3

1.2 Successive Approximations to a Modelfor Short-Term Memory*

GEORGE SPERLING

about 4 or 5 of the letters (Cattell, 1885). Thesimplest model for the action of reproducingvisually presented letters might be organizedinto two main components: (1) a visual mem-

Introduction

One major difficulty in studying human into two main components: (1) a visual mem-memory is that we have not yet learned how ory containing the letters (called visual infor-to obtain systematic physiological information.This means that the only technique availablefor the study of human memory is to presentthe subject with a variety of memory tasks, and

Marion storage) and (2) a translation compo-nent, which can translate a visual image of theletters into a series of motor actions; namely,copying the letters onto a piece ofpaper (Figure

then to record his actions and the accuracy of 1). The limited memory span of the subjecthis performance. From these observations, we might be represented in the model by progres-try to abstract the functions or operations that sive deterioration— a fading into illegibility—ofthe subject performs on the to-be-remembered the contents of visual storage. While the sub-stimulus in order to produce the observed per-formance.

ject is writing, the contents of his visual mem-ory are decaying, so that when he finally comes

Because of the complexity and subtlety of to write the fifth or sixth letter his visual mem-human behavior, it seemed desirable to us to

confine ourselves initially to a simple memoryory of the stimulus no longer is legible.

Without elaborating further on the diffi-task. A subject looks at a row ofrandom letters culties of Model 1, we can reject it immediatelyand then writes them down. If this situation for one basic reason: before the subject beginswere understood, perhaps the principles could to write the letters, his visual image of the let-be generalized to more complex tasks.

Models LightPattern

WrittenLetters-+-oOutputMODEL 1

When a row of letters is exposed briefly, i.e.,for l/20th sec, an adult subject can reproduce

Figure 1 Model 1. The large box represents thesubject. Arrows indicate the direction of informationflow. The components are visual information storage(VIS) and a translator. The translator converts aninput (the memory of a letter) into an output (aseries of motor actions) which result in a written

* Originally presented at the symposium on "MemoryandAction," XVIII International Congress of Psychology, Mos-cow, 1966. It was published inActa Psycbofogia, 1967,vol. 27,pp. 285-292. The first person plural is used to refer at varioustimes to Mrs. Susan A. Speetli, Mrs. Mary W. Helms, andDr. Roseanne G. Speelman, whose able assistance is acknowl-edged thereby. representation of the letter.

32

» ■

SPERLING-SUCCESSIVE APPROXIMATIONS FOR SHORT-TERM MEMORY J3

]

.

rM * ■

i

IC

Ed

>r-

10-

heIv,ire

:ct

ofIb-n-

les

fri-llyMs

et-

:nIts

iiit

the:ionage

an(a

:tcn

ters has already disappeared. (The measurement

of the duration of visual storage is describedbelow.) Having shown that letters are not

stored visually until they are reproduced, wemust now determine the form in which theyare stored.

MODEL 2

Occasionally a subject, when he is writingdown letters, can be heard to mumble the let-ters as he is writing them. His tendency to saythe letters aloud can be emphasized by playingloud noise into his ears. Noise itself does not

seem to alter performance in any other signifi-cant way. We have used this technique, to-

gether with a microphone placed near thesubject's mouth, to record the actual letters thesubject is saying. We also recorded automat-ically whenever the subject was writing. Themost interesting results with this techniquearc obtained when the subject is required to

wait (e.g., for 20 sec) after the stimulas expo-sure before writing the letters. He repeats (re-hearses) the entire letter sequence several timeswith a pause between each repetitior. duringthe interval. Then, at the time of writing eachletter, he also may speak it simultaneously.

Rehearsal suggests an obvious memorymech-anism. The subject says a letter, hears himselfsaying it, and then remembers the auditoryimage. As the auditory image fades, he repeatsit to refresh it. Most of our subjects do notvocalize during recall, but they all concur instating that they rehearse subvocally. Therefore,we assume that the sound-image of a letter can

LightPattern

enter auditory memory directly from subvocalrehearsal without the necessity ofactually beingconverted into sound and passing into the ex-ternal world. These relations are illustrated inFigure 2.

The auditory nature of subvocal rehearsalcan be emphasized by playing distractingspeechinto one's ears during rehearsal. The speechseems to emanate from one set of locations inspace (the ears) while one's rehearsal is heardas an internal voice speaking from the center

of the head. External sound also can be used asa clock against which to measure the rate ofsubvocal rehearsal. Another method ofmeasur-ing the rate of subvocal rehearsal is to ask sub-jects to rehearse a sequence of letters subvocally10 times and to signal when through. This maybe compared to a vocal rehearsal of the samesequence. All these indirect measures ofthe rate

of subvocal rehearsal indicate that, while it maybe slightly faster than vocal rehearsal, it is basi-cally the same process (cf. Landauer, 1962). Themaximum possible rate is about 6 letters persecond but, in memory experiments, maximumrates of about 3 letters per second are moretypical.

The existence of auditory memory in visualreproduction tasks also may be inferred fromthe deterioration in performance which occurswhen the stimulus letters sound alike (B, C, D,etc.) We have studied a large variety of tasksin which stimuli were presented visually orauditorily and found almost the same rule to

apply to both modalities of presentation. Whenthe memory load is small (about 2.5 letters inan auditory task, 3 letters in a visual task) it

WrittenLetters

FIGURE 2 Model 2. VIS = visual information storage, AIS = auditory information storage,T = translator.

SHORT-TERM VISUAL STORAGE MODELS AND EVIDENCE31 I

'iII

I

*

makes little difference to performance whetherthe stimulus letters sound alike or sound differ-ent. Additional letters beyond the minimalnumber are remembered only about halfas wellwhen they sound alike as when they sound dif-ferent. This dependence of performance on thesound of letters—even in a task which nomi-nally involves only looking and writing—is ofpractical as well as of theoretical importance(Conrad, 1963; Sperling, 1963).

According to Model 2, stimulus letters firstare retained in visual storage. They are re-hearsed, one at a time (i.e., converted from avisual to an auditory form), and then remem-bered in auditory storage. Subsequently theymay be rehearsed again and again as requireduntil they are written down. The limits on per-formance may arise either from the limitedduration visual storage (so that some letters de-cay before they can be rehearsed) or from thelimited capacity of the rehearsal-auditor)' stor-

age loop, depending on the stimulating condi-tions.

Attractive as Model 2 seems, it is inadequatefor the following reason: it is possible to gen-erate an image in visual storage which has aduration of definitely less than .1 sec and fromwhich 3 letters can be reported. This would re-quire a rehearsal rate of over 30 letters persecond, which clearly is completely beyond thecapabilities of the rehearsal processes describedfor Model 2. Before considering Model 3, weneed to examine in more detail some propertiesof visual information storage.

SHORT DURATION VISUALIMAGESWhen the contents of an image in visual

storage exceed four or five items, the)- can bemeasured by a sampling technique which re-quires the subject to report only a part of thecontents (Sperling, I960). For example, by thistechnique it was shown that the visual imageinduced by a Vzotb. sec exposure may contain asmany as 18 unrelated letters, and that as manyas 10 items may still remain 2 sec after the ex-posure. The visual image of shortest durationthat we have measured by this technique wasproduced by a stimulus exposure of M>oth sec,

preceded and followed by bright white fields.Immediately after the exposure, 14 letters werecontained in visual storage; within 1/2 sec theyhad vanished (Averbach and Sperling, I960).To produce and to measure really short dura-tion images, however, different methods arerequired.

In a letter-noise stimulus sequence, a second,interfering, stimulus (visual "noise") is exposedimmediately on termination ofthe letccr stimu-lus. The duration of the letter images can beestimated by comparing them to an auditorysignal. Two different methods were used. In thefirst method two clicks were produced at theears of the subject. He then adjusted the in-terval between the clicks until the auditoryinterval was judged equal to the visual dura-tion. In the second method, the subject heardonly one click at a time. He adjusted this clickto occur so that it coincided subjectively withthe onset of the visual image. After this judg-ment was complete, he made another adjust-ment of the click to coincide with the termina-tion of the visual image. The measured intervalbetween clicks—taken to be the duration of thevisual image—was the same by both methods.The apparent image duration of the letters in aletter-noise sequence is zero for extremely briefexposures (e.g., less than 10 msec) and thenincreases linearly with increasing exposure dura-tion for durations exceeding about 20 msec(Figure 3a).

When stimuli of 5 letters, followed by noise,are exposed for various durations, the accuracyof report increases with exposure duration asshown in Figure 3b. The most interesting aspectof these data is revealed by analyzing separatelythe accuracy of report at each of the 5 locations(Figure 3c). The accuracy of report at each lo-cation reported increases continuously as afunction of exposure duration. For this subject,the order of the successive locations which arereported correctly is generally left-to-right (I to

V ), except that location Vis reported correctlyat shorter exposures than location IV. Othersubjects have different idiosyncratic orders, e.g.,I, V, 111, 11, IV. By definition, in a purely serialprocess the «,h location is not reported betterthan chance until the exposure duration at

35SPERLING-SUCCESSIVE APPROXIMATIONS FOR SHORT-TERM MEMORY

sures, may be interpreted as evidence of anessentially parallel process for letter-recogni-tion.1 This process gives the illusion of beingserial because the different locations mature at

different rates (cf. Glezer and Nevskaia, 1964;

Sperling, 1963). These findings are taken into

account in Model 3 (Figure 4).

MODEL 3In Model 3, the scan-rehearsal component of

Model 2 is subdivided into three separate com-

ponents. The first of these is a scan componentwhich determines—within a limitedrange— the

sequence of locations from which informationis entered into subsequent components. The

extent to which the subject can vary his orderof scanning is a current research problem. Invery brief exposures, the variation in scanningmay be limited to changing the rate ofacquisi-tion at different locations—information process-ing beginning simultaneously at all locations.On the other hand, the overall rate of informa-tion flow through the scanner must be limited.

The second new component is the recogni-tion buffer-memory. It converts the visual im-

Exposure Duration (msec)■

;

Figurk 3 (a) The apparent duration ofthe letters in

a letter-noise presentation. Abscissa is the exposureduration of the letters. Ordinate is the duration ofan

interval between clicks which was judgedequal to thevisual letters. Data for one typical subject, (b) The

total number of letters reported correctly. Five letters

were presented, one in each location, I to V (c) Thepercent of letters reported correctly, shown individ-ually for each location, I to V.

age of a letter provided by the scanner into a

"program of motor-instructions," and stores

these instructions. This program of motor in-structions, when it is executed by the rehearsalcomponent, constitutes rehearsal. The impor-

i An alternative interpretation is that the scan is serial but

the order varies from trial to trial and/or there is great van-

ability in the processing time per item. We tentatively con-

sider that interpretation to be unlikely because (a) the modalscan pattern is highly repeatable from session to session and(b) parallel processing of other aspects of the stimulus is

occurring, e.g., of its orientation, overall length, brightness,

which the >i~A location is reported with max-

imum accuracy is exceeded. The observationthat all locations begin to be reported at betterthan chance levels even at the briefest expo- etc.

iL

j:t

b

WrittenLetters:t

J:r recognition buffer-memoryIt

I

(c)

FIGURE A Mode! 3. (See Figure 2) R-buffer =

t

I36 SHORT-TERM VISUAL STORAGE MODELS AND EVIDENCE

capable of being verbalized or acted upon.Within the limits ofthe tasks for which Model3 was proposed, we can identify the contents

of the scan component with the contents ofconsciousness. This is because the scan com-ponent contains the information upon whichactions are performed.

tant idea embodied in the recognition buffer-memory is that the program of motor-instruc-tions for a rehearsal can be set up in a veryshort time (e.g., 50 msec for 3 letters) com-pared to the time necessary to execute it (e.g.,500 msec for 3 letters).

The recognition buffer is efficient partly be-cause the programs for rehearsing several letterscan be set up in it simultaneously. However,the major gain in speed derives from the as-sumption that setting up a program to rehearsea letter is inherently a faster process than exe-cuting the program, i.e., rehearsing the letter.In fact, the biological organization of motor

systems is extremely hierarchical. Thus the pro-gram in the recognition buffer could be a pro-gram to call a program, etc., and the ultimaterepresentation at the top of such a pyramidcould be called quickly.

There are several inferences to be drawn fromthis identification. When contents of visualmemory are not scanned before they fade away,they never become conscious. And, we are un-conscious of all contents of our auditory mem-ory except those being scanned. Another infer-ence is that if the contents of a memory cannot

be scanned, they are not accessible to conscious-ness.2 The untransformed contents ofthe recog-nition buffer-memory are not accessible to scan-ning and therefore never the objects of con-sciousness. This makes it indeed a mysteriouscomponent; it cannot be observed directlycither from within or from without! However,this inaccessibility should not surprise us. It isaxiomatic that in any system which examinesitself there ultimately must be some part ofthemechanism which is inaccessible to examina-tion from within. The recognition buffer-mem-ory is such a part in the human memorymechanism.

i

The rehearsal component executes the re-hearsal, which then is entered and rememberedtemporarily in auditory storage. The memoryof the rehearsal in auditory storage is scanned,the auditory image is converted to motor-instructions in the recognition buffer, and asecond rehearsal is executed. This loop con-tinues until the response is called for and theletters are written down. Iknow almost nothingabout the translation of the memory ofa letterto its written representation except that itoccurs, and therefore must berepresented in themodel. It has been represented in parallel withrehearsal because writing a letter so often is ac-companied by vocalization.

i

iSummary

Experimental data are considered from asimple task in which an observer looks at lettersand then writes them down. Three models arcproposed. Model 1 consists of only two com-ponents: a visual memory for the letters and amotor translation component to enable copyinga visual memory onto paper. Model 1 is inade-quate because the visual image is shown not to

persist until the time ofreproduction. Model 2

corrects this deficiency by incorporating the

! j

Consciousnessin the Memory Models

One can know the contents of the conscious-ness of another individual only insofar as theyare expressed by his behavior, particularly byhis verbal behavior. In the models, this struc-

ture would induce us to look for evidence ofconsciousness at the level of the rehearsal unit.However, one also must admit that a personwho is unable to speak or act may still retainconsciousness. The critical aspect of the con-tents of consciousness is that they normally are

'2 In conceptualizing Sternberg's experiments, it is usefulto assume that therecognition buffer-memory can be scannedfor a minor aspect of its content, e.g., whether it containsanitem which was entered much more recently than the others.A dotted line has been drawn from therecognition buffer to

the scan component to indicate thepossibility of this kind ofscan.

TT^

„.:

37MACKWORTH-THE VISUAL IMAGE AND POST-PERCEPTUAL IMMEDIATE MEMORY

i «**''

. '

1

'■

la 1f v

I ;:.-t-'-i,-

Js'CE

iron,

[oddtentsc> ofjcom-hich

fromfisuali-.vav.1c un-pem-tnfer-Innotcious-ecog-scan-

■ i)n--

fcrii msteethLever,

15 h is

mines[of thepnina-■mem-;morv

torn aletters

[pis areh com-

"* and a

ppying

1 not to

odel 2ag the

fIS usefulC scanned

ir.r ,in> .inbe others.buffer to

:i

possibility of subvocal rehearsal of the stimulusletters and an auditory memory for the re- Referenceshearsal. However, Model 2 cannot account forperformance with extremely short duration £ Q 1%1 ,n c cherry

images because of the limit on the maximum (Rd lnfomilt;on theory, London: Butterworth,

rehearsal rate. The critical improvement in 196_.2U.Model 3is a more detailed specification of scan- Cattell, J. McK., 1885. Philos. Stud., 2, 635-650.

form of memory which is inherent in the proc- of USSRj w> yu_714ess of recognition itself. Model 3 accounts for Landauer> T. X., 1962. ft^/. A^ 646.

these data and incidently gives rise to some in- Sperling, G, i960. Psychol. Monogr., 74, No. 11,

terestine inferences about the nature of con- (Whole No. 498)tcrc:,uuS Sperling, G., 1963. Human Factors, 5, 19-31.sciousness.

1.3 The Relation between the Visual Image

and Post-perceptual Immediate Memory*

JANE F. MACKWORTH

,vcral authors have suggested that incoming It ££?%£££%stimuli are stored in

„ory span and the rate of read-Hmmary very brief stor gc follo^edby a sde

Brencr (1940) has shown that the

worth, UV, spcnuig, y ; 1 »shanes show the shortest memory span. Mack-

rllf. nreliminarv storage of visual information snapes snow ; 1

£JSButL not clearly distinguish worth (^) - rcportccl .at grates of

between this and the classical concept 01mm. read fastest. Sampsondiate memory. He showed that much m ame ordy g g^information was £ "^°J ilLtional digits were used asvisual image than the 5 oul orn 1> c

port, and Averbach and Sperling (.uoi; sug

ab», 0.2, sec Mackwor.t, (1963) showed te «* fa=. The to P^,„e duea.lc, of =»» * "'; "f f"!So,, twee,, .he visual In.age eep*,was constant at about 1-.! sec so that the num immediateber of items -ported depended on the rate at

A hypothesis iswhich they could be identified. for rclationships>

*>.nw/ ofVerbal Learning andVerbal Behavior, 1963, vol. experiments are described which investi-

LSK^^^SS^Sit^. ff" relationship between the interference