1
RELATIONAL LEARNING
Back to RealismApplied to / Home PagePavlovian Journal of Biological Science, Vol 17, No 2, pp.69-79
Direct_Contin80.doc
Direct and Continuous Measurement of Relational Learning in Human Pavlovian
Conditioning
John J. Furedy, Ph.D., Jane M. Arabian, Ph.D., Edda Thiels, B.A.,
and Leonard George, B.A.
University of Toronto
Toronto, Canada
Abstract—Three experiments were conducted employing a continuous measure of conditional stimulus/unconditional stimulus (CS/US) contingencies as perceived by the subject (i.e., subjective contingency or SC). It is argued that direct measurement of relational learning, as indexed by SC, can lead to a better understanding of Pavlovian conditioning processes. The first two experiments applied this approach to a methodologic controversy, raising the debate from a procedure-based argument to testing what the subject actually learns about CS/US relationships. While the issue was not resolved, testable hypotheses for future research were generated from the data. The third experiment contrasted the contingency stimulus-stimulus (S-S) account of Pavlovian conditioning with an earlier stimulus-response (S-R) continguity-reinforcement account. In this experiment, both SC and skin resistance were measured. Evidence for the existence of both cognitive-propositional and response-learning processes in conditioning was obtained.
1
RELATIONAL LEARNING
The recent "paradigm shift" (Segal and Lachman 1972) in psychology from stimulus-response (S-R) behaviorism to cognitive approaches has been evident in the specific area of human Pavlovian autonomic conditioning. For example, in a symposium on conditioning and the cognitive processes, held in 1971, the proceedings of which were published two years later (Lockhart 1973), there was considerable emphasis on the role of "cognitive factors," "relational learning," and "awareness of the [conditional stimulus-unconditional stimulus] CS-US contingency" in human Pavlovian autonomic conditioning. Although some contributions to the symposium emphasized the limits of cognitive control (Furedy 1973), the more predominant approach is represented by statements like the conclusion of another contribu-
This paper is based on a paper presented at the meeting of the Pavlovian Society, Budapest, Hungary, August 1980.
Address reprint requests to: Dr. J. J. Furedy, Department of Psychology, University of Toronto, Toronto M5S 1A1 Canada.
tion, according to which "human classical conditioning is mediated by an expectancy of, and preparation for, the UCS" (Dawson 1973, p. 85). There is, indeed, an implicit epiphenomenalism in the recent literature on conditioning, which is similar to the earlier behaviorist epiphenomenalism. According to that earlier tradition, relational learning or awareness was considered to be a mere by-product or epiphenomenon of "true" autonomic conditional responding and was therefore not worthy of serious consideration. Hence human subjects in a conditioning experiment tested according to S-R behaviorist traditions were seldom asked about their cognitions.
The current epiphenomenalism has simply switched terms, so that now "a naive observer might be excused for concluding that autonomic researchers now consider cognitive factors to be the only possible factors in conditioned autonomic behavior" (Furedy 1973, p. 108). The new exclusive emphasis on cognitive factors is also exemplified by the current dominance of contingency models of Pavlovian conditioning, as put forward by Rescorla (1967) and Wagner and
1
RELATIONAL LEARNING
Rescorla (1972). Specific responses to the CS, like the autonomic galvanic skin response (GSR) or the skeletal amount-of-lever-press-suppression in the CER paradigm are then seen as manifestations or indices of the underlying cognitive relational learning, contingency-analysis process.
The approach taken in this and previous (e.g., Furedy and Schiffmann 1973) papers at the Toronto laboratory has been to try to measure relational learning directly, i.e., as a phenomenon in its own right, by assessing the subjective CS/ US contingency. Assessment by a postexperimental rating scale {e.g., Furedy and Schiffmann 1971) or by a more sensitive procedure which seeks to assess continuously changes in subjective contingency (SC) throughout the experiment (Furedy 1973) has been employed. The latter method involves subjects indicating their moment-to-moment changes in belief about the likelihood of US occurrence (i.e., SC) by means of a rotary lever, the position of which is recorded continuously just as changes in skin resistance are recorded.
One aspect of Pavlovian conditioning that has been revealed by this focus on directly and separately measuring relational learning is the presence of certain marked instances of fractionation between the process measured by the SC index and the autonomic process measured by such indices as the GSR. Whether SC was measured after (e.g., Schiffmann and Furedy 1972) or during (e.g., Schiffmann and Furedy 1977) the experiment, fractionation was manifested by: (a) the finding that, whereas the cognitive SC measure differentiated between zero and negative CS/US contingency, i.e., between what Rescorla (1967) has called, respectively, the "truly random" and "explicitly unpaired" control CS, the autonomic conditional responses (CRs) did not differentiate between the two contingencies (cf. Furedy and Schiffmann, who employed the GSR; Schiffmann and Furedy 1972, who examined peripheral vasomotor responding; and Szalai and Furedy 1978, who measured heart-rate deceleration); (b) the finding that the extent of SC discrimination between the CS associated with the US (CS+) and the control CS (CS-) was not correlated with the extent of autonomic CR discrimination (e.g., Furedy and Schiffmann 1974).
The more sensitive, within-experiment SC measure has also made it possible to distinguish between objective (procedural, experimenter-arranged) CS/US contingency and subjective (subject-perceived) CS/US contingency. For example, although the SC measure generally shows the relational learning of subjects to be veridical
and quite sensitive, as in the case of the negative vs. zero contingency discrimination cited above, there can also be cases of nonveridical, yet highly reliable, instances of relational learning. Schiffmann and Furedy (1977) used an arrangement where the USs were delivered at intervals of as short as ten seconds, together with CS-presentations where the CS- was a signal for a safety period (no US) of at least 29 seconds. In terms of objective stimulus contingencies, both the US and the CS- are negative (safety) signals, but the CS- is "more negative" than the US since it predicts longer no-US periods. However, the SC measure showed that, in terms of subjective relational learning, the US was perceived as markedly more negative than the CS- (Schiffmann and Furedy 1977, Figure 1).
This paper presents three experiments that continue the approach of directly measuring relational learning. The first two experiments apply the approach to a methodologic controversy which heretofore has been conducted in strictly procedural (objective-contingency) terms. The approach advances and sharpens the enquiry beyond strictly procedure-based arguments to ones that deal with what the subject actually learns rather than what the experimenter has arranged in the Pavlovian conditioning experiment.
The third experiment was designed to empirically contrast the contingency S-S account of Pavlovian conditioning (e.g., Rescorla and Wagner 1972), with an earlier S-R contiguity-reinforcement one (e.g., Jones 1961). The accounts lead to opposite predictions rather than the weaker pattern of predictions, more common in psychologic research, where one account predicts a difference and the other does not.
Actual Relational Learning About Purportedly Random Control CSs: Experiments I and II
The above-mentioned finding of a failure in the autonomic system to differentiate between zero and negative CS/US contingency is contradicted by only one human autonomic study in which there was some evidence for this sort of differentiation (Prokasy et al. 1973). One feature which was unique to that experiment was that all three CSs were delivered to each subject in a totally within-subject design. Specifically, all subjects received a CS +, a CS that was negatively correlated with the US (i.e., CS-), and a CS that was said to be random with respect to the US (i.e., RS). Within-subject designs seem ideal for assessing the relationships between RS, CS-, and CS + in autonomic conditioning because of the increased sensitivity and avoidance
1
RELATIONAL LEARNING
1
RELATIONAL LEARNING
Fig. 1. Relational learning (vertical axis: mean subjective contingency, percent full scale) in a within-subject design as a function of whether the purportedly random stimulus ("RS") is generated by the "independence" (Experiment I) or the "overlap" (Experiment II) method. The horizontal axis represents seconds after CS onset.
1
RELATIONAL LEARNING
of such problems as the law of initial values. However, Furedy, Poulos, and Schiffmann (1975a) questioned the validity of Prokasy et al.'s (1975) results on the grounds that their RS was actually excitatory, i.e.,like a CS +. On the other hand, Prokasy et al. (1973, p. 152) had questioned the validity of such earlier studies from the Toronto laboratory (e.g., that of Furedy and Schiffmann 1971), on the grounds that the nominal RS in those studies was actually inhibitory, i.e., like a CS-.
A summary of the procedural debate is that it turns on the question of whether an RS should be generated by equating occasions when the RS does and when the RS does not overlap with the US (Prokasy et al/s "overlap" method) or by ensuring independence between RS and US onsets (Furedy et al's "independence" method). Beyond this summary, however, no further advance has been made in resolving the problem. The last paper from the Toronto laboratory asserted that the overlap method of generating RSs led to "logical problems" (Furedy et al. 1975b), such as being forced to misclassify a CS+ in a trace and delay-without-overlap conditioning paradigm (Kimble 1961) as a CS-. On the other hand, the last paper from the Utah laboratory asserted that the independence method led to an inhibitory CS, as shown by "actuarial data" (Prokasy 1975b). It bears emphasis that, although technical, the debate has considerable
empirical significance, if only because without its resolution it is impossible to empirically evaluate the influential position advanced by Rescorla (1967) and accepted by many Pavlovian conditioners that the "proper" control for Pavlovian conditioning is the "truly random" (RS) rather than the "explicitly unpaired" (CS-) arrangement.
However, it is at least arguable from a psychology-of-learning point of view that the more relevant issue is not the procedural one of which arrangement leads to an objectively random stimulus but rather the question of how the subject perceives the two sorts of nominally random stimuli. The continuous within-subject SC measure was employed in the present two experiments to answer this question. The RSs were generated either by use of the overlap (Experiment I) or independence (Experiment II) method.
Materials and Methods
Subjects. Twenty-four and ten college-aged subjects were used for Experiments I and II, respectively. Subjects received course credit for writing a brief account of the experiment after it was over.
Apparatus. The indicator for continuously measuring SC (cf. Furedy 1973, pp. 110-111) was a rotary lever mounted on a scale positioned in front of the seated subject. The horizontal
1
RELATIONAL LEARNING
left, vertical, and horizontal right positions were marked, respectively, -100, 0, and +100. The subject was told that these positions represented, respectively, certainty that the noise would occur, complete uncertainty, and certainty that the noise would not occur. The output from this dial was recorded on an E & M Physiograph in an adjoining room; the chart paper ran at a continuous speed of 1 mm/second. The USs and CSs were the same as those used by Prokasy et al. (1973): a 115-dB sound-pressure-level (spl) noise source from a standard white-noise generator delivered through headphones for the US and three white lights arranged in a column on a box located adjacent to the SC dial for the CS + , CS-, and RS. The sequence of lights and noises was controlled by a combination of paper tape and relays located in another room.
Procedure. In Experiment I, the purpose was to duplicate the computer-generated arrangements as described by Prokasy et al. (1973, pp. 147-8) and as further detailed in a protocol sent by Prokasy (1976). First, the 46-minute session was divided in 540 5.1-second periods; the five-second CS occurrence was randomly assigned to these periods with a probability of .25. The 134 CS events were then randomly allocated to the CS + , CS- or RS categories. Next, the session was again divided into 2700 1.02-second periods, and US occurrence was randomly assigned to these with a probability of .033, there being 95 such 0.2-second US (noise) occurrences. Finally, the two adjustments made by Prokasy et al. (1973) were made, namely: (a) any US falling within 13 seconds of CS- onset was eliminated,* and (b) any US falling within 13 seconds of CS+ onset was moved to CS-\- offset, i.e., a CS-US interval of five seconds. The sequence containing the period with the first 50 US presentations was transcribed to paper tape because Prokasy et al. (1973) reported data for only this part of their experiment.
Before the start of the experiment, subjects were informed that there would be lights and noises and were instructed to use the SC dial to indicate "moment-to-moment" changes in belief about the likelihood that the noise would or would not occur, as in previous studies using this SC measure {e.g., Furedy and Schiffman
* It is this adjustment that, according to critics of the overlap method of generating RSs, results in the RS becoming excitatory (cf. Furedy et al. 1975a, p. 100). On the other hand, as indicated by Prokasy (1975a), the method does result in an approximately equal number of occasions on which the RS and the US do and do not physically overlap with each other, i.e., an equal overlap-nonoverlap ratio for the RS.
1973). All subjects then received the above-mentioned sequence of stimuli. During the sequence of stimuli, the experimenter was in an adjacent room, and the subject was continuously visible through a one-way mirror.
In Experiment II, the aim was to adapt the "independence" method of generating RS employed in two-group (e.g., Furedy 1971) and three-group (e.g., Schiffmann and Furedy 1977) designs to the single-group, within-subject design used by Prokasy et al. (1973). The CS and US durations were the same as in Experiment I (five and 0.2 seconds, respectively); a randomly ordered sequence of 15 CS + , 15 CS-, and 15 US-alone trials was created. In this arrangement, the CS+ was always followed by a US occurring at CS+ offset (i.e., a five-second CS-US interval), and the intertrial intervals were randomly assigned duration of 30, 40, or 50 seconds (so that a CS- was never followed by a US for at least 29 seconds). Then, the experimental session was divided into 191 ten-second intervals, and 15 five-second RSs were randomly assigned to these periods. The exact placement of each RS within a given ten-second interval was determined by dividing the interval into 20 0.5-second subintervals and distributing RS onset randomly among these.* In other respects, the procedure for these subjects was identical to that used in Experiment I.
Results. The index of relational learning, SC, was measured second by second from 0 to five seconds following stimulus (CS + , CS-, and RS) onset. The results are summarized in the left (Experiment I) and right (Experiment II) panels of Figure 1, averaged over trials and subjects. The SC response topography indicated that SC level at second 5 (the point of noise onset for CS + ) could be used as a measure of relational learning. In terms of that index, statistical analyses confirmed the main trends suggested by Figure 1, namely that: (a) in Experiment I (left panel), both CS+ and RS responding exceeded that to CS- (/ (23) = 9.1 and 7.3, respectively, P < .001), but that between CS + and RS did not differ (/ (23) = 1.8); (b) in Experiment II
* This procedure resulted in an independence or lack of correlation between RS and US occurrence so that "the probability of a US is the same given either the presence or absence of the_CS" (Rescorla 1969, p. 25), i.e., p(US/RS) = p(US/RS). However, as has been detailed elsewhere (cf. Prokasy 1975a, especially Table 1), the procedure results in far fewer physical RS/US overlaps than instances wherein the RS and US do not occur together. On these grounds, it has been argued that the independence method of generating RSs results in the nominal RS becoming inhibitory (cf. Prokasy 1975a, Prokasy et al. 1973).
1
RELATIONAL LEARNING
(right panel), CS+ responding exceeded that to RS and CS- (t (9) = 35.3 and 35.7, respectively, P < .001), whereas that to RS and CS- did not differ (t = .31). These statistical outcomes clearly indicate that the nominal RS in Experiment I generated by the overlap procedure, was perceived as being as excitatory as the CS + (as argued by Furedy 1975a, b) and that the nominal RS in Experiment II, generated by the independence procedure, was perceived as being as inhibitory as the CS- (as argued by Prokasy et al. 1975a, b).
In addition to these main outcomes, statistical analyses were performed to determine the reliability of other trends suggested by Figure 1. The first of the questions raised by inspection of the figure is whether the intertrial interval SC levels, as defined by the 0-second (stimulus-onset) values, were the same prior to each stimulus. These 0-second values did not differ between the CS + , RS, and CS- conditions in either experiment, F < 1, suggesting that subjects were unable to predict the nature of the succeeding stimulus.
These results, therefore, also indicate that the 0-second SC values can be validly compared with the five-second SC values to assess the significance of the effect of each stimulus on relational learning. These tests, on the SC difference between seconds 0 and 5, showed that: (a) in Experiment I, CS + and RS were both perceived as significantly positive (t (23) = 6.0 and 6.4, respectively; P < .001), but the perceived negativity of CS- failed to reach significance (t (23) = 1.0); (b) in Experiment II, the perceived positivity of CS + and the negativity of RS and of CS- were significant in all three cases (/ (9) = 27.0, 10.5, and 10.2, respectively;/5 < .001).
Finally, the development of these individual-stimulus SC effects over the session was studied by dividing the trials in the experiments into three blocks. In Experiment I, the number of trials was not divisible by three, so that the first, second, and third blocks contained, respectively, seven, seven, and eight trials with each of the three stimuli (CS + , RS, and CS-). Oneway analyses of variance, with blocks as the factor and the algebraic SC difference between 0-second and five-second values as the dependent variable, did not show any significant development over blocks for CS + , RS, or CS- (F (2.60) = 3.0, 2.3, and 1.4). In Experiment II, the trials were blocked into three 15-trial blocks of five of each of the three stimuli. Analyses of variance showed that both the perceived positivity (in the case of CS + ) and the perceived negativity (in the case of RS and CS-) significantly increased over blocks of trials {F (2.27) - 10.1, 6.2, and
< .001), RS, and
6.7, respectively, for CS-{P < .01)).
Discussion
The results of central concern are those that relate to the relational learning induced by the two nominal RSs. In this respect, the outcomes are clear. Both the criticisms by the Toronto laboratory {e.g., Furedy et al. 1975a) of the Utah laboratory overlap method of generating RSs (cf. results of Experiment I) and those by the Utah laboratory {e.g., Prokasy 1975a) of the Toronto laboratory's independence method of generating RSs {cf. results of Experiment II) appear to be strongly confirmed by the data: the RS generated by the overlap method (Figure 1, left panel) was perceived to be as excitatory as the CS + , and the RS generated by the independence method (Figure 1, right panel) was perceived to be as inhibitory as the CS-.