The Perception of Similarity

Similarity is one of the central problems of psychology. It underlies object recognition and categorization, which are crucial to much of modern cognitive research, not to mention survival in the real world. It underlies transfer of learning, errors of memory, perceptual organization, social bonding, and many other experimental problems one might choose almost at random from the psychological literature. Distinguished thinkers from Aristotle to modern psychologists such as Shepard (e.g. 1987), Tversky (e.g. 1977), and Nosofsky (e.g. 1992) have wrestled with it. It is no surprise that similarity plays a key role in the subject matter of several of the other chapters in this book, such as in discussions of object and pattern recognition (Cook, 2001, Kirkpatrick, 2001, Chase & Heinemann, 2001), categorization (Huber, 2001; Young & Wasserman, 2001), and attention (P. Blough, 2001, Shimp, Herbranson, & Fremouw, 2001). What is similarity and how is it measured in birds? This chapter is a brief review of some answers to this question, together with examples from existing research.

Similarity is a relationship that holds between two perceptual or conceptual objects. The discussion here will be restricted to similarity considered as the perceptual resemblance of objects to one another. Other chapters in this book are concerned with more abstract relationships (e.g. Young & Wasserman, 2001). Here, the term “object” will mean any reasonably unitary stimulus, including real objects, such as a stone or a tree, but also a patch of light on a screen, a moving dot, or a bird call.

Similarity as perceptual resemblance is a psychological construct in somewhat the same way that such sensory attributes such as hue, brightness, or pitch are constructs or intervening variables. Like such attributes, similarity depends heavily on the physical characteristics of stimulus objects, but this dependence is complex, and in the case of similarity the details of the physical-psychological relationship are usually unknown. Fortunately, similarities may be specified independently of any physical measures of the stimuli involved; that is, one may determine the similarities among a group of objects from behavioral responses to the objects without specifying anything about them except which behavioral measures go with which object (see multidimensional scaling). However, as will be illustrated shortly, one important use of similarity structures is to help to identify psychophysical relations that are difficult to determine in other ways.

For animals similarity is derived from such measures as discrimination errors, generalization response rates, transfer responses, and discriminative reaction times, but similarity cannot be directly equated with such measurements. For example, as will be illustrated below, a generalization gradient does not define the similarity of test stimuli to a training stimulus, though a gradient may contribute to such a definition. (See generalization gradients in the measurement section.)

The reasons for taking the trouble to go beyond physical measures of stimuli and raw behavioral response data to construct scales of similarity are somewhat analogous to those for going beyond physical measures to construct scales of psychological attributes. Measures of hue, brightness, or pitch are used to construct stimuli and to state psychological results because they enter into much simpler relations with behavior than do wavelength, intensity, and frequency. They also provide generality, for after they are determined in one task they may, if used with care, help organize and predict the results of other tasks that share the same stimuli. Likewise with similarity. The most striking example of a broad generalization of this sort is Shepard’s Universal Law of Generalization, which states that the probability of a response learned to a stimulus S decays exponentially with dissimilarity between a test stimulus and stimulus S (e.g. Shepard, 1987).

One reason for caution in applying similarity relations across situations is that variables may affect the behavioral measures through which similarity is determined without affecting similarity. For example, if pigeons get food in the presence of one random set of photos and do not get food in the presence of another set of photos, the birds may respond equally to all the “food” items and withhold response to all the “no-food” items. This presumably does not mean that pigeons find the photos within each group perceptually similar to each other, and dissimilar to those in the other group.

On the other hand, similarity between objects is not solely dependent on the characteristics of those objects. It is also affected by other present and immediately past stimuli, as well as long-term experience with related objects. For humans, a notorious case in point is the effect of experience on similarity among phonemes. A well-known example is that native English speakers find spoken “L” and “R” quite distinct, whereas to native speakers of Japanese they sound extremely similar. Because one of the advantages of work with non-human animals is the possibility of controlling past experience with stimuli, an interesting and challenging task is the study the conditions under which such “perceptual learning” occurs, and in separating such perceptual learning from of other sorts of influences.

Until recently, students of animal behavior have often been casual in their assumptions about stimulus properties, including similarity. Typically, stimuli were implicitly assumed to have certain similarity relations, which might mean that similarity is assumed to be inherent in the stimulus objects (red keys are “physically” similar to orange keys) or it might mean that similarity depends on how the stimuli look to the experimenter (red is similar to orange because it looks that way to me). The first assumption ignores the findings of psychophysics. The second is anthropocentric (or, more exactly perhaps, solopsistic), and easily falsified by

examples: similar (to us) white flowers may look quite different to bees because of their differing reflection of ultraviolet light. Of course, individual humans differ in sensory capacity (e.g. - colorblind vs. normal, see Figure 1), but more interestingly their perception is affected by training (a trained musician hears things inaudible to an untrained person, or in the phoneme example cited above).

The remainder of this chapter consists mainly of two parts having to do first with some conceptual and theoretical issues surrounding similarity, and second with the measurement of similarity in birds.