*_

\

i

101

a

ELEMENTARY PERCEIVER AND MEMORIZER:REVIEW OF EXPERIMENTSEdward FeigenbaumDepartment of Business AdministrationCenter for Research in Management .cienceUniversity of

California,

Berkeley

Herbert A. SimonGraduate School of Industrial AdministrationCarnegie Institute of Technology

RBCm EXPERIMENTS WITH THE EPAM SIMULATION OF VERBAL LEAROTG

-Ph. momentary P»rceiver and Memorizer (EPAM) is a programwhich

in one or more sense modes.

„.!,.-.

statements of hypotheses about in-The EPAM programs «J^^L^^^SotLr simulations,formation V^^^^^u^^Smm* cy which the consequencesthe computer runs vith EPAM (i.e. , the me » conditions)ing the phenomena of verbal learning.

'«. Wer 1. .WW -"J -»" £?.£$ Si^ce-Eu.g-u ss-i.-r^i^stS ~—» —eating the direction which the research has taken.

mm 1. . co«puter progr«»rltten »» «"' ""^""^S, of5J2 eo-,c,

ment.

I

102

Symposium on Simulation Models____» EPAM Simulation*

■__»" EPAM program has teen described in detail elsewhere (Feigen-baum, 1959, 19^1, Itens 1 and 2 in the Bibliography). We shall sum-marise it luite briefly here:

The EPAM processes perform the following four principalfuncafcrns:

fa), recognize an external stimulus as one about which some in-foimrion has already been memorized;

* (

I

103

Simulation of Cognitive Processes

(j\ - Discriminating test at a nodeI I » Image at a terminal

\

Pig. 1-A typical EPAM discrimination net

I C = Image ani cue at a terßlnaX

__=

Empty terminal

104

Symposium on Simulation Models

RAW STIMULUS

\

RESPONSE OUTPUT

Fig. 2— EPAM performance process for producing theresponse associated with a stimulus

Simulation of Cognitive Processes

105

EPAM responds to a stimulus (see Figure 2) by sorting it in thediscrimination net, finding the associated response cue, sorting thatcue in the same net, finding its image and using the response image toproduce the response.

EPAM does not simulate the initial sensory and perceptual pro-cesses (the processes in Figure 2 labeled "perceive Features ofStimulus") that scan an external stimulus and extract from it encodedinformation that is used by the memorizing and responding processes.The EPAM program takes up at the point where the initial perceptualencoding has already been accomplished. Its "stimuli" are encodingsof the external stimuli. Thus, the EPAM stimuli corresponding toletters of the alphabet hold information about whether a letter con-tains a curved segment, a straight segment, a closed loop, and so on.Each letter is discriminated from the others, as the discriminationnet grows, by its possession of a unique combination of simple topo-logical and metrical properties of this type.

Different encodings are used to represent stimuli in the differ-ent sensory modes, and in certain sub-modes. Thus, there may be a netfor discriminating among aural phonemes by encodings that representelementary phonemic characteristics; another net for discriminatingamong visually presented syllables of letters; still another net fordiscriminating among objects in terms of single characteristics ofshape, etc. We will be concerned, on the sensory side, with thesethree specific modes: aural phonemic, visual literal, and visualobject modes.

On the response side, too, several modes are represented by dif-ferent encodings. In the learning of reading, EPAM uses three re-sponse modes, paired, respectively, with the three stimulus modes.There is an oral response mode, whose images may be interpreted asthe signals that activate the muscles used to produce spoken language.There is a written response mcde, representing signals that activatemuscles used in printing letters. There is a pointing response mode,which, as its name implies, signals the selection of an object from aset of objects by pointing. In the present EPAM program, the responsemodes are represented only in a rather rudimentary

form,

since thesystem has been elaboratedprincipally on the stimulus side.

Verbal Learning Experiments with EPAM

To facilitate experimentation with the model, an Experimenterprogram was written. The Experimenter is quite distinct from EFAM(the subject). It contains a simulated memory-drum apparatus commonlyused in learning experiments; timing procedures; stimulus materials ofvarious

kinds;

error-checking- and recording procedures; and so on.

The experiments which we have run are of the following types:

l) Serial learning of nonsense syllables*

"For further details, see C. I. Hovland ■ Human Learning and Retention, Handbook of Experimental Psychology (Wiley: l9sl)-

106

Symposium on Simulation Models2) Learning of nonsense syllables in associate-pairs*

3) Serial learning of nonsense syllables of two lists,with a retest of the first list following thelearning of the second list (also two-list learningof paired-associate syllables).*

k) Learning a set of associate-pairs (as in 2 above),vith the stimuli in one sense mode, and the re-sponses ln another.

5) Multi-mode chaining of associations (reading). 'Thespoken nana ls associated with a visual object.The written name of the object is associated withthe spoken name. This learning ls done for a numberof objects. Without further learning, a readingtest for identification of the visual objects from ,the written name ls given. \

Bi of Results of Experiments

1. Stimulus ai_d Response Generalization^ If Sia associatedwith R, ST TsTmllar-to S) is presented, and a subject responds withR, this is a stimulus generalization error. If Sis associated withR, R' is a response similar to R, and the presentation of S resultsin R' as response, this is a response generalization error. General-ization errors are commonplace in the behavior of both human subjectsand EPAM. Stimulus generalization is most easily seen in the firstfew trials of the learning of a second list of items after a firstlist has been learned. Stimulus items in the second list which aresimilar to items of the first list are responded to with the approp-riate response items of the first list. Response generalization ismost easily seen in the same kind of experiment. The effect of thesecond-list learning may be to store in the memory responses similarto first-list responses. Since the information to discriminate thesimilar responses will not have been stored during the first-listlearning (there was no need to do it at the time the first list waslearned) the similar responses may be confused and the wrong one maybe produced.

2. Oscillation. If one looks at the pattern of success andfailures over-trials for any one syllable in a serial list learnedby a human subject, one finds that this pattern is highly irregular:

failures,

followed by some successes, then a

failure,

then successes,further

failure,

etc. This is a feature of the behavior of EPAM. Itis caused by the interference of later-learned items with earlier-learned items and associations in a growing association memory. Thephenomenon is intimately related to the forgetting phenomenon to bedescribed under 5 below.

"Ibid.

107

Simulation of Cognitive Processes

3 Effects of Intra-List Similarity_of Items. Lists of highlysimilar iteSs"a?e"mch harder for human subjects to learn than listsoFlw intra-list similarity. This firmly established empirical re-sult is also a prominent feature of EPAM's behavior. In a recent ex-syllable lists of high and low intra-listClarity originally used by Underwood in his experiments on thephenorenon), the low-similarity list was learned to criterion lnei*ttrlS£ whiie the high-similarity list was leaded to criterion ineighteen trials. Discrimination is a very time-consuming process

in

DAM and the learning of high-similarity lists entails a greatdeal

S! Scrlminluon between items than the learning of low-similaritylists.

k Item Duplication. EPAM cannot learn a serial list of itemsin which thS «£?«« Ipiears twice. A hu__*n subject can learn such" l2t thougn not as easily as a standard list.

In responding EPAM

Ci£ e«SSfto oscillate between the different assoc istes of he re-lated item. Human subjects learn to distinguish between differentof the duplicated item &*>*«__**".." „ qtiDoaed to the occurrence "late lo the He. The cit.iSL aie^S. rtt_ the Kiel 1. «erlou. "* »«■

Symposium on Simulation Models

108

to the storage of more cue information, and the association is there-by reestablished.

c) There is no separate mechanism in EPAM to account for for-getting. The phenomenon is purely a consequence of a simple type ofassociation memory structure and some basic processes for discrimin-ation and association of symbols.

In EPAM, oscillation and retroactive inhibition phenomena areseen to bt intimately related. Oscillation is caused by the intra-list interference of items; retroactive inhibition is caused, in pre-cisely the same way, by the inter-list interference of items.

A number of hypotheses relating the degree of intra-list simi-larity to the amount of retroactive inhibition have emerged from EPAMexperiments. They are as follows: \

a) The learning of a list with low intra-list similarity ofitems, following the learning of a list of high intra-list similarityof items, produces low retroactive inhibition.

b) The reverse procedure (high- similarity list following low-similarity list) produces high retroactive inhibition.

c) If tyro high-similarity lists are learned, there is low retro-active inhibition.

d) The learning of two low-similarity lists produces an amountof retroactive inhibition intermediate between that produced in thea) and b) situations.

Psychologists have identified another type of forgetting calledproactive inhibition, the interference of items earlier learned withitems later learned. In a typical experiment, a control group learnsa list B, then is retested on B. An experimental group learns a listA; then learn3a list B; then is retested on B. Any decrement in thelist B retention in the experimental group is called proactive in-hibition. There is no proactive inhibition observable in the behaviorof EPAM. This phenomenon, we believe, derives from long-term memoryprocesses which go beyond the scope of the present model. There is asimple heuristic argument which will account for the phenomenon inEPAM-llke terms, but the problem needs further study and will be re-ported on elsewhere in the near future.

EPAM does, however, exhibit some proactive effects during learn-ing. The effect is on the difficulty of learning a given list ofitems. A given list of nonsense syllables is more difficult for EPAMto learn if it is learned after another list has just been learned.In one experiment, using lists of high intra-list similarity, a givenlist was learned in fifteen trials when learned as the first list ofa pair; it was learned in eighteen trials when it occurred as thesecond list of the pair. Processing in a small net (with only alittle information stored in it) is easier than processing in a larger

*

Simulation of Cognitive Processes

109

net (containing much stored information).

6. Reading. EPAM has simulated an elementary form of reading.The learning phase of the experiment consisted of two paired-associatetasks. In the first, EPAM learned associations between the character-istics of visually-presented objects (a dog, a cat, a car, and a ball)and the aurally-presented names of these objects (DAWG, KAT, KAHR,BAWL) In the second, EPAM learned associations between the visually-presented names (DOG,

CAT, CAR,

BALL) and the aurally-presented names.In the performance phase of the experiment, the visual-symbolic nameswere given, and EPAM successfully (on the first trial) pointed tothe correct objects in the set of visual objects it had learned. Itdid this by going throu^i a "mediating response," an internal re-sponse in the aural mode. Since humans learn a great many associa-tions between sound-nanes and visual objects early in life, and sincewe almost always learn the visual-symbolic name of an object in as-sociation with its aural name (because we, or our teacher, pronouncesit), we may be using internal aural "mediating" responses a great dealof the time in reading.

A detailed treatment of the EPAM reading experiment is given inItem 3 of the Bibliography.

One of the virtues of the simulation approach is that it gl^esthe theory builder reasonably precise inforsation on the sufficiency(or insufficiency) of his theory. We have formulated and have begunto attack some of the problems vith the EPAM model revealed by theexperimentation described above. We shall merely mention the majorproblems here.

The preceding section that described experiments with item du-plication on serial lists raised the question of symbol-associationsversus token-associations. 3y an internal symbol we shall mean thememorized internal encoding of an external stimulus configuration.By an internal token we shall mean a symbol occurrence manufacturedfor use in a particular instance (e.g., the token might be a completecopy of the symbol, a partial copy, etc.). Symbols are basic entitiesto the EPAM learning system; they contain the information necessaryfor recognition of stimuli, output of responses, etc. Symbols occurrepeatedly in many different learning contexts. To allow them to betied together in association vith each other (as is done in thepresent EPAM model) is to allow the learning mechanism to be tiedinto knots, for each symbol can be associated only once with anothersymbol. The problem of alloying associations to and from a symbol ina variety of contexts is solved by postulating a process that createstokens of symbols for association in particular contexts. Tokens thenbecome the associated entities. A token can be informationally quitesparse. All the information that is really necessary in the tokenis that information just sufficient to retrieve the actual symbol fromthe discrimination net in which the symbol is stored.

Problems Raised by the EPAM Experiments

Symposium on Simulation Models

»

110

The experiments with serial learning and paired-associatelearning have led us to believe that the EPAM processes do not useenough "serial" properties in serial learning. At present EPAM treatsserial learning as a special kind of paired-associate learning. Wenow believe serial learning and paired-associate learning to be quitedifferent information processing tasks— that serial ordering andserial cues make serial learning much easier than the same learningdone in pure paired-associate style. Perhaps this is because humansdeal so much, in their symbol manipulation, with serial strings ofsymbols, and therefore have learned efficient serial processingmechanisms. The modification of EPAM currently in progress will havebasic processes specifically designed to learn serial orderings ofstimuli .

Finally, we have been forced to deal vith the problem of thelearning of meaningful stimuli. Objects are "meaningful" if they (ortheir sub-objects) can be recognized as being associated with objectsalready learned. Meaningful objects are learned more easily thannon-meaningful objects because the former do not have to be learnedcompletely from elementary properties (or symbols), but can be learnedas structures of associated objects previously learned. A zebra,when first seen, is really quite a meaningful object. It is a horsewith stripes. The emerging new EPAM system incorporates processes forbuilding up complex objects from previously learned sub-objects. Themechanism seems to be a simple means by vhich previous learning canbe brou#_t to bear in a useful vay on current learning.

EPAM Bibliography

The references given belov are detailed expositions of the modeland various experiments summarized in this paper.

1. Feigenbaum, E. A. , "An Information Processing Theory of VerbalLearning," The RAND Corporation Paper, P-1817,

October,

1959.2. Feigenbaum, E. A., "!Ihe Simulation of Verbal Learning Behavior,"

Proceedings of the Western Joint Computer

Conference,

Vol. 19,1961, pp. 121-132.

3. Feigenbaum, E. A. and H A.

Simon,

"Performance of a Reading Taskby an Elementary Perceiving and Memorizing Program," The RAMDCorporation, P-2358, July, 2361.

k. Feigenbaum, E. A. and H. A.

Simon,

"Forgetting in an AssociationMemory," Proceedings of the 1961 National Conference of theAssociation for Computing Machinery, Vol. 16, pp. 202-205.