Are theories of learning necessary? – Skinner

Autor: B. F. Skinner (1950)
Origem: First published in Psychological Review, 57, 193-216.
Fonte: http://psychclassics.yorku.ca/Skinner/Theories/ (o texto integral encontra-se disponível na fonte)

Certain basic assumptions, essential to any scientific activity, are sometimes called theories. That nature is orderly rather than capricious is an example. Certain statements are also theories simply to the extent that they are not yet facts. A scientist may guess at the result of an experiment before the experiment is carried out. The prediction and the later statement of result may be composed of the same terms in the same syntactic arrangement, the difference being in the degree of confidence. No empirical statement is wholly non-theoretical in this sense, because evidence is never complete, nor is any prediction probably ever made wholly without evidence. The term “theory” will not refer here to statements of these sorts but rather to any explanation of an observed fact which appeals to events taking place somewhere else, at some other level of observation, described in different terms, and measured, if at all, in different dimensions.

Three types of theory in the field of learning satisfy this definition. The most characteristic is to be found in the field of physiological psychology. We are all familiar with the changes that are supposed to take place in the nervous system when an organism learns. Synaptic connections are made or broken, electrical fields are disrupted or reorganized, concentrations of ions are built up or allowed to diffuse away, and so on. In the science of neurophysiology statements of this sort are not necessarily theories in the present sense. But in a science of behavior, where we are concerned with whether or not an organism secretes saliva when a bell rings, or jumps toward a gray triangle, or says bik when a cards reads tuz, or loves someone who resembles his mother, all statements about the nervous system are theories in the sense that they are not expressed in the same terms and could not be confirmed with the same methods of observation as the facts for which they are said to account.

A second type of learning theory is in practice not far from the physiological, although there is less agreement about the method of direct observation. Theories of this type have always dominated the field of human behavior. They consist of references to “mental” events, as in saying that an organism learns to behave in a certain way because it “finds something pleasant” or because it “expects something to happen.” To the mentalistic psychologist these explanatory events are no more theoretical than synaptic connections to the neurophysiologist, but in a science of behavior they are theories because the methods [p. 194] and terms appropriate to the events to be explained differ from the methods and terms appropriate to the explaining events.

In a third type of learning theory the explanatory events are not directly observed. The writer’s suggestion that the letters CNS be regarded as representing, not the Central Nervous System, but the Conceptual Nervous System (p. 421), seems to have been taken seriously. Many theorists point out that they are not talking about the nervous system as an actual structure undergoing physiological or bio-chemical changes but only as a system with a certain dynamic output. Theories of this sort are multiplying fast, and so are parallel operational versions of mental events. A purely behavioral definition of expectancy has the advantage that the problem of mental observation is avoided and with it the problem of how a mental event can cause a physical one. But such theories do not go so far as to assert that the explanatory events are identical with the behavioral facts which they purport to explain. A statement about behavior may support such a theory but will never resemble it in terms or syntax. Postulates are good examples. True postulates cannot become facts. Theorems may be deduced from them which, as tentative statements about behavior, may or may not be confirmed, but theorems are not theories in the present sense. Postulates remain theories until the end.

It is not the purpose of this paper to show that any of these theories cannot be put in good scientific order, or that the events to which they refer may not actually occur or be studied by appropriate sciences. It would be foolhardy to deny the achievements of theories of this sort in the history of science. The question of whether they are necessary, however, has other implications and is worth asking. If the answer is no, then it may be possible to argue effectively against theory in the field of learning. A science of behavior must eventually deal with behavior in its relation to certain manipulable variables. Theories — whether neural, mental, or conceptual — talk about intervening steps in these relationships. But instead of prompting us to search for and explore relevant variables, they frequently have quite the opposite effect. When we attribute behavior to a neural or mental event, real or conceptual, we are likely to forget that we still have the task of accounting for the neural or mental event. When we assert that an animal acts in a given way because it expects to receive food, then what began as the task of accounting for learned behavior becomes the task of accounting for expectancy. The problem is at least equally complex and probably more difficult. We are likely to close our eyes to it and to use the theory to give us answers in place of the answers we might find through further study. It might be argued that the principal function of learning theory to date has been, not to suggest appropriate research, but to create a false sense of security, an unwarranted satisfaction with the status quo.

Research designed with respect to theory is also likely to be wasteful. That a theory generates research does not prove its value unless the research is valuable. Much useless experimentation results from theories, and much energy and skill are absorbed by them. Most theories are eventually overthrown, and the greater part of the associated research is discarded. This could be justified if it were true that productive research requires a theory, as is, of course, often claimed. It is argued that research would be aimless and disorganized without a theory to guide it. The view is supported by psychological texts that take their cue from the logicians rather than empirical science and [p. 195] describe thinking as necessarily involving stages of hypothesis, deduction, experimental test, and confirmation. But this is not the way most scientists actually work. It is possible to design significant experiments for other reasons and the possibility to be examined is that such research will lead more directly to the kind of information that a science usually accumulates.

The alternatives are at least worth considering. How much can be done without theory? What other sorts of scientific activity are possible? And what light do alternative practices throw upon our present preoccupation with theory?

It would be inconsistent to try to answer these questions at a theoretical level. Let us therefore turn to some experimental material in three areas in which theories of learning now flourish and raise the question of the function of theory in a more concrete fashion.

The Basic Datum in Learning

What actually happens when an organism learns is not an easy question. Those who are interested in a science of behavior will insist that learning is a change in behavior, but they tend to avoid explicit references to responses or acts as such. “Learning is adjustment, or adaptation to a situation.” But of what stuff are adjustments and adaptations made? Are they data, or inferences from data? “Learning is improvement.” But improvement in what? And from whose point of view? “Learning is restoration of equilibrium.” But what is in equilibrium and how is it put there? “Learning is problem solving.” But what are the physical dimensions of a problem — or of a solution? Definitions of this sort show an unwillingness to take what appears before the eyes in a learning experiment as a basic datum. Particular observations seem too trivial. An error score falls; but we are not ready to say that this is learning rather than merely the result of learning. An organism meets a criterion of ten successful trials; but an arbitrary criterion is at variance with our conception of the generality of the learning process.

This is where theory steps in. If it is not the time required to get out of a puzzle box that changes in learning, but rather the strength of a bond, or the conductivity of a neural pathway, or the excitatory potential of a habit, then problems seem to vanish. Getting out of a box faster and faster is not learning; it is merely performance. The learning goes on somewhere else, in a different dimensional system. And although the time required depends upon arbitrary conditions, often varies discontinuously, and is subject to reversals of magnitude, we feel sure that the learning process itself is continuous, orderly, and beyond the accidents of measurement. Nothing could better illustrate the use of theory as a refuge from the data.

But we must eventually get back to an observable datum. If learning is the process we suppose it to be, then it must appear so in the situations in which we study it. Even if the basic process belongs to some other dimensional system, our measures must have relevant and comparable properties. But productive experimental situations are hard to find, particularly if we accept certain plausible restrictions. To show an orderly change in the behavior of the average rat or ape or child is not enough, since learning is a process in the behavior of [p. 196] the individual. To record the beginning and end of learning or a few discrete steps will not suffice, since a series of cross-sections will not give complete coverage of a continuous process. The dimensions of the change must spring from the behavior itself; they must not be imposed by an external judgment of success or failure or an external criterion of completeness. But when we review the literature with these requirements in mind, we find little justification for the theoretical process in which we take so much comfort.

The energy level or work-output of behavior, for example, does not change in appropriate ways. In the sort of behavior adapted to the Pavlovian experiment (respondent behavior) there may be a progressive increase in the magnitude of response during learning. But we do not shout our responses louder and louder as we learn verbal material, nor does a rat press a lever harder and harder as conditioning proceeds. In operant behavior the energy or magnitude of response changes significantly only when some arbitrary value is differentially reinforced — when such a change is what is learned.

The emergence of a right response in competition with wrong responses is another datum frequently used in the study of learning. The maze and the discrimination box yield results which may be reduced to these terms. But a behavior-ratio of right vs. wrong cannot yield a continuously changing measure in a single experiment on a single organism. The point at which one response takes precedence over another cannot give us the whole history of the change in either response. Averaging curves for groups of trials or organisms will not solve this problem.

Increasing attention has recently been given to latency, the relevance of which, like that of energy level, is suggested by the properties of conditioned and unconditioned reflexes. But in operant behavior the relation to a stimulus is different. A measure of latency involves other considerations, as inspection of any case will show. Most operant responses may be emitted in the absence of what is regarded as a relevant stimulus. In such a case the response is likely to appear before the stimulus is presented. It is no solution to escape this embarrassment by locking a lever so that an organism cannot press it until the stimulus is presented, since we can scarcely be content with temporal relations that have been forced into compliance with our expectations. Runway latencies are subject to this objection. In a typical experiment the door of a starting box is opened and the time that elapses before a rat leaves the box is measured. Opening the door is not only a stimulus, it is a change in the situation that makes the response possible for the first time. The time measured is by no means as simple as a latency and requires another formulation. A great deal depends upon what the rat is doing at the moment the stimulus is presented. Some experimenters wait until the rat is facing the door, but to do so is to tamper with the measurement being taken. If, on the other hand, the door is opened without reference to what the rat is doing, the first major effect is the conditioning of favorable waiting behavior. The rat eventually stays near and facing the door. The resulting shorter starting-time is not due to a reduction in the latency of a response, but to the conditioning of favorable preliminary behavior.