Animal Learning & Behavior1975, 1'01.3 (3),212-216

An analysis of positive general transferin discrimination learning in the rat

GEOFFREY HALLUniversity of Sussex, Brighton, England

As a result of learning a single discrimination problem, rats and pigeons will show positive transfer tototally new discriminations. It has been suggested that this transfer effect reflects the deveiopment of aprocess of general attentiveness. However, alternative explanations of the transfer effect are availablewhich do not require any radical departure from traditional discrimination learning theory. The threeexperiments reported here demonstrate positive general transfer in rats trained on a simultaneousdiscrimination and tested under conditions which rule out these alternative explanations, Theexperiments also attempt to specify the nature of the processes underlying the general transfer effect.

When animals solve a discrimination problem, theymay learn more than just those habits which enablethem to show a preference for one stimulus overanother. With primate subjects, general transfer effectshave been amply demonstrated in experiments onlearning set formation. More recently it has beensuggested that the equivalent of a learning set may beformed by rats and by pigeons as a result of theirlearning a single simple discrimination. For example,Honig (1969) has shown that pigeons which are firsttrained on a successive free-operant discriminationbetween two stimuli and are then trained to respond to anew and different stimulus show strong control by thisstimulus in a final generalization test. Control subjectsgiven pseudodiscrimination (PD) training in the firststage (training in which the periods of reward andnonreward are not correlated with the training stimuli),or given single-stimulus training, show only weakstimulus control. Thomas, Freeman, Svinicki, Burr, andLyons (1970) have replicated and extended this findingand interpret their results as evidence for a mechanismof "general attentiveness." Discrimination training, byraising an animal's level of general attentiveness, is heldto enhance the readiness with which the animal willattend to and learn about all other stimuli,

Support for this interpretation has been sought inexperiments which look at the effects of discriminationtraining with one set of stimuli upon the learning of asecond problem with totally new stimuli (anextra-dimensional shift), It has been found (Eck, Noel, &Thomas, 1969; Keilitz & Frieman, 1970; Frieman &Goyette, 1973) that birds given initial truediscrimination (TD) training learn the second problemmore rapidly than birds given initial PD training or birdsgiven pretraining with just a single stimulus. Thomas,Miller, & Svinicki (1971) have found this effect in ratstrained initially with bright and dim houselights andthen shifted to a discrimination between tones.

Work supported by a grant from the United Kingdom ScienceResearch Council.

The hypothesis to be examined in the experimentsreported below is that animals given true discriminationtraining are capable of learning that stimulus differencescan serve as predictors of reinforcement and that theywill transfer their learning to discriminations involvingnew stimuli. Although the results just cited areconsonant with this hypothesis, they cannot supplyconclusive proof. There are two major difficulties. First,it may be possible to explain the findings withoutrecourse to any concept which requires the animals totransfer their learning from one set of stimuli to acompletely different set. It is known, for instance, thatdiscrimination training, in contrast to the various controltreatments, reduces the control exerted by irrelevantstimuli present in the training situation (Wagner, Logan,Haberlandt, & Price, 1968). These stimuli, which willstill be present in the test situation, might otherwise actto mask the presence of control by stimuli that theexperimenter is manipulating (Turner & Mackintosh,1972) or might interfere with new learning. It seemsparticularly likely that the differential neutralization ofirrelevant cues produced by TD and PD trainingaccounts, at least in part, for differences found in theextradimensional shift.

Secondly, even if we accept that these results mightimply a general transfer effect of the type observed inthe formation of learning sets, it is by no means clearthat the effect is a positive or helpful one. Mackintosh(1973) has argued in favor of a process of "learnedirrelevance"; he has suggested that animals may learnthat stimuli are not correlated with reinforcement andthat such learning will be a hindrance should the animalssubsequently be required to learn that the stimuli arecorrelated with reward. It is possible that animals givenPD or single-stimulus training might learn the irrelevanceof the stimuli and that this learning might transfer tonew stimuli. (There is certainly experimental evidence toshow that rats given PD training learn anextradimensional shift less rapidly than untreatedcontrols; Mandler, 1966; Bainbridge, 1973). It follows

212

TRANSFER IN DISCRIMINATION LEARNING IN RAT 213

that a difference between a group of animals given TDtraining and its control group might be the result of aretardation of learning in the latter rather than afacilitation in the former. It is difficult to think of acontrol treatment that is not open to this criticism. Theuse of an untreated control is ruled out since anyadvantage shown by a trained group over an untrainedcontrol could be put down to the extra habituation tothe experimental situation that the former had received.

The present experiments are designed to overcomethese difficulties and to demonstrate unambiguously apositive general transfer effect. They do not try todisconfirm the theories outlined above; it is acceptedthat positive and negative general transfer effects are apossibility and that the neutralization of irrelevant cuesby TD training is a reality. Rather, it is argued that theexistence of these various effects renders inconclusivethose experiments which have previously been taken asevidence for positive general transfer. As a first step,therefore, we must arrange a situation in which adifference between groups on an extradimensional shiftcannot be explained away as due to a retardation in thecontrol group. The experiments reported here do this bycomparing the shift performance of an experimentalgroup given prior discrimination training with that of anuntrained control group. Of course, since theneutralization of irrelevant cues will put theexperimental group at an advantage in the test phase, asuperiority of the experimental over the control groupcould be solely due to this effect and would not provethe hypothesis to be examined in Experiment I-thehypothesis that discrimination training produces a"learned relevance" effect which transfers from one setof stimuli to another. Accordingly, the strategy of theexperiment is to provide as a test situation not a simpleextradimensional shift but one in which learnedrelevance will result in negative transfer. If it is foundthat the experimental group is retarded with respect tothe control group in solving the test problem (in spite ofthe various advantages that discrimination training mightbe expected to bestow), evidence for a strong generaltransfer effect would be provided.

EXPERIMENT I

The experiment consisted of two stages, initial train­ing followed by a test problem. This latter was aprobability learning task with black and white stimuli,one stimulus being the positive on 70% of the trials andthe negative on the remaining 30% of the trials. Whenrats are trained by a noncorrection procedure on such aproblem, they come to choose the "majority stimulus"on all trials (i.e., they "maximize"). One group ofsubjects (the control group) was given no training (apartfrom basic pretraining) before being presented with theprobability problem, the experimental group was givenprior training on a consistently reinforced orientation

discrimination, and the performance of the two groupswas compared on the probability problem. The secondgroup should benefit from any advantage that derivesfrom its extra experience of handling, of the trainingsituation, and so on, and it will also benefit from anysuppression of irrelevant cues that might occur. If thesesubjects learn, in addition, that tl: sre are stimuli whichreliably predict the occurrence of reward, such learningmight be expected to deter them from adopting amaximizing strategy; a strategy which requires them tocontinue to respond to one stimulus despite the fact thatthey are sometimes not rewarded for doing so. Thisexpectation is derived from Goodnow's (1955)observation that the failure of human subjects tomaximize on a probability learning task stems, in part,from their belief that 100% success is possible. In anexperiment with rats, Solomon (1962) has providedsome evidence in favor of this notion by showing thatrats raised in a constant early environment are less likelyto maximize than rats raised in an environment thatvaries in many respects from day to day.

MethodSubjects. The subjects were 16 male hooded rats. They were

maintained on a schedule of food deprivation, food beingavailable for only the 2 h each day that followed experimentaltreatment.

Apparatus. The apparatus was a jumping stand. It consisted ofa goaIbox with two adjacent apertures and with small landingplatforms 15 cm wide and 8 ern deep fixed in front of eachopening. A vertical partition 12 cm deep prevented animals fromstepping from one landing platform to the other. The ratsjumped from a stand shaped like a small elevated Y maze; itsoverall length was 20 em and its arms, 9 em long and 7 em wide,directly faced the goalbox apertures. The goalbox doors, 15 emsquare, were themselves the stimulus objects. In pretraining, twoidentical gray doors were used. During training on theorientation discrimination, the doors bore alternating black andwhite stripes, 1 cm wide, which ran horizontally (Stimuli H) orvertically (V). For the probability problem, the doors were black(B) and white (W).

Procedure. The animals were given 55 pre training trials spreadover 9 days. During the first 15 trials, a single gray goalbox doorwas locked in place, randomly on the left and on the right. Thesubjects could therefore cross from the stand and enter thegoalbox through one of the apertures and could eat from thebaited foodcup. On the remaining 40 pretraining trials, two graydoors were in place, one of them unlocked. The subjects learnedto push down the unlocked door and to enter the goalbox. Thegap between the stand and the goalbox, which on the 1st day ofpretraining was 2 em, was gradually increased, reaching 17 ern bythe end of pretraining. This gap was sufficiently wide that theanimals had to jump across it. If an animal jumped to the lockeddoor, it was returned to the stand and allowed to jump again andthus, during pretraining, each animal was rewarded each time itwas put into the apparatus.

Ten trials of discrimination training were given each day.Noncorrection procedure was used; after a correct response theanimal gained access to the goalbox, which contained four 45-mgfood pellets, while after an incorrect response, the animal wasdetained for 10 sec on the landing platform in front of thelocked goalbox door before being returned to its holding cage.The positive, or majority, stimulus appeared equally often on theleft and on the right, its position being determined according toa modified Gellerrnann sequence. The interval between trials was

21.+ HALL

EXPERIMENTS II ANDIII

Table 1Experiment I: Group Mean Scores ~~Training and Transf~

Group

then the transfer of this learning would help them tomaximize on a probability problem. The experimentsreported below provide a test of this new hypothesis.

MethodThe subjects in each experiment were 16 naive male hooded

rats maintained as in Experiment I. The apparatus andpretraining procedures were exactly as in that experiment. InExperiment II, eight subjects were trained initially on the Hagainst V discrimination, half with H positive and half with Vpositive. A correction procedure was used. After an incorrectchoice, the rat was detained for 10 sec on the landing platform,was then replaced on the stand and allowed to jump again. If therat repeated its incorrect choice, it again was put back on thestand and gently pushed toward the correct goalbox door,guidance that was invariably enough to elicit a jump to thecorrect alternative. This training procedure yields two scores,initial or first-jump errors, and total errors made in reachingcriterion. To reach criterion, a subject was required to make nomore than two errors over 2 days with the last day's responses allcorrect. After reaching criterion on the orientation problem, theExperimental group was given 2 days of overtraining and wasthen transferred to the alternation problem with B and Wstimuli. The stimulus that was to be positive on the first trial of

270.00124.00

Control

102.5045.62

141.2560.87

b.penmen tal

TrainingTriabFrrors

TranskrTrialsErrors

These experiments follow the basic design ofExperiment I. In each of them, an experimental group ofsubjects is given discrimination training, and itsperformance on a test problem is compared with that ofan untrained control group. As in Experiment I, the testproblem was designed to be one in which two sets offactors are played off one against the other; with thistest problem, the neutralization of irrelevantcues shouldhelp learning while the transfer of a tendency toapproach one stimulus and to avoid the other shouldproduce negative transfer. Again, in order todemonstrate the general transfer effect in which we areinterested, it is necessary to find an overall negativetransfer effect. The test problem chosen required theanimals to alternate between stimuli, black and whiteeach being rewarded on alternate trials. If animals givendiscrimination training on an orientation problemtransfer a tendency to approach one stimulus of a pairand to avoid the other, then they should perform poorlyon this problem.

Experiments II and III were identical in design anddiffered only in details of the training technique.

about 5 min. Eight animals chosen at random \\·C[C trained onthe probability problem, four subjects with B as the majoritystimulus and four with W. Response to the majority stimulus wasrewarded on seven trials each day, response to the minoritystimulus on the remaining three. Training was continued for 400trials unless the problem was solved e~rlie r. The criterion oflearning was 18 choices of the majority stimulus over 2 davs,with the last 10 responses all being made to the majoritystimulus. The remaining eight subjects were trained initially onthe orientation discrimination; half with H positive and half withV positive. After reaching cirterion on this problem. they weregiven 2 days of overtraining and were then transferred to theprobability problem, half of each subgroup having B as theirmajority stimulus and half having W.

ResultsTable 1 presents trials and errors to criterion for both

stages of the experiment; the scores include performanceon the 2 days over which the criterion was reached.During probability learning, a response to the minoritystimulus was scored as an error. The experimental groupclearly learned the probability problem more readilythan the control group, making fewer errors(Mann-whitney U test, U = 13, P < .05) and needingfewer trials to reach criterion (U = II, P < .03). Twosubjects in the control group failed to learn theprobability problem, but all other subjects maximized.No animal showed any tendency to minimize(consistently to choose the minority stimulus).

DiscussionRats that have learned an orthodox simultaneous

discrimination readily solve a probability problem. Thereis no sign that experience of a consistentstimulus-reinforcer relationship hinders the learningof adisc rimina t ion in which the stimulus-reinforcerrelationship is only probabilistic. There seems to bethree possible accounts of this finding. First is thepossibility that there is no "learned relevance" effect;that there is no source of general transfer indiscrimination training apart from the beneficial effectsproduced by experience of handling and by theneutralization of irrelevant cues. Secondly, it remainspossible that discrimination training does indeed inducein the rats a tendency to seek out consistentstimulus-reinforcer relationships but that the size of thenegative transfer effect produced in this way isinsufficient to overcome the positive effects that weknow to be present. In view of the large differencebetween the experimental and control groups seen onthe test problem, both these possibilities seem unlikely.Therefore, a third worth considering is thatdiscrimination training produces some general transfereffect, other than those already mentioned, which helpsprobability learning. For example, if duringdiscrimination training, the rats learn that, of the twostimuli, one is a better predictor of reward than theother (or, in different terminology, that one stimulus ofthe pair is to be approached and one to be avoided),

TRANSFER IN DISCRIMINATION LEARNING IN RAT 215

Table 2Experiments II and III: Group Mean Scores

in Training and Transfer

Experiment II Experiment III

Experi- Experi-mental Control mental ControlGroup Group Group Group

TrainingInitial Errors 27.00 23.50Total Errors 30.00 24.50

TransferInitial Errors 87.12 75.50 86.75 71.25Total Errors 100.00 84.00 98.62 73.63

each day was chosen at random, and errors made on this trialwere not scored. The eight subjects in the control group weretrained from the outset on the alternation problem. All subjectswere given 20 days of alternation training.

The correction procedure used in Experiment II was anattempt to facilitate learning of the alternation problem, whichis a difficult one. We cannot expect differences between groupsto show up if most animals fail to make any progress towardsolution. In Experiment III, another change was made to speedthe learning of the test problem. The intertrial interval wasreduced from 5 min to 10 sec. In all other respects,Experiment III was an exact replicate of Experiment n.

ResultsThe results of both experiments are given in Table 2.

The direct comparison of the data of these twoexperiments may be open to question because they werecarried out at different times with rats drawn fromdifferent populations, but such a comparison suggeststhat massed training leads to more efficient leaming ofthe oreintation problem than does spaced training. Lesssurprising is the marked superiority of these subjectsover those trained by noncorrection on the sameproblem in Experiment I (Table 1).

On the transfer test, the experimental groupsperformed less well than the control groups. InExperiment II, the groups differed both in their initialerror scores (D = 9, p < .02) and in their total errorscores (D = 11, P <.03); in Experiment III, thesedifferences were again statistically significant (D= 1.5,P < .001, and D = 0, p < .001). Although reliable, thedifferences between groups were small; all subjectsclearly found the alternation problem difficult, and noneshowed signs of achieving consistently accurateperformance.

In order to identify "hypotheses" or consistentresponse patterns, performance over the first 5 days ofthe transfer test was analyzed in more detail. Inparticular, it was hoped that the experimental groupsmight show a greater tendency to respond consistentlyto one stimulus (B or W) than the control groups.Accordingly, the number of days on which each animalmade at least 8 of its 10 response to one stimulus wasscored. The mean scores for both groups inExperiment II was .75; in Experiment III, a difference

was found in the expected direction but was notstatistically significant (D = 29), the experimental grouphaving a mean of 1.37 and the control group of .87. Asimilar analysis of position responding provided littleevidence for the suppression of position habits in theexperimental group. Calculating the number of days onwhich at least 8 of the 10 responses were to one positionyielded mean scores of 1.75 and 2.25 for theexperimental and control groups in Experiment II and1.50 and .87 for these groups in Experiment III. Neitherdifference was statistically reliable (D =32 and D =27).

DiscussionRats trained on an orthodox simultaneous

discrimination between horizontal and vertical stimuliare hindered in learning a problem which requires themto alternate between black and white stimuli. There isthus some transfer effect produced by discriminationtraining which is sufficiently strong that it can overcomethe known beneficial effects of such training and resultin overall negative transfer. Although the transferproduced in this special test situation is negative, thesame process will be active when the transfer test is asimple extradimensional shift, and, in this case, itpresumably acts to help new learning. An effect of thissort, although often postulated, has not previously beenadequately demonstrated since the experimental designsusually used have not been able to exclude the action ofother simple transfer effects.

Can we be more specific about the nature of thepositive general transfer? Traditionally, it has beenreferred to as the result of a "set to discriminate" or ofheightened "general attentiveness" but these labels seemuninformative descriptions rather than explanations.Thus, to say that discrimination training produces a setto discriminate does not help us to explain why thistraining helps the learning of a probability problem butretards performance on alternation. (A similar argumentcan be used against attempts to explain the presentresults in terms of the development of simple observingresponses).

The hypothesis outlined above, that discriminationtraining teaches the animal that of two stimuli one is tobe preferred over the other, has the virtue of explainingthe facilitation of probability learning and theretardation seen on alternation. But it creates twoproblems of its own. First, a number of alternativestatements of the hypothesis are available. We could say,for example, that discrimination training teaches theanimal to approach one stimulus and to avoid the otheror, again, simply that discrimination training teaches theanimal not to alternate between stimuli. The extent towhich these are real alternatives rather than mererewordings requires consideration. And secondly, a moreformal specification of the mechanism by which thisproposed general transfer effect operates is clearlyneeded.

216 HALL

REFERENCESBainbridge. P. L. Learning in the rat: Effect of early experience

with an unsolvable problem. Journal of Comparative andPhvsiological Psychology. 1973.82.301-307.

Eck. K. 0., Noel, R. C., & Thomas. D. R. Discrimination learningas a function of prior discrimination and nondifferentialtraining. Journal of Experimental Psychology. 1969. 82.156-162.

Frieman, J.. & Goyette, C. H. Transfer of training acrossstimulus modality and response class. Journal of ExperimentalPsv ch ology , 1973, 97, 235-241.

Goodnow, J. J. Determinants of choice-distribution intwo-chotce situations. American Journal of Psychology. 1955.68,106-116.

Honig, W. K. Attentional factors governing the slope of thegenera1ization gradient. In R. ),1. Gilbert and N. S. Sutherland(Eds.), Animal discrimination learning. London: AcademicPress. 1969.

Keilitz, 1., & Frieman, J. Transfer of training following errorlessdiscrimination learning. Journal of Experirnental Psychology.1970,85,293-299.

).Iackintosh, N. J. Stimulus selection: Learning to ignore stimulithat predict no change in reinforcement. In R. A. Hinde and J.

Stevenson-Hinde (Eds.j, Constraints On learning. London:Academic Press. 1973.

Mandler, J. ).1. Behavior changes during ove rtraining and theireffects on reversal and transfer. Psycho norni c MonographSupplements, 1966, 1,187-202.

Solomon. S. Effects of variations of rearing, drive level, andtraining procedures on performance in probability learningtasks. Psychological Reports, 1962. 10. 679-689.

Thomas. D. R .. Freeman. F., Svinicki, J. G., Burr, D. E. S., &Lyons. J. Effects of extradimensional training on stimulusgeneralization. Journal of Experimental PsychologyMonograph, 1970,83.1·21.

Thomas. D. R.• Miller, J. T., & Svinicki, J. G. Nonspecifictransfer effects of discrimination training in the rat. Journal ofComparative and Phy siolo gical Psychology. 1971. i4, 96-107.

Turner. C .. & Mackintosh, N. J. Stimulus selection and irrelevantstimuli in discrimination learning by pigeons. Journal ofComparative and Phsvi olo gical Psychology, 1972.78.1-9,

Wagner. A. Roo Logan. F. A., Hab erland t , K.. & Price. T.Stimulus selection in animal discrimination learning. Journalof Experimental Psychology. 1968. 76, 171-180.

(Received for publication October 9,1974:revision accepted January 30, 1975.)