..

f

WESTERN JOINT COMPUTER CONFERENCE

PAPER PRESENTED ATTHE JOINT IRE-AIEE-ACM COMPUTER CONFERENCE

LOS ANGELES, CALIF., MAY 9-11, 1961

SponsorsTHE INSTITUTE OF RADIO ENGINEERS

Professional Group on Electronic Computers

THE AMERICAN INSTITUTE OF ELECTRICAL ENGINEERSCommittee on Computing Devices

THE ASSOCIATION FOR COMPUTING MACHINERY

© 1961 byNational Joint Computer Committee

Reproduction of all or part of this paper is permissible when acknowl-edgement is givento theauthor and National JointComputer Committee.The ideas and opinions expressed herein are solely those of the author,and are not necessarily representative

of,

or endorsed by, the WJCCCommittee or the NJCC Committee.

Published byWESTERN JOINT COMPUTER CONFERENCE

121

t

%

3.2

THE SIMULATION OP VERBAL LEARNING BEHAVIOR*

E. A. FeigenbaumUniversity of California

Berkeley, Californiaand

The RAND CorporationSanta Monica, California

Summary

An information processing model ofelementary, human symbolic learning isgiven a precise statement as a computerprogram, called Elementary Percelverand Memorizer (EPAM). The program simu-lates the behavior of subjects in exper-iments involving the rote learning ofnonsense syllables. A discriminationnet which grows is the basis of EPAM'sassociative memory. Fundamental infor-mation processes include processes fordiscrimination, discrimination learning,memorization, association using cues,and response retrieval with cues . Manywell-known phenomena of rote learningare to be found in EPAM's experimentalbehavior, Including some rather complexforgetting phenomena. EPAM is programmedin Information Processing Language V.

H. A. Simon has described somecurrent research in the simulation ofhuman higher mental processes and hasdiscussed some of the techniques and prob-lems which have emerged from thisresearch. The purpose of this paper isto place these general issues in the con-text of a particular problem by describingin detail a simulation of elementaryhuman symbolic learning processes .

The information processing modelof mental functions employed is realizedby a computer program called ElementaryPerceiver and Memorizer (EPAM). TheEPAM program is the precise statement ofan information processing theory of verballearning that provides an alternativeto other verbal learning theories whichhave been proposed.** It is the result

*I am deeply indebted to Herbert A.Simon for his past and present colla-boration in this research. This researchhas Deen supported by the Computer Sci-ences Department, The RAND Corporation,and the Ford Foundation. I wish to ex-press appreciation for the help andcritical comments of Julian Feldman,Allen Newell, J. C. Shaw and Fred Tonge.

♦♦Examples of quantitative (or quasi-quantitative) theories of verbal learningare those of Hull, et.al. [l] , Gibson [2].and Atkinson [3] .

of an attempt to state quite preciselya parsimonious and plausible mechanismsufficient to account for the rotelearning of nonsense syllables. Thecritical evaluation of EPAM must ulti-mately depend not upon the interest whichit may have as a learning machine, butupon Its ability to explain and predictthe phenomena of verbal learning.

I should like to preface my dis-cussion of the simulation of verballearning with some brief remarks aboutthe class of information processingmodels of which EPAM is a member.

a. These are models of mentalprocesses, not brain hardware.They are psychological modelsof mental function. No physio-logical or neurological assump-tions are made, nor is anyattempt made to explain infor-mation processes in terms ofmore elementary neural pro-cesses .

b. These models conceive of thebrain as an information pro-cessor with sense organs asinput channels, effector organsas output devices, and withinternal programs for testing,comparing, analyzing, re-arranging, and storing Infor-mation.

c. The central processing mechanismis assumed to be serial; i.e.,capable of doing only one (ora very few) things at a time.

d. These models use as a basicunit the information symbol;i.e., a pattern of bits which isassumed to be the brain'sinternal representation ofenvironmental data .

c. These models are essentiallydeterministic, not probabilisticRandom variables play no funda-mental role in them.

1223.2

t

THE BASIC EXPERIMENT

Early in the history of psychology,the psychologist invented an experimentto simplify the study of human verballearning. This "simple' experiment isthe rote memorization of nonsensesyllables in associate-pairs or seriallists.

The Items to be memorized are gen-erally three-letter words having con-sonant letters on each end and a vowelin the middle . Nonsense syllables arechosen In such a way that the three-letter combinations have no ordinaryEnglish meaning. For example, CAT isnot a nonsense syllable, but XUM is.*

In one basic variation, the rotememory experiment is performed as follows:

For successive trials thegsyllables are reordered randomly.This style of carrying out theexperiment is called paired-associates presentation.

The other basic variant of theexperiment is called serial-anticipationpresentation. The nonsense syllables(say, 10 or 12 items) are arranged ina serial list, the order of which isnot changed on successive trials . Whenhe is shown the nth syllable, the subjectis to respond with the (n+l)st syllable.A few seconds later, the (n+l)st syllableis shown and the subject is to respondwith the (n+2)nd syllable, and so on.The experiment terminates when the subjectis able to correctly anticipate all ofthe syllables.

Numerous variations on this experi-mental theme have been performed.*The phenomena of rote learning are wellstudied, stable, and reproducible.For example, In the typical behavioraloutput of a subject, one finds:

a. Failures to respond to a stimulusare more numerous than overterrors .

b. Overt errors are generallyattributable to confusion bythe subject between similarstimuli or similar responses.

c. Associations which are givencorrectly over a number oftrials sometimes are thenforgotten, only to reappearand later dissappear again.This phenomenon has been calledoscillation.**

d. If a list x of syllables orsyllable pairs is learned tothe criterion; then a list yis similarly learned; andfinally retention of list x istested; the subject's abilityto give the correct x responsesis degraded by the interpolatedlearning. The degradation iscalled retroactive inhibition.The overt errors made in the

*For an extended treatment of thissubject, see Hovland, C. 1., "HumanLearning and Retention." [4]

**By Hull [5] . Actually he called it"oscillation at the threshold of recall,"reflecting his theoretical point of view.

a. A set of nonsense syllables Ischosen and the syllables arepaired, making, let us say, 12pairs .

b. A subject is seated in front ofa viewing apparatus and thesyllables are shown to him, onepair at a time.

c . First, the left-hand member ofthe pair ( stimulus Item) isshown. The subject tries to saythe second member of the pair(response item) .

d. After a short interval, theresponse item is exposed so thatboth stimulus and response itemsare simultaneously in view.

c. After a few seconds, the cyclerepeats Itself with a new pairof syllables. This continuesuntil all pairs have beenpresented (a trial) .

f. Trials are repeated, usually untithe subject is able to givethe correct response to eachstimulus. There is a relativelyshort time interval betweentrials.

*People will defy an experimenter'smost rigorous attempt to keep the non-sense syllables association-free. Listsof nonsense syllables have been prepared,ordering syllables on the basis of theirso-called 'association value," In orderto permit the experimenter to control"meaningfulness . "

1233.2

retest trial are generallyIntrusions from the list y. Thephenomenon disappears rapidly.Usually after the first retesttrial, list x has been relearnedback to criterion.

c. As one makes the stimulussyllables more and more similar,learning takes more trials .

The Information Processing Model

This section describes the processesand structures of EPAM.

EPAM is not a model for a particularsubject. In this respect it is to becontrasted with the binary choice modelsof particular subjects which Mr. Feldmanis presenting in this session. The factis that individual differences play onlya small part In the results of the basicexperiment described above.

It is asserted that there are certainelementary information processes which anindividual must perform if he is todiscriminate, memorize and associateverbal items, and that these infor-mation processes participate in all thecognitive activity of all individuals ."*

It is clear that EPAM does not yetembody a complete set of such processes .It is equally clear that the processesEPAM has now are essential and basic .

*Some information processing modelsare conceived as models of the mentalfunction of particular subjects: e.g. ,Feldman's Binary Choice Model [6] . Otherstreat the general subject as EPAM does.Still others are mixed in conception,asserting that certain of the processes ofthe model are common for all subjects whileother processes may vary from subject tosubject; e.g., the General Problem Solverof Newell, Shaw and Simon [7] . Alterna-tively, information processing modelsmay also be categorized according to howmuch of the processing is "hard core"(i.e., necessary and invariant) as opposedto "strategic" (i.c, the result ofstrategy choice by control processes). Isuggest the obvious: that models ofstrategies for information processingwill tend to be models of the generalsubject. As exemplars, Lindsay's ReadingMachine [BJ , a "hard core" model, treatsthe general subject; Wlckelgren's modelof the conservative Focusing Strategyin concept attainment (Wickelgren [9];Bruner,

Goodnow,

and Austin [10] ) , a purestrategy model, can predict only thebehavior of particular subjects.

Overview: Performance and Learning

Conceptually, EPAM can be brokendown into two subsystems, a performancesystem and a learning system. In theperformance mode, EPAM produces responsesto stimulus items. In the learning mode,EPAM learns to discriminate and associateitems .

The performance system is thesimpler of the two. It is sketched inFig. 1. When a stimulus Is noticed, aperceptual process encodes it, producingan internal representation (an inputcode) . A discriminator sorts the Inputcode in a discrimination net (a tree oftests and branches) to find a storedimage of the stimulus . A response cueassociated with the image is

found,

andfed to the discriminator. The discrimina-tor sorts the cue in the net and findsthe response image, the stored form ofthe response. The response image is thendecoded by a response generator letter byletter in another discrimination netinto a form suitable for output. Theresponse is then produced as output.

The processes of the learning systemare more complex. The discriminationlearning process builds discriminationsby growing the net of tests and branches.The association process builds associa-tions between images by storing responsecues with stimulus images . Theseprocesses will be described fully in duecourse .

The succeeding sections on theinformation processing model give adetailed description of the processes andstructures of both systems .Input to EPAM: Internal Representationsof External Data

The following are the assumptionsabout the symbolic input process when anonsense syllable is presented to thelearner. A perceptual system receivesthe raw external information and codesit into internal symbols . These internalsymbols contain descriptive informationabout features of external stimulus.For unfamiliar 3-letter nonsense symbols,It is assumed that the coding is done interms of the individual letters, forthese letters are familiar and are well-learned units for the adult subject.*

*The basic perception mechanism Ihave in mind is much the same as that ofSelfridge [ll] and Dinneen, whose computerprogram scanned letters and perceivedsimple topological features of theseletters .

1243.2

The end result of the perception processis an internal representation of the non-sense syllable —a list of internal symbols(1.c., a list of lists of bits) con-taining descriptive information aboutthe letters of the nonsense syllable.Using Minsky' s terminology [l2 ], this isthe "character" of the nonsense syllable.

I have not actually programmed thisperception process. For purposes ofthis simulation, I have assigned codedrepresentations for the various lettersof the alphabet based on 15 differentgeometrical features of letters . Forpurposes of exploring and testing themodel, at present all that is reallyneeded of the input codes is:

a. that the dimensions of a lettercode be related in some reason-able way to features of realletters .

b. that the letter codes be highlyredundant, that is, includemany more dimensions than Isnecessary to discriminate theletters of the alphabet.

To summarize, the internal represen-tation of a nonsense syllable is a listof lists of bits, each sublist of bitsbeing a highly redundant code for aletter of the syllable.

Given a sequence of such inputs,the essence of the learner's problem istwofold:

first,

to discriminate eachcode from the others already learned,so that differential response can bemade; second, to associate informationabout a "response" syllable with theinformation about a "stimulus" syllableso that the response can be retrievedif the stimulus is presented.

Discriminating and Memorizing: GrowingTrees of ""Images"

I shall deal with structure firstand reserve my discussion of processfor a moment .

Discrimination net. The primaryinformation structure in EPAM is thediscrimination net. It embodies in itsstructure at any moment all of the dis-crimination learning that has takenplace up to a given time. As an infor-mation structure it is no more than afamiliar friend: a sorting tree ordecoding network. Fig. 2 shows a smallnet . At the terminals of the net arelists called Image lists, in whichsymbolic information can be stored. Atthe nodes of the net are stored programs,

called tests, which examine character-istics of an input cede and signal branchleft or branch-right. On each imagelist will be found a list of symbolscalled the image . An image is a partialor total copy of an input code. I shalluse these names in the followingdescription of net processes.

Net Interpreter. The discriminationnet is examined and altered by a numberof processes, most important of which isthe net interpreter. The net interpretersorts an input code in the net andproduces the image list associated withthat input code. This retrieval processis the essence of a _pur_e ly as sociabivememory : the st imulus information itselfleads to the retrieval of "the informationassociated" with~that~ stimulus . "The netInterpreter is a very simple process.It finds the test in the topmost nodeof the tree and executes this program.The resulting signal tells it to branchleft or branch right to find thesucceeding test. It executes this,tests its branches again, and repeatsthe cycle until a terminal is found.The name of the image list is produced,and the process terminates. This is thediscriminator of the performance systemwhich sorts items in a static net .

Discrimination Learning. The dis-crimination learning process of thelearning system grows the net. Initiallywe give the learning system no dis-crimination net but only a set of simpleprocesses for growing nets and storingnew images at the terminals.

To understand how the discriminationand memorization processes work, let usexamine in detail a concrete examplefrom the learning of nonsense syllables.Suppose that the first stimulus-responseassociate-pair on a list has been learned(Ignore for the moment the question ofhow the association link is actuallyformed.) Suppose that the first syllablepair was DAX-JIR. The discriminationnet at this point has the simple two-branch structure shown in Pig. 3-Because the syllables differ in theirfirst letter, Test 1 will probably be atest of some characteristic on which theletters D and J differ. No more testsare necessary at this point.

Notice that the image of JIR whichis stored is a full image. Full responseimages must be stored--to provide theInformation for producing the response;but only partial stimulus images needbe stored--to provide the informationfor recognizing the stimulus . How muchstimulus image information is required

1253.2

the learning system determines for itselfas it grows its discrimination net, andmakes errors which it diagnoses as inade-quate discrimination.

To pursue our simple example, supposethat the next syllable pair to be learnedls PIB-JUK. There are no storageterminals in the net, as it stands, forthe two new Items. In other words, thenet does not have the discriminativecapability to contain more than two items.The input code for PIB is sorted by thenet interpreter. Assume that Test 1sorts it -down the plus branch of Fig. 3<As there are differences between theincumbent image (with first-letter D)and the new code (with first-letter P) anattempt to store an image of PIB at thisterminal would destroy the informationpreviously stored there.

Clearly what Is needed is the abilityto discriminate further. A match fordifferences between the incumbent imageand the challenging code is performed.When a difference is

found,

a new test iscreated to discriminate upon this differ-ence. The new test is placed in the netat the point of failure to discriminate,an image of the new item is created, andboth images --incumbent and new- -arestored in terminals along their appropriatebranches of the new test, and. the conflictIs resolved.*

*With the processes just described,the discrimination net would be growneach time a new item was to be added tothe memory. But from an information pro-cessing standpoint, the matching and net-growing processes are the most time-consuming in the system. In general, withlittle additional

effort,

more than onedifference can be detected, and more thanone discriminating test can be added to

■

the net. Each redundant test placed inthe net gives one "empty" image list.At some future time, if an item is sortedto this empty image list, an Image can bestored without growing the net. There isa happy medium between small nets whichmust be grown all the time and large netsreplete with redundant tests and a waste-ful surplus of empty image lists. Experi-mentation with this "structural parameter"has been done and it has been foundthat for this study one or two redundanttests per growth represents the happymedium. However, I would not care tospeak of the generality of this parti-cular result.

The net as it now stands Is shown inFig. 4. Test 2 is seen to discriminateon some difference between the lettersP and D.

The input code for JUK is now sortedby the net Interpreter. Since Test 1cannot detect the difference between theinput codes for JUK and JIR (under ourprevious assumption), JUK is sorted tothe terminal containing the image ofJIR. The match for differences takesplace. Of course, there are no first-letter differences. But there are dif-ferences between the incumbent image andthe new code in the second and thirdletters .

Noticing Order. In which lettershould the matching process next scanfor differences? In a serial machinelike EPAM, this scanning must take placein some order. This order need not bearbitrarily determined and fixed. Itcan be made variable and adaptive. Tothis »nd EPAM has a noticing order forletters of syllables, which prescribesat "any "moment a letter-scanning sequencefor the matching process . Because it isobserved that subjects generally considerend-letters before middle-letters, thenoticing order Is initialized as follows:

first-letter,

third-letter, second-letter. When a particular letter beingscanned yields a

difference,

this letteris promoted up one position on thenoticing order. Hence, letter positionsrelatively rich in differences quicklyget priority in the scanning. In ourexample, because no first-letter differ-ences were found between the image, ofJIR and code for JUK, the third lettersare scanned and a difference is found(between R and X) . A test is createdto capitalize on this third-letterdifference and the net is grown asbefore. The result is shown in Fig. 5«The noticing order is updated; third-letter, promoted up one, is at the head.

Learning of subsequent itemsproceeds in the same way, and we shallnot pursue the example further.

Associating Images: Retrieval Using Cues

The discrimination net and itsinterpreter associate codes of externalobjects with internal image lists andimages. But the basic rote learningexperiment requires that stimulusinformation somehow lead to response

1263.2

Information and a response. The discri-mination net concept can be used for theassociation of internal images with eachother (i.e., response with stimulus) withvery little addition to the basicmechanism.

An association between a stimulusimage and a response image is accomplishedby storing with the stimulus image some ofthe coded information about the response.This information is called the cue . A cueis of the same form as an input code, butgenerally contains far less Informationthan an input code. A cue to an associa-ted image can be stored in the discrimi-nation net by the net interpreter to re-trieve the associated image.

If,

for ex-ample, In the net of Fig. 3 we had storedwith the stimulus image the letter J asa cue to the response JIR, then sortingthis cue would have correctly retrievedthe response image. An EPAM internalassociation is built by storing with thestimulus image information sufficient toretrieve the response image from the netat the moment of association.

The association process determineshow much information is sufficient bytrial and error. The noticing order forletters is consulted, and the first-priority letter is added to the cue . Thecue is then sorted by the net Interpreterand a response image is produced. Itmight be the wrong response image; for ifa test seeks information which the cue doesnot contain, the interpreter branches leftor right randomly (with equal probabili-ties) at this test.* During association,the selection of the wrong response IsImmediately detectable (by a matchingprocess) because the response input codeis available. The next-priority letter Isadded to the cue and the process repeatsuntil the correct response image is re-trieved. The association is then con-sidered complete.

Note two important possibilities.First, by the process just described, acue which Is really not adequate toguarantee retrieval of the response imagemay by happenstance give the correctresponse image selection during associa-tion. This "luck" usually gives rise toresponse errors at a later time.

s is the only use of a randomvariable in EPAM. We do not like it. Weuse it only because we have not yet dis-covered a plausible and satisfying adapt-ive mechanism for making the decision.The random mechanism does, however, givebetter results than the go-one-way-all-the-time mechanism which has also beenused.

Second,

suppose that the associationbuilding process does its job thoroughly.The cue which it builds is sufficient toretrieve the response image at one parti-cular time, the time at which the two iteItems were associated.

If,

at somefuture time, the net is grown to encompassnew images being added to the memory,then a cue which previously was sufficientto correctly retrieve a response imagemay no longer be sufficient to retrievethat response image. In EPAM, associationlinks are "dated," and ever vulnerableto interruption by further learning.Responses may be "unlearned" or "forgotten"temporarily, not because the response in-formation has been destroyed in the memory,but because the information has been tem-porarily lost in a growing network. If anassociation failure of this type can bedetected through feedback from theenvironmental or experimental situation,then the trouble ls easily remedied byadding additional response informationto the cue. If not, then the responsemay be more or less permanently lost inthe net. The significance of thisphenomenon will perhaps be more easilyappreciated in the discussion of resultsof the EPAM simulation.

Responding: Internal and External

A conceptual distinction is madebetween the process by which EPAM selectsan internal response image and the pro-cess by which it converts this imageinto an output to the environment .

Response retrieval. A stimulus Itemis presented. This stimulus input codeis sorted in the discrimination net toretrieve the image list, in which the cueis found. The cue is sorted in the netto retrieve another image list containingthe proposed response image. If there isno cue, or if on either sorting pass anempty image list is selected, no responseIs made .

Response generation. For purposesof response generation, there is a fixeddiscrimination net (decoding net),assumed already learned, which transformsletter codes of internal images intooutput form. The response image is de-coded letter by letter by the netInterpreter in the decoding net forletters .The Organization of the Learning Task

The learning of nonsense symbols bythe processes heretofore described takestime. EPAM is a serial machine. There-

127

i

3.2

fore,

the individual items must be dealtwith In some sequence. This sequenceis not arbitrarily prescribed. It is theresult of higher order executive pro-cesses whose function is to control EPAM'sfocus of attention. These macroprocesses,as they are called, will not be describedor discussed here. A full exposition ofthem is available in a paper byFeigenbaum and Simon. [l3]

Stating the Model Precise!Computer Program for EPA]

The EPAM model has been realizedas a program in Information ProcessingLanguage V [l4] and is currently beingrun both on the Berkeley 704 and the RAND7090. Descriptive information on thecomputer realization, and also thecomplete IPL-V program and data structuresfor EPAM (as it stood In

October,

1959)are given in an earlier work by theauthor [15] .

IPL-V, a list processing language,was well suited as a language for theEPAM model for these key reasons:

a. The IPL-V basic processes dealexplicitly and directly with list struc-tures . The various information structuresin EPAM (e.g., discrimination net, Imagelist) are handled most easily as liststructures. Indeed, the discriminationis, virtually by

definition,

a liststructure of a simple type.

b. It ls useful in some places, andnecessary In others, to store with somesymbols information descriptive of thesesymbols. IPL-V s description list anddescription list processes are a goodanswer to this need.

c. The facility with which hier-archies of subroutine control can bewritten in IPL-V makes easy and un-complicated the programming of the kindof complex control sequence which EPAMuses .

Empirical Explorations with EPAM

The procedure for exploring thebehavior of EPAM is straightforward. Wehave written an "Experimenter" programand we give to this program the parti-cular conditions of that experiment asinput at the beginning of an experiment.The Experimenter routine then puts EPAMqua subject through its paces in thatparticular experiment. The completerecord of stimuli presented and responsesmade is printed out, as in the final net.Any other information about the pro-cessing or the state of the EPAM memory

can also be printed out.

A number of simulations of thebasic paired-associate and serial-anticipation experiments have been run.Simulations of other classical experi-ments in the rote learning of nonsensesyllables have also been run. Thecomplete results of these simulationexperiments and a comparison betweenEPAM's behavior and the reported be-havior of human subjects will be the sub-ject of a later report. However, somebrief examples here will give an indi-cation of results expected and met.

a. Stimulus and response generali-zation. These are psycho-logical terms used to describethe following phenomenon. IfX and X' are similar stimuli,and V is the correct responseto the presentation of X; thenif V is given in response tothe presentation of X', this iscalled stimulus generalization.Likewise, if V and V' are simi-lar responses, and V' Is givenin response to the presentationof X, this is called responsegeneralization. Generalizationis common to the behavior ofall subjects, and is found inthe behavior of EPAM. It isa consequence of the respondingprocess and the structure of thediscrimination net. For those"stimuli" are similar in theEPAM memory whose input codesare sorted to the same terminal;and one "response" is similarto another if the one is storedin the same local area of thenet as the other (and henceresponse error may occur whenresponse cue Information isinsufficient) .

b. Oscillation and RetroactiveInhibition. We have describedthese phenomena in an earliersection.

Oscillation and retroactiveinhibition appear In EPAM'sbehavior as consequences ofsimple mechanisms for discrim-ination, discrimination learning,and association. They were inno sense "designed into" thebehavior. The appearance ofrather complex phenomena suchas these gives one a littlemore confidence in the credi-bility of the basic assumptionsof the model.

1283.2

c

*

These two phenomena are discussedtogether here because in EPAMthey have the same origin. Asitems are learned over time,the discrimination net grows toencompass the new alternatives.Growing the net means addingnew tests, which in turn meansthat more Information will beexamined in all objects beingsorted. An important class ofsorted objects Is the set ofcues . Cue information sufficientat one moment for a firm associa-tion may be Insufficient at alater moment. As describedabove, this may lead to responsefailure . The failure is causedentirely by the ordinary processof learning new items. In thecase of oscillation, the newitems are items within a singlelist being learned. In the caseof retroactive inhibition, thenew items are items of the secondlist being learned In the samediscrimination net. In bothcases the reason for the responsefailure is the same. Accordingto this explanation, the phenom-ena are first cousins (an hypo-thesis which has not been widelyconsidered by psychologists).

In the EPAM model, the termInterference is no longer merelydescrlptive--it has a preciseand operational meaning. Theprocess by which later learninginterferes with earlier learningis completely specified.

Forgetting. The usual explan-ations of forgetting use in oneway or another the simple andappealing Idea that stored infor-mation is physically destroyedIn the brain over time (e.g., thedecay of a "memory trace," orthe overwriting of old infor-mation by new

information,

asin a computer memory) . Suchexplanations have never dealtadequately with the commonplaceobservation that all of us canremember, under certain condi-tions, detailed and seeminglyunimportant information aftervery long time periods haveelapsed. An alternative explan-ation, not so easily visualized,is that forgetting occurs notbecause of information destruc-tion but because learnedmaterial gets lost and inaccess-ible in a large and growingassociation network.

EPAM forgets seemingly well-learned responses. This for-getting occurs as a directconsequence of later learningby the learning processes.Furthermore, forgetting is onlytemporary: lost associationscan be reconstructed by storingmore cue information. EPAMprovides a mechanism for explain-ing the forgetting phenomenonin the absence of any infor-mation loss . As far as weknow, it is the first concretedemonstration of this type offorgetting in a learning machine .

Conclusion: A Look Ahead

Verification of an informationprocessing theory is obtained by simu-lating many different experiments andby comparing in detail specificqualitative and quantitative features ofreal behavior with the behavior of thesimulation. To date, Mr. Simon and Ihave run a number of simulated experiments.As we explore verbal learning

further,

more of these will be necessary.

We have been experimenting with avariety of "sense modes" for EPAM,corresponding to "visual" input and"written" output, "auditory" input and"oral" output, "muscular" inputs andoutputs . To each mode corresponds aperceptual input coding scheme, and adiscrimination net. Associations-across-nets, as well as the familiar associations-within-nets, are now possible. Internaltransformations between representationsin different modes are possible. Thus,EPAM can "sound" in the "mind's ear"what it "sees" in the "mind's eye," justas all of us do so easily. We have beenteaching EPAM to read-by-association,much as one teaches a small child beginningreading. We have only begun to explorethis new addition.

The EPAM model has pointed up afailure shared by all existing theoriesof rote learning (including the presentEPAM) . It is the problem of whetherassociation takes place between symbolsor between tokens of these symbols .For example, EPAM cannot learn a seriallist in which the same item occurstwice. It cannot distinguish betweenthe first and second occurrence of thethe item. To resolve the problem wehave formulated (and are testing)processes for building, storing, andresponding from chains of token associa-tions .

129

*

3.2

i

References

1. Hull, C. L., C. I. Hovland, R. T.Ross, M. Hall, D. T. Perkinsand F. B. Fitch, Mathematico-deductlve Theory of Rote Learning,New Haven, Connecticut, YaleUniversity Press, 1940.

2.

Gibson,

E. J., "A Systematic Appli-cation of the Concepts ofGeneralization and Differentiationto Verbal Learning," Psychol. Rev.,Vol. 47, 1940, pp. 196-229.

3. Atkinson, R.

C,

"An Analysis ofRote Serial Learning in Termsof a Statistical Model," IndianaUniversity, Doctoral Dissertation,1954.

4.

Stevens,

S.

S.,

cd., Handbook ofExperimental Psychology, New York,Wiley, 1951.

5. Hull, C. L., "The Influence ofCaffeine and Other Factors onCertain Phenomena of Rote Learning,"J. Gen. Psychol., Vol. 13, 1935,pp. 249-273-

6. Feldman, J., "An Analysis of Predic-tive Behavior in a Two ChoiceSituation," Carnegie Institute ofTechnology, Doctoral Dissertation,1959.

7- Newell, A., J. C.

Shaw,

and H. A.

Simon,

"Report on a General ProblemSolving Program," InformationProcessing: Proceedings of theInternational Conference onInformation ProcessTng, UNESCO ,TFarTs, June, 1959, pp. 256-264.

8. Lindsay, R., "The Reading MachineProblem," CIP Working Paper #33,Carnegie Institute of Technology(Graduate School of IndustrialAdministration), Pittsburgh, i960.

9. Wlckelgren, W., "A SimulationProgram for Conservative Focusing,"unpublished manuscript, Universityof California, Berkeley, January,1961.

12. Minsky, M., "Steps Toward ArtificialIntelligence," Proceedings ofthe IRE, Vol. 49, No. 1, January,1961, PP. 8-30.

13- Feigenbaum, E. and H. A.

Simon,

"A Theory of the Serial PositionEffect," CIP Working Paper #14,Carnegie Institute of Technology(Graduate School of IndustrialAdministration), Pittsburgh, 1959.

14. Newell, A., F. M. Tonge, E. A.Feigenbaum, G. H. Mealy, N.

Saber,

B. P.

Green,

and A. K.

Wolf,

In£orTOatio^^roce_ss_ing Language VManual XSect ions I and I TJ]Santa Monica, California, The RANDCorporation, P-1897 and P-1918.

15. Feigenbaum, E., An InformationProcessing Theory "oT~" VerbalLearning'," Santa Monica, CaliforniaThe RAND Corporation, P-1017,

October,

1959.

10. Bruner, J.

S.,

J. J.

Goodnow,

andG. A. Austin, A Study of Thinking,New York, Wiley, 1956".

11. Selfridge, 0.

G.,

"Pattern Recogni-tion and Modern Computers,"Proceedings of the 1955 WesternJoint Computer

Conference,

IRE 7March, 1955.

1303.2

RAW STIMULUS

EPAM STIMULUSINPUT CODE

RESPONSE IMAGE

RESPONSE OUTPUT

Fig. I — EPAM performance processfor producing the responseassociated with a stimulus

131

i

i

3.2

(t) = Discriminating test at a nodeI = Imaae at a terminnlImage at a terminal

lL2j = 'mage and cue at a terminalI I = Empty terminal

Fig. 2— A typical EPAM discrimination net

STIMULUS RESPONSEDAK JIR

Fig. 3 — Discrimination net after the learningof the first two items. Cues are not shown.

Condition: no redundant tests added.Test 1 is a first-letter test.

1323.2

Fig. 4 — Discrimination net of Fig. 3 after thelearning of stimulus item, Plß.

Test 2 is a first letter test

>■

Fig. 5— Discrimination net of Fig. 4 afterthe learning of the response item, JUK.

Test 3 is a third-letter test