Journal of Experimental Psychology: Animal Behavior Processes 1980, Vol. 6, No.

2, 112-125

Stimulus Similarity and Retroactive Interference and Facilitation in Monkey Short-Term Memory
Douglas L. Medin
University of Illinois at Urbana-Champaign

Thomas J. Reynolds
Rockefeller University

John K. Parkinson
University of Illinois at Urbana-Champaign Four experiments examined the effects of similarity between the sample and an interpolated stimulus in a modified delayed-matching-to-sample (DMTS) paradigm. The basis trial sequence was as follows: (a) a sample was presented, and the response to it was either rewarded or nonrewarded, (b) an interpolated stimulus was presented, and the response to it was either rewarded or nonrewarded, and (c) after a delay interval a choice test was given between the initial sample and a new stimulus, with the sample being correct if it had been initially rewarded and incorrect if it had been initially nonrewarded. When the sample and the interpolated stimulus were associated with opposite outcomes (sample rewarded, interpolated stimulus not rewarded or sample not rewarded, interpolated stimulus rewarded), retroactive interference was observed to increase as the similarity of the two stimuli increased. When the two stimuli were either both rewarded or both norewarded, retroactive facilitation was observed so long as the sample and interpolated stimuli were similar along either the relevant or an irrelevant dimension. Finally, test probes involving the interpolated stimulus and a new stimulus revealed asymmetrical interactions such that the interpolated stimulus influenced performance on the initial stimulus more than the initial stimulus altered performance on the interpolated stimulus. Taken as a whole, the results are inconsistent with the idea that alternative stimuli are encoded independently of each other. At best, independence may hold only for cases in which the stimuli are not similar to each other. From the earliest beginnings of what is now called classical interference theory (MeGeogh, 1932, 1952; Melton & Irwin, 1940; Underwood, 1948, 1957), stimulus similarity was assumed to play an important role in retroactive interference ( R I ) . The transfer and retroaction surface of Osgood (1953, p. 532), for example, provided both a summary of research on similarity effects and a framework for later research on retroactive interference in human memory (e.g., Wickelgren, 1965, 1966). One might, therefore, have expected that the recent upsurge of interest in animal memory (e.g., D'Amato, 1973; Honig & James, 1971; Medin, Roberts, & Davis, 1976) would have included a corresponding emphasis on stimulus similarity effects, Rather than focusing on similarity relations in RI, however, theoretical treatments have assumed that interfering stimuli primarily T T a l t e r performance by reducing or eliminating
. . . . , , . . , rehearsal of the to-be-remembered mformation in short-term memory (STM) (e.g., Roberts & Grant, 1978; Wagner, Rudy, & ,,,,., i n m \ TM i u Whitlow, 1973). This emphasis is by no

This research was supported by U.S. Public Health Service Grants MH 2S134 and MH 32489 and by a U.S. Public Health Service postdoctoral fellowship MN 07SS1 to the second author. Requests for reprints should be sent to Douglas L. Medin, Psychology Department, University of

Illinois, Champaign, Illinois 61820.

means arbitrary; for example, Grant and

Copyright 1980 by the American Psychological Association, Inc. 0097-7403/80/0602-0112$00.75

112

STIMULUS SIMILARITY AND STM Roberts (1976) tested pigeons in a clelayedmatching-to-sample paradigm (DMTS) with interpolated stimuli in order to examine the effects of similarity, cue salience, familiarity, stimulus complexity, and degree of illumination. A small set of stimuli consisting of four colors and four forms was used, and within a day only two stimuli appeared as the sample ; for example, on color test days red and green served as sample and test stimuli, and the interpolated stimulus could be either red or yellow (high similarity) or a diagonal or cross pattern (low similarity). Grant and Roberts reported that degree of illumination was the only significant source of retroactive interference; similarity had virtually no overall effect. Although Grant and Roberts found that similarity was not a prominent factor controlling retroactive interference in their pigeons, interactions between particular sample and interpolated stimuli suggested that similarity indeed played an important role. Specifically, very strong RI was observed when the sample was green and the interpolated stimulus was yellow or when the sample was red and the interpolated stimulus was blue. In contrast, little RI (or even modest retroactive facilitation) was found when green was followed by a blue or when red was followed by a yellow. How might one interpret the Grant and Roberts results on similarity and RI ? It seems that one could either emphasize the relation between the interpolated stimulus and the initial stimulus or the relation between the interpolated stimulus and the choice test. Consider the latter possibility. If a blue stimulus is more similar to green than red and if a yellow stimulus is more similar to red than green, then the pattern of interactions observed by Roberts and Grant would apparently follow from principles of stimulus generalization. According to this interpretation, one could argue that the initial sample and interpolated stimulus were encoded independently, did not influence each other, and indeed that the same pattern of results would have occurred if the initial sample stimulus was eliminated and only the blue or yellow stimulus had

113

been presented prior to the choice test between red and green. This interpretation is consistent with the Roberts and Grant trace strength competition model which assumes that the alternative stimuli are independent. The other main possibility is that the appearance of a sample stimulus and a later interpolated stimulus cannot be treated as independent events and that the particular interaction can produce patterns of both retroactive interference and retroactive facilitation that cannot be explained by appealing to relations between an interpolated stimulus and the choice test. The theoretical and experimental contrasts can best be seen with a concrete example, to which we now turn. Experiment 1 The design of Experiment 1, shown schematically in Figure 1, serves to bring out some of the theoretical contrasts of interest. Stimuli for each trial sequence were drawn from a large pool of stimulus objects. Therefore, although Figure 1 accurately depicts the logical relations among the alternative trial sequences, the particular stimuli embodying these conditions varied from trial to trial. Characteristics of interpolated stimuli differentiated the conditions. The five main conditions consisted of (a) no interpolated stimulus (control), (b) an interpolated stimulus identical to the sample (A-A), (c) an interpolated stimulus matching the sample along the dimension that is relevant on the choice test but differing from the sample and test stimuli in either color or form (A-C), (d) an interpolated stimulus differing from the sample and test stimuli in both color and form (A-D), and (e) an interpolated stimulus identical to the test stimulus which had not appeared as the sample (A-B). Sample and interpolated stimuli within one trial were either both rewarded ( + ) or both nonrewarded ( ) . The correct choice on the test was consistent with the reward status of the sample; that is, if the sample was rewarded, it was correct on the choice test, but if the sample was not rewarded, it was incorrect on the choice test. The task required attention to both the color

114

D. MEDIN, T. REYNOLDS, AND J. PARKINSON

and form of the sample since on half of the trials color was relevant and on half form was relevant. The main comparison of theoretical interest is Condition A-A versus Condition A-C. Since on A-C trials the interpolated stimulus matches the sample on the relevant dimensions and does not match the alternative on either color or form, there should be more generalization from the interpolated stimulus to the sample than to the alternative stimulus. Given that the same outcome is always associated with the sample and interpolated stimulus, performance should be facilitated, relative to the control condition. The A-A trials contain this same source of facilitation, but, in addition, the interpolated stimulus matches both stimuli along the irrelevant dimension. Whether matching along a constant, irrelevant dimension should impair or improve performance is a theoryspecific question, but most theories predict impairment. We distinguish between two classes of generalization theories. In one class are those formulations that assume that overall generalization is a simple, additive function of the generalization from individual component dimensions (e.g., Spence's 1936 discrimination learning theory). In contrast, other theories assume that generalization along each dimension is not independent of differences along other dimensions, and thus that component dimensions interact (e.g., Medin, 1975, 1976; Medin & Schaffer, 1978; Spiker, 1963, 1970). To begin with, we focus on predictions derived from analysis of the relation between interpolated stimuli and choice objects. Additive generalization theories lead to the prediction that A-C trials should produce performance at least as good as that associated with A-A trials. On A-A and A-C trials there will be the same difference in generalization tendency to the correct and incorrect choice stimulus arising from the relevant dimension, and the only question concerns the contribution of generalization along the irrelevant dimension. When the sample and interpolated stimuli both are rewarded, some response tendency should be added to

each choice stimulus; when the sample and interpolated stimuli both are not rewarded, some response tendency should be subtracted from each choice stimulus. To observe effects of these conditions on performance, we need to specify some choice rule mapping these generalization tendencies onto choice responses. Choice rules that ignore constant, irrelevant cues (see Flagg & Medin, 1973, for a review) will, of course, predict that Condition A-C will produce performance equivalent to Condition A-A, since the same difference in generalization will be present in each case. Choice rules that take absolute strengths into account, such as rules employing a threshold concept (e.g., Spence, 1956, pp. 204206),present a somewhat more complicated picture. If performance is a monotonically increasing or decreasing function of absolute strength with the difference in strength between choices held constant, then the relative ease of Conditions A-A and A-C will interect with reward and nonreward. If A-A yields better performance than A-C when both sample and interpolated stimuli are rewarded, then the ordering should be reversed when the stimuli are not rewarded. If performance first increases with absolute strength and for higher values then decreases, any difference between A-A and A-C trials should interact with reward and nonreward. The only case in which A-A could be uniformly better than A-C would be if performance first decreased with increases in absolute strength and later increased with absolute strength and A-C trials produced strengths that fell near that minimum point. To our knowledge, no current theory employs such a choice rule, and we tentatively conclude that additive generalization theories imply that A-A trials will not yield better performance than A-C trials. Predictions of interactive generalization theories that focus on the relation between interpolated stimuli and choice test objects (e.g., Medin, 1975 ; Medin & Schaffer, 1978) depend on the specific choice rule employed. In these formulations, generalization along each stimulus dimension is not independent of differences along other dimensions. According to interactive theories, generaliza-

STIMULUS SIMILARITY AND STM

115

tion along the relevant dimension in the A-C condition will be reduced by the differences between the interpolated stimulus and the test stimuli along the irrelevant dimension. If a ratio rule is used (as in Medin & Schaffer, 1978), this reduction in generalization will affect the correct and incorrect choice stimuli equally, and there will be no difference predicted between A-A and A-C trials. If choices are a function of differences rather than ratios (as in Medin, 1975), then Condition A-A should yield better performance than Condition A-C. It should be noted, however, that Medin's (1975) theory specifically rejects the idea that two alternative stimuli appearing in the same experimental context (such as the sample and interpolated stimuli) are encoded independently. For the moment, we propose that accounts of retroactive interference based solely on the relation between the interpolated stimulus and the choice test stimuli generally predict that A-C trials will produce performance comparable with that seen on A-A trials. In contrast, theories that focus on the relation between the sample and interpolated stimuli predict that A-A trials will be associated with better performance than A-C trials. Our rationale is based on an analysis of stimulus processing in relation to reinstatement or retrieval cues (e.g., Spear, 1973, 1976; Wagner et al., 1973). Suppose that when the sample is presented and associated with some outcome, the stimulus and outcome are rehearsed together in shortterm memory. Further assume that the longer the stimulus-outcome pair is rehearsed the better will be performance. Normally, posttrial events, such as the lowering of the Wisconsin General Test Apparatus (WGTA) screens, tend to disrupt rehearsal (see Motiff, Dekock, & Davis, 1969), but when the interpolated stimulus is presented, it may act as a retrieval cue for the sample stimulus and its associated outcome (reward or nonreward) and result in additional rehearsal of the stimulus-outcome pair. The more similar an interpolated stimulus is to the initial stimulus, the more effective the interpolated stimulus would be in reactivating the initial stimulus-outcome pair. A

red triangle should be a better retrieval cue for a red triangle than a green triangle, and therefore Condition A-A should yield better performance than Condition A-C. A brief summary is in order. If retroactive interference from an interpolated stimulus results solely from generalization between the interpolated stimulus and the choice test stimuli, then Condition A-C should be at least as good as Condition A-A. If retroactive interference also reflects contributions from the relation between sample and interpolated stimuli, than A-A should be better than A-C. Of secondary interest are comparisons between A-D and Control trial sequences. One might expect A-D trials to yield lower performance than control trials, to the extent that poststimulus processing of the sample is disrupted by the appearance of an irrelevant, interpolated stimulus. However, the effect might well be small in that performance in this paradigm presumably reflects contributions from both short- and long-term memory, with the mixture increasingly biased toward long-term memory as the retention interval lengthens. In Experiment 1 the shortest retention interval was 21 sec (12 sec between sample and interpolated stimulus, plus the 3 sec that it takes on the average for monkeys to respond to the interpolated stimulus, plus the minimum 6-sec delay interval between the interpolated stimulus and the choice test). Under these circumstances performance may be primarily based on long-term memory. Finally, the A-B trial sequences were included as a follow-up to earlier experiments (Reynolds & Medin, 1979) aimed at assessing the conditions under which monkeys could learn to choose the first of two sample stimuli if both were given on the choice test. Reynolds and Medin did not observed above chance performance on this condition after more than 600 trials. The A-B test in the present study was more demanding, in that Reynolds and Medin used a procedure in which initial sample stimulus was always correct on choice tests whereas in the present study, the sample stimulus could be correct or incorrect on the choice test, depending

116
FORM SAMPLE CONTROL A-A--AB A - CAB A-D-AB A-B--AB

D. MEDIN, T. REYNOLDS, AND J. PARKINSON

RELEVANT TEST

A-21,30,48 sec.--

k-IZ sec.A-6.l5,33sec.L--l2sec.A-6,15,33!
: t

^--I2sec. k--l2seo.[!}6,l5,33sec.V&
COLOR
SAMPLE

lR

RELEVANT
TEST

CONTROL A-A-AB A-C-AB

A - D--AB

A-2l.30,488ec.-

hr- IZsec.Ar6,l5,33sec.-^
tx- I28ec.r b,-- I2sec.(

A-B-AB

--I2seo.^

Figure 1. Schematic illustration of the design of Experiment 1. (A sample stimulus was followed by either no interpolated stimulus [Control] or by interpolated stimuli that varied in their similarity to the sample and alternative test stimulus.)

on its reward status when it appeared as

the sample.

Method
Subjects. The subjects were four pigtailed monkeys (Macaco nemestrina), two rhesus monkeys (Macaco mulatto), and one cynomolgus monkey (Macaco fascicularis). All were 5-8-yr-old jungleborn females with previous discrimination and DMTS experience. They were maintained on a 12: 12 hr light/dark cycle initiated at 0600 hours were fed laboratory chow daily after testing. Apparatus and materials. The monkeys were tested in a darkened room whose outside sounds were masked by white noise. The WGTA was lighted by two 15-W fluorescent bulbs, and the three food wells on the gray form board were spaced IS cm center to center. The stimuli were 10 geometric shapes cut from .61-cm plywood (star, arrow, semicircle, triangle, inverted u, inverted v, x, hourglass, house, bar) in each of the 10 colors (red, green, yellow, blue, black, white, orange, purple, brown, tan). The objects were equated subjectively for size by the experimenters such that they appeared to be of equal overall area, and each was presented flat on the food tray (the objects were not mounted on a base) in a single orientation throughout testing.

Procedure. The experimental design (shown schematically in Figure 1) consisted of the factorial combination of S trial types X 2 relevant dimensions (color or form) X 2 reward conditions (samples rewarded or not rewarded) X 3 delay intervals (21, 30, 48 sec) for a total of 60 trial sequences. On control trials, a single object was presented over the center food well, and the monkey displaced it for either a raisin reward ( + ) or no reward ( ) . Following a delay, a choice test was given with the sample and a new object covering the side food wells. If the sample had been rewarded, it was correct and rewarded on the test; if it had not been rewarded, the new stimulus was correct and rewarded. On interference trials, the sample object was presented over the center food well, and the monkey displaced it for either a raisin reward ( + ) or no reward ( ). Following a 12-sec interstimulus interval, an interpolated stimulus was given, being associated with the same reward/no-reward condition as the initial sample, that is, the two presentations were either both rewarded or both nonrewarded. After a delay, a choice test was given with the first sample and a new stimulus covering the side food wells. The correct, rewarded object on the test was consistent with the reward status of the initial sample. Each stimulus and test presentation ended with the retraction of the form board and the lowering of the opaque screen of the WGTA. The four interference conditions shown in Figure 1 were distinguished by the similarity relation among the samples, interpolated stimuli, and test objects: in Condition A-A, it was identical to the sample; in A-C, it differed on the irrelevant dimension only; and in A-D, it differed on both the relevant and irrelevant dimensions. Finally, the interpolated stimulus in the A-B condition differed from the first sample on the relevant dimension only, but unlike other trial sequences, the interpolated stimulus appeared along with the sample on the choice test. These S basic trial types (control, 4 interferences) X 2 reward conditions X 3 delays were presented with either color or form as the relevant dimension. When color was relevant, the first sample and the two choice objects were all the same form; when form was relevant, they were all the same color. The monkeys were tested S days/week, 30 trials/ day, for 48 days. In each session the trials were presented in one of eight quasi-random orders assembled such that half of the trials had color relevant and half, form relevant, and the samples were rewarded and nonrewarded equally often. Further, each of the 60 trial sequences appeared once every two sessions, and the position of the correct object on a test was randomized but balanced within a day, subject to the constraint that the correct object occupy the same position no more than four trials in a row. The 72 objects for

STIMULUS SIMILARITY AND STM each session were randomly drawn without replacement from the 100-object stimulus pool, with all the monkeys receiving the same stimuli on any given day. The intertrial interval was a constant 15 sec, correct responses were rewarded with a raisin, and a noncorrection procedure was employed.

117

.90-

o
UJ

o o

.80-

Results The main conditions were associated with substantial differences in performance, but color-versus-form tests and rewarded samples versus nonrewarded samples did not alter performance. Results for the main conditions of interest are shown in Figure 2. Both A-A and A-C trials produced facilitation relative to control (no interpolated stimulus), and consistent with the idea that interpolated stimuli affect processing of sample stimuli, Condition A-A was associated with better performance than Condition A-C. The A-D condition was clearly no worse and perhaps slightly better than the control condition. The differences interacted somewhat with delay; the crucial A-A, A-C difference being more apparent at the longer delays. Finally, the A-B condition produced performance that was consistently below chance (48%, 43%, 40% correct at delay intervals of 21, 30, 48 sec). Since a preliminary analysis of 8-day practice blocks indicated that there were no significant practice effects, the data were collapsed across days and a Delays X Reward Condition X Trial Types X Color Versus Form-relevant analysis of variance was conducted. Data from A-B trials were excluded since this condition did not bear directly on the interpretations of RI and stimulus generalization under consideration. The overall analysis of errors indicated that the effects of delays, F(2, 12) = 6.18, MSe = 7.90, p < .05, trial types, F(3, 18) = 22.08, MSe = 4.63, p < .001, and the interaction of Delays X Trial Types, F(6, 36) = 2.40, MSe 3.64, p < .05, were significant. No other effect or interaction was statistically significant. Individual comparisons employing a Newman-Keuls test and the .05 confidence level indicated that Condition A-A was better than Conditions A-C, A-D, and Control and that Condition A-C in turn yielded better
.70-

21

30
INTERVAL (SEC.)

48

SAMPLE-TEST

Figure 2. Proportion correct as a function of sample test interval for the main conditions in Experiment 1. (See Figure 1 for identification of groups.)

performance than A-D and Control sequences were not reliably different. Detailed analyses of the Conditions X Delay interaction revealed the following pattern of differences : At the 21 -sec delay, Conditions A-A and A-C were reliably better than A-D and Control; at the 30-sec delay, Condition A-A was better than the other three conditions; at the 48-sec delay, A-A performance was better than A-D and Control performance. Discussion A full repetition of the sample stimulus was more effective than repetition of just the relevant stimulus dimension. This result is consistent with models that assume that the initial sample stimulus and an interpolated stimulus are not independent. Specifically, the interpolated stimulus may act as a retrieval cue for the sample representation and thereby allow additional processing of the sample stimulus. An interpolated stimulus different in both color and form from the choice stimuli neither facilitated nor impaired performance compared with control trials. The persistently poor performance of A-B trial sequences indicates that the monkeys were unable to base their performance on the first of two sample stimuli when both appeared on the test. This result is mildly surprising since some investigators have argued that DMTS involves a temporal discrimination

118
SAMPLE CONTROL A

D. MEDIN, T. REYNOLDS, AND J. PARKINSON

TEST 15,30 sec.-

DIFFERENT-*^

Figure 3. Schematic illustration of the design of Experiment 2. (The interpolated stimulus always matched the sample and the alternative stimulus in color.)

(e.g., D'Amato, 1973), and on that basis one would expect that either the second or the first of two sample stimuli could control performance. One possible objection to the main contrast between the A-A and the A-C conditions is that an exact repetition of the sample stimulus may have some special (perhaps configurational) properties. The next experiment extended the generality of the findings of Experiment 1 to the case in which the interpolated stimulus was never a repetition of the initial sample. Experiment 2 The key contrast in Experiment 2 was between the control condition and a condition in which the interpolated stimulus matched the sample on the irrelevant dimension. If the sample was a red triangle, for example, and the choice test was to be between the red triangle and a red square (i.e., color irrelevant), then the interpolated stimulus might be a red circle (see Figure 3). On the average, the interpolated stimulus should generalize equally to the two choice stimuli, since the interpolated stimulus matches the choice stimuli in color and is not more similar to one choice object than the other in form. If the interpolated stimulus influences performance solely in terms of generalization to the test stimuli, interpolated stimuli either will not help or may impair performance. The full design of Experiment 2 is sche-

matized in Figure 3. Control trials have no interpolated stimulus and provide a baseline to evaluate the other trial types. The interpolated stimulus on the other trial sequences differs from the choice stimuli in form but matches them in color. On Same trials the sample and interpolated stimuli are associated with the same outcome (both rewarded or both not rewarded), and on Different trials the sample and interpolated stimulus are associated with different outcomes (one rewarded, one nonrewarded). As in Experiment 1, the correct choice on the test was determined by the reward status of the sample. In addition, both the interval between the sample and interpolated stimulus and the delay between the initial sample and the test were varied. Any outcome associated with the interpolated stimulus should generalize equally to the sample and alternative-choice test stimulus. By the same logic developed in Experiment 1, additive generalization models based on extant choice rules predict either that the interpolated stimulus will have no effect or that its effect will interact with whether or not the interpolated stimulus is rewarded and will thereby produce facilitation for one outcome and impairment for the other (at this level of generality one cannot specify which outcome will produce which effect, only that reward and nonreward should produce opposite effects). Interactive generalization models, whether based on ratios or differences, predict the same lack of effect so long as one considers only generalization from the interpolated stimulus to the test stimuli. If an interpolated stimulus is viewed as a retrieval cue, it could provide the opportunity for additional processing of the initial sample and its associated outcome. On Same trials this should only be beneficial, but on Different trials the memory representation for the sample stimulus may be modified by the (different) outcome paired with the interpolated stimulus. This should interfere with performance. This interference may possibly derive from the sample and interpolated stimulus being rehearsed simultaneosuly in short-term memory. Tech-

STIMULUS SIMILARITY AND STM

119

nically, this state of affairs might arise either if the interpolated stimulus reactivated the representation of the sample stimulus or if the sample stimulus was still being rehearsed at the time the interpolated stimulus was presented. We prefer the former possibility because it leads more directly to the idea that similarity between the interpolated stimulus and the sample is a critical variable altering performance. According to the above line of thinking, we should observe RI on Different trials but retroactive facilitation on Same trials. Davis and Fitts (1976) studied forgetting in monkeys, using a direct analogue of the Same versus Different contrast shown in Figure 3. They reported that RI was substantially greater in the Different condition than in the Same condition. Since their experiments had a somewhat different purpose, they used stimuli having no special similarity relations (magazine pictures) and did not include a control condition having no interpolated stimulus. Method
Subjects. The subjects were the four pigtailed, two rhesus, and the one cynomolgus monkeys used in Experiment 1. Approximately 2 mo and a move to a new laboratory intervened between Experiment 1 and Experiment 2 during which time the monkeys participated in DMTS studies using junk objects. In these experiments proactive interference was studied as a function of intertrial interval and interstimulus interval, with some samples rewarded and some not rewarded. Apparatus and materials. The monkeys were again tested in a WGTA with 252 commonly used and manufactured (junk) objects that were each painted 1 of 14 different colors. Both the WGTA and the form board were black rather than gray. The stimuli were always the same color within a trial, but they differed in color between trials. The stimulus pool was exhausted, and the stimuli were randomly re-sorted every 3 days. Procedure. The experimental conditions comprised a factorial design with certain missing cells using the variables of trial type, interstimulus interval, reward, and delay interval. Trial sequences were of three main types as indicated in Figure 3: Control, Same, or Different. Control trials involved a single sample presentation that was rewarded or nonrewarded, followed by a delay of 15 or 30 sec, and a choice test between the sample and a new stimulus. On Same trials either 6 or 12 sec after the sample was responded to, an in-

terpolated stimulus matching the sample in color but not form was presented and associated with the same reward outcome as the sample. When the interstimulus interval was 6 sec, delays of either 6 or 21 sec preceded the choice test so that the sample-to-test interval matched that of Control trials. For the 12-sec interstimulus interval, a 15sec delay interval was used to match the longer control trial delay. Different trials mirrored Same trials exactly except that the outcome associated with the interpolated stimulus was always the opposite of that associated with the sample on that trial. The correct object on the choice test was consistent with the outcome associated with the sample, that is, if the sample was rewarded, it was correct on the choice test, and if it was not rewarded, it was incorrect on, the choice test. The combination of trial types, sample reward status, interstimulus interval, and delay interval produced 16 unique presentation sequences, which were presented twice each session in one of 6 quasi-random orders. The position of the correct object on a test was randomized, but balanced within a day, subject to the constraint that the correct object occupy the same position no more than four trials in a row. Testing continued 32 trials each day, 5 days a week, for 4 wk. The intertrial interval was a constant 15 sec, correct responses were rewarded with a raisin, and a noncorrection procedure was employed.

Results The main results are shown in Figure 4. Compared with the control condition, interpolation of a stimulus sharing the value along the constant, irrelevant dimension had a dramatic effect. When reward conditions for the sample and interpolated stimulus differed (Different trials), retroactive interference was strong (performance on Different trials was significantly above chance but very poor ) ; when reward conditions for the sample and interpolated stimulus matched, retroactive facilitation was observed, especially at the longer delay interval. This pattern of results is exactly that predicted by the idea that the initial sample presentation and the interpolated stimulus do not act as independent events. The results cannot be explained in terms of generalization from the interpolated stimulus to the choice stimuli. The interval between the sample and the interpolated stimulus had a negligible effect, and forgetting with delay interval is apparent only on control trials. Statistical tests confirmed this pattern of

120

D. MEDIN, T. REYNOLDS, AND J. PARKINSON Discussion

io
LU

.70Control Different-12 * Different- 6

|.50 +

a.
15 30 SAMPLE-TEST INTERVAL (SEC.)

Figure 4. Proportion correct as a function of

sample-test interval for the main conditions in Experiment 2.

retroactive interference and facilitation. Because the experiment was not a complete factorial design, two separate analyses of variance were performed. The first excluded the interpolated stimulus trials with a 12-sec interstimulus interval and used the factors of trial type, delay interval, and sample reward versus nonreward. This analysis indicated that the effect of trial type was significant, F(2, 12) = 30.21, MSe = 16.5, p < .01, and that the effect of delay, F ( l , 6) = 5.44, MSe = 6.4, .05 < p < .10, and the Delay X Trial Type interaction, F(2, 12) = 3.73, MSe - 15.0, .05 < p < .10, approached significance. Individual comparisons among trial types with a Newman-Keuls test at the .05 confidence level indicated that Different trials yielded worse performance than either Same or Control trials and that Same and Control trial difference was short of significance (.05 < p < .10). A second analysis of variance excluded the data from short delay intervals and considered the control condition as part of a factorial analysis involving trial types, interstimulus interval, and sample reward versus nonreward. This procedure followed the recommendation of Winer (1962, p. 267). The analysis of variance indicated that only the effect of trial type, F(2, 52) = 6.35, MSe (pooled) = 55.46, p < .01, was significant. Individual comparisons revealed that Same trials produced better performance than either Different or Control trials and that differences between the latter two trial types fell short of significance (.05 < p < .10).

The results are inconsistent with the idea that the sample and interpolated stimulus presentations are processed independently. An interpolated stimulus sharing the value of the sample on the irrelevant dimension (in this experiment, color) produced either retroactive facilitation or retroactive interference, depending upon whether the outcome associated with the interpolated stimulus was the same or different from that of the sample. Exactly this pattern of results is predicted by the idea that the interpolated stimulus acts as a retrieval cue to reactivate and further influence the representation of the sample stimulus. Further, the lack of delay effects for the interpolated stimulus conditions despite forgetting on control trials is also consistent with the reactivation process, that is, the effective sample test interval would be shortened if the interpolated stimulus acted as a retrieval cue. The results are most consistent with the idea that sample and interpolated stimuli do not function as independent events influencing DMTS performance. It is possible to argue that the difference in performance on Same-versusDifferent trials represents some general confusion that does not depend on the similarity of the sample and the interpolated stimulus. That is to say, an animal might remember which object was presented but not whether it was rewarded. If the monkeys tended to confuse the respective outcomes associated with the sample and interpolated stimulus, performance on Different trials would suffer. A third experiment was run in an attempt to show that the Same-versusDifferent effect depends on the similarity of the sample and the interpolated stimuli. Experiment 3 The third experiment was designed to test the prediction of retrieval models that similarity of the sample and interpolated stimulus should interact with Same-versusDifferent trial type. In particular, high similarity should facilitate performance more on Same trials and impair performance more

STIMULUS SIMILARITY AND STM

121

on Different trials, relative to low stimulus similarity. For the high similarity condition, the interpolated stimulus differed from the sample stimulus in both color and form, as in the A-D condition in Experiment 1. The stimuli for the choice test were always of the same color and different form. Generalization models considering only generalization from the interpolated stimulus to the choice stimuli predict that neither variable should have much effect, since generalization from the interpolated stimulus to the two choice stimuli should be approximately equal. Method
Subjects and apparatus. The four pigtailed and one cynomolgus monkeys used in the first two experiments again served as subjects. Experiment 3 immediately followed Experiment 2 and employed the same apparatus. Stimuli. A pool of 288 junk objects that were each painted 1 of 14 different colors was used. Each day's testing required 96 objects, and when the stimulus pool was exhausted, the stimuli were randomly re-sorted and used again. Procedure. The experimental design consisted o f a 2 X 2 X 2 X 2 factorial with trial types (Same versus Different), reward status (sample rewarded versus nonrewarded), delay (6, 21 sec), and similarity (high versus low) of the interpolated stimulus to the sample as factors. On each trial the sample was presented and rewarded or nonrewarded, 6 sec later an interpolated stimulus was presented and rewarded or nonrewarded, and after a delay of 6 or 21 sec a choice was given between the sample and a new stimulus of the same color but a different form. If the sample had been rewarded, it was correct on the test; if it had not been rewarded, the other stimulus was correct. On Same trials the sample and interpolated stimulus were associated with the same outcome (both rewarded), and on Different trials they were associated with contrasting outcomes. The interpolated stimulus was always different in form from the other objects used within the trial. On high-similarity trials the interpolated stimulus was the same color as the other objects, and on low-similarity trials the interpolated stimulus differed in both color and form. Each of these 16 distinct trial sequences appeared twice each day in one of six quasi-random orders for 20 days. The position of the correct object on a test was randomized, balanced within a day, and subject to the constraint that the correct object occupy the same position no more than four trials in a row. The intertrial interval was a constant 15 sec, correct responses were rewarded with a raisin, and a noncorrection procedure was employed.

Table 1 Proportion Correct on Same and Different Trials as a Function of the Similarity of the Interpolated Stimulus to the Other Stimuli Trial type Similarity High
Low

Same
.16 .20

Different
.42 .39

Results Overall, subjects averaged 18% errors on Same trials and 36% errors on Different trials. As Table 1 shows, high similarity facilitated performance on Same trials and impaired performance on Different trials. This is just the interaction predicted on the basis of the idea that the initial sample and the interpolated stimulus are not processed independently. An analysis of variance with the factors of similarity, delay, trial type, and reward revealed that the main effect of trial type (Same versus Different) was significant, F(l, 14) =84.0, MSe = 20.0, p < .01, as was the theoretically important Trial Type X Similarity interaction, F(l, 4) = 9.2, MS? = 3.27, p < .05. The effect of delay approached significance, F(l, 4) =4.66, MSf = 3.68, .05 < p < .10. No other effect or interaction was significant. Discussion The results provide yet another demonstration that sample and interpolated stimuli do not act independently in influencing delayed matching performance. Monkeys made twice as many errors when the sample and interpolated stimulus were associated with different rather than matching reward outcomes. Further, these effects depended on the similarity of the interpolated and sample stimuli. When the similarity was high, performance was better on Same trials and worse on Different trials than when similarity was low. This interaction between similarity and trial type was produced when the interpolated stimulus had no more formal similarity to the sample stimulus than

122

D. MEDIN, T. REYNOLDS, AND J. PARKINSON when the stimulus pool was exhausted, the stimuli were randomly re-sorted and used again. Procedure, The experimental design involved the factors of trial types (Same versus Different), probe type (first sample tested versus second sample tested), reward status (probe rewarded as sample versus probe not rewarded as sample), and delay (short versus long). On each trial the first sample was presented and rewarded or nonrewarded, 6 sec later a second sample was presented and rewarded or nonrewarded, and after a further delay of 6 or 15 sec, either the first or the second stimulus appeared, along with a new stimulus on the choice test. If the sample had been rewarded, it was correct on the test; if it had not been rewarded, the other stimulus was correct. On Same trials the first and second sample stimuli were associated with the same outcome, and on Different trials the two samples were associated with contrasting outcomes. The two sample stimuli and the new stimulus on a given trial were always of the same color, but color could vary between trials. Each of these 16 distinct presentation sequences appeared in one of six quasi-random orders twice each day. Testing continued 5 days a week for 18 days. The position of the correct object on a test was randomized, balanced within a day, and subject to the constraint that the correct object occupy the same position no more than four trials in a row. The intertrial interval was a constant IS sec, correct responses were rewarded with either a raisin or a small piece of apple, and a noncorrection procedure was employed. Results Performance as a function of trial type, probe type, and delay is shown in Table 2. The interference produced by Different outcomes is obvious, being more clearly evidenced when the first sample was probed than when the second sample was probed. In other words, the interaction between the two sample stimuli was asymmetrical, with the first of two sample stimuli being influenced by the reward status of the second stimulus more than the second was influenced by the first. A 2 x 2 X ' 2 x 2 analysis of variance was conducted with the factors of trial type, probe type, reward status, and delays.1 The
1 Technically, Delay is not a proper factorial variable because the total delay between the to-beprobed sample and the choice test depended on whether the first or the second sample was probed. It is more convenient, however, to treat delay as a factorial variable than to compute separate analyses of variance, and the main experimental results do not hinge on this practice.

to the new stimulus for the choice test. The idea that the influence of an interpolated stimulus can be described solely on the basis of generalization to the choice stimuli cannot account for this pattern of results. Although the interaction between similarity and trial type was quite small, it was statistically reliable, and we have subsequently replicated this effect as part of a larger study, using a different group of monkeys (Medin, Note 1). Experiment 4 A major unresolved question is whether the interactions between the sample and the interpolated stimulus are undirectional. That is to say, if the interpolated stimulus was paired with a new object on the choice test, would interactions between the processing of the initial sample and the interpolated stimulus produce proactive interference and facilitation similar to the retroactive effects occurring when the first sample is probed on a choice test ? Predictions from the idea that the interpolated stimulus acts as a retrieval cue leading to access and further processing of the sample stimulus depend on further details concerning just what is reactivated and rehearsed (e.g., just the representation of the initial stimulus or both the representation of the initial stimulus and its associated outcome). In the one study bearing on this issue, Davis and Fitts (1976) found that the reward status of the second of two stimuli affected performance on the first stimulus substantially, but they noted little if any influence of the reward status of the first stimulus on performance based on the second. The Davis and Fitts study used unrelated stimuli, and it is possible that proactive influences might be greater for more similar stimuli. Experiment 4 tested this possibility. Method
Subjects and apparatus. The four pigtailed and one cynomolgus monkeys from the earlier experiments again served as subjects. Experiment 4 immediately followed Experiment 3 and used the same apparatus. Stimuli. The pool of 288 junk objects each painted 1 of 14 different colors was again employed. Each day's testing required 96 objects, and

STIMULUS SIMILARITY AND STM

123

main effects of probe type, F(l, 4) = 11.42, MSe = 18.50, p < .05, and trial type, F(l, 4) = 43.33, MSe = 10.44, p < .01, were significant, with tests using the second sample producing better performance than tests with the first sample, and Same trials being better than Different trials. The interaction of probe type with trial type, F(l, 4) = 35.58, MSe = 5.94, p < .01, and the three-way interaction involving trial type, probe type, and delay, F(l, 4) = 10.21, MSe = 6.70, p < .05, were also reliable. Newman-Keuls tests revealed that the difference between Same and Different outcomes was significant when the second sample was probed as well as when the first sample was probed (p < .05). The three-way interaction in Table 2 appears to arise from the fact that performance dropped only slightly with delay on Same or Different trials when the second stimulus was probed but dropped considerably on Same trials and actually improved with delay on Different trials when the first stimulus was probed. Discussion Performance on the second of two samples was influenced by whether or not its outcome matched that of the first sample, which indicates that the interactions between sample stimuli produced both proactive and retroactive effects. The interaction of the sample stimuli, however, was asymmetrical, with retroactive effects being substantially larger than the proactive effects. The idea that the second stimulus acts as a retrieval cue to access and provide an opportunity for further rehearsal of the first stimulus is generally compatible with these results. To handle the asymmetrical interference effects, one could offer the assumption that when memory for the initial sample stimulus is reactivated, the outcome associated with it may also be reactivated and that this outcome has some (small) probability of becoming associated with the second sample. Another possibility is that these modest proactive effects might arise on trials when animals remember that the two sample stimuli were associated with different outcomes but have forgotten which outcome was associated with which stimulus. The results of Experiment 3 imply that such

Table 2 Proportion Correct for the Two Probe Types and the Same Versus Contrasting Sample Outcomes First sample probed Delay Short Long Same
.86 .77

Second sample probed Same

.82 .81

Different
.55 .63

Different
.79 .75

an explanation might need to take into account the similarity of the initial stimulus to the second (or interpolated) stimulus. General Discussion These experiments show a consistent, clear pattern of interactions between sample and interpolated stimuli. As might be expected, when the sample and interpolated stimulus are both rewarded or both not rewarded, facilitation is observed if both stimuli have the same value along the relevant dimension (Experiment 1, Conditions A-A and A-C versus Control). Less obvious is the finding that this facilitation is greater if the sample and interpolated stimulus match on a constant, irrelevant dimension as well (Experiment 1, Condition A-A versus A-C). Perhaps most surprising of all is the finding that retroactive facilitation and retroactive interference (if the sample and interpolated stimulus are associated with contrasting outcomes) are observed when two stimuli share only a constant, irrelevant dimension and differ along the relevant dimension (Experiments 2 and 3). The present results are inconsistent with the idea that when two stimuli are presented consecutively as samples, they are encoded independently and that any patterns of interference or facilitation can be described in terms of a simple, additive function of generalization along component stimulus dimensions. The first experiment provided evidence that component stimulus dimensions cannot be treated as additive, a result consistent with a large body of results in the domain of discrimination learning (e.g.,

124

D. MEDIN, T. REYNOLDS, AND J. PARKINSON

Spiker, 1963, 1970; Meclin, 1975, 1976). Experiments 2 and 3 showed that the difference between performance when the sample and interpolated stimulus have the same outcome and performance when they have contrasting outcomes is greater when the two stimuli match each other on a constant irrelevant dimension than when they do not. Taken as a whole, the results of these experiments suggest that the underlying mechanisms producing nonindependence of stimulus events are closely tied to retrieval and rehearsal processes. The idea that an interpolated stimulus acts as a retrieval cue that may lead to further processing of the initial stimulus is consistent with the results of Experiment 4 which showed that whether or not the reward status of the first and second stimulus matched influenced performance more on test probes involving the first stimulus than on test probes involving the second stimulus. Although the ties to the discrimination learning literature give some assurance that the present analysis will have some generality, it is worth considering whether the present results have implications for DMTS procedures in which sample presentations are not rewarded and rewards are available only on choice tests (e.g., D'Amato, 1973). The idea that a stimulus may act as a retrieval cue is not restricted to situations in which sample presentations are associated with reinforcement. One might, therefore, expect to observe retroactive facilitation similar to that reported here in other situations. To optimize the likelihood of detecting such an effect, it might be necessary to use more than a few alternative samples in order to reduce between-trial interference. Another possibly relevant consideration is that in our experiments using a WGTA, each stimulus and test presentation ends with retraction of the form board and lowering of the opaque screen which should tend to interfere with short-term memory processes. This "interruption" may set the stage for interpolated stimuli to be more necessary and effective as retrieval cues. If the initial sample presentation was always being actively rehearsed when an interpolated stimulus was presented,

one can imagine that only retroactive interference, rather than facilitation, would be observed. In any event, the present experiments suggest that accounts of animal memory might benefit from reconsideration of the old idea that stimulus similarity plays an important role in interference (and facilitation). The idea that the alternative stimuli appearing on a trial can be treated as independent may hold only for the case in which the stimuli are totally dissimilar. Reference Note
1. Medin, D. L. Stimulus similarity and interactions in monkey short-term memory. Unpublished manuscript, 1978.

References
D'Amato, M. R. Delayed matching and short-term memory in monkeys. In G. H. Bower (Ed.), The psychology of learning and motivation (Vol. 7). New York: Academic Press, 1973. Davis, R. T., & Fitts, S. S. Memory and coding processes in discrimination learning. In D. L. Medin, W. A. Roberts, & R. T. Davis (Eds.), Processes of animal memory. Hillsdale, N.J.: Erlbaum, 1976. Flagg, S. F., & Medin, D. L. Consant irrelevant cues and stimulus generalization in monkeys. Journal of Comparative and Physiological Psychology, 1973, 5, 339-345. Grant, D. S., & Roberts, W. A. Sources of retroactive inhibition in pigeon short-term memory. Journal of Experimental Psychology: Animal Behavior Processes. 1976, 2, 1-16. Honig, W. K., & James, P. H. R. (Eds.). Animal memory. New York; Academic Press, 1971. McGeogh, J. A. Forgetting and the law of disuse. Psychological Review, 1932, 39, 352-370. McGeogh, J. A. The psychology of human learning. New York: Longmans, Green, 1952. Medin, D. L. A theory of context in discrimination learning. In G. H. Bower (Ed.), The psychology of learning and motivation (Vol. 9). New York: Academic Press, 1975. Medin, D. L. Animal models and memory models. In D. L. Medin, W. A. Roberts, & R. T. Davis (Eds.), Processes of animal memory. Hillsdale, N. J.: Erlbaum, 1976. Medin, D. L., Roberts, W. A., & Davis, R. T. (Eds.). Processes of animal memory. Hillsdale, N. J.: Erlbaum, 1976. Medin, D. L., & Schaffer, M. M. Context theory of classification learning. Psychological Revieiv, 1978, 85, 207-238. Melton, A. W., & Irwin, J. M. The influence of degree of interpolated learning on retroactive

STIMULUS SIMILARITY AND STM inhibition and overt transfer of specific responses. American Journal of Psychology, 1940, 53, 173203. Motiff, J. P., Dekock, A. R., & Davis, R. T. Concealment of stimuli during delay in the delayedresponse problem. Perceptual and Motor Skills, 1969, 29, 788-790. Osgood, C. E. Method and theory in experimental psychology. New York: Oxford University Press, 1953. Reynolds, T. J., & Medin, D. L. Strength versus temporal order information in delayed matching to sample performance by monkeys. Animal Learning and Behavior, 1979, 7, 294-300. Roberts, W. A., & Grant, D. S. An analysis of light-induced retroactive inhibition in pigeon short-term memory. Journal of Experimental Psychology: Animal Behavior Processes, 1978, 4, 219-236. Spear, N. E. Retrieval of memory in animals. Psychological Review, 1973, 80, 163-194. Spear, N. E. Retrieval of memories. In W. K. Estes (Ed.), Handbook of learning and cognitive processes (Vol. 4). Hillsdale, N J.: Erlbaum, 1976. Spence, K. W. The nature of discrimination learning in animals. Psychological Review, 1936, 43, 427-449. Spence, K. W. Behavior theory and conditioning, New Haven, Conn.: Yale University Press, 1956.

125

Spiker, C. C. The hypothesis of stimulus interaction and an explanation of stimulus compounding. In L. P. Lipsitt & C. C. Spiker (Eds.), Advances in child development and behavior (Vol. 1). New York: Academic Press, 1963. Spiker, C. C. An extension of Hull-Spence discrimination learning theory. Psychological Review, 1970, 77, 496-515. Underwood, B. J. Retroactive and proactive inhibition after five and forty-eight hours. Journal of Experimental Psychology, 1948, 38, 29-38. Underwood, B. J. Interference and forgetting. Psychological Review, 1957, 64, 49-60. Wagner, A. R., Rudy, J. W., & Whitlow, J. W. Rehearsal in animal conditioning. Journal of Experimental Psychology, 1973, 97, 407-426. Wickelgren, W. A. Acoustic similarity and retroactive interference in short-term memory. Journal of Verbal Learning and Verbal Behavior, 1965, 4, 53-61. Wickelgren, W. A. Phonemic similarity and interference in short-term memory for single letters. Journal of Experimental Psychology, 1966, 71, 396-404. Winer, B. J. Statistical principles in experimental design. New York: McGraw-Hill, 1962. Received March 12, 1979 Revision received August 31. 1979