Feature Representations In Cognitive Psychology Research Paper

1. Introduction

Cognitive psychologists have often attempted to characterize the entities of a domain in terms of a relatively small set of such features, with different entities having different combinations of these features. To take a simple example, consider capital letters. Each capital letter can be described uniquely by some combination of seven basic features: vertical, horizontal, or oblique lines; continuous or discontinuous curves; and right or acute angles between the junctions of lines. For instance, the letter R has a vertical line, two horizontal lines, an oblique line, a discontinuous curve, and three right angles. In contrast, the letter O has a single feature—a continuous curve.

Being able to characterize the entities of a domain in this manner would be an important step in specifying how entities are identified. For example, suppose that in a person’s long-term memory, capital letters were represented as combinations of basic features like those described above. Then by using processes that identified such features and matched them to ones represented in long-term memory, a person could identify a letter. For example, upon detecting an oblique line and a continuous curve, a person would determine that these features matched those of the representation of Q and identify the letter as a Q. Evidence suggests that the human visual system has feature detectors (i.e., neurons). For example, some neurons respond to lines of different orientations and to junctions of lines (i.e., corners) (Hubel and Wiesel 1965).

Identifying entities is the starting point for many important, everyday cognitive tasks. In this example, identifying the letter would be an important step in identifying a word, which in turn would be an important step in reading. As another example, by identifying an object, we gain access to a large amount of stored information that is useful in interacting with that object. For instance, by recognizing something as a bear you can then access stored knowledge which lets you anticipate things about the bear that may not be immediately present (e.g., that the bear may approach and attack you).

In many domains, the representation (and sub-sequent processing) of features is considerably more complex than implied by the example of capital letters. In contrast to capital letters, it is not always clear what features characterize an entity. In addition, multiple factors often determine how a feature is identified. Further, relations between features can be important for identifying entities. Finally, people not only identify features but must also learn new ones. In the remainder of this research paper, I discuss these and other issues related to feature representation and processing.

2. Context-Dependent Feature Representations

The features that we use to represent entities in the world are often influenced by the surrounding context (what other information is present at encoding, our expectations, etc.). A classic example is illustrated by a study involving the rat-man figure (see Fig. 1). Bugelski and Alampay (1961) found that when preceded by several pictures of human faces people tend to see the rat-man figure as a man, but when preceded by several pictures of animals they tend to see the ratman figure as a rat. In these cases, people represent the same perceptual input in terms of different features. For example, the small circles in the figure are seen as ears or as eyeglasses depending on the preceding pictures. As another example, whether a white area on an X-ray indicates a lung tumor, a bone, or an artifact of the procedure can depend on a variety of contextual factors, including the radiologist’s knowledge of the patient’s case history, the location of bones, whether the X-ray is a chest X-ray, and the likelihood that the X-ray was underexposed. These and many other examples suggest that the representation of a feature is not just based on perceptual input but also on what is in the head of the perceiver. As result, the features associated with a perceptual input can vary depending on the perceiver’s prior knowledge.

3. Relations Between Features

Feature representations often capture relations between other features. These relations are necessary in order to identify an entity accurately. For example, a bucket and a coffee cup can both be described by the same two basic features (an open cylinder and a handle). Thus, it would not be possible to distinguish these objects if they were only represented by such features. Note, though, that the spatial relationship between the two features differs in the two objects. That is, the handle of the bucket is attached to the rim of the open cylinder, whereas the handle of the coffee cup is attached to the side of the open cylinder. Thus, people could distinguish these objects if their featural representations included such spatial relationships. Biederman (1987) describes a model of object recognition that represents objects in terms of part-like features (called geons), and the relations between them. As another example, the words ‘team’ and ‘meat’ are made up of the same basic features, called phonemes (i.e., the smallest significant sound units in a language). However, the temporal relationship between these phonemes is different. For example, the ‘t’ sound is pronounced first in ‘team,’ but last in ‘meat.’ Thus, people could distinguish these spoken words if their featural representations included such temporal relationships.

4. Construction Of New Features

In a number of areas of cognitive psychology, the researcher and subject agree implicitly on the features which characterize the cognitive task being investigated. For example, in studying how people learn about new categories of entities, researchers often present subjects with stimuli whose features are well-specified, familiar, and unambiguous. The question of interest is how people determine which features are important for distinguishing one category from another. However, learning does not just involve selecting relevant features. In any domain, people must sometimes figure out the features themselves (i.e., acquire new features; see Biederman and Shiffrar 1987, Lesgold et al. 1988, Schyns et al. 1998, Wisniewski and Medin 1994, for examples).

To illustrate this point, consider a study by Schyns and Rodet (1997) in which subjects learned about categories of novel stimuli called ‘Martian cells’ (see Fig. 2). One group of subjects first learned a category whose members had x blobs followed by a category with y blobs, and then a category with xy blobs (a fusion of the separate x and y blobs; see Fig. 2a). A second group learned the categories in the reverse order (see Fig. 2b). These different orders affected how subjects determined the features of the category with xy blobs. The first group perceived the xy blob as a conjunction of the x and y features. They had previously learned that x was diagnostic of the x blobs category, and that y was diagnostic of the y blobs category. Thus, previously learning the x and y features biased the first group to segment the xy blob into these two features. In contrast, the second group perceived the xy blob as a single, unitary feature xy. Because the initial category that they learned was the one with xy blobs, their perception of the xy blob was not influenced by having learned features x and y.

5. Local vs. Global Feature Representations

In many areas of cognitive psychology, researchers use symbols to represent features that correspond closely to language, and which tend to bottom out at the level at which the language bottoms out (Solomon and Barsalou in press). For example, in the English language, people typically use the word ‘stripes’ to refer to the stripes of zebras, barber poles, the American flag, and so on, rather than a more detailed term. In turn, researchers represent features using similar, language-like terms. Often, these representations take the form of attribute-value pairs. For example, stripes could be represented as the attribute texture. Do researchers believe that such representations have no further specification of meaning? They probably do not, as intuition strongly suggests that there is something more to such representations. Instead, such a feature representation may be a shorthand term for the commonalities shared by its referents (which researchers usually leave unspecified). For example, one possible commonality among the stripes of entities is that they have a color which contrasts with the background color of the entity. Thus, researchers generally assume that a feature has a similar representation across the entities to which it applies (i.e., a global representation).

However, evidence suggests that feature representations are also more fine-grained and specific or local to their associated entities. For example, Solomon and Barsalou (in press) tested the local representation view in a series of property verification tasks. Subjects saw an object (e.g., a bus) followed a property name (e.g., seat) and had to verify whether the property was true of the object. On a subsequent trial, verifying that property for a different object was facilitated if the property had the same form in the second object. For example, having previously verified that seat is true of bus, subjects more quickly verified that seat was true of truck. However, this benefit was not obtained when verifying that seat was true of bicycle. This finding implies that the representation of seat in truck and bus is similar, but differs from that in bicycle. As another example, many properties are relative. For instance, a large mouse and a large house are not very similar in size. A large mouse is large compared to other mice, but a large house is large compared to other houses. Thus, relative properties also have local representations.

6. Determining The Appropriate Features

As much of the discussion above implies, a central issue in feature representation is determining the appropriate features of an entity. In principle, one can describe any entity by an infinite number of features (e.g., a German shepherd dog has 4 legs, but it also has fewer than 6 legs; it has a tail, but also a tail of a particular length; it is found on the planet Earth, but not at the bottom of the ocean; it has a nose, but it also has a nose of a particular color and shape, and so on). Clearly, there must be constraints on what counts as a feature (Murphy and Medin 1985). Many of these constraints are provided by the perceptual system (see Barsalou 1999, Schyns et al. 1998, for discussion of perceptual mechanisms that extract features). For example, the perception of an object part is influenced by discontinuities and points of maximum curvature on its surface (e.g., Hoffman and Richards 1984).

Other constraints are related to the categories to which an entity belongs. For example, a small, umbrella-shaped entity growing in a forest might be categorized as a mushroom. The representation of mushroom might contain the features stem and umbrella-shaped cap so as to distinguish it from other plants. On the other hand, a mushroom belongs to a more specific category such as Death Cap—the most poisonous mushroom. Its representation might contain the features green-tinged cap and thick stem, so as to distinguish it from other mushrooms. Thus, the features associated with a category may be partially constrained by its contrasting categories. Further, the same aspect of an entity may have multiple feature representations (e.g., cap, green-tinged cap).

7. Final Comment

Ultimately, any theory of how features are represented must address the symbol grounding problem (Harnard 1990). That is, the representations of features that are in our heads must make contact with what they represent—what is out in the world. Much of the utility of representations is in allowing us to interact successfully with the world. For example, the representation of the green-tinged cap must allow mushroom pickers to identify actual green-tinged caps of mushrooms so that they can avoid eating them. As another example, a relatively abstract feature such as ‘plays music’ for a CD player is not particularly useful unless it indicates more specifically how one can use a CD player to play music.

Bibliography:

1. Barsalou L W 1999 Perceptual symbol systems. Brain and Behavioral Sciences 22: 577–609
2. Biederman I 1987 Recognition-by-components: A theory of image understanding. Psychological Review 94: 115–47
3. Biederman I, Shiffrar M M 1987 Sexing day-old chicks: A case study and expert systems analysis of a difficult perceptuallearning task. Journal of Experimental Psychology: Learning, Memory and Cognition 13: 640–45
4. Bugelski B R, Alampay D A 1961 The role of frequency in developing perceptual sets. Canadian Journal of Psychology 15: 205–11
5. Harnad S 1990 The symbol grounding problem. Physica D 42: 335–46
6. Hoffman D D, Richards W 1984 Parts of recognition. Cognition 18: 65–96
7. Hubel D H, Wiesel T N 1965 Receptive fields and functional architecture in two nonstriate visual areas (18 and 19) of the cat. Journal of Neuropsychology 28: 229–89
8. Lesgold A M, Rubinson H, Feltovich P, Glaser R, Klopfer D, Wang Y 1988 Expertise in complex skill: Diagnosing X-ray pictures. In: Chi M T H, R Glaser Farr M (eds.) The Nature of Expertise. Erlbaum, Hillsdale, NJ, pp. 322–51
9. Murphy G L, Medin D L 1985 The role of theories in conceptual coherence. Psychological Review 92: 289–316
10. Schyns P G, Rodet L 1997 Categorization creates functional features. Journal of Experimental Psychology: Learning, Memory and Cognition 23(3): 681–96
11. Schyns P G, Goldstone R L, Thibaut J P 1998 The development of features in object concepts. Behavioral and Brain Sciences 21: 1–54
12. Solomon K O Barsalou L W (in press) Representing properties locally: Evidence for the perceptual grounding of concepts. Manuscript submitted for publication
13. Wisniewski E J, Medin D L 1994 On the interaction of theory and data in concept learning. Cognitive Science 18: 221–81