Over the past several years, our research has been directed toward the development a behavioral account of those human performances typically attributed to the intercession of cognitive functions. The general approach in this program has been to construct these performances from overt components, somewhat in the style of the Columban simulations reported by Skinner and Epstein. The logic of these simulations would argue that if such performances can be constructed entirely from explicit components, then the need for hypothesized mediators to carry the account is eliminated. In that endeavor the research was successful, but the work also showed that some elaboration of the operant vocabulary was needed with respect to the description of stimulus control.
The fact is, that the behavioral approach suffers a profound weakness in its lack of an adequate account of how non-arbitrary relations between stimuli control behavior. By non-arbitrary is mean consistent relations such as same as (identity) larger than, before/after, etc. Almost certainly this deficiency exists because the notion of discriminative stimulus (SD) control was formulated to account for relations between stimuli and responses in tasks where the stimulus characteristics of the response were irrelevant . Thus, the rela tionship between the SD and the stimulus characteristics of the rat's bar was never an issue. An SD was said to have gained control of a response when it was shown to control the rate of the response.
As illustrated in Figure 1, it is changes in response
rate, as a function of changes in some dimension of the SD
(e.g. pitch or color), that define the basic phenomena of stimulus control
such as discrimination and generalization, as well as peak shift and behavioral
contrast. In this context the SD has unquestionably
been a useful concept. Thus in the typical generalization gradient, as
illustrated in this figure, it is rate that is plotted against a physical
But extending the SD concept to account for performances in tasks where one stimulus specifies the selection of another, has been problematic because it requires that stimulus selection be accounted for in terms of response rate.
What needs to be made explicit here is that in both cases selection of the square is only incidental. It is a by-product of the increased strength of the selection response. The conditional stimulus, or sample, does not actually specify any properties of the comparison stimulus, rather, the conditional stimulus merely causes an increase in the rate of a selection response in the neighborhood of the comparison. Thus, the SD account applies where any two stimuli are arbitrarily paired. Nothing in the account actually recognizes the identity relation between the stimuli in the second example.
And so, as illustrated in this next figure, training subjects to select the square comparison in response to the square sample, and the triangle comparison in response to the triangle sample, provides no basis for a pointing response to the circle comparison in response to the circle sample. Thus, if subjects merely learn pointing responses during training, there can be no basis for generalization during the test.
As a result, behavioral explanations of generalized matching have invariably resorted to nonbehavioral entities such as if-then conditional relations, or conceptual rules to explain what it is subjects learn in tasks where behavior is based on a consistent, generalizable relation between stimuli. This systemic inadequacy in the vocabulary of stimulus control is the fundamental problem facing current efforts in behavior analysis to account for the types of behavior typically attributed to cognitive processes. But it is not an insurmountable problem, and perhaps not even a difficult one. It turns out, that the manner in which one stimulus specifies another can easily be described with a simple addition to the notion of stimulus control.
Figure 4 shows how this may be done in a generalized identity
matching task. This research involved non-vocal retarded children, and
these data were reported some years ago, but it provides a clear example.
All features of the behavior were directly trained. Initially, during handsign
tact training, subjects learned the handsigns for the shapes illustrated
at the top of the figure . Thus subjects were trained to tact each shape
with the handsigns shown.
Next, in the matching to sample task shown below, subjects were trained to respond to each sample with the appropriate handsign. This is labeled on the figure as the sample tact. They also learned to rehearse this handsign over a delay interval until the array of comparisons appeared, and they continued to rehearse the handsign while perusing the comparisons. Next, and this is critical, subjects learned to tact one of the comparisons without modifying the rehearsed handsign, and then, as shown in the last picture, they pressed on the comparison that allowed this tact -- thereby selecting it. It is clear that this performance necessitates an identity match.
The top of Figure 5 illustrates what happened here. After a tact of distinctive topography was made to the sample shape, the sample itself was removed. One might say that the distinctive handsign topography now represented the sample shape. Self-echoic rehearsal insured that this topography would not be modified, and as a result, when the subject attempted to make a tact to a comparison, the only comparison that could be tacted, if the rehearsed topography was to be maintained, was the comparison identical in shape to the sample. As a result, when the topography was emitted in response to this comparison, it was in fact under joint self echoic and tact control. That is, the topography was both a self-echoic repetition of the rehearsed topography, and JOINTLY an accurate handsign tacting that comparison. This converging control is illustrated by the converging arrows.
The bottom of the figure shows the essential elements. Notice, there is still stimulus control over the selection of the comparison, but it is not SD control. The comparison is not selected due to an increase in the rate of a pointing or pushing response. Rather, the correct comparison is specified by the handsign because there is only one comparison stimulus which can be accurately tacted while also allowing the subject to maintain the self-echoic rehearsal of the sample handsign. Accurately tacting any other comparison would require the subject to relinquish rehearsal of the sample handsign.
In essence, we see that it is the handsign response topography that dictates which stimulus to select. Thus, the only comparison that allowed a repetition of the maintained handsign topography was the two dots. Now, since the handsign topography was itself controlled by the sample, the sample thereby specified a particular comparison, and so here one stimulus, the sample, through the topography of a mediating response, actually does specify the properties of another stimulus: the comparison.
The stimulus to be selected is thus determined before the selection response is evoked. It is not the comparison stimulus itself, but rather the occurrence of joint control that evokes the selection response. To see this, look again at the top of the figure . We see that subjects select a comparison by pushing on it. Now if they pushed on it because the comparison allowed the handsign to occur under joint echoic-tact control, then this pushing response would be a descriptive autoclitic because it reported on the nature of control over other verbal behavior, namely the handsign. We'll come back to this point later
Now, of course, all of this is of interest only if it provides an account of generalization, because it is the generalization of matching to novel stimuli that shows that responding is indeed controlled by the relation between the stimuli rather than by their particular features.
As illustrated in Figure 6, the current account provides a straightforward explanation of how the relational matching performance, developed here, would foster generalization. The top of the figure illustrates four novel shapes. Below them is a new handsign trained to one of these shapes. And below that, is a matching performance with a novel stimulus. It is clear that the matching performance involves only one unique behavior, and that is the unique handsign topography required for the new stimulus. All of the other features, such as tacting the sample, followed by the self-echoic rehearsal, and its repetition as a comparison tact, these are common to all stimuli. And so it was that when the novel handsigns were trained to each of these novel stimuli, generalized identity matching appeared from the very first generalization test trial; good evidence that subjects were responding on the basis of the actual identity relation. The lower part of the figure shows this for one new shape -- the handsign was trained for the shape and then generalized matching was tested with that shape. And so here is a behavior typically attributed to the possession of a generalized concept, the identity relation, entirely constructed here from behavioral relations.
There was, however, something more to these data, as shown in Figure 7 In the course of producing generalization, the handsigns for the novel shapes seemed to spontaneously acquire what has been called a symbolic function. That is, a symmetrical relation between the novel handsigns and their associated shapes seemed to appear without explicit training. Both of these relations are shown in the figure . For the sake of clarity, the originally trained shape - handsign relation has an arrow, with an O for original in it, and the symmetrical handsign-shape relation has a thinner arrow with an S in it. Now, as mentioned above, in preparing subjects for the generalization test, they were trained in the shape-handsign relation with the novel shapes -- that is, in the original relation shown at the top. But in the course of the generalization test, they seemed to spontaneously show the symmetric handsign-shape relation.
This is illustrated in the top of the next figure by the relation entitled Selection by Symmetry. Thus, in making the sample tact, the subject performed the originally trained shape - handsign relation. But by the end of the delay interval, there remains only the handsign -- which then guides comparison selection. Thus comparison selection depends on the untrained and symmetric handsign-shape relation. But there is another possibility: There may be no symmetry. The new relation may actually be a reappearance of the originally trained shape-handsign relation. Thus, as we see in the bottom of the figure, since comparison selection was under joint control, then the comparison was selected on the basis of the original shape-handsign relation. That is, subjects rehearsed the sample handsign and selected the comparison which allowed the handsign as an appropriate tact. Thus, there was no symmetry. The apparently spontaneous symmetrical relation was a secondary, and derived process, constructed, once again, from simpler behavioral relations.
Figure 9 illustrates this process in a more typical situation. Assume naive subjects are taught the novel names for each of the shapes and colors illustrated at the top of the figure . The subject thus has object-word relations. The subject is now required to demonstrate the symmetrical word-object relation. That is, given the sample "gray trap", the subject must now select the corresponding object. To assume such a selection is the result of symmetrical object-word and word-object relations would mean that in the course of scanning the comparisons, a pointing response, although never explicitly trained, is somehow evoked when this comparison is encountered after the subject hears its name. On the other hand, under joint control, there is no need to appeal to symmetry. Selection of the correct comparison would only depend on the subject maintaining the original object-word relation. Take a moment to find the comparison specified at the bottom of the figure . Notice that you scan the comparisons until one is encountered which may be described by a repetition of the sample.
As illustrated in Figure 10, this repetition, is just another repetition of the originally trained object-word relation. The comparison is thus located without actually emitting the symmetric relation. The implication here is that in general, what may seem to be symmetrical word - object relations may actually reflect only the originally trained relation acting under joint control.
There is one other interesting feature of the behavior illustrated here: It provides an account of how objects are recognized from their description It does this by distinguishing stimulus recognition from stimulus selection. The next figure illustrates this with a task similar to one I have been using to study how children recognize objects from their description.
In the top of this figure is the traditional explanation of how an object is selected in response to its description. That is, it is phrased in terms of a conditional discrimination. In this account, of course, the selection response, pointing, is simply evoked by the comparison when it is encountered in the presence of the spoken sample. There is thus no way to explain how selection is controlled by grammatical relations between the elements of the description. In this case two and black which modify square; gray, which modifies circle; and the preposition under describing the relation between the circle and squares. And as a result, current accounts typically ascribe the interpretation of descriptions to receptive, and typically cognitive processes within the subject.
One way to solve these problems is illustrated in the lower figure. The illustration describes a recognition process that precedes stimulus selection. Thus the sample here "Two black squares under a gray circle", may be said to describe what is to be sought. And as illustrated, while rehearsing the topography of this description as a self-echoic, the comparisons are scanned until one is encountered whose description, evoked as a tact, enters into joint control with the topography of the self-echoic. This might be labeled as the moment of recognition. The subject has found a comparison that matches the description in the sense that he can repeat the description both as an self-echoic, and jointly, as a tact.
As shown in the next figure, what the subject does next,
would be an autoclitic report of this joint echoic-tact control. The subject
might point to the stimulus, or might report verbally that he has
found the sought-after object. Either way, the behavior would be an autoclitic
report of prior joint control over verbal behavior.
This provides a simple behavioral account that differentiates recognition from selection, and also suggests that a description differs from an SD in that a description, by specifying the response topography of the self-echoic, thereby specifies the stimulus to select.
Such a notion nicely circumvents the problem of how the grammatical relations within a description can control a selection response. For as we see here, the role of grammar lies within the expressive behavior of the subject rather than within any additional receptive capabilities. Comprehension of a description is thus the product of specifiable behavior rather than the workings of an innate process. And so, to the extent the subject can emit the relevant grammatical relations within a tact and thereby produce the same topography that he is rehearsing as an echoic, he can recognize the stimulus as fulfilling the description. Thus, despite the fact that the term recognition implies cognition, it seems that it too may be constructed from behavioral relations.
And so to conclude, I would merely like to suggest that if the examples discussed here are any indication, it would seem that in the realm of complex performances response topography plays a far more important role in determining stimulus selection than does response rate. Appreciating this appears to provide a basis for reducing many apparently abstract performances to simple behavioral relations.