Three Disputes

Three Disputes
A. Exemplars versus Prototypes
I will not attempt anything like a complete review of the exemplar–
prototype debate here. In outline, the two positions are as follows. Prototype
theory (Hampton, 1979, 1995; Rosch, 1975; Rosch & Mervis, 1975; Smith &
Medin, 1981) argues that concepts are summary representations of the
central tendency of a category. For example, I have a concept of telephones
that includes information, such as their usual shapes, sizes, functions, inner
workings, colors, and parts. This information is a summary representation
in that it stands for the category1
as a whole. This representation does not
describe a specific object but rather sets out the likely properties of all the
objects in the category (Rosch & Mervis, 1975). Thus, I might store the fact
that telephones have buttons that you press or dials that you turn, even
though no individual telephone (that I have seen) has both. Exemplar theory
denies that such general representations are created. Instead, it argues that
my concept of telephones is the set of memories I have of specific telephones.
1
I generally use category to refer to the actual set of objects and concept to refer to the
mental representation of those objects. Thus, one would pedantically say that I have a concept
of dogs that picks out the category of dogs. However, I avoid such pedantry where possible and
sometimes use category ambiguously when I want to refer to both the concept and category, as
the two generally do go together.
2 Gregory L. MurphySo, if I am trying to decide if something is a telephone, I do not consult
information about telephones in general but instead consult memories of
past telephones I have known.
These theories diVer greatly in the processes they posit for how concepts
are learned, in how concepts are represented, and in how categorization and
other conceptual processes take place. These diVerences are so great that
one must imagine that the theories can readily be distinguished. In one
sense, they can be. Studies by Medin, Nosofsky, Kruschke, and others have
found that exemplar models generally do better than prototype models in
carefully controlled category-learning experiments, in which mathematical
instantiations of the models could be compared (for reviews, see Murphy,
2002; Nosofsky, 1992; Ross & Makin, 1999). In a number of comparisons,
exemplar models fit the data as well as or just slightly better than prototype
models; however, in other comparisons, prototype models were totally
unable to explain the results. For example, prototypes are incompatible with
nonlinearly separable categories, yet people are able to learn them (Estes,
1986; Medin & Schwanenflugel, 1981).
Despite this overall advantage of exemplar models in traditional category-
learning experiments, a number of researchers in the field have been slow to
adopt this type of model. In cognitive development, virtually all research
seems to assume a summary representation of a category. Researchers who
address the interaction of concepts with world knowledge almost all speak
in terms of prototypes, perhaps because it is diYcult to represent world
knowledge in terms of exemplars (Murphy, 2002). Those who study the
interaction of word meaning and concepts (e.g., conceptual combination,
language production) all take a prototype view. Why haven’t the experi-
mental demonstrations of exemplar model superiority filtered down to these
other domains?
B. Knowledge in Category Learning
A major development in the study of concepts has been the attempt to
integrate conceptual knowledge with more general knowledge of the
concept’s domain. For example, it has been argued that learning a concept
of a new animal, perhaps seen at the zoo, involves knowledge of other
animals and of biology in general. Such knowledge serves to aid the
induction process so that a rich representation of the animal can be
constructed based on relatively little experience. For example, when I see a
new mammal in the African exhibit at the zoo, I might assume that it
breathes, gives birth to live young, can stand high temperatures, and so on,
even if I have not observed these properties directly. This use of background
knowledge is found not only in category learning, but in induction,
Ecological Validity and the Study of Concepts 3conceptual combination, and other conceptual processes (Heit, 1997;
Murphy, 1993, 2002).
Numerous demonstrations show that prior knowledge influences concep-
tual processes. Nonetheless, the main theories of concepts have not taken up
this influence. For example, even recent comparisons of prototype and
exemplar models use models that do not incorporate any knowledge, and
they are tested on literally meaningless categories (e.g., Nosofsky &
Johansen, 2000; Smith & Minda, 2000). The problem with this is that
eVects found with abstract, meaningless categories are often not found or
are even reversed when the categories are meaningful. For example,
nonlinearly separable (NLS) categories are much harder to learn than
linearly separable (LS) categories when the categories’ features are related to
one another meaningfully (Murphy & Kaplan, 2000; Wattenmaker, Dewey,
Murphy, & Medin, 1986), but this is not true for abstract categories.
Categories formed by disjunctive rules are usually diYcult to learn, but they
are not if the rule is consistent with prior knowledge (Pazzani, 1991). When
empirical factors have been pitted against consistency with knowledge in
classification tests, knowledge has often been found to be more important
(Keil, 1989; Wisniewski, 1995).
In short, demonstrations of the use of real-world knowledge might have
been expected to change the nature of the experiments done in the concepts
field. But rather than changing the direction of the field as a whole, the main
eVect has been to create two parallel tracks of research: one investigating
structural eVects in abstract categories and one exploring the influence of
knowledge in meaningful categories. Although it is certainly possible that
both will make contributions to our understanding of concepts, it also seems
possible that one of the approaches is not right, or at least not as useful as
the other. However, it is unclear on what basis that judgment should be
made, as each tradition now has a set of empirical findings to point to as
documenting the importance of its own questions.
C. Studies of Children versus Studies of Adults
The final conflict in the field is more empirically oriented. As described
elsewhere briefly (Murphy, 2001), researchers in cognitive development
seem to have a very diVerent idea of how concepts are structured than
researchers in adult concepts. The typical experiment on concept learning in
adults requires subjects to categorize items one at a time, over and over,
until they get all the items correct. The diVerence in learning times for
diVerent concepts is often the critical data used to evaluate diVerent
theories. The concepts in such studies are often extremely diYcult to learn.
For example, Medin and Schwanenflugel (1981) constructed two categories
4 Gregory L. Murphywith four items apiece, and subjects attempted to learn them over the course
of 16 repetitions of the set of items. By the end of the learning, about a third
of the subjects still had not mastered the categories. That is, they had not
learned to say ‘‘category 1’’ to four items and ‘‘category 2’’ to the other four.
Lamberts (2000) did not stop subjects after a set number of blocks but
instead allowed them to continue until they had learned to categorize nine
items perfectly. His subjects took an average of 38 blocks to do so, which
seems extremely long. Once one starts looking for it, such poor performance
in category-learning experiments is not unusual. Indeed, it occurs in some of
my own studies (e.g., Murphy & Wisniewski, 1989). In some designs, the
categories are probabilistic, and subjects cannot score more than 80%
correct (e.g., Maddox, 1995).
The reason for this poor performance is not hard to find. Basically, the
categories used in such experiments are poorly structured (Smith & Minda,
1998; Smith, Murray, & Minda, 1997). The categories have few features that
are common to most members, and some of the members are very similar
to members of the other categories. Such experiments seem to have an
underlying assumption that the categories people learn are quite diYcult
and therefore that considerable experience is necessary to learn the detailed
structure of the category.
However, studies of concept and word learning in children do not seem to
share this assumption. This is best illustrated through the phenomenon of
fast mapping (Carey, 1978; Markson & Bloom, 1997). In fast mapping, a
child is shown one or two exemplars of a category, which are named two or
three times. The naming may be overt (e.g., ‘‘Look at the koba’’) or it may
be indirect (e.g., ‘‘Can you hand me the chromium one?’’). No further
explanation or definition of the new name is given. The surprising aspect of
this phenomenon is that children learn words under these circumstances and
can remember them at least to some degree a month later, with no
intervening exposure.
The memory for the word is of interest in its own right, but for the present
purposes, the critical aspect of the fast mapping situation is its assumptions
about the categories, underlying common words. In contrast to the adult
literature, the notion of fast mapping seems to assume that the child can
figure out the category picked out by the new word based on a single
example. Clearly, these researchers do not believe that learning of real
categories requires exposure to a large number of category exemplars, to be
studied with great eVort over the course of, say, 38 blocks. If they did, they
would not show one exemplar and then test learning. (From now on, I will
flout convention somewhat by using fast mapping to refer to the procedure
of teaching a category by presenting one or two exemplars rather than to the
learning process.)
Ecological Validity and the Study of Concepts 5The reason for the empirical discrepancy between adult and child studies
is also not hard to find. In the fast-mapping study, the child views one or
two exemplars and then is typically tested on the same exemplar or one that
is extremely similar. There are no category members that are not similar to
the learning item. The assumption is apparently that category members are
generally uniform. But this assumption is completely diVerent from that of
many adult studies, which have very weak category structure. In some cases,
the exact same stimulus can be in diVerent categories (e.g., Gluck & Bower,
1988).
Two possible interpretations of the diVerence between adult and child
studies suggest themselves. One interpretation is that children are very
bright creatures who require minimal exposure to learn categories.
Unfortunately, they grow up into very dull college sophomores who cannot
learn categories without blocks and blocks of learning—a striking failure of
our educational system. The other alternative is that one of these types of
study is not right—its assumptions about category structure do not
correspond to reality. Or, less optimistically, both are wrong.
D. Summary of the Three Disputes
What is surprising about these three cases is not that there are disagree-
ments. Theoretical disagreements are part and parcel of science (although
one might argue that psychologists have a tendency to prolong such
disagreements beyond what is healthy). What is surprising is that very basic
questions about the nature of concepts and the acquisition process remain
unresolved. We have reached the 25th anniversary of the main proposal of
the exemplar theory of concepts (Medin and SchaVer, 1978). After intensive
research resulting in dozens of articles, however, the field does not seem any
more in consensus than it was 25 years ago. Are concepts complex, poorly
structured entities that take much experience and eVort to learn, or are they
simple things that you can get pretty well after a single example? That seems
to me to be an extremely basic question about the topic we are studying.
How can we go forward at all without knowing the answer to it? And if
prior knowledge strongly influences learning and concept use, what are we
to think of these models and theories that simply ignore it?
It would be easiest to attribute these conflicts to the shortsightedness of
researchers who do not wish to address data that are problematic for their
own approaches. No doubt this is part of the explanation. However, I will
argue that these conflicts come about in large part because of a lack of
agreement on what the basic questions are that the psychology of concepts
should answer. Without some agreement on these questions, it will not be
possible for researchers to have a basis on which to resolve their diVerences.
6 Gregory L. MurphyI