Machine Generalization and Human Categorization: An Information-Theoretic View (original) (raw)

Explaining basic categories: Feature predictability and information

Psychological Bulletin, 1992

The category utility hypothesis holds that categories are useful because they can be used to predict the features of instances and that the categories that tend to survive and become preferred in a culture (basic-level categories) are those that best improve the category users' ability to perform this function. Starting from this hypothesis, a quantitative measure of the utility of a category is derived. Application to the special case of substitutive attributes is described. The measure is used successfully to predict the basic level in applications to data from hierarchies of natural categories and from hierarchies of artificial categories used in category-learning experiments. The relationship of the measure to previously proposed indicators of the basic level is discussed, as is its relation to certain concepts from information theory. Categorization is one of the most basic cognitive functions. Why is the ability to categorize events or objects important to an organism? An obvious answer to this question is that categories are important because they often have functional significance for the organism. Another familiar answer is that grouping objects into categories allows for efficient storage of information about these groups of objects. One purpose of this article is to explore connections between these two answers regarding the utility of categories. The idea that categories serve certain functions for the organism raises the possibility that some categories fulfill these functions better than others. The clearest evidence that certain natural categories are "better" than others stems from the work on "basic-level" categories (Mervis & Rosch, 1981; Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1976). A basic-level category is one that is preferred by people over its superordinate and subordinate categories. For example, when shown a picture of a particular object, most people will identify it as a chair rather than as furniture or a kitchen chair. From this and other evidence, chair is considered to be a basic-level category for most people. A variety of empirical phenomena demonstrates the superiority of basic-level categories (Mervis & Rosch, 1981). As suggested above, when people are shown an object, they tend to name it at the basic level (Rosch et al, 1976). In recognition tasks, people recognize basic-level objects faster than either subordinates or superordinates (Jolicoeur, Gluck, & Kosslyn, 1984; Rosch et al., 1976). Basic-level names generally have arisen earlier in the development of languages (Berlin, Breedlove, & Raven, 1973), and basic categories are used earlier in the naming and other behavior of young children (Anglin, 1977; Brown, Many people have provided us with suggestions, criticisms, and comments on this work over the last 6 years. However, we must especially thank Pat Langley and Doug Fisher for their early and continuing encouragement. Doug Fisher's own research has contributed to our understanding of many of the issues discussed here. We also thank three anonymous reviewers for their insightful and detailed feedback.

Object categories and expertise: Is the basic level in the eye of the beholder?

Cognitive Psychology, 1991

Classic research on conceptual hierarchies has shown that the interaction between the human perceiver and objects in the environment specifies one level of abstraction for categorizing objects, called the basic level, which plays a primary role in cognition. The question of whether the special psychological status of the basic level can be modified by experience was addressed in three experiments comparing the performance of subjects in expert and novice domains. The main findings were that in the domain of expertise (a) subordinate-level categories were as differentiated as the basic-level categories, (b) subordinate-level names were used as frequently as basic-level names for identifying objects, and(c) subordinatelevel categorizations were as fast as basic-level categorizations. Taken together, these results demonstrate that individual differences in domain-specific knowledge affect the extent that the basic level is central to categorization. 8 1991 Ara&mir Prccr. Inr In a series of important experiments, Rosch, Mervis. Gray, Johnson, and Boyes-Braem (1976) established that a basic level of abstraction has special significance in human categorization (also see Brown, 1958). The basic level was shown to be the most inclusive level at which a generalized shape of category exemplars is identifiable and imaginable. In addition, basic categories elicit similar motor programs and basic-level category labels are the first names learned by children. Based on their analysis of structure at the basic level, Rosch et al. (1976) predicted that basic-level categories would be the classifications made when objects are first perceived. Rosch et al. (1976) demonstrated the special status of basic-level categories for object identification in a free-naming and a categoryverification task. In the naming task, subjects were presented with a series of pictures in rapid succession and were asked to write down the We would like to thank our subjects for their enthusiastic support of this project, Gregory Murphy for his expert advice, Robert Mauro for the use of his lab, and

A 'Normal'Category-Specific Advantage for Naming Living Things

Neuropsychologia, 1999

`Artefactual' accounts of category-speci®c disorders for living things have highlighted that compared to nonliving things, living things have lower name frequency, lower concept familiarity and greater visual complexity and greater within-category structural similarity or `visual crowding'. These hypotheses imply that de®cits for living things are an exaggeration of some `normal tendency'. Contrary to these notions, we found that normal subjects were consistently worse at naming nonliving than living things in a speeded presentation paradigm. Moreover, their naming was not predicted by concept familiarity, name frequency or visual complexity; however, a novel measure of visual familiarity (i.e. for the appearance of things) did signifcantly predict naming. We propose that under speeded conditions, normal subjects and nonliving things harder to name because their representations are less visually predictable than for living things (i.e. nonliving things show greater within-item structural variability). Finally, because nonliving things have multiple representations in the real world, this may lower the probability of finding impaired naming and recognition in this category.

Object categorization: Reversals and explanations of the basic-level advantage

Journal of Experimental Psychology: General, 2007

People are generally faster and more accurate to name or categorize objects at the basic level (e.g., dog) relative to more general (animal) or specific (collie) levels, an effect replicated in Experiment 1 for categorization of object pictures. To some, this pattern suggests a dual-process mechanism, in which objects first activate basic-level categories directly and later engage more general or specific categories through the spread of activation in a processing hierarchy. This account is, however, challenged by data from Experiment 2 showing that neuropsychological patients with impairments of conceptual knowledge categorize more accurately at superordinate levels than at the basic level-suggesting that knowledge about an object's general nature does not depend on prior basic-level categorization. The authors consider how a parallel distributed processing theory of conceptual knowledge can reconcile the apparent discrepancy. This theory predicts that if healthy individuals are encouraged to make rapid categorization responses, the usual basic Ͼ general advantage should also reverse, a prediction tested and confirmed in Experiment 3. Implications for theories of visual object recognition are discussed.

How children know the relevant properties for generalizing object names

2002

Abstract Young children's novel word extensions indicate that their animal categories, like those of adults, are characterized by multiple similarities among instances; whereas their artifact categories, again like those of adults, are characterized more simply by commonalities among instances in shape. Three experiments shed light on the nature and development of a mechanism that enables children to organize novel lexical categories differently for different kinds of objects.

Parts and the basic level in natural categories and artificial stimuli: Comments on Murphy (1991)

Memory & Cognition, 1991

Natural taxonomies consist of categories that vary in level of abstraction. Categories at the basic level, such as chair and apple, are preferred in a broad range ofsituations (Rosch, Mervis, Gray, Johnson, & Boyes-Broem;. Several studies have revealed qualitative differences between the basic level and other levels. For example, Tversky and Hemenway (1984) presented evidence that parts proliferate at the basic level; they proposed that parts link the appearance of category members with their {unctions. Although not taking issue with these findings, investigated whether parts are necessary or sufficient for a basic level: In an attempt to demonstrate that parts are not necessary, Murphy used artificial stimuli that did not capture the essential features ofnatural taxonomies. These discrepancies preclude any conclusions based on his studies. Murphy's data also do not support his claim that parts are not sufficient for a basic leoel: Finally, it is unlikely that pursuing questions ofnecessity or sufficiency will produce insights into human categorization.

On the nature of implicit categorization

Psychonomic Bulletin & Review, 1999

Current categorization models disagree about whether people make a priori assumptions about the structure of unfamiliar categories. Data from two experiments provided strong evidence that people do not make such assumptions. These results rule out prototype models and many decision bound models of categorization. We review previously published neuropsychological results that favor the assumption that category learning relies on a procedural-memory-based system, rather than on an instance-based system (as is assumed by exemplar models). On the basis ofthese results, a new categorylearning model is proposed that makes no a priori assumptions about category structure and that relies on procedural learning and memory. There is much recent evidence that human category learning relies on multiple systems (e.g.