Animal Learning & Behavior1982,10 (1),1-6

Retrieval of memory in the pigeonby context manipulationsDAVID R. THOMAS and ALAN R. McKELVIE

University ofColorado, Boulder, Colorado 80309In Experiment I, 12 pigeons were given eight sessions of VI single stimulus training with a

color in a particular context followed by eight sessions of similar training with a line angle inanother context. On the next day, half of the subjects were tested for wavelength and angularitygeneralization in each of the two contexts, a procedure that was thus consistent with trainingfor one dimension and inconsistent for the other. The subjects made significantly more responsesto each training stimulus under the consistent context condition, but there was no differencein absolute or relative generalization slopes. In Experiment 2, 12 pigeons were trained as inExperiment I, but during generalization testing they were exposed to both contexts sequen-tially. Under the consistent context condition, the subjects responded more to the two trainingstimuli and yielded sharper absolute and relative wavelength generalization gradients: Underthe inconsistent context condition, responding to the training wavelength was substantiallydisrupted. Thus, under appropriate testing conditions, contextual control over both the amountand the selectivity of responding can be demonstrated.

Spear (1971, 1978) has proposed that success atretrieval of a memory is a function of the extent towhich the subject notices, during retention testing,ambient contextual stimuli that were present but in-consequential to the target learning task and werestored as attributes of the target memory. The role ofcontextual retrieval cues in human memory has fre-quently been demonstrated. Rand and Wapner (1967)reported that the retention of a list of nonsense syl-lables was better when the subject was tested in thesame postural position as during original learning(i.e., standing up or lying down). Greenspoon andRanyard (1957) found that the color and locationof the experimental room functioned as retrieval cuesin a similar manner.

A typical example from the animal memory liter-ature is a proactive interference (PI) experiment reoported by Spear (1971). Rats trained on a passiveavoidance task followed by an active avoidance tasktypically showed recency in a delayed test, behavingin accordance with the active avoidance contingency.If a tone had been present during passive avoidancetraining and was present in the subsequent retentiontest, however, then behavior more appropriate to thepassive avoidance task was exhibited in the retentiontest.

In our laboratory, we (Thomas, McKelvie, Ranney,& Moye, 1981) recently completed a study concep-tually similar to the Spear (1971) experiment. Pigeons

Requests for reprints should be addressed to David R. Thomas,Department of Psychology, Campus Box 345, University ofColorado at Boulder, Boulder, Colorado 80309. This researchwas supported by NSF Research Grant BNS-7801407. It was re-ported by Alan McKelvieat the April 1981 meetings of the RockyMountain Psychological Association in Denver.

Copyright 1982 Psychonomic Society, Inc.

first learned a successive wavelength discrimination,538 om S+ (reinforced)vs. 576 om S- (extinguished),followed by a reversal of this discrimination (i.e.,576 nm S+, 538 nm S-). For a control group, ageneralization test administered in extinction a daylater revealed recency (i.e., the gradients all peakedat the Problem 2 S+. For one experimental group,however, as in Spear's (1971) experiment, Problem 1had been learned in a distinctive context (Context 1,either houselight + white noise or no houselight +1,(XX)-Hz tone), and generalization testing was carriedout in this context. The result was substantial PI; nogradients showed recency. Some peaked at the Prob-lem 1 S+, but most showed peaks at both S+ valuesor a peak intermediate betweenthem. Thus, Context 1was a sufficiently effective retrieval cue to preventretrieval of Problem 2 but not, in most cases, topermit successful retrieval of Problem 1. A secondexperiment, in which subjects were tested alternatelyin the contexts in which the two problems had beenlearned, demonstrated that successful retrieval of thememory of both problems was possible. Under thistest condition, each subject yielded a gradient appro-priate to the problem learned in each context (i.e.,the gradient peaked at 538 nm in Context 1 and at576 nm in Context 2).

The degree of control over behavior by context inThomas et a1. 's (1981) Experiment 2 was quite strik-ing, resembling that which would have been expectedhad the subjects been trained on an explicit condi-tional discrimination with the context as the super-ordinate cue. Note, however, that such conditionaldiscriminations are usually difficult to establish inpigeons when traditional methods involving repeatedalternations of the conditional (superordinate) stimuli

00904996/82/010001-06$00.85/0

2 THOMAS AND McKELVIE

are employed (cf. Richards, 1979). In the Thomaset al. experiment, there was only one alternation intraining between the two contexts; furthermore, theassociation between Context 2 and Problem 2 wasformed in a single training session. It is of interestto determine which aspect of Thomas et aI.'s pro-cedure accounted for the success of the retrievalprocess in their experiment. One possible factor wastheir use of a reversal paradigm. Perhaps the "sur-prise" (Kamin, 1969) of the reversal manipulationcalled attention to the change in contextual condi-tions, thereby increasing their significance. It is alsopossible that the use of discrimination training inthe Thomas et al. study was a critical factor. Thomas(1970) has argued that discrimination training enhancesattention to cues, such as the context, that are incon-sequential to the discrimination problem itself.

In the Thomas et al, (1981) study, the success ofretrieval was reflected primarily in the obtaining ofa generalization gradient with a peak of respondingat the appropriate S+ value. In the present experi-ments, single stimulus training rather than discrim-ination training was used. Thomas and Lopez (1962)showed that relative generalization gradients obtained24 h after the completion of single stimulus trainingwere reliably flatter than those obtained in an im-mediate (i.e., l-min-delayed) test. They attributed thisflattening to some forgetting of the value of the train-ing stimulus. Interestingly, there was no reliable re-duction in number of responses to the training stim-ulus over the 24-h period in their experiment. Thisfinding may merely indicate that generalization slopeis a more sensitiveindex of forgetting than is absoluteresponse strength to the training stimulus. However,Newlin and Thomas (in press) haveargued that the twomeasures reflect different processes (i.e., those ofdimensional stimulus control and excitatory stimuluscontrol, respectively). As evidence for this position,Newlin and Thomas presented data from severalexperiments indicating that the two measures mayvary in opposite directions with manipulations of thesame independent variable (e.g., brief vs. continuousstimulus exposures). The present experiments enabledus to study the way in which measures of excitatoryand dimensional stimulus control are affected by thesame retrieval cue manipulations. Since forgettingfollowing single stimulus training is reflected in aflattening of the generalization gradient and (in somecases) in a reduction of responding to the trainingstimulus, successful retrieval of the memory of train-ing should be reflected in changes in the oppositedirection (i.e., a sharpening of generalization and anincrease in responding to the training stimulus).

There are several reasons why relative general-ization slope has become the typical measure of di-mensional stimulus control. The transformation ofeach subject's gradient into a percentage of total re-sponses equates the contribution of each subject to

the group average. It also eliminates the large indi-vidual differences in levelof responding, which oftenmake statistical significance of differences in gener-alization slope difficult to demonstrate. It is generallythe case that absolute and relative gradients revealparallel effects of a given experimental manipulation.For example, both are sharpened by discriminationtraining and flattened by nondifferential training(cf. Thomas, 1970). This parallel is not always pres-ent, however. Thomas and King (1959) reported adeprivation level study in which, within each depri-vation condition, the high-responding animals gaverelative gradients that were indistinguishable fromthose of the low responders. Their absolute gradientswere, of course, substantially sharper. Given the pos-sibility that the context manipulation might affectabsolute and relative gradients differently, both ab-solute and relative gradients will be reported anddiscussed in this paper.

EXPERIMENT 1

In Experiment 1, pigeons were given single stim-ulus training to peck at a green keylight for variableinterval reinforcement in a particular context. Theythen received singlestimulus training with a horizontalline stimulus in a different context. For parposes ofretention testing, the birds were divided into twogroups, one tested in Context 1 and one tested inContext 2. The generalization tests were conducted,in extinction, 1 day after the completion of training,and each test included stimuli from both wavelengthand line-orientation dimensions. Thus, for each birdthe test context was consistent with one of the twotraining problems and inconsistent with the other.The experimental design was within subjects withregard to the assessment of memory for the two train-ing problems, but between subjects with regard tothe use of a single test context with each bird. Testingwith both dimensions provides a control for any non-associative effectsof the contextualcues(e.g., masking)because response strength should be greater and thegeneralization gradient should be sharper only whentesting is carried out in the consistent context con-dition. If any particular set of contextual stimuli weremore effective in retrieving the memory of its asso-ciated training or, on the other hand, were more ef-fective in masking the expression of stimulus control,such effects would be revealed by the counterbal-ancing of contextual stimuli across conditions, asdescribed below. Presumably, what is critical is therelationship between each context and its associatedproblem, rather than the context per se.

MethodSubjects. The subjects were 12 experimentally naive pigeons

obtained from a local supplier and maintained at 7S07o of theirad-lib weights. They were housed in individual cages in a colony

RETRIEVAL OF MEMORY 3

Figure 1. Absolute and relative generalIzation gradients obtainedin Experiment 1, averaged across three blocks of testing (see textfor details).

across contextual stimuli (HLT and HLN) since thesehad no effect. Absolute and relative generalizationgradients were calculated for each block and thenaveraged to minimize the effects of extinction on therelative gradients. The absolute gradients in the toppanels of Figure 1 suggest that the context condition,consistent vs. inconsistent, had an effect. For bothdimensions, the gradient was higher and sharper underthe consistent context condition.

Let us consider first the measure of excitatory stim-ulus control (i.e., the number of responses to thetraining stimuli). The mean number of responses to555 nm was 59 in Context I (consistent) vs. 26 inContext 2; for the O-deg line, the corresponding valueswere 55 in Context I (inconsistent) vs. 72 in Context 2.A mixed-design analysis of variance (group x testdimension) performed on the number of responses tothe S+ indicated that this difference, in favor of theconsistent context condition, was statistically signif-icant[F(I,lO) =5.12, p < .05].

The upper panels of Figure I suggest that dimen-sional stimulus control (i.e., the slope of the absolutegeneralization gradients) also differed with respect tocontext. However, analysis of variance disconfirmedthis result. Separate mixed-design ANOVAs (group

...-. CONSISTENTCONTEXT

0---0 INCONSISTENTCONTEXT

o 15 30 45 60s+LINE ANGLE (deg.)

490 538 555 576 606S+

WAVELENGTH (nm)

f/)wf/)z00-sn 30w0::I- 20zwU0::W 100-Z

4 THOMAS AND McKELVIE

Figure 1. Absolute and relative generalization gradients ob-tained In Experiment 1, averaged across three blocks of testing (seetext for detaUs).

textual stimuli (HLT and HLN) since these had noeffect. Absolute and relative generalization gradientswere calculated for each block and averaged acrosstest blocks. As in Experiment I, test context had asignificant effect on level of responding to the train-ing stimuli. The mean number of responses to 555 omwas 47 in Context 1 vs, 14 in Context 2; for the O-degline, the corresponding values were 34 in Context 1vs. 50 in Context 2. A two-way repeated measuresANOVA (context x test dimension) performed onthe responses to S+ indicated that significantly moreresponding occurred to the training stimulus in theconsistent than in the inconsistent condition [contextmain effect: F(l,ll) =40.2, p < .01]. It is clear fromFigure 2 that the use of a consistent context resultedin both higher and sharper absolute generalizationgradients. In contrast with the results of Experiment 1,the two-way repeated measure ANOVAs (context xstimuli) performed separately for each dimensionshowed that the absolute gradient slopes of the con-sistent vs. inconsistent conditions were significantlydifferent [Fs(4,44)=12.9 and 4.4 for wavelength andline angle, respectively; ps < .01]. Indeed, respondingto the wavelengths under the inconsistent contextcondition was severely disrupted; it was low, spo-radic, and largely unsystematic.

...-. CONSISTENTCONTEXT

0---0 INCONSISTENTCONTEXT

490 538 555 576 606 0 15 30 45 60s+ s+

WAVELENGTH (nm) LINE ANGLE (deg.l

60

50enw

~ 40oa..en 30wa::2 20~w:!:

10

enwen240oa..enw 30a::I-

~ 20oa::

~ 102

~w:!:

EXPERIMENT 2

MethodSubjects. The subjects were 12 experimentally naive pigeons

maintained as in Experiment 1.Apparatus. The apparatus was the same as that in Experiment 1.Procedure. The training procedure was the same as that in Ex-

periment 1. Again, for half of the subjects Context 1 (associatedwith wavelength) was HL T, and for half it was HLN. The con-ditions for Context 2 were similarly counterbalanced.

Generalization testing in extinction was carried out 24 h afterthe last training session. Test stimuli were presented for 30 sec,separated by S-sec blackouts. For all subjects, testing began witha block of 10 wavelengths, followed by a block of 10 line angles,followed by a block of wavelengths, and so on. The context waschanged after every two test blocks, and the test started with Con-text 1 for half of the subjects and with Context 2 for the otherhalf. Testing continued until each subject experienced three blocksfrom each dimension in each context condition (i.e., a total of12blocks).

x stimulus) performed on the responses obtainedduring the wavelength and line angle tests yielded nogroup X stimulus interactions [Fs(4,40)=1.57 and1.56 for wavelength and line angle, respectively; ps=.2]. As is typical in free-operant experiments, indi-vidual variation in response levels under each con-dition was so great that the differences apparent inthe figure did not achieve statistical significance.

To the extent that there is some difference, theresults suggest that the consistent context conditionacts as a constant multiplier of responding to all stim-ulus values. This relationship is made clear in thelower panels of Figure 1, in which it can be seen thatfor each dimension the relative generalization gra-dients obtained under consistent and inconsistentcontext conditions are virtually superimposed. Clearly,the results indicate that Context 1 was associatedwith the wavelength problem and Context 2 was as-sociated with the line angle problem. The memory oftraining was retrieved more effectively in the con-sistent than in the inconsistent context, but the effectsare seen only in the excitatory measure (i.e. absoluteresponse level).

In the Thomas et al. (1981) study, it was shownthat the effectiveness of the context as a retrieval cuecould be substantially enhanced by the use of a gener-alization test procedure in which both contexts werepresented sequentially. Experiment 2 was designedto determine whether a similar result would be ob-tained in the experimental paradigm used here.

Results and DiscussionFigure 2 presents the mean generalization gradients

obtained in this experiment. The curves in the upperpanels are the absolute generalization gradients; therelative generalizationgradients are in the lower panels.The wavelength generalization gradients appear inthe left panels, while the angularity gradients are inthe right panels. The data were pooled across con-

The lower panels of Figure 2 indicate that the ef-fectiveness of context as a retrieval cue was greaterfor the colors than for the lines. Again, the effect ofcontext on angularity gradients was multiplicative, sothe relative gradients are superimposed. In the case ofwavelength, however, there was a substantial differencein the slope of the relative generalization gradients.A two-way repeated measures ANOVA (context xstimuli) indicated that the gradient obtained under theconsistent context condition was significantly sharperthan that obtained under the inconsistent context con-dition [context X stimuli interaction: F(4,44)=6.9,p < .00IJ.

In general, the effect of the context condition wasgreater in Experiment 2, in which subjects were ex-posed to both contexts during a test session, than ithad been in Experiment 1, in which testing was carriedout under a single context condition. Furthermore,in Experiment 2 the effect of the context was muchgreater with the wavelength than with the line anglestimuli. There are several possible reasons for this.No attempt had been made to equate the two sets oftest stimuli for discriminability, and the gradientsalong the angularity dimension were rather flat underall conditions. Thus, there was little room to showfurther flattening. Furthermore, at the time of thetest, the subjects had seen the O-deg line on the pre-ceding day but had not seen the 555-nm wavelengthvalue for 9 days. It seems reasonable that remindercues would be more effective under the latter con-dition.'

Again, as in Experiment I, the measures of excit-atory and dimensional stimulus control were in agree-ment in indicating that retrieval of the memory fortraining was more successful under the consistentcontext condition. This time, however, both the re-sponse strength and the absolute generalization slopedifferences were significant.

The finding that retention performance is betterin the consistent context than in the inconsistent con-text can be interpreted in several different ways. Theconsistent context may facilitate retrieval, the incon-sistent context may interfere with retrieval, or botheffects may occur. To determine which of these pos-sibilities occurred, a "neutral" context is requiredfor comparison purposes, but the definition of whatconstitutes a neutral context presents a logically for-midable problem. A novel context is unsuitable be-cause it leads to a disruption of operant keypecking,thus providing no measure of dimensional stimuluscontrol. An alternative definition of a neutral contextis that it is a context that is demonstrated to be in-effective. Thus, with regard to the measure of relativegeneralization slope, both contexts in Experiment 1meet this definition. The best estimate of the measureof relative generalization slope under a neutral con-text condition is obtained by pooling the measuresobtained under the two (ineffective) context condi-

RETRlEVALOF MEMORY 5

tions of Experiment 1. For wavelength, the resultantmean is 33% of total responses to 555 nm. This valuefalls between the values of 43070 for the consistentcontext condition and 26% for the inconsistentcontextcondition in Experiment 2. Although it is admittedlyspeculative, we would suggest, based upon this com-parison, that under the conditions of Experiment 2,Context 1 facilitated the retrieval of the memory ofthe training wavelength, whereas Context 2 inter-fered with the retrieval of this memory.'

The effect of the consistent context in facilitatingthe retrieval of memory was anticipated, based on theresults of our previous research. We did not expect,however, the substantial interference we observed inthe inconsistent context condition. Many of the indi-vidual subjects' gradients obtained under this conditionwere essentially flat or peaked at a nontraining value,suggesting an inhibitory process. Note that, unlikethe Thomas et al. (1981) study, in which a reversalwas used, the two "competing" memories in thisstudy were orthogonal (i.e., peck at a color, and peckat a line). Thus, there is no clear-cut basis for inter-ference between them. Why the presence during test-ing of a retrieval cue for one of the memories shouldso completely disrupt expression of the memory asso-ciated with the alternative cue (under the condition inwhich both cues are experienced within the session) isunclear at present. Perhaps a useful way to concep-tualize the effect is that the inconsistent context ren-ders the memory of the target problem inaccessible inits presence (cf. Tulving & Pearlstone, 1966, for adiscussion of the distinction between availability andaccessibility in human verbal memory).

It is important to note that a critical determinantof the effectiveness of context as a retrieval cue is theprocedure employed-specifically, whether the re-trieval cue for one memory or the cues for both (po-tentially conflicting) memories are used within a ses-sion. Given that the subject has the opportunity tocompare different retrieval cues during testing, neithera reversal paradigm nor the use of discriminationtraining is required to produce evidence of very sub-stantial control over behavior by ambient contextualstimuli used as retrieval cues.

It may be appropriate in closing to raise the questionof whether the results of these experiments requirethe postulation of a construct of memory for theirinterpretation. Taken alone, the resultsof Experiment 1do not. Butter (1963) trained pigeons to peck at avertical line of a particular color and then varied boththe angle and the color in generalization testing. Inthe presence of a nonvertical line, the wavelengthgradient was multiplicatively flattened, a findingparallel to that produced by the inconsistent contextin Experiment 1. The differences between the resultsof our two experiments, however, seem to require theconstructs of memory and its retrieval. Note thattraining was identical in the two experiments; thus,

6 THOMAS AND McKELVIE

differences in the success with which subjects mani-fested their learning in test performance lend them-selves to an interpretation postulating a retrievalprocess.

REFERENCES

BUTIER, C. M. Stimulus generalization along one and two dimen-sions in pigeons. Journal of Experimental Psychology, 1963,65, 339-346.

GREENSPOON, J., & RANYARD, R. Stimulus conditions and retro-active inhibition. Journal of Experimental Psychology, 1957,53,55-59.

KAMIN, L. J. Predictability, surprise, attention and conditioning.In B. Campbell & R. Church (Eds.), Punishment and aversivebehavior. New York: Appleton-Century-Crofts, 1969.

NEWLIN, R. J., & THOMAS, D. R. On the acquisition and measure-ment of stimulus control in pigeons. Animal Learning & Be-havior, 1982, in press.

RAND, G., & WAPNER, S. Postural status as a factor in memory.Journal of Verbal Learning and Verbal Behavior, 1967, 6,268-271.

RICHARDS, R. W. Stimulus control following training on a condi-tional discrimination. Animal Learning & Behavior, 1979, 17,309-312.

SPEAR, N. E. Forgetting as retrieval failure. In W. K. Honig &P. H. R. James (Eds.), Animal memory. New York: AcademicPress, 1971.

SPEAR, N. E. The processing of memories: Forgetting and reten-tion. Hiilsdale, N.J: Erlbaum, 1978.

THOMAS, D. R. Stimulus selection, attention, and related matters.In J. H. Reynierse (Ed.), Current issues in animal learning.Lincoln: University of Nebraska Press, 1970.

THOMAS, D. R., & KING, R. A. Stimulus generalization as afunction of the level of motivation. Journal of ExperimentalPsychology, 1959.57,323-328.

THOMAS, D. R. & LOPEZ,L. J. The effect of delayed testing ongeneralization slope. Journal ofComparative and PhysiologicalPsychology, 1962,44,541-544.

THOMAS, D. R., McKELVIE, A. R., RANNEY, M .. & MOYE,T. B.Interference in pigeons' long-term memory viewed as a retrievalproblem. Animal Learning & Behavior, 1981,9,581-586.

TULVING, E., & PEARLSTONE, Z. Availability vs. accessibility ofinformation in memory for words. Journal of Verbal Learningand VerbalBehavior, 1966,5,381-391.

NOTES

1. Note that Thomas and Lopez (1962) had found that gener-alization gradients obtained after a I-week delay were no flatterthan those obtained after a l-day delay. This may reflect a flooreffect in the generalization measure, rather than the absence of anyfurther forgetting. The "retrievability" of a memory may providean alternative index of the extent to which it had been forgottenin the first place.

2. Note that all subjects in both experiments were trained by thesame experimenter, in the same apparatus, under identical pro-cedures, and during the same time period. In some sense, then, thecross-experiment comparison may be viewed as a comparisonbetween groups in the same experiment. The groups differed onlyin test length and test procedure, which is precisely the variable ofmajor importance in this study.

(Manuscript received July 28,1981;revision accepted for publication October 27, 1981.)