Perception and Its Development

A Tribute To Eleanor J. Gibson

A. D. Pick

• Publisher: Psychology Press
• Year: 2019
• Print ISBN: 9780898594096
• eBook ISBN: 9781317722533
• Topic: Psychology
• Subtopic: Experimental Psychology
• Index: Psychology

We have acquired important new knowledge about the nature and development of perception in recent years, and the insights of Eleanor Jack Gibson have had a prominent role in guiding the search for that knowledge.  The purpose of this volume is to honor her continuing conrbution to our understanding of perception. First published in 1979. Routledge is an imprint of Taylor & Francis, an informa company.

I

LEARNING AND GENERALIZATION

1	Stimulus Generalization

Donald A. Riley

Marvin R. Lamb

University of California, Berkeley

For years psychologists studying learning have focused on the formation of associations via operant and classical conditioning procedures. The precise nature of these associations has been an important, and often elusive, issue. One way of addressing this question is to ask about the parameters of the stimuli that enter into associations. What stimuli present on a conditioning trial become associated? In what ways might these stimuli be represented internally? How might these internal representations get translated into responses? We may begin to answer questions of this sort by examining how the effects of training on one stimulus transfer to other stimuli.

In classical conditioning, for example, one stimulus (the US) occurs often in the presence of another stimulus (the CS) and rarely in its absence. When a change in behavior is observed upon the establishment of this predictive relationship, it is inferred that an association has been formed between the CS and the US. If, however, a stimulus similar to the CS is presented, the newly learned response is likely to occur. This effect has been called stimulus generalization. Such generalized responses rarely occur with as high a probability or as rapidly or vigorously as responses to the training stimulus, however. Instead, stimulus generalization tends to decrease as the similarity between the test stimulus and the CS decreases, resulting in what is called a gradient of stimulus generalization. These relations between training and test stimuli are of interest because they are central to our conceptions of learning.

In the history of thinking about stimulus generalization, two explanations have emerged: strength theory and stimulus classification theory. Strength theory assumes that as training proceeds, associations increase in strength. When, after some training, a test with stimuli other than those used in training reveals a tendency to respond, the associative strength is said to have generalized. That is, the excitatory strength gained by the training stimulus is thought to have spread to other stimuli as well. In stimulus classification theory, learning is conceptualized as the progressive identification of the defining properties of the stimulus that predicts reinforcement. As training progresses, the organism becomes more effective at this task, and the probability, speed, and vigor of the response increase. When tested with other stimuli, the organism may make mistakes for a variety of reasons. That is, the subject may incorrectly classify the test stimulus as a member of the class, “training stimulus.” In this chapter we describe the development of these two conceptions of stimulus generalization from their inception to their present state and examine a small set of experiments that seem to bear critically on the relative merits of these two alternative views.

THE DEVELOPMENT OF STRENGTH THEORY

Strength theory began with Pavlov’s development of a behaviorally based account of the formation of an association resulting from the contiguous presentation of two stimuli. In the course of his investigations, Pavlov found that other stimuli were also affected by this associative process, and it is from the analysis of this fact that the topic of stimulus generalization developed.

Pavlov’s analysis of transfer of training along dimensions has been discussed at length in a number of different volumes and need not be described in detail here (see Mackintosh, 1974; Pavlov, 1927; Riley, 1968). Briefly, Pavlov observed that when an animal was trained in a classical conditioning task to salivate to a conditioned stimulus of a given brightness, size, or frequency, then other stimuli on the same dimension also tended to elicit the same response. The magnitude of the response, however, varied with the distance on the appropriate dimension from the conditioned stimulus. Pavlov designated this phenomenon “generalization of stimuli,” and others later referred to the orderly change in responding as a function of similarity of the test stimulus to the CS as a “gradient of stimulus generalization.” Conditioning, according to Pavlov, consisted of the development of an excitatory event traveling from the locus in the cortex corresponding to the CS to the locus corresponding to the US. With repeated pairings of the CS and US, the excitatory strength of the CS increased. In addition, the excitatory strength thus accruing to the CS was thought to spread to other nearby points in the cortex, with increasingly distant points receiving increasingly less excitatory strength. Thus, when a stimulus similar but not identical to the CS was presented, excitation from that stimulus would spread to the locus of the CS, arousing in turn that location; and as a consequence, it would also arouse the location of the US, evoking the conditioned response. Pavlov assumed that for each observed gradient of stimulus generalization, there was an isomorphic gradient of excitation in the cerebral cortex. This scheme, assuming spatial isomorphism between sensory events and cortical events, fell into disfavor as a result of a substantial amount of research reviewed elsewhere (see Riley, 1968), but the notion of generalization as an active process persisted.

When Hull (1943) took the facts of conditioning as a source of postulates from which to deduce more complex phenomena of performance and learning, one of the central facts of conditioning was that of stimulus generalization. Hull accepted as a fact of habit formation that the response associated with the CS would also be elicited by other similar stimuli according to a law of similarity. Although Hull did not follow Pavlov in assuming that generalization reflected the spread of excitation across the cerebral cortex, his writings and those of his students viewed generalization as a process associated with conditioning. As excitatory strength (or habit strength) developed in accord with the laws of conditioning, this excitation was said to “generalize” from the CS to other stimuli sharing common dimensions with it. Thus Spence (1937), in his analysis of transposition in terms of generalization of excitation and inhibition, wrote: “We shall assume there is a generalization of this acquired excitation tendency to stimulus objects of similar size and that this generalization follows a gradient [p. 433].”

One implication of the notion of generalization as an active process is that the spread of excitation can be to points that are known to be discriminable from the CS. Pavlov demonstrated that test stimuli were, in fact, discriminated from the CS early in the generalization test; only after the first few test trials were there any indication that excitation had accrued to stimuli other than the CS. Hovland (1937) ran a series of auditory conditioning experiments with human subjects. Following training, the subjects were tested wtih stimuli that were 25, 50, or 75 just noticeable differences (JND) from the CS. The fact that these increasingly remote test stimuli elicited increasingly smaller conditioned responses was taken to mean that the generalization of excitation was progressively smaller as distance from the CS along the sensory dimension increased. Although the point was not made explicitly, it should be clear that the method by which these points were selected implied that all test stimuli were highly discriminable. These precautions suggest that these investigators recognized that the stimulus generalization gradient could occur, not because of the spread of an excitatory process, but because of discriminatory failures. Experiments such as Hovland’s and an earlier one by Bass and Hull (1934), also with humans and in which vibrotactile stimuli were used at four different points on the body surface from the shoulder to the calf of the leg, seemed to preclude the possibility that generalized responses were confusion errors.

Another early distinction provided by students of Hull was the distinction between primary and secondary stimulus generalization. Primary stimulus generalization was used to describe generalization in situations where there was no obvious reason to assume that the response to the different test stimuli was based on learned stimulus equivalence. Secondary stimulus generalization was intended to apply in situations where learned stimulus equivalence had either been demonstrated or was the most reasonable assumption. One of the earliest demonstrations of secondary stimulus generalization by Shipley (1935), working in Hull’s laboratory, showed (Hull, 1943):

a flash of light followed by a tap of a padded hammer against the cheek below the eye, thus conditioning lid closure to the light flash. Next the same subject was repeatedly given an electric shock on the finger. This evoked not only a sharp finger withdrawal from the electrode but lid closure as well. Finally the flash of light was delivered alone. It was found in considerable proportion of the subjects of both experiments that during this latter maneuver the light evoked finger retraction, even though the former had never been associated with either the shock or the finger retraction. The interpretation is that light evoked the lid closure and proprioceptive stimulation produced by this act (or some other less conspicuous act conditioned at the same time) evoked the finger retraction [p. 192].

The important point in this description lies in its assumption of a presumed second type of stimulus generalization—this one based on learned equivalence. Again, the assumption is clear that the subjects responded to two different stimuli in a similar way, not because they confuse them, but because of some other process. In this case, both stimuli activate a second process that controls the final output.

Thus, in the Hullian tradition, both primary generalization and secondary generalization were assumed to be processes by which response strength spread to test stimuli recognizably different from the training stimulus according to the similarity, physical or learned, between the test and training stimuli. The greater interest in primary stimulus generalization was at least in part because of the ready-made ordinal scales that separated several points from the training stimulus along a sensory dimension.

Both phenomena—primary and secondary generalization—were used as explanatory principles to account for more complex effects. Among these was Gibson’s (1940) extension of stimulus generalization from classical conditioning paradigms to voluntary response tasks in humans, Spence’s (1937) analysis of transposition in terms of interacting gradients of excitation and inhibition, and Hull’s (1943) assumption of the summation of generalized positive habits to explain how stimulus-response associations increase in strength even though the same stimulus is never experienced twice (the stimulus learning paradox). The use of stimulus generalization as an explanatory principle had also been informally invoked by Pavlov (1927) to account for the facilitating effects of performing an easy discrimination on the ability subsequently to perform a difficult discrimination on the same dimension. This list of phenomena have in common the fact that all the behaviors described are in some sense under the control of stimuli that differ from the nominal training stimulus, and for all, there is an explanation that assumes generalization of excitation. It is interesting that the assumption that primary stimulus generalization results from the spread of a habit has been regularly questioned during the past 20 years, and today it is not taken very seriously, even by former proponents.

During the past 10 years, strength theory has undergone transformation largely because of the development of the Rescorla-Wagner (1972) model of conditioning. Like Hull’s theory of learning, this model assumes that excitatory strength approaches a maximum exponentially, but the model also specifies that the conditionability of a stimulus that is part of a compound is determined by the strength of the entire compound and not by the current associative strength of the stimulus itself. A number of consequences of this assumption have been explored by Rescorla, Wagner, and their students. Of particular interest to us, however, is the implication of this assumption for phenomena of stimulus generalization. In particular, if the overall strength of a stimulus compound is low, then an increase in excitatory strength of the compound will increase the strength of all the members of the compound regardless of the initial strength of the individual members. As we shall see, these assumptions about the role of the compound in determining the conditionability of the components of the compound result in some interesting predictions. Another facet, however, of this thinking has been a new interest in considering a “stimulus” as an aggregate of elements, each of which may have its own excitatory strength.

THE DEVELOPMENT OF STIMULUS CLASSIFICATION THEORY

The assault on the “neo-Pavlovian” view of stimulus generalization was launched by Lashley and Wade in 1946. They argued, and variations were subsequently introduced by Prokasy and Hall (1963), that a process of stimulus generalization was unnecessary to account for what, under other circumstances, would be called errors. An animal that has been trained to respond positively to a certain stimulus is shown another stimulus and asked if it is the same one. If the animal responds affirmatively, it is an error. Errors, according to this view of stimulus generalization, occur for a variety of reasons. One is that the animal has not associated different stimulus values with different outcomes. Generalization gradients, Lashley and Wade argued, arise only after the animal has learned a discrimination on the dimension in question. They provided evidence for this position by showing that rats did not show a generalization gradient in a choice test following single stimulus training. But this amounted to an assertion that all previous demonstrations of stimulus generalization were in some way the product of inadvertent discrimination procedures, and other investigators quickly produced evidence that test procedures that did not permit such an interpretation would indeed produce generalization gradients (Grice, 1948). One might also argue that when generalization gradients are observed, they are the result of associations formed prior to the experiment itself. Several experiments conducted with pigeons reared and trained with monochromatic light have found evidence for generalization gradients even under these conditions and so have not supported this speculation (see Riley, 1977). Thus, stimulus generalization gradients seem to occur even in the absence of an opportunity for secondary generalization to develop.

There was, however, more to the attack on the Hullian interpretation of stimulus generalization gradients. Lashley and Wade also suggested that monotonically declining gradients could reflect variable stimulus difference thresholds. This argument assumes that an animal recognizes a class of stimuli to which it responds with an appropriate learned response and another class not associated with that response. If the threshold shifts from trial to trial, then a gradient should emerge. Prokasy and Hall (1963) argued similarly that even in cases of generalization gradients extending across highly discriminable differences, the gradient might be due to insensitive testing procedures that would result in the animal’s failing to detect these stimulus differences. Thus, these interpretations of generalization gradients assume that the gradient occurs, not because of a gradient of habit strength, but because of variations by the subject in the classification of stimulus events.

Modern versions of stimulus classification theory assume that the animal, on the basis of different outcomes such as different reinforcement schedules, classifies stimuli into different categories and responds appropriately. Thus in a simple case, a pigeon might be required to peck at a key on the left in the presence of one stimulus and at a key on the right in the presence of another stimulus. The advantage, in this analysis, of having two clearly defined alternative responses lies in knowing precisely what response is to be given to each stimulus. If, as is often the case, the response alternatives are “respond” (e.g., peck) and “don’t respond,” the second class is unspecified. It consists of everything except pecking the key. Nevertheless, both procedures have been used.

The only quantitative analyses of stimulus generalization in terms of stimulus classification theory have used ideas from the theory of signal detection. A signal detect...

Table of contents

• Cover
• Half Title
• Title Page
• Copyright Page
• Table of Contents
• Foreword: A Note on E. J. G. by J. J. G.
• Introduction
• PART I: LEARNING AND GENERALIZATION
• PART II: DEVELOPMENT OF SPATIAL PERCEPTION
• PART III: PERCEPTION OF PATTERN AND STRUCTURE
• PART IV: PERCEPTION OF MEANING
• PART V: EXPLORATION AND SELECTIVITY IN PERCEPTUAL DEVELOPMENT
• Afterword
• Author Index
• Subject Index