.
This does not mean that experimental studies should be abandoned, but it does imply
that we should pay close attention to general considerations concerning the plausi-
bility of rule- and similarity-based models in normal thought. Computational con-
straints provide an important class of such general considerations.
Specifically, the debate between rule- and similarity-based processes in cognitive
science can draw on the insights and generalizations derived from experience in
AI and machine learning of attempting to use each approach in practical contexts.
The lessons from computation have been little recognized in psychology. However,
we suggest, these lessons provide a vital complementary source of evidence in
evaluating the plausibility of rule- and similarity-based accounts of human cogni-
tion.

4.2.1. Are theories possible?

Where theories – collections of rules – are available, they can be remarkably
effective, as evidenced by spectacular predictive and explanatory successes in many
areas of science. The central challenge for a rule-based approach, however, is
actually determining rules which adequately capture our common-sense knowledge
of the world. The problem has proved to be very hard. It has required enormous
intellectual effort even to provide adequate axioms for set theory and arithmetic12.
Very few aspects of scientific knowledge have been more than partially formalized,
and constructing theories for common-sense knowledge appears still more difficult.
One problem is that common-sense knowledge does not appear to break up into
separate domains. Thus, trying to provide rules for parts of knowledge seems to lead
inevitably to the endless task of capturing the whole of human knowledge. This is
what Fodor calls the isotropy of common-sense knowledge (Fodor, 1983). Another
problem is that common-sense rules almost always have exceptions (Reiter, 1980),
which raises enormous technical and conceptual difficulties (McDermott, 1987;
Oaksford and Chater, 1991, 1993). Moreover, even given a set of rules, there
remains the problem of how these rules are applied to specific instances. This
too appears to depend on vast amounts of background knowledge, in ways that
are not at all well-understood (Oaksford and Chater, 1991; Pickering and Chater,
1995).
For these reasons, AI has not been able to provide a feasibility proof of the claim
that knowledge is represented in terms of rules by building general purpose rule-
based systems. Although ‘expert-systems’ based on rules have been developed for
highly specialized domains, e.g. DENDRAL (Feigenbaum, 1977), MYCIN (Short-
liffe, 1976) or ASSESS (Dayal et al., 1993), ‘scaling-up’ to real world materials has
not been achieved. Thus, it seems that a ‘pure’ rule-based approach to cognition is
unlikely to be viable.
12
Frege’s formalization of set theory, which appeared directly to reflect basic intuitions turned out to be
inconsistent; and rnoreover Gödel showed that a complete, consistent axiomatization of arithmetic is
impossible (Boolos and Jeffrey, 1988). Both results suggest that formalization of human knowledge may
also encounter unexpected difficulties in other domains.
222 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

4.2.2. The power of similarity-based reasoning

Reasoning by similarity to past instances can be applied even in domains which
are little understood and where no theory is available, where theories are partial or
there are competing theories (as in law) (Ashley, 1990; Porter et al., 1990). Addi-
tionally, re-using entire past ‘problem-solutions’ in domains such as problem-sol-
ving or planning, can lead to faster processing than continually reasoning from first
principles (Schank, 1982; Kolodner, 1991, 1992). The ability to cope with ‘partial’
theories is of central psychological importance, because complete common-sense
theories may not be feasible, as outlined above. Indeed, even currently popular
‘theory-based views’ in psychology, which emphasise the role of general knowledge
in categorization, claim only that partial theories are brought to bear, thus leaving an
important role for similarity-based reasoning (Hahn and Chater, 1997).
Furthermore, similarity-based reasoning can be highly effective. For example, the
simplest such algorithm, nearest neighbor, has excellent asymptotic classification
accuracy in comparison with other inference methods (Cover and Hart, 1967).13 The
prospect of general and effective reasoning has fuelled enormous interest in simi-
larity-based reasoning in AI.
In practice, however, the situation is not quite so ideal. For example, nearest
neighbor methods typically require vast numbers of past instances to achieve
good performance on complex problems. Moreover, similarity-based methods are
dramatically impaired by redundant or irrelevant features, which have as much
effect on similarity computation as the crucial ones, causing inaccurate performance
(Wettschereck and Aha, 1995) and slow learning (Langley and Sage, 1994).14 This
has prompted research into so-called ‘knowledge-poor’ feature-selection algorithms
(Aha and Bankert, 1994; Wettschereck and Aha, 1995), to choose or preferentially
weight relevant features. No such method can learn optimal weight settings for all
tasks, however, (Mitchell, 1990; Wettschereck et al., 1995) and, where there are
only few past instances, invoking background knowledge is the only option. Thus,
sophisticated CBR systems in AI (Branting, 1989; Ashley, 1990) rely on massive,
knowledge-based preprocessing. The situation gets worse, where past cases are so
sparse that their solutions require significant adaptation, such as in case-based
planning (Kolodner, 1991, 1992) – a step which also introduces other forms of
inference. Thus, ‘scaled-up’ similarity-based accounts must be supplemented with
accounts of how background knowledge affects feature-weighting, which is cur-
rently poorly understood (Hahn and Chater, 1996). It seems certain that such an
account will also have to integrate other, non-similarity-based, forms of inference,
and that a ‘pure’ similarity-based account of cognition, like a pure ‘rule-based’
account is not viable.

13
Asymptotically, the single nearest neighbor algorithm has (assuming smoothness) a probability of
error which is less than twice the Bayes probability of error and thus less than twice the probability of error
of any other decision rule, non-parametric or otherwise, based on the infinite sample set (Cover and Hart,
1967).
14
The number of instances needed for nearest neighbor to reach a given level of accuracy grows
exponentially with the number of irrelevant features (Langley and Sage, 1994).
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 223

4.2.3. Knowledge revision

People can clearly learn both from experience and from being told about the
world. These two sources of information fit very differently with the two classes
of model.
Rule-based models have the potential of adding new general information directly
into the rule-base (i.e. as if the system has been told new information) and can
interact productively with existing rules. However, this very generality means that
learning from experience is very difficult, because there are typically many alter-
native changes to the rule-base that can capture new ‘data.’ This is analogous to the
problem of theory induction or revision in the light of new data, which has notor-
iously resisted formal treatment. This problem has been profoundly problematic in
AI; we already noted that ‘expert systems’ have been developed successfully only
within very restricted domains, but even here, they are not the result of automated
learning procedures. Rather, they are based on ‘knowledge engineering,’ i.e. human
compilation of relevant domain knowledge into a computer accessible form. Only
very limited versions of the problem of rule-induction have been addressed in
machine learning, such as the induction of simple logical conjunctions from
instances described as simple property lists (Langley, 1996) although more complex
structures have increasingly been studied (Muggleton, 1992). Finally, rule-based
systems face the problem of dealing with inconsistency between rules (for instance,
as a result of ‘noisy data’ or of exceptions) which can potentially lead to complete
inferential anarchy15.
Similarity-based models, by contrast, can learn from experience simply by adding
new instances to the data base. However, learning from being told information
which covers large areas of the domain at a stroke, which is so effective for rule-
based systems, is not possible for an instance-based system. On the other hand, by
reasoning ‘locally’ from nearby instances, rather than from the global predictions of
an entire system of rules, instance-based systems neither need to be globally revised,
nor do they run the risk of logical inconsistency and the resulting inferential chaos.
The complementary strengths of both types of system suggest integration of both.

5. Conclusions: integrating rules and similarity

In this paper, we have explicated the core distinction between rule- and similarity-
based generalization, based on how representations of novel items are matched to
stored representations. This core distinction is what is necessarily implied wherever
the terms are contrasted without further specification. In doing so, we have resolved
three difficulties with the initial intuitive distinction: we have shown that apparent
mimicry arguments do not apply, we have provided clear criteria for deciding
intuitively unclear cases and we have provided a clear target for empirical investi-
gation. Moreover, we have provided an organizing framework which positions both
15
In classical logic, all propositions and their negations can be derived from an inconsistent set of rules.
Non-classical, ‘para-consistent’ logics, which seek to contain inference from contradictions, have there-
fore become a major topic of research (Touretzky, 1986; Smolenov, 1987).
224 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

rule- and similarity-based generalization in a way that shows the alternatives to both
and allows visualisation of the effects of adopting further constraints on ‘rule’ or
‘similarity’ for the empirical problem of distinguishing between them.
We have also investigated the power of various experimental tests which have
frequently been used to distinguish ‘rules’ or ‘similarity’ without further specifica-
tion. This has revealed that these tests are not individually decisive, although con-
vergent evidence from several sources may be compelling. We have also argued,
however, that computational considerations drawn from AI provide valuable addi-
tional support in evaluating the plausibility of either account. We draw from AI the
moral that pure rule- and similarity-based mechanisms appear not to be computa-
tionally viable for solving real-world problems and that neither viewpoint accounts
for the human ability to learn both by example and from instruction. Both types of
computational consideration suggest that the psychological concern with deciding
between the two viewpoints may be misguided. Instead, it may be crucial to under-
stand how the two can be integrated, combining the strengths of both.
This view is reflected in an increasing interest in hybrid systems within AI (Riss-
land and Skalak, 1991; Rissland et al., 1993). It also sits well with the not uncommon
finding of both rule and similarity effects in recent experimental work on category
learning reported in Section 4.1.1 above. Furthermore, given the difficulties of
finding complete theories from which all desired instances can be deduced, it is
also the most suggestive interpretation (Hahn and Chater, 1997) of experimental
evidence in support of the theory-based view of conceptual structure (e.g. Medin and
Wattenmaker, 1987). Finally, the need for interaction between the two processes is
suggested by considering the structure of the law, next to science the most elaborate
and explicit system we have developed for dealing with everyday life. The law
displays both instance- and rule-based reasoning in the form of precedent and
statute. While legal systems differ regarding the relative weight they place on
each of these factors (e.g. the Anglo-American tradition emphasises similarity to
past cases and the continental tradition emphasises rules), the ‘blend’ of both is
common to all western legal systems.
These considerations suggest that rules and similarity both have their respective
roles, not just side by side, with similarity covering some domains and rules others,
or ‘doubling up’ in parallel (Sloman, 1996), but in an active interplay within a single
task. The idea that rules and similarity might operate together is frequently sug-
gested, even by advocates of mental rules (e.g. Smith et al., 1992; Marcus et al.,
1995); and where real-world inference has been subjected to psychological explana-
tion (Pennington and Hastie, 1993), a complex interplay of a variety of types of
inference has been implicated. This suggests a shift of emphasis in future research,
from pitting rules against similarity toward experimental and computational inves-
tigation of the potential interplay of rules and similarity in cognition.

Acknowledgements

The authors would like to thank Jacques Mehler, Steven Sloman and three anon-
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 225

ymous reviewers for their detailed and valuable comments on an earlier version of
this manuscript, and Andreas Schöter, Andrew Gillies and Martin Redington for
helpful discussion. Ulrike Hahn was funded by ESRC Grant No. R004293341442.
Nick Chater was partially supported by ESRC Grant No. R000236214. The research
reported in this article is based on Ulrike Hahn’s doctoral dissertation and was, in
part, carried out while the authors were at the Department of Experimental Psychol-
ogy, University of Oxford.

References

Aamodt, A., Plaza, E., 1994. Case-based reasoning: foundational issues, methodological variations, and
system approaches. AI Communications 7, 39–59.
Aha, D., 1997. Editorial for the special issue: lazy learning. Artificial Intelligence Review 11, 7–10.
Aha, D., Bankert, R., 1994. Feature selection for case-based classification of cloud types: an empirical
comparison. In: Proceedings of the AAAI-94 Workshop on Case-Based Reasoning.
Aha, D., Kibler, D., Albert, M., 1991. Instance-based learning algorithms. Machine Learning 6, 37–66.
Allen, S., Brooks, L., 1991. Specializing the operation of an explicit rule. Journal of Experimental
Psychology: General 120, 3–19.
Altmann, G., Dienes, Z., Goode, A., 1995. On the modality independence of implicitly learned gramma-
tical knowledge. Journal of Experimental Psychology: Learning, Memory and Cognition 21, 899–
912.
Anderson, J., 1983. The Architecture of Cognition. Harvard University press, Cambridge, MA.
Andrews, R., Diederich, J., Tickle, A., 1995. A survey and critique of techniques for extracting rules from
trained artificial neural networks. Knowledge-Based Systems 8, 373–389.
Ashley, K., 1990. Modeling Legal Argument – Reasoning with Cases and Hypotheticals. MIT Press,
Cambridge, MA.
Barsalou, L., 1990. On the indistinguishability of exemplar memory and abstraction in category repre-
sentation. In: Srull, T.K., Wyer, R.S. (Eds.), Advances in Social Cognition, Vol. III, Content and
Process Specifity in the Effects of Prior Experiences. Erlbaum, Hillsdale, NJ, pp. 61–88.
Bates, E., Elman, J., 1993. Connectionism and the study of change. In: Johnson, M. (Ed.), Brain Devel-
opment and Cognition. Blackwell, Oxford.
Berry, D., Broadbent, D., 1984. On the relationship between task performance and associated verbalizable
knowledge. The Quarterly Journal of Experimental Psychology 86a, 209–231.
Berry, D., Broadbent, D., 1988. Interactive tasks and the implicit-explicit distinction. British Journal of
Psychology 79, 251–271.
Boolos, G., Jeffrey, R., 1988. Computability and Logic, third ed. Cambridge University Press, Cambridge.
Braine, M., 1978. On the relation between the natural logic of reasoning and standard logic. Psychological
Review 85, 1–21.
Branting, K., 1989. Integrating generalizations with exemplar-based reasoning. In: Proceedings of the
Eleventh Annual Meeting of the Cognitive Science Society Ann Arbor, Michigan. Erlbaum, Hillsdale,
NJ, pp. 139–146.
Branting, K., 1991. Integrating Rules and Precedents for Classification and Explanation. Ph.D. thesis,
University of Texas at Austin.
Brooks, L., Vokey, J., 1991. Abstract analogies and abstracted grammars: comments on Reber (1989) and
Mathews et al. (1989). Journal of Experimental Psychology: General 120, 316–323.
Brooks, R., 1991. Intelligence without representation. Artificial Intelligence 47, 139–159.
Brown, M., Filer, N., 1995. Beauty vs. the beast: the case against massively parallel retrieval. In: First
United Kingdom Case-Based Reasoning Workshop. Springer Verlag, in press.
Bullinaria, J., 1994. Internal representations of a connectionist model of reading aloud. In: Proceedings of
the Sixteenth Annual Meeting of the Cognitive Science Society. Erlbaum, Hillsdale, NJ.
226 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

Bullinaria, J., Chater, N., 1995. Connectionist modelling: implications for cognitive neuropsychology.
Language and Cognitive Processes 10, 227–264.
Chater, N., Oaksford, M., 1996. The falsity of folk theories: implications for psychology and philosophy.
In: O’Donohue, W., Kitchener, R. (Eds.), Psychology and Philosophy: Interdisciplinary Problems and
Responses. Sage, London.
Cheng, P., Holyoak, K., 1985. Pragmatic reasoning schemas. Cognitive Psychology 17, 293–328.
Chomsky, N., 1980. Rules and representations. The Behavioral and Brain Sciences 3, 1–61.
Chomsky, N., 1986. Knowledge of Language: Its Nature, Origin, and Use. Prager, Westport, CT.
Collins, A., 1992. On the paradox Kripke finds in Wittgenstein. Midwest Studies in Philosophy XVII, 74–
88.
Cosmides, L., 1989. The logic of social exchange: has natural selection shaped how humans reason?
Studies with the Wason selection task. Cognition 31, 187–276.
Cost, S., Salzberg, S., 1993. A weighted nearest neighbour algorithm for learning with symbolic features.
Machine Learning 10, 57–78.
Cover, T., Hart, P., 1967. Nearest neighbour pattern classification. IEEE Transactions on Information
Theory 13, 21–27.
Davies, M., 1995. Two notions of implicit rule. In: Tomberlin, J. (Ed.), Philosophical Perspectives, Vol. 9,
AI, Connectionism, and Philosophical Psychology. Ridgeview, Atascadero, CA.
Dayal, S., Harmer, M., Johnson, P., Mead, D., 1993. Beyond knowledge representation: commercial uses
for legal knowledge bases. In: Proceedings of the Fourth International Conference on Artificial
Intelligence and Law. ACM, New York, NY.
Delosh, E., 1993. Interpolation and extrapolation in a functional learning paradigm. Purdue Mathematical
Psychology Program, Purdue University.
Dreyfus, H., 1992. What computers still can’t do - a critique of Artificial Reason (third ed.). MIT Press,
Cambridge, MA.
Dulaney, D., Carlson, R., Dewey, G., 1984. A case of syntactical learning and judgement: how conscious
and how abstract? Journal of Experimental Psychology: General 113, 541–555.
Ervin, S., 1964. Imitation and structural change in children’s language. In: Lenneberg, E. (Ed.), New
Directions in the Study of Language. MIT Press, Cambridge, MA.
Feigenbaum, E., 1977. The art of Artificial Intelligence: themes and case studies of knowledge engineer-
ing. In: Proceedings of IJCAI-77.
Fodor, J., 1983. Modularity of Mind. Bradford Books, London, UK; MIT Press, Cambridge, MA.
Forrester, N., Plunkett, K., 1994. The inflectional morphology of the Arabic broken plural: a connectionist
account. In: Proceedings of the Sixteenth Annual Meeting of the Cognitive Science Society. Erlbaum,
Hillsdale, NJ.
Funnell, E., 1983. Phonological processing in reading: new evidence from acquired dyslexia. British
Journal of Psychology 74, 159–180.
Gentner, D., 1983. Structure-mapping: a theoretical framework for analogy. Cognitive Science 7, 155–
170.
Gentner, D., 1989. The mechanisms of analogical learning. In: Vosniadou, S., Ortony, A. (Eds.), Simi-
larity and Analogical Reasoning. Cambridge University Press, Cambridge, UK.
Gentner, D., Forbus, K.D., 1991. MAC/FAC: a model of similarity-based retrieval. In: Proceedings
of the Fifteenth Annual Meeting of the Cognitive Science Society. Erlbaum, Hillsdale, NJ, pp.
504–509.
Gentner, D., Markman, A., 1994. Structural alignment in comparison: no difference without similarity.
Psychological Science 5, 152–158.
Ginet, C., 1992. The dispositionalist solution to Wittgenstein’s problem about understanding a rule:
answering Kripke’s objections. Midwest Studies in Philosophy XVII, 53–88.
Ginsberg, M., 1987. Readings in Nonmonotonic Reasoning. Morgan Kaufmann, San Mateo, CA.
Glushko, R., 1979. The organization and activation of orthographic knowledge in reading aloud. Journal
of Experimental Psychology: Human Performance and Perception 5, 674–691.
Goldstone, R., 1994a. The role of similarity in categorization: providing a groundwork. Cognition 52,
125–157.
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 227

Goldstone, R., 1994b. Similarity, interactive activation, and mapping. Journal of Experimental Psychol-
ogy: Learning, Memory, and Cognition 20, 3–28.
Goodman, N. (1972). Seven Strictures on Similarity. In: Problems and Projects. Bobbs Merill, Indeana-
polis.
Goswami, U., Bryant, P., 1990. Phonological Skills and Learning to Read. Erlbaum, Hillsdale, NJ.
Hadley, R., 1993. The ’explicit/implicit’ distinction. Technical report CSS-IS TR93–02. Simon Frasier
University, Burnaby BC, Canada.
Hahn, U., 1996. Cases and Rules in Categorization. Ph.D. thesis, University of Oxford, UK.
Hahn, U., Chater, N., 1996. Understanding similarity: a joint project for psychology, case-based reason-
ing, and law. Artificial Intelligence Review, in press.
Hahn, U., Chater, N., 1997. Concepts and similarity. In: Lamberts, K., Shanks, D. (Eds.), Knowledge,
Concepts, and Categories. Psychology Press/MIT Press, Hove, UK.
Hahn, U., Nakisa, R., Plunkett, K., 1997. The dual-route model of the English past-tense: another case
where defaults don’t help. In: Proceedings of the GALA ’97 Conference on Language Acquisition.
Haugeland, J., 1985. Artificial Intelligence: The Very Idea. MIT Press, Cambridge, MA.
Hayes, P., 1979. The naive physics manifesto. In: Michie, D. (Ed.), Expert Systems in the Micro-electro-
nic Age. Edinburgh University Press, Edinburgh.
Herbig, B., Wess, S., 1992. Ähnlichkeit und Ähnlichkeitsmasse. In: Fall-basiertes Sehliessen – Eine
Übersicht. SEKI Working papers SWP-92–08,, University of Kaiserslautern, Germany.
Herrnstein, R., 1990. Levels of stimulus control: a functional approach. Cognition 37, 133–166.
Inhelder, B., Piaget, J., 1958. The Growth of Logical Reasoning. Basic Books, New York.
Johnson-Laird, P., Byrne, R., 1991. Deduction. Lawrence Erlbaum, Hillsdale, NJ.
Jordan, M., 1986. An introduction to linear algebra and parallel distributed processing. In: Rumelhart, D.,
McClelland, J. (Eds.), Parallel Distributed Processing:explorations in the Microstructure of Cognition,
Vol 1: Foundations. MIT press, Cambridge, MA.
Koh, K., Meyer, D., 1991. Function learning: induction of continuous stimulus-response relations. Journal
of Experimental Psychology: Learning, Memory, and Cognition, 17, 811–836.
Kolodner, J., 1991. Improving human decision making through case-based decision aiding. AI Magazine,
52–68.
Kolodner, J., 1992. An introduction to case-based reasoning. Artificial Intelligence Review 6, 3–34.
Komatsu, L., 1992. Recent views of conceptual structure. Psychological Bulletin 112, 500–526.
Kripke, S.A., 1982. Wittgenstein on Rules and Private Language. Blackwell, Oxford.
Kruschke, J., 1992. ALCOVE: an exemplar-based connectionist model of category learning. Psycholo-
gical Review 99, 22–44.
Lamberts, K., 1995. Categorization under time pressure. Journal of Experimental Psychology: General
124, 161–180.
Langley, P., 1996. Elements of Machine Learning. Morgan Kaufmann, San Francisco, CA.
Langley, P., Sage, S., 1994. Oblivious decision trees and abstract cases. In: Working Notes of the AAAI-
94 Workshop on Case-Based Reasoning. AAAI Press.
Marcus, G., Brinkmann, U., Clahsen, H., Wiese, R., Woest, A., Pinker, S., 1995. German inflection: the
exception that proves the rule. Cognitive Psychology 29, 189–256.
McCarthy, R., Warrington, E., 1986. Phonological reading: phenomena and paradoxes. Cortex 22, 868–
884.
McDermott, D., 1987. A critique of pure reason. Computational Intelligence 3, 151–160.
McDowell, J., 1984. Wittgenstein on following a rule. Synthese 58, 325–363.
Medin, D.L., Wattenmaker, W., 1987. Category cohesiveness, theories, and cognitive archaeology. In:
Neisser, U. (Ed.), Concepts and Conceptual Development: Ecological and Intellectual Factors in
Categorization. Cambridge University Press, Cambridge, UK.
Medin, D., Schaffer, M., 1978. Context theory of classification learning. Psychological Review 85, 207–
238.
Mitchell, T., 1990. The need for biases in learning generalizations. In: Shavlik, J., Dietterich, T. (Eds.),
Readings in Machine Learning. Morgan Kaufmann, San Mateo, CA.
Muggleton, S., 1992. Inductive Logic Programming. Academic Press, New York.
228 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

Myllymäki, P., Tirri, H., 1993. Massively parallel case-based reasoning with probabilistic similarity
metrics. In: First European Workshop on Case-Based Reasoning. Springer Verlag, Berlin.
Nakisa, R.C., Plunkett, K., Hahn, U., 1998. A cross-linguistic comparison of single and dual-route models
of inflectional morphology. In: Broeder, P., Murre, J. (Eds.), Cognitive Models of Language Acquisi-
tion. MIT Press, Cambridge, MA, in press.
Nakisa, R., Hahn, U., 1996. Where defaults don’t help: the case of the German plural system. In:
Proceedings of the 18th Annual Meeting of the Cognitive Science Society. Erlbaum, Mahwah, NJ,
pp. 177–182.
Newell, A., 1963. The chess machine. In: Sayre, K., Crosson, F. (Eds.), The Modeling of the Mind. Notre
Dame University Press, South Bend, IN.
Newell, A., 1991. Unified Theories of Cognition. Cambridge University Press, Cambridge, UK.
Newell, A., Simon, H., 1972. Human Problem Solving. Prentice-Hall, Englewood Cliffs, NJ.
Newell, A., Simon, H., 1990. Computer science as empirical enquiry: symbols and search. In: Bodenz, M.
(Ed.), The Philosophy of Artificial Intelligence. Oxford University Press, Oxford, UK.
Nisbett, R., Wilson, T., 1977. Telling more than we can know: verbal reports on mental processes.
Psychological Review 8, 231–259.
Nosofsky, R., 1984. Choice, similarity and the context theory of classification. Journal of Experimental
Psychology: Learning, Memory and Cognition 10, 104–114.
Nosofsky, R., 1988. Exemplar-based accounts of relations between classification, recognition, and
typicality. Journal of Experimental Psychology: Learning, Memory and Cognition 14, 700–708.
Nosofsky, R., Clark, S., Shin, H., 1989. Rules and exemplars in categorization, identification,
and recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition 15, 282–
304.
Nosofsky, R., 1992. Exemplars, prototypes, and similarity rules. In: Healy, A., Kosslyn, S., Shiffrin, R.
(Eds.), From Learning Theory to Connectionist Theory: Essays in Honor of William K. Estes. Erl-
baum, Hillsdale, NJ.
Oaksford, M., Chater, N., 1993. Mental models and the tractability of everyday reasoning. Behavioural
and Brain Sciences 16, 360–361.
Oaksford, M., Chater, N., 1991. Against logicist cognitive science. Mind and Language 6, 1–38.
Pavlov, I., 1927. Conditional Reflexes. Oxford University Press, London, UK.
Pennington, N., Hastie, R., 1993. Reasoning in explanation-based decision making. Cognition 49, 123–
163.
Pickering, M., Chater, N., 1995. Why cognitive science is not formalized folk psychology. Minds and
Machines 5, 309–337.
Pinker, S., Prince, A., 1988. On language and connectionism: analysis of a parallel distributed processing
model of language acquisition. Cognition 28, 73–193.
Plunkett, K., Marchman, V., 1991. U-shaped learning and frequency effects in a multi-layered perceptron:
implications for child language acquisition. Cognition 38, 43–102.
Porter, B., Bareiss, R., Holte, R., 1990. Concept learning and heuristic classification. Artificial
Intelligence 45, 229–263.
Posner, M., Keele, S., 1970. Retention of abstract ideas. Journal of Experimental Psychology 83, 304–
308.
Putnam, H., 1974. The ‘corroboration’ of theories. In: Schilpp, P. (Ed.), The Philosophy of Karl Popper,
Vol. 1. Open Court Publishing.
Pylyshyn, Z., 1984. Computation and Cognition. MIT Press, Cambridge, MA.
Quine, W., 1960. Word and Object. MIT Press, Cambridge, MA.
Reber, A.S., 1989. Implicit learning and tacit knowledge. Journal of Experimental Psychology: General
118, 219–235.
Redington, M., 1996. What is learnt in Artificial Grammar Learning? Ph.D. thesis, Department of Experi-
mental Psychology, University of Oxford.
Redington, M., Chater, N., 1994. The guessing game: a paradigm for artificial grammar learning. In:
Proceedings of the Sixteenth Annual Meeting of the Cognitive Science Society. Erlbaum, Hillsdale,
NJ.
 U. Hahn, N. Chater / Cognition 65 (1998) 197–230 229

Redington, M., Chater, N., 1996. Transfer in artificial grammar learning: a re-evaluation. Journal of
Experimental Psychology: General 125, 123–138.
Reed, S., 1972. Pattern recognition and categorization. Cognitive Psychology 3, 382–407.
Reiter, R., 1980. A logic for default reasoning. Artificial Intelligence 13, 81–132.
Rips, I., 1994. The Psychology of Proof. MIT Press, Cambridge, MA.
Rissland, E., Skalak, D., 1991. CABARET: rule interpretation in a hybrid architecture. International
Journal of Man-Machine Studies 34, 839–887.
Rissland, E., Skalak, D., Friedman, M., 1993. BankXX: A program to generate argument through case-
based search. In: Proceedings of the Fourth International Conference on Artificial Intelligence and
Law, ACM, New York, NY.
Rosch, E., Mervis, C., Gray, W., Johnson, D., Boyes-Braem, P., 1976. Basic objects in natural categories.
Cognitive Psychology 8, 382–439.
Ross, B., 1984. Remindings and their effects in learning a cognitive skill. Cognitive Psychology 16, 371–
416.
Ross, E., 1987. This is like that: the use of earlier problems and the separation of similarity effects. Journal
of Experimental Psychology: Learning, Memory and Cognition 13, 629–637.
Ross, B., Kennedy, P., 1990. Generalizing from the use of earlier exemplars in problem solving. Journal of
Experimental Psychology: Learning, Memory, and Cognition 16, 42–55.
Rumelhart, D., McClelland, J., 1986. On learning past tenses of English verbs. In: Rumelhart, D., McClel-
land, J. (Eds.), Parallel Distributed Processing, Vol 2: Psychological and Biological Models. MIT
press, Cambridge, MA.
Rumelhart, D., Todd, P., 1993. Learning and connectionist representations. Attention and Performance,
pp. 3–30.
Rumelhart, D., Zipser, D., 1985. Feature discovery by competitive learning. Cognitive Science 9, 75–112.
Schank, R., 1982. Dynamic Memory: A Theory of Learning in Computers and People. Cambridge
University Press, Cambridge, UK.
Searle, J., 1980. Rules and causation. Behavioral and Brain Sciences 3, 1–61.
Seidenberg, M., McClelland, J., 1989. A distributed, developmental model of word recognition and
naming. Psychological Review 96, 523–568.
Seifert, C., 1989. Analogy and case-based reasoning. In: Proceedings: Case-Based Reasoning Workshop.
Morgan Kaufmann, San Mateo, CA.
Selfridge, O., 1959. Pandemonium: a paradigm for learning. In: Office, L.H.S. (Ed.), Symposium on the
Mechanization of Thought Processes.
Shallice, T., 1988. From Neuropsychology to Mental Structure. Cambridge University Press, Cambridge,
UK.
Shanks, D., 1995. Rule induction. In: The Psychology of Associative Learning. Cambridge University
Press, Cambridge.
Shanks, D., John, M.S., 1994. Characteristics of dissociable human learning systems. Behavioral and
Brain Sciences 17, 367–395.
Shepard, R., 1957. Stimulus and response generalization: a stochastic model relating generalization to
distance in psychological space. Psychometrika 22, 325–345.
Shepard, R., 1980. Multidimensional scaling, tree-fitting, and clustering. Science 210, 390–399.
Shieber, S.M., 1986. An Introduction to Unification-Based Approaches to Grammar. Center for the Study
of Language and Information, Stanford, CA.
Shortliffe, E., 1976. Computer-based Medical Consultations: MYCIN. Elsevier, New York.
Sidman, M., Tailby, W., 1982. Conditional discrimination vs. matching to a sample: an expansion of the
testing paradigm. Journal of the Experimental Analysis of Behavior 37, 5–22.
Sloman, S., 1996. The empirical case for two systems of reasoning. Psychological Bulletin 119, 3–22.
Smith, E., Langston, C., Nisbett, R., 1992. The case for rules in reasoning. Cognitive Science 16, 1–40.
Smith, E., Medin, D., 1981. Categories and Concepts. Harvard University Press, Cambridge, MA.
Smolenov, H., 1987. Paraconsistency, paracompleteness and intentional contradictions. Journal of Non-
classical Logic 4, 5–36.
Touretzky, D., 1986. The Mathematics of Inheritance Systems. Morgan Kaufman, Los Altos, CA.
230 U. Hahn, N. Chater / Cognition 65 (1998) 197–230

Touretzky, D., Hinton, G., 1988. A distributed connectionist production system. Cognitive Science 12,
423–466.
Tversky, A., 1977. Features of similarity. Psychological Review 84, 327–352.
Vaughan, W., 1988. Formation of equivalence sets in pigeons. Journal of Experimental Psychology:
Animal Behavior Processes 14, 36–42.
Vokey, J., Brooks, L., 1992. Salience of item knowledge in learning artificial grammars. Journal of
Experimental Psychology: Learning, Memory, and Cognition 18, 328–344.
Vokey, J., Brooks, L., 1994. Fragmentary knowledge and the processing specific control of structural
sensitivity. Journal of Experimental Psychology: Learning, Memory and Cognition 18, 1504–1510.
Wason, P., 1968. Reasoning about a rule. Quarterly Journal of Experimental Psychology 20, 273–281.
Westermann, G., Goebel, R., 1994. Connectionist rules of language. In: Proceedings of the Seventeenth
Annual Meeting of the Cognitive Science Society. Erlbaum, Hillsdale, NJ, pp. 236–241.
Wettschereck, D., Aha, D., 1995. Weighting features. In Proceedings of the First International Conference
on Case-Based Reasoning.
Wettschereck, D., Aha, D., Mohri, T., 1995. A review and comparative evaluation of feature weighting
methods for lazy learning algorithms. Technical report AIC-95–012, Navy Center for Applied
Research in AI, Washington DC.
Young, R., O’Shea, T., 1981. Errors in children’s subtraction. Cognitive Science 5, 153–177.