Learning and Cognition—Issues and Concepts: Concept Learning

Concepts form the building blocks of thought and the present review demonstrates that concept learning is dynamic and complex. Various theories of concept representation and learning are reviewed, including classical, prototype, exemplar, and theory theories. Research suggests that the nature of conceptual organization changes as a function of expertise and cultural experience. Furthermore, research shows that individual goals influence the construction of concepts and that predisposed constraints on concept formation are specific to content domain. To illustrate the interaction between various factors in concept learning we examine the domains of folkbiology and mathematics and consider developmental and cross-cultural research. Learning and Cognition—Issues and Concepts: Concept Learning In what follows, we use concept to refer to a mental representation and category to refer to the set of entities or examples picked out by the concept. It is generally accepted that instances of a concept are organized into categories. Almost all theories about the structure of categories assume that, roughly speaking, similar things tend to belong to the same category and dissimilar things tend to be in different categories. For example, robins and sparrows both belong to the category bird and are more similar to each other than they are to squirrels or pumpkins. Similarity is a pretty vague term, but most commonly it is defined in terms of shared properties or attributes. Although alternative theories assume concepts are structured in terms of shared properties, theories differ greatly in their organizational principles. The Classical View The classical view assumes that concepts have defining features that act like criteria or rules for determining category membership. For example, a triangle is a closed geometric form of three sides with the sum of the interior angles equaling 180 degrees. Each of these properties is necessary for an entity to be a triangle, and together these properties are sufficient to define triangle. A fair amount of research has examined people’s knowledge about object categories like bird, chair, and furniture and this evidence goes against the classical view. Not only do people fail to come up with defining features but also they do not necessarily agree with each other (or even with themselves when asked at different times) on whether something is an example of a category. Philosophers and scientists also have worried about whether naturally occurring things like plants and animals (so-called “natural kinds”) have defining features. The current consensus is that most natural concepts do not fit the classical view. The Probabilistic View The major alternative to the classical view is the probabilistic view which argues that concepts are organized around properties that are characteristic or typical of category members but crucially, they need not be true of all members. That is, the features are only probable. For example, most people’s concept of bird may include the properties of building nests, flying, and having hollow bones, even though not all birds have these properties (e.g., ostriches, penguins).The probabilistic view has major implications for how we think about categories. First, if categories are organized around characteristic properties, some members may have more of these properties than other members. In this sense, some members may be better examples or more typical of a concept than others. For example, it has been found that the more frequently a category member’s properties appeared within a category, the higher was its rated typicality for that category. For instance, robins were rated to be very typical birds and penguins are rated as very atypical birds. A second implication is that category boundaries may be fuzzy. Nonmembers of a category may have almost as many characteristic properties of a category as do certain members. For example, whales have a lot of the characteristic properties of fish, and yet they are mammals. Third, learning about a category cannot be equated with determining what the defining features are because there may not be any. Typicality: Central Tendency vs. Ideality. Is typicality only based on central tendency? Although typicality effects are robust (and problematic for the classical view), other research shows that the underlying basis for typicality effects may vary with both the kind of category being studied and with the population being studied. While the internal structure of taxonomic categories is based primarily on the central tendency (or the average member) of a category, the internal structure of goal-derived categories such as “things to wear in the snow” is determined by some ideal (or the best possible member) associated with the category. The best example of snow clothing, a down jacket, was not the example that was most like other category members; instead it was the example with the maximum value of the goal-related dimension of providing warmth. One might think that ideals will only come into play when the category of interest lacks the natural similarity structure that characterizes common taxonomic categories such as bird, fish, and tree. However, for tree experts (people who know a lot about trees such as landscapers, parks workers and taxonomists), the internal structure of the category tree is organized around the positive ideal of height and the negative ideal of weediness. The best examples of tree are not trees of average height but trees of extraordinary height (and free of “weedy” characteristics like having weak limbs, growing where they aren’t wanted, and being susceptible to disease). Indeed, research does suggest that people who have considerable knowledge in a domain tend to base typicality judgments on ideals and not the number of typical features. For example, for Itzá Maya adults living in the rainforests of Guatemala the best example of bird is the wild turkey which is culturally significant, prized for its meat, and strikingly beautiful. The fact that U.S. tree experts based typicality on ideals suggests that it’s not just that the Itzá have a different notion of what typicality means. It has also been found that found that Native American and European American fishermen’s typicality judgments were based on ideals though those ideals differed somewhat across groups. Prototype vs. Exemplar Theories. If categories are not represented in terms of definitions, what form do our mental representations take? One suggestion about how concepts are represented is known as the family resemblance principle. The general idea is that category members resemble each other in the way that family members do. A simple summary representation for such a family resemblance structure would be an example that possessed all the characteristic features of a category. The best example is referred to as the prototype. In a prototype model of categorization, classifying a new example is done by comparing the new item to the prototype. If the candidate example is similar enough to the prototype for a category, it is classified as a member of that category. More detailed analyses, however, show problems with prototypes as mental representations. Prototype theory implies that the only information abstracted from categories is the central tendency. A prototype representation discards information concerning category size, the variability of the examples, and correlations among attributes, and people can use all three of these types of information. An alternative approach, which is also consistent with the probabilistic view, assumes that much more information about specific examples is preserved. This approach appropriately falls under the general heading of exemplar theories. Exemplar models assume that people initially learn some examples of different concepts and then classify a new instance on the basis of how similar it is to the previously learned examples. The idea is that a new example reminds the person of similar old examples and that people assume that similar items will belong to the same category. For example, suppose you are asked whether large birds are more or less likely to fly than small birds. You probably will answer “less likely,” based on retrieving examples from memory and noting that the only non-flying birds you can think of are large (e.g., penguin, ostrich). Quite a few experiments have contrasted the predictions of exemplar and prototype models. In head-to-head competition, exemplar models have been considerably more successful than prototype models. Why should exemplar models fare better than prototype models? One of the main functions of classification is to allow one to make inferences and predictions on the basis of partial information. Relative to prototype models, exemplar models tend to be conservative about discarding information that facilitates predictions. For instance, sensitivity to correlations of properties within a category enables finer predictions: From noting that a bird is large, one can predict that it cannot sing. In short, exemplar models support predictions and inferences better than do prototype models. More recent research has pointed to three major limitations of these simple forms of prototype and exemplar models: 1. they have narrowly focused on categorization and have paid little attention to how other conceptual functions such as communication and inference may affect concept representation and learning, 2. they view learning as a passive accumulation of statistical information rather than active learning that may reflect particular learner goals, and 3. they pay little attention to how theoretical notions and causal reasoning organize learning. With respect to the second point, we have just reviewed evidence from a number of populations indicating that typicality is driven by ideals and that later learning builds on earlier learning. If category ideals tend to be learned first then they will have an important role in the development of categories, and modelers are beginning to shift to this more active view of learning. With respect to the role of theories, there is evidence that using (abstract) similarity relations may be likely to be a strategy of last resort, used only when more relevant information is unavailable. Let’s examine the theory view in a bit more detail. The Theory View A number of researchers have argued that the organization of concepts is knowledgebased (rather than similarity-based) and driven by intuitive theories about the world. The idea that concepts might be knowledge-based rather than similarity-based suggests a natural way in which concepts may change—namely, through the addition of new knowledge and theoretical principles. There is also good evidence that these theories help determine which abstract and observable features learners pay attention to. We have a different set of categories for mental disorders now than we had 100 years ago, in part because our knowledge base has become more refined. Often knowledge of diseases develops from information about patterns of symptoms to a specification of underlying causes. For example, the advanced stages of syphilis were treated as a mental disorder until the causes and consequences of this venereal disease were better understood. Recently, it has been shown that clinical psychologists organize their knowledge of mental disorders in terms of rich causal theories and that these theories (and not the atheoretical diagnostic manual they are supposed to use) guide their diagnostic classification and reasoning.