Avian Visual Cognition

Visual Categorization in Pigeons

Ludwig Huber
Institute of Zoology, University of Vienna

avcrule.gif (935 bytes)

Next Section: Exemplar View




This is a highly subjective review of the literature concerning the mechanisms of visual categorization in pigeons. In the first half of the review, it is suggested that the way pigeons sort complex sets of visual stimuli into experimenter-defined categories does not require conceptual abilities and is therefore accessible to simpler lines of analysis such as associative learning theory. A few experiments, conducted in our laboratory using carefully constructed stimulus sets, are described in order to illustrate this point. In the second half of the chapter, it will be shown that pigeon categorization studies suffer from overly simplistic assumptions concerning the perceptual aspects of natural categorization. Both the nature of natural visual classes and the capacity of the pigeon's perceptual system to exploit these classes are underrepresented in the psychologist's laboratory. In a recent series of experiments, we found that pigeons were able to detect the diagnostic class characteristics in the surface domain of natural images.

avcrule.gif (935 bytes)
Chapter Outline & Navigation
I.   Introduction 
       Three types of representation

II.  Exemplar View
Learning by rote instead of abstracting a symmetry concept

III. Feature Learning
Feature analysis versus feature learning
         Determining the controlling features
         A test of the linear feature model
         Elemental versus configural theories of conditioning
         Limits of the feature account

IV. A Prototypical Point of View

V. A Synthetic Approach Using Natural Stimulus Classes
Complex visual classes
        The role of texture and two-dimensional shape in images of human faces
        The role of low-level features in images of human faces

VI. Conclusion

VII. References

I.  Introduction

"A pigeon pecks rapidly at a small photograph of Harvard Yard containing trees, buildings, people, sky. After a few seconds, a hopper of grain appears and the pigeon eats. Now the scene changes to a treeless Manhattan street. The bird emits a few desultory pecks, then turns away and paces about. After a minute or so, a picture of a leafy suburban garden appears and the bird begins pecking again."

 (Shettleworth, 1998) 

The above quote describes a now famous experiment in which pigeons demonstrated their ability to discriminate complex classes of stimuli (Herrnstein, 1979). In this and other experiments of its kind-- commonly called category or concept discriminations--pigeons are presented with large sets of pictorial stimuli; ranging from simple figures to more complex visual arrays including color photographs and even real objects (Cabe, 1976; Lumsden, 1977; Delius, 1992; Watanabe, 1993). The birds are trained, by means of operant procedures, to respond differentially to the stimuli according to an experimenter-defined class rule (e.g. "tree" and "non-tree" in the above example). This type of experiment has received considerable attention not only because the results suggest that pigeons possess an ability that transcends the discrimination of simple stimulus dimensions such as wavelength, intensity, or frequency (Honig & Urcuioli, 1981; Mostofsky, 1965), but also because they imply that the pigeons' classification behavior is mediated by abstract, or conceptual, rules, and therefore resembles the cognitive solution accomplished by humans.

Categorization--as it is reviewed here--lies intermediate between discrimination of elementary stimuli and the linguistic manipulation of classes of objects, events or ideas by using symbolic representations and by attaching to them verbal names. Categorization can be viewed as the ability to treat similar, but not identical, things as somehow equivalent, by sorting them into their proper categories and by reacting to them in the same manner (Rosch, 1978; Medin & Smith, 1984; Harnard, 1987; Neisser, 1987). Considering the vast amount of information arriving at the perceptual systems of mobile organisms and the few behavioral output patterns possible in non-human animals, categorization may be conceptualized as an adequate solution to this "informational bottleneck" (Delius, Siemann & Jitsumori; 2000). The drastic information reduction is a quite fundamental principle of cognitive economy and therefore widely dispersed among species. The evolutionary pressures to minimize processing requirements in small information-processing systems such as non-human animals are self-evident (Cook, Wright & Kendrick, 1990). Nevertheless, the specific solutions found by a wide range of species to compress the amount of information to be retained vary at different levels, from the pure perceptual (selective attention), to the level of learning or representation and, finally, to the level of reasoning (see Marler, 1982; Premack, 1976). 

At the upper end, language and--more generally--the ability to form symbolic representations widens the bottleneck in a countless number of ways. Humans can think about an indefinite variety of things. Although the basic principles of categorization shade into human cognition, and many "implicit" solutions found by humans to artificial categorization problems are best explained in simple associative terms (Mackintosh, 1995), language clearly involves far more than mere categorization. This is even true for the language competences of great apes (Ristau & Robbins, 1982). It would lead to a regrettable underestimation of their competencies if they were interpreted as simple categorization and, at the same time, to an overestimation of the distinctiveness of those processes that give rise to the bottleneck solutions found in many other species. While language depends on categorization, categorization does not depend on language (Herrnstein, 1984). We therefore need an adequate model system to investigate the "middle" range of categorization phenomena--lying between simple discriminations and the formation of symbolic representations. The pigeons' (Columba livia) obvious lack of language competencies, but extraordinary perceptual capacities render it an ideal species for investigating categorization in its more general dimensions (Lea, 1984; Wasserman, 1991; Mackintosh, 1995).

Since the pioneering experiments of Herrnstein and Loveland in 1964, psychologists have repeatedly demonstrated that simple creatures like the pigeon can categorize stimulus classes containing instances so variable that we cannot physically describe even the class rule. These animals seem readily capable of categorizing images from natural scenes; among the reported stimulus classes are aerial photographs (Skinner, 1960; Lubow, 1974), images from people (Herrnstein & Loveland, 1964), pigeons (Poole & Lander, 1971;Click here to see examples Watanabe, 1991), trees and bodies of water (Herrnstein, Loveland & Cable, 1976), oak leaves (Cerella, 1979), chairs, cars, humans, and flowers (Bhatt, Wasserman, Reynolds & Knauss, 1988; click here to see examples of their stimuli), birds and other animals (Roberts & Mazmanian, 1988), and pictures of a geographic location (Wilkie, Willson & Kardal, 1989). But the rather strange flexibility of pigeons has been demonstrated by using defective pharmaceutical capsules or diodes (Cumming, 1966; Verhave, 1966), letters of the alphabet (Morgan et al., 1976; Lea & Ryan, 1983, 1990), line drawings of cartoon characters (Cerella, 1980), squiggles (Vaughan & Greene, 1984), dot patterns (Watanabe, 1988), schematic faces (Huber & Lenz, 1993, 1996) and real faces (Jitsumori & Yoshihara, 1997; Troje, Huber, Loidolt, Aust & Fieder, 1999), color slides of paintings by Monet and Picasso (Watanabe, Sakamoto & Wakita, 1995) and excerpts from famous pieces of classical music (Porter & Neuringer, 1984).

In sum, the quite impressive list of readily solved categorization problems by pigeons indicates a robust perceptual/cognitive ability. This conclusion is justified even if we take into account the surprising failures of pigeons to achieve a satisfactory solution in seemingly similarly difficult tasks. For instance, there are three unpublished experiments by Herrnstein (conducted 1970-1972) in which pigeons failed to discriminate photographs of food cups, bottles, and wheeled vehicles. Such failures only prove that pigeons are not limitlessly malleable by reinforcement on the stimulus side (Lea & Ryan, 1990; Lea, Lohmann & Ryan, 1993). Obviously more irritating in the above list of pigeon categorization experiments is the inability of the experimenter to give a straightforward answer to the most interesting question of how the pigeons were able to succeed in all these complex experiments. Although the birds readily learned to discriminate between the pictures shown during classification training and also were able to generalize to other instances of the categories, this tells us little or nothing about the perceptual or cognitive mechanisms underlying it.

Consider, for instance, Herrnstein and Loveland's (1964) groundbreaking demonstration of pigeons being able to quickly sort color slides showing similar natural scenes according to the "higher-order" concept "human being."  This first experiment of category discrimination remained typical in two respects, its procedure and Click here to see examples implication. Half of the slides contained human forms and half did not, but otherwise the slides looked comparable (click here to see some actual stimulus examples from their experiment). In the presence of a human slide, pecking was intermittently reinforced with brief access to food. In the presence of a non-human slide, pecking earned no reinforcement. The pecking behavior in the presence of the two types of stimuli became increasingly different as the pigeons learned the discrimination; they  frequently pecked in the presence of human slides and completely withheld pecking in the presence of non-human slides. 

Figure 1.  Human or Non-Human?

The most interesting aspect of this experiment was that the stimuli were not collected by any very exact criteria but according to the natural language concept "human being present"--in the experimenter's best judgement. In fact, positive and negative instances varied in a large number of visual dimensions. Even the positive feature (human being) varied in position, in number (from a single person to groups of various sizes), in appearance (clothed, semi-nude, or nude; adults or children; men or women; sitting, standing, or lying; black, white, or yellow), in lighting and coloration, and so on. Each experimental session consisted of the successive presentation of 40 positive and 40 negative instances in semi-arbitrary sequence. Over 1200 slides were shown to the pigeons. In sum, the birds were required to detect human beings in photographs constituting "a class of visual stimuli so diverse that it precludes simple characterization" (Herrnstein & Loveland, 1964, p. 549).

Unsurprisingly, the anthropocentric approach led to an interpretation of the experiment--and hence a title of the paper--in terms of "complex visual concept in the pigeon". The outcome of the experiment had been contrasted with the well-known ability of pigeons to discriminate stimuli differing in size, shape, or color. There was no hint to suggest that the quick mastery of the concept task arose from some trivial and unsuspected visual clue in the slides, or from some non-visual property of the procedure. The authors concluded that the subjects entered the experiment  with some already existing general concept "human".

Current evidence suggests that this interpretation was premature because we do not know whether pigeons have concepts and use them to solve category problems (Watanabe, Lea & Dittrich, 1993; Monen, Brenner & Reynaerts, 1998). During the early period of pigeon categorization, testing the subjects with novel stimuli--as soon as they have arrived at a satisfactory level of training performance--was considered as the critical operation in order to test conceptual categorization. Experimenters relied upon such transfer tests because successful generalization of the novel stimuli indicates that the subjects have not memorized the training pictures together with their psychological consequence. However, evidence for anything more interesting than pure picture memorization is not per se indicative of the acquisition or use of a "complex visual concept". Indeed, Greene (1983)--in replicating the people/non-people experiment--provided clear evidence that the pigeons had not only attended to the concept-relevant features of the people instances; their classification behavior had also been controlled by irrelevant features included in the "background".

In fact, perceptual mechanisms alone may explain the results of these experiments. Although it is difficult to prove that conceptual behavior is not involved in classification tasks, I agree with Chater and Heyes (1994) that the idea of a concept has not yet been sufficiently decoupled from natural language to make this possible. Furthermore, there is at present no coherent account of what animal concepts might involve (clusters of features or anything more abstract or knowledge-based). I do not want to continue a fruitless debate here (but see also Huber, 1995, 1998, and my recent account to this in Huber, 1999). I only want to emphasize that the early study of visual categorization in pigeons has a past that is strongly rooted in the experimental analysis of human conceptualization. In contrast, the understanding of the "stimulus problem in nature" (Herrnstein, 1990; Fetterman, 1996) and the analysis of the actual performance of pigeons in categorization experiments has been poorly developed in the early period of animal categorization.

In the late seventies the focus of pigeon categorization studies shifted towards the classification of much simpler, albeit more carefully specified, sets of stimuli (Cerella, 1979; Lea & Harrison, 1978; Morgan, Fitch, Holman & Lea, 1976). Taking into account the explanations offered by classic theories of discrimination and generalization, the consensus grew that pigeons were seldom doing anything more complex than associating a large number of pictures, and/or the features that they contained, with reward (Mackintosh, 2000). The main question concerned the type of representation that was acquired during classification learning. Categories may be temporary clusters of simpler mental categories (usually called features) that recur together in the encounters of the sensorium with the external world and so, by association, are stored together. Or they may be the perceived pictures of all or many instances that maintain their configural structure in the pigeons' vast memory stores. Finally, categories may be prototypes that reflect the central tendency of any encountered class of stimuli, and involve abstract class rules such as those that underlie family resemblance.

In this review, I will focus on the research that my colleagues and I have conducted with pigeons since 1991. Using carefully constructed sets of stimuli we started to test the three main models of human categorization (that we adapted to the pigeon in the Skinner box) and went forward to prove whether our main findings were applicable to more natural stimulus classes (without losing control over their feature content and distribution). I hope to show that today we have very good reason to doubt that pigeon categorization is "a secret" (Herrnstein, 1985, p. 129), "utterly mysterious" (Marler, 1982, p. 87), or "shrouded in mystery" (Premack, 1983, p. 357).

Three Types of Representation

Despite the common tenet that categorization is not a logical device, but a matter of assessing similarity, there is little agreement concerning the level of abstraction from which the descriptors of open-ended categories are obtained. After several decades of experimentation, simulation, and theorizing, essentially three views have taken their place in human psychological research: the exemplar, the feature, and the prototype view. (Click here to learn more about a fourth classical view of categorization).

These three theories can be applied to perceptual categorization in animals and have, in fact, tacitly guided its analysis. For example, Herrnstein's (1990) functional approach extended the model troika only by adding pure discriminations to the "lower" end, and relational concepts to the "upper" end. In between these two points, learning by rote corresponds to the exemplar view, open-ended categorization to the feature view, and the abstraction of concepts to the prototype view. If the match between human models of categorization and Herrnstein's (1990) functional approach is not perfect, then differential emphasis can be laid on the details of processing, representation and learning. However, as a conceptual framework for guiding discussion of our own, as well as other closely related studies, the match is sufficiently precise. The next three sections examine the research related to each of these different views of categorical representation.

Next Section: Exemplar View