Concepts I: The Classical View

In response to my most recent plea for requests, Clark wrote:

I'd be interested in how cognitives looks at the problem of reference. Does cognitive science tend to be internalist or externalist and how does it deal with the "world out there."

I started to try to address this, but began to realize that each time I did so, I had to assume more knowledge of cognitive theories of concepts in readers than is probably reasonable. So, in order to get to a discussion of reference, I want to start with a discussion of theories of concepts. It may even become clear how cognitive scientists, or at least those cognitive scientists who study concepts, deal with reference and "the world out there" by the time we're done with the background. If not, then I can talk about that more directly afterwards.

Before I get started, I should note that the terms "concepts" and "category" are painfully ambiguous in the cognitive scientific literature. Bertrand Russell once remarked¹:

What, exactly is meant by the word ‘category’, whether in Aristotle or in Kant and Hegel, I must confess that I have never been able to understand.

He might as well have substituted classical, prototype, and exemplar theories for Aristotle, Kant and Hegel, because the use of term "category," along with "concept," is very difficult to pin down in cognitive science. Because the terms "category" and "concept" are used pretty much interchangably in the literature, I'm going to do the same, but I'm not going to make any attempt to try to define either term, so you'll just have to deal with the ambiguity.

Cognitive scientists who study concepts are generally concerned with three issues: how concepts are represented, how we classify instances (which I will call exemplars from here on out) as belonging to a concept, and how we use concepts in reasoning. Of course, cognitive scientists aren't the first to address these issues. Since the days of Plato and (more importantly for our purposes) Aristotle, the nature of concepts, in their representations, how we recognize exemplars of particular concepts, and how we reason with concepts, has been one of the most studied issues in philosophy. And since the days of Aristotle, one view of concepts has dominated. Under that view, which is now commonly referred to as the classical or definitional theory of concepts, concepts are defined by a set of necessary and sufficient features. In this post, I'm going to discuss the classical theory of concepts, and it's many (many, many) problems. In subsequent posts, I'll talk about the theories that have replaced it, starting with the two main types of similarity-based theories (prototype and exemplar theories), and then moving on to more recent theories that attempt to account for some of the deficiencies of similarity-based accounts. If, by the end of this series of posts, you haven't become fully aware of the incredibly muddled state of concept research, then I will have failed to do my job.

Concepts: The Classical View

For both Plato and Aristotle, concepts were defined by their essences. While Plato's focus on ideal forms doesn't exactly lend itself to scientific theories of concepts, many theories of concepts in the first half of the 20th century (such as those one might find among the logical positivists) were similar to Aristotle's, in that they treated concepts as being defined by a set of empirically-discoverable (though sometimes not directly observable) necessary and sufficient features. This view of concepts is actually very powerful, and provides a simple account of several features of concepts that any theory must explain. For instance, it provides a straightforward explanation for how we separate members of a category from non-members. Members of a category are just those exemplars that exhibit the necessary and sufficient features that define the category, and any exemplars that do not exhibit those features are not members of the category. It also provides for an intuitive account of concept formation. We form concepts by encountering many examplars, and extracting the features that unequivocably divide these exemplars into separate classes. Finally, as researchers studying semantic networks in the 1960s and 70s showed², the classical theory provides a nice way to construct taxonomic relationships between concepts. Members of a taxonomy are related by their necessary and sufficient features, with more subordinate members being defined by the same features that define their superordinates, along with an additional set of features that distinguishes them from other superordinates.

It was the power and intuitive appeal of the classical theory that allowed it to survive relatively unchallenged for more than 2 millenia. However, in the 1950s, the classical view began to come under attack from philosophical circles, and by the end of the 1970s, there were few cognitive scientists who still held it. The criticisms are many. In 1951, Quine famously criticized the classical account because of its attachment to the analytic-synthetic distinction. Around the same time, Wittgenstein's criticisms of definitions were also being published. As everyone knows, he used GAME as an example of a concept that admits no definition. Subsequent research has shown that concepts like GAME are not exceptions to the general rule that concepts are defined by their necessary and sufficient features, but that in fact, definitions may not exist for any real-world concepts. Take, for example, the concept BACHELOR, which has often (even by Quine) been used as an example of a concept which is clearly defined by a set of necessary and sufficient features. For BACHELOR, those features are male and unmarried, and thus the definition of BACHELOR is "an unmarried male". Yet, if we adopt the classical view, and treat this as the concept's representation, we immediately begin to run into problems. First of all, unmarried male children are not bachelors, and neither are people who've been married and are no longer, so we must amend the definition by inserting the word "adult" between "unmarried" and "male," and perhaps change the wording a bit to make it clear that bachelors are adult males who've never been married. Then, we have to come up with a qualification in the definition that allows us to exclude many other types of individuals. For instance, are middle-aged gay men who have been in monogomous relationships for years bachelors? What about priests? Most of us probably feel that neither of these types of individuals classify as bachelors, and if we thought a little, we could probably come up with a whole host of other types of individuals that fit the definition, but that we wouldn't classify as bachelors. In the end, we're going to end up with a really long set of disjunctive rules to keep out all the non-bachelors who fit with our first definition of bachelors. Unfortunately for adherents of the classical view, the same is probably true of every other everyday concept as well.

There may be other philosophical problems for the classical view, as well. As Eric Margolis and Stephen Laurence have argued³, the classicical view is really just an instance of a descriptivist theory of reference, within which concepts refer to descriptions of real-world instances. For this reason, they believe that the classical view is subject to the same criticisms that Kripke⁴ and Putnam⁵ use against descriptive theories of reference. The primary criticism is that it is possible to discover that our descriptions, or definitions, are wrong. I'm not so sure that this is a problem for cognitive psychological versions of the classical theory, however. Since psychologists argue that the discovery of the essences, or definitions, of concepts is a learning process, it stands to reason that our representations of concepts can be erroneous. Even if we do no have all the information about the class to which a concept refers, we could still represent the concept with a definition, and when we discover that parts of our definition are erroneous, we can revise our representations.

Even if adherents of the classical view can get around Kripke and Putnam's arguments, the lack of definitions, and Quine's problem of analyticity, are pretty damning. However, the philosophical problems are nothing compared to the psychological ones, and there are many of those. First, when queried, people are rarely able to produce definitions of the concepts they employ⁶, and when they do produce definitions, they often differ greatly both across subjects and for the same subject over time⁷, making it difficult to explain how people are able to communicate or reason about concepts over time, if they represent concepts as definitions. In most cognitive versions of the classical theory, definitions are represented in the form of rules, with Boolean disjunctions representing the set of necessary and sufficient features. However, researchers have shown that people are not conscious of the rules that define many concepts, and that some concepts may not be represented by any sort of rule⁸. Furthermore, if there are definitions, but they must account for a large number of problem instances (e.g., the gay males and priests in the BACHELOR example), then definitions will become increasingly difficult to learn. In fact, by the time you add all the Boolean disjunctions necessary to define a concept as simple as BACHELOR, you've probably made the defininition unlearnable.

Experiments designed to test the classical theory have failed to show evidence of definitions as well. For instance, when participants are presented with one concept (like BACHELOR), their reaction times are no slower for that concept than for concepts that make up part of their definition (e.g., MALE)⁹. Since, under the definitional view, the complexity of a concept should be a function of the complexity of the concepts that are used to define it, we would expect reaction times to be slower for BACHELOR than MALE. Perhaps the most damaging empirical finding, however, was that of typicality effects. Elanor Rosch and her colleagues¹⁰ conducted several experiments demonstrating that some exemplars (e.g., robins as a member of the category bird) are treated as better members of a category (i.e., more typical) than others (e.g., penguins or kiwis... sorry Richard). The persistence and predictive power of typicality poses a serious problem for any definitional account. If concepts are represented only by their definitions, then exemplars are either exhibit those features or do not, and thus are either members of a category or not. This was one of the positive qualities of classical theories above, but after Rosch's research, it makes the classical view difficult to retain. Any additional information added to the concept to account for typicality effects would have to be something other than a definition.

If the existence of typicality effects weren't enough, further research in the 1970s demonsrated that category membership is often not a binary relation (an exemplar is either a member of a category or it is not). Instead, category boundaries tend to be fuzzy, and there is a great deal of both inter-subject and intra-subject disagreement about category membership, about whether a particular exemplar is a member of a category¹¹. Once again, this goes against the all-or-none membership that the classical theory predicts.

So, by the beginning of the 1980s, the consensus among cognitive scientists was that it was time to get rid of the classical theory. Many, especially Rosch and her colleagues, had been developing alternative theories since the mid-70s, which treated categories not as definitions, but as family-resemblance structures, and category membership was determined by determining the similarity between an exemplar and the features used to represent the concept. The period between 1975 and 2000 can probably be called the age of the similarity-based views (which is not to imply that they are not still prominent, just that there weren't really any alternatives until recently). There are two general types of similarity-based views, prototype theories and exemplar theories. In the next post, I'll discuss prototype theories. After that, exemplars.

¹ Russel, Bertrand (1946). A History of Western Philosophy. London: George Allen & Unwin.
² Collins, A.M., & Quillian, M.R. (1969). Retrieval time from semantic memory. Journal of Verbal Learning and Verbal Behaviour, 8, 240-247.
³ Margolis, E., & Laurence, S. (2003). Concepts. In Stich, S. P. (Ed.), Blackwell Guide to Philosophy of Mind.
⁴ Kripke, S. (1980). Naming and Necessity. Cambridge, MA: Harvard University Press.
⁵ Putnam, H. (1975). The Meaning of 'Meaning'. In K. Gunderson (Ed.), Language, Mind and Knowledge. Minneapolis: University of Minnesota Press.
⁶ McNamara, T.P., & Sternberg, R.J. (1983). Mental models of word meaning. Journal of Verbal Learning and Verbal Behavior, 22, 449-474.
⁷ Rosch, E. (1975a). Cognitive representations of semantic categories. Journal of Experimental Psychology: General, 104, 192-233.
⁸ Maddox, W.T., Filoteo, J.V., Lauritzen, J.S., Connally, E., & Hejl, K.D. (In Press). Disontinuous Categories Affect Information-Integration, but not Rule-Based Category Learning. Journal of Experimental Psychology: Learning, Memory, and Cognition.
⁹ Fodor, J. A., Garrett, M., Walker, E., & Parkes, C. (1980). Against Definitions. Cognition, 8, 263-367.
¹⁰ Starting with Rosch, E. (1973). On the internal structure of perceptual and semantic categories. In T. E. Moore (Ed.), Cognitive Development and the Acquisition of Language. New York, Academic Press.
¹¹ McCloskey, M., & Glucksberg, S. (1978). Natural Categories: Well-defined or fuzzy sets? Memory and Cognition, 6, 462-472. Barsalou, L. (1989). Intraconcept similarity and its implications for interconcept similarity. In S. Vosniadou & A. Ortony (Eds.), Similarity and Analogical Reasoning.