I's).:hoIogicai Review ~t 1988 by the American Ps).:boIO8icaI A8IOciaIioo. Inc.
1988, Vol. 95. No.4. 528-551 0033-295X/88/$OO.75
Judgmentsof Frequencyand Recognition Memory
in a Multiple-Trace Memory Model
L.
Douglas Hintzman
University of Oregon
The multiple-trace simulation model, MINERVA 2, was applied to a number of phenomena found in
experiments on relative and absolute judgments of frequency, and forced-choice and yes-no recogni-
tion memory. How the basic model riate simulations; attempts to modify the
model to deal with additional phenomena ~e also described. Questions related to the representa-
tion of frequency are addressed, and the model is ~uated and compared with related models of
frequency judgments and recognition memory.
Although memory for specificevents(episodic memory) and marily concernedwith similarity, repetition, and retrieval. The
memory for abstract concepts (generic memory) seem quite secondsectiondescribes how the model accountsfor severalex-
different intuitively, experimental evidencefor different under- perimental results that havebeen reported in the literature on
lying syStemsis sparse(see McKoon, Ratcliff, & Dell, 1986; memory for frequency and recognition memory. In the third
Ratcliff & McKoon, 1986; Tulving, 1986).One suggestionhas section, new experimentsare presentedthat test predictions of I
beenthat the two systemsare affecteddifferently by repetition, the model concerning similarity and recognition memory. The
with multiple occurrencesestablishing multiple traces in epi- fourth sectiondescribesa slightly more elaborateversionof the
sodic memory but strengtheninga single representationin ge- model that includes an intertrace resonanceprocess(Hintz-
neric memory. A primary purpose behind the simulation man, I 986b) and applies this model to further resultson mem-
v
model, MINER A 2 (Hintzman, 1984),is to test this notion indi- ory for frequency.The fifth sectionbriefly di~ attempts to
rectly by attempting to account for performance in both epi- deal with additional phenomena by constructing special ver-
sodic and generic memory (Hintzman, 1978) tasks using the sions of the model, with varying success. Finally, the general
same multiple-trace mechanism" Application of the model to discussion evaluatesthe model and compares it with related
generic memory has focused on concept learning, as repre- medels of the sametasks, and addresses issuesconcerning the
sentedin the laboratory by the schemaabstraction, or classifi- representationand encoding of frequency information and its
cation learning task (Hil)tzman, 1986b)"The presentarticle de- relation to recognition memory.
scribeshow the model can be applied to memory for presenta-
tion frequency-a quintessentiallyepisodic memory task-and
to recognition memory, which is treated as a special case of The Model
memoryfor frequency. " "
'.
MINERVA 2 IS pn " marily concern edWIth long-tenD
" or secon d -
2,
The model, MINERVA ISan outgrowth of theoretIcal Ideas .
although there ISalsoassumed t 0 be a tem r
. ".
regardingeffectsof repetItIon on memory that havebeenstated ary memory,. "" po ary
I . I I h (H "
essngorous y e sew ere Intzman, 1976 ; H Intzman & BIock ,
" buffer store or pnmary memory, whosefunction ISto commu-
. I
mcateWith seco ndary memory.Asam ultIpe-tracem odel
" "
.
1971 H ' tz G d &G Id 1981) Beca th ". ,MIN-
, In man, ran y, 0, . use e Insplra- tha h . cedevent ISrepresen In ted "
tI" fi th "deas fi " " hi h b. ERV 2 assumes t eac expenen
A "
. on or eseI
d fi came romf expenments In w c su ~ects
", " memory by Its own tr ace. From a theor etJ perspective, sec-
' " caJ .
jU ge rom memory aspects 0 an Item s presentation-most
. I I " fi dary " t collect! . f isodi .c memory
. . " .
partlCUar y, Its requencY-lt IS Important In evaluatIng th e
" on memory ISseenas a vas
tr aces,the majon ty 0f whi ch were fi
"'
on 0 ep
m
od h
e to esta IS how weII It deals WIth data 0btalned In
I "bl " . '" OrIned ou~ +~; expen- de the "
"
t
men tal con tex. A s will be shown Iat er, however. con text uaI in- "
frequencY-judgment
Th . tI" expenments. "
fth t am I " fi II Th fi formatIon .. .
. specifiedIn the retrieval cue can greatly suppress the
e organlza on 0 e presen
, c e ISas 0 ows: e rst actIvati" on 0 f tracesfi ed In nonspecl"fied COIl x" Th us, In
. te ts '
od I, bas" hi h
o
th ' " OrIn
sectIonpresents erne s ICassumptions,w c are pn- ." tal
pnnClpIe, the ellt 0f extr aexpenmen tr aceson expenmen-
" "
"""'"
tal performancecan be reducedarbitrarily closeto zero, simply
by increasingthe amount of contextual information in the re-
was by
Thisresearch supported NationalScience FoundationGrants trieval cue. To makesimulation man~~ble, therefore,extraex-
BNS- and
7824987 BNS-8403258" perimental traces were ignored in the present simuiatiOllS, on
Correspondence this be to
concerning articleshould addressed Doug- the assumption that they would haveonly negligible effectson
lasL. Hintzman, of University Oregon,
Department Psychology, of Eu- performanceon the experimental tasks"
gene,Oregon 97403. For mathematical simplicity, a specific event is represented
528
(j
.
, -
. MULnPLE-TRACE MODEL 529
!
Probe The similarity of a given trace, i, to the probe is given by
N
Si = L ~Ti,j/Ni, (I)
i-i
where Pj is the value of featurej in the probe, Ti,j is the value
of featurej in trace i, and Ni is the number of featuresrelevant
to the comparisonof the probe and trace i, (Featurejis relevant
if either Pj ", 0 or Ti,j'" 0; thus, Ni = N - Zi, where Zi is the
number of features for which both Pj = 0 and Ti,j = 0.) The
numerator of this function is a versionofTversky's (1977) simi-
larity metric. Si behaves much like a Pearsonr, being zero when
trace i is orthogonal to the probe and + I when the two are iden-
tical, ValuesofSi approaching -I are mathematically possible,
but they are extremely unlikely and would have no particular
theoretical meaningin the work presentedhereo
The degreeto which a trace is activatedis a positively acceler-
ated function of its similarity to the probe. The simulations re-
ported hereusedthe activation function,
Ai = Sr, (2)
~A (Ect.. Int...lty)
The nonlinearity of the Ai function allows retrieval to be quite
of in
Figure1. Activation traces secondary memoryby a probe.(The selective:In principle, all secondary memory traces are acti-
of by
levelof activation eachtrace,Ai, is determined its feature-by- of
vated by the probe, but the response secondarymemory asa
featuresimilarityto the probe,Echointensityis the sumof theAi val- whole is dominated by thosetracesthat most closelymatch the
ues.) the
probe. Note that the expressionfor Ai preserves sign ofSi,
so that a trace can have negativeactivation. The relation be-
tween and Si is shownin Figure 2. The ~ative
Ai rangeof the
2
in MINERVA as a vector of feature loadings having the values function in the presentsimulations is roughly that contained in
+ I, 0, and -I. The array labeled probe at the t~ of Figure I the unshadedarea of the grapho
showshow an event is representedo There are N features,j = The simultaneousactivation of all tracesby a probe produces
I . , . N ordered from left to right, and every feature is assigned an echo that has two pr~es, intensity and content.The in-
a value in the vector representingeachevent,A value of 0 indi- tensity of the echo is found by summing the activation levelsof
catesthat, for the event in question, the indicated feature is ei- all traces:
ther irrelevant or unknownoOne could view the elementsof the M
vector as connections, linking the feature nodes at a lower level I = L Ai, (3)
with a single-eventnode at a higher level. Valuesof + I and -I i-I
couldo as
then be interpreted, respectively, excitatory and inhibi- where M is the number of traces in secondary memory. The
tory hn~. .. 0 more tracesthere are that match the probe and the more closely
Encoding an event entails ~ng the event vector Into sec-
ondary memory, representedby the large box in Figure 10In
the model, each individual feature is stored with probability
L the learnl
" ng rate and so encodingmay be im l"ect , . If a
O
., 1.0 00."',",.0,"'.00."0,0," .:.
::0::-::'::'::0::0:: .
particularfeature not stored, valueentered thetrace
is the into ~ :::~:~::~:::~:~:~:::~:~::::
:~: .
is O.The parameterL is applied independentlyto each feature
in every eventoThus, when 0 A} = L Pr{IA = k}. [Pr{IB> k} + .5.Pr{IB = k}],
i
where IA and IB are the intensities produced by the A and B
items, respectively, and k indexesthe intervals.
Applying this rule to the distributions of Figure 4 yielded the
-.5 .0 .5 1.0 1.5 forced-choice data shown in the main panel of Figure 5. The
general pattern is that performance improves with increasing
Echo Intensity differencebetweenthe ~ but
and smaller frequencies; uthe
differenceis held constant,performancedeclinesasthe two (re-
echointensity
Figure4. Typical for
distributions testprobes
having
fre- .. This descri tJ . wm cal f freq
' ..:~
. f0-5 quenCles Increase. p on 15 ',7Y' 0 uency~
quenCles 0 . .
cnJDlnatJonaccuracydata. For compansoo purposes,theInset
. . . .
'.
of Figure 5 showsdata from
Gold (1983),who tested an experiment by Hintzman and
with one
subjects two instructions: to
interval widths of .067. (d) The entire procedure was repeated choosethe item with the ~ frequency and one to choosethe
1,000 times for 1,000 simulated subjects.The resulting distri- item with the smaller frequency.Becausethe results suggested
butions of intensity valuesare shownin Figure 4. It is clear from that the two different wordings may haveinduced ~posite re-
Figure 4 that the mean and variance increasewith frequency. sponse biases,the data from the two conditions havebeencom-
~
The model's structure constrains relations among intensity bined in Figure 5.
distributions, but parameter settings affect the distributions'
quantitative characteristics. Exploration of the effects of four
main parametersof the model revealthe following: 40
I. With a rise in the learning rate, L, differences among
meansincreaseand variancesdecrease
Equation I).
Ni
2. As N increases, also increases,
(due to increasingNI in
and so Si and Ai become
3D
h
D
more stable.Thus, although meansare not affectedby increas- 40
ing N, variancesdecline. ,
Hi.urn..
. (1983)
Gold
3. The greaterare Pr{ + I} and Pr{ -I}, the larger15Ni.Thus,
of Pr{
theeffects increasing + I} andPr{-I} aresimilarto those S 30
ofincreasingN. t
- 4. When Pr{ + I} = Pr{ -I}, as in all simulations reported UJ
here, randomly generateditems tend to be orthogonal to one ~ 1D
another. the ratio of Pr{+ I} to Pr{-I} drifts away
As from I ~ 20
in either direction, items become more similar to one another C
on average. Mean I values rise and so do variances-the latter ~
causinga generalincreasein overlapamongdistributions.
Implicit in thesegeneralizationsis the fact that there is con-
e
Q.
10
siderable trade-off among parameters. The primary determi-
nant of performanceon both frequency-judgmentand recogni-
tion tasks is the overlap among distributions, and this can be 0
influenced by manipulating Lor N, or even Pr{ + I}, Pr{ -I}, 0 1 2 3 4
and Pr{O}. As a practical matter, then, none of the abovepa-
of
rametersare identifiable in the sense havingdistinctive effects Sma Iler Frequency
on task performance.This being the case,for most of the follow-
ing simulations the value of N was set to a convenient value at frequency-discrimination
Figure5, Simulated as
accuracy a functionof
the outset, and preliminary parameter adjustments ~ car- thesmallerandthelaIlerofthe~comparedfrequencies(parameter=
ried out only with L. l~ (Inset:
frequency). data.)
experimental
~
;:;;;:
532 DOUGLAS L. HINTZMAN
AbsoluteFrequencyJudgments 1.1 a 1kIok1ey, 1984
To predict numerical frequencyjudgments, one can assume .1
that the echo intensity scaleis partitioned by severalcriterion .1
c. a
values, Thus,if I > cs,the testitem is assigned frequency 1.0
judgment of 5; if C4 0), and therefore one the model was run on a frequency-discriminatioo task nearly
criterion. Recognition confidence ratings can be modeled by identical to the ooe that produced the data shown in Figure 5.
setting several of
criteria, for different degrees confidence,in the A total of 500 subjectsweresimulated on the task using a learn-
region in which the distributions for frequency = 0 and fre- ing rate of L = .60; subsequently, samewasdone using L =
the
quency = I overlap. .30. The learning rates were such that in the secondrun, the
In general,the model is consistentwith analysesof recogni- number of features stored ~ about ooe half that in the first
tion memory based on signal detection theory (e.g., Banks, run. This difference can be used to simulate f~ng, under
1970). The noise (frequency = 0) distribution originates in the the assumptioo that ooe cause of forgetting is trace decay. A
activation of traces of list items by new probes, or lures. Typi- learning rate of .30 is equivalent to learning with L = .60, foI-
cally, an individual trace will be only slightly activated by a lure, lowed by forgetting with probability .50, in which a "forgotten"
a
but becauseecho intensity is the sum of the Ai values, new to
featurevalueof + 1 or -1 reverts 0 (cf. Hintzman, 1986b).
producean intensityhigh enough
testitem will sometimes to the
Thelargepanelof Figure7 shows f~ng for
curves sev-
that
suggest theitem'straceis in secondarymemory. Thereare eralrepresentatm cooditionsfrom thesimulationrun.
severalinteresting consequences the way the noise distribu-
of There are well-known ~ in comparing forgettingcurves
by 2.
tion is produced MINERVA Two obviousones,explored variable(Loftus,
that fall at differentlevelsof the dependent
.
..~I
I MULTIPLE-TRACEMODEL 533
1.00 tions differently(cf. Hintzman, 1969;Hintzmanet al., 1981;
.90 '-- Hintzman& Stern,1984); evenwithin the sameexperiment,
~ ~2-D test instructions may affect the ordering (Hintzman & Gold,
~
.so ~~~4-1 1983). Third, the ordering of comparisons involving different
.70 ~ ~=: frequencies,such as 1-0 versus4-2, is certain to be very sensi-
1.00 2-1 in of
tive to subtlechanges the variancesand shapes the underly-
5-2 ~ .&O Hlntzm81, Stem. 1984 ing distributions. An example, demonstrated in the next sec-
3-1"""- 50 tion, is that different comparisons are affected differently by list
'0 .90 I~DIATE IELAYED length. Fourth and finally, the ordering of conditions in Figure
~
~ 4-1 7 wasobtained a valueof L, anddifferent
using constant order-
0 ~ /\ ings can be obtained under the plausible assumption that atten-
U .80 2-0 tion, and therefore L, declines systematically across repetitions.
C 3-2 although the discrepancies
For thesereasons, betweenorderings
.g " ..::::~ 4-2 and of 6 it
in themainpanel theinset Figure areworthnoting,
& .70 1-0 is not clear that they signify any fundamental problems with
0 2-1 the model.
...
Q. .&0
List-Length Effects
Although list length has been found to affect recognition
.50 memory in severalexperiments(e.g.,Bowles& Glanzer, 1983;
60% 30% Gillund & Shiffrin, 1984; Legge,Grosmann, & Pieper, 1984;
. Strong, 1912), studies of its effect on memory for frequency do
Features In Memory not seem to have been done. To explore the influence of list
Figure7.Simulated curves forced-choice
forgetting for (1-
recognition length on recognition memory and frequencydiscrimination in
and
0 and2-0 curves) frequency (all
discrimination o~). (Insetcon- MINERVA2, four simulation runs were compared. One con-
data.)
tainsexperimental sisted of the data plotted in Figure 5. In that simulation, four
replications of the frequencies 1-5 were stored in secondary
1985; Underwood, 1954, 1964). One solution is to generatea m~mory, for a total list length of60. The three additional simu-
family of forgetting curves encompassinga range of levels for lations used the same parameters,except that the numbers of
each of the conditions to be compared, so that different forget- replications stored were 1,2, and 3, yielding list lengths of 15,
i or
ting rates can appear as the convergence crossingof the two 30, and 45, respectively.Enough subjects were simulated in
setsof curves. This method was used by Hintzman and Stern eachof thesenew runs to give2,000 observationsper data point
(1984) to compare forgetting rates in forced-choicerecognition representative
(vs. 4,000 for list length = 60). Several conditions
and frequency discrimination. In their Experiment 2, Hintz- havebeenselectedfor display in Figure 8.
man and Stern (1984) had subjects make five kinds of fre-
~
quency-discrimination judgments, two of which involved fre-
quency = 0 items andwas
memory Testing donecorrespondedto 2 weeks
tests. thereforeeither min or recognition-
10 after 100:: ...~ - -
5-1
presentation of the list. The inset of Figure 7 shows the data. 95 ====~
Therewasno reliabledifference forgetting between
in rate rec- ~ :::::::::::: 2-D
(comparethe 2-0 curve with that of 4-1, for example,and 1-0
and
ognitiondecisions decisions involving higherfrequencies
. 0
... 90
5-2
with 4-2). () 85
Hintzman and Stern(1984) interpreted the data assuggesting & 4-2
that "the increments traces) by successive
(or left repetitions
are ! 80
all lost at the same rate" (p. 412). That statement accurately i 1-0
describes the decay process underlying the simulated data of
Figure 7; and becausethe simulated data mimic the human
0
~ 75 -~---
data in showingtwo essentiallyparallel setsof curves,the simu- a. 2-1
lation supportsthe interpretation givenby Hintzman and Stern 70
(1984). ;
2
It is apparent in Figure 7 that MINERVA did not order the 65
various conditions exactly as the human subjectsdid, and some 15 30 45 &0
comment is in order as to what this may mean. First, note that
Hintzman and Stern (1984) warned againstcomparisonsof the List Length
levelsof their curves(as~posed to the forgetting rates)because
counterbalancingacrosslevelswas incomplete. Second,differ- Figure8. Simulated of recognition
effects list lengthon forced-choice
ent frequency-discrimination experiments often order condi- and
(1-0 and2-0 curves) frequency (all
discrimination o~).
'-
..
~
534 DOUGLASL. HINTZMAN
is
Thebasicobservation that list lengthhada muchstronger 8 R_. 1974
effecton recognition
decisions and2-0) thanon discrimi-
(1-0
nations among nonzero frequencies.The apparent r~n
that recognition is
accuracy stronglyaffected thevarianceis
by of i
>-
..
6
4 '"
"Ii'
~Ai for nontarget
traces,
and this varianceincreases
linearly 1
Discrimination
withlistlength. ~requenci~
among ~eat~than ?: ~ 2 ,.,2-"'li'
zero, in contrast, depends more heaVIly on varIatIon In the .in.30
goodnessof encoding of the traces of the items being com- c a
. ~ 012 4 .
pared-a factor that list length doesnot affect. Nontarget traces -
do havean effect, but their influence decreases with increasing £
and =
frequency evenat frequency I is pr~rtionally small. ,.g.20
performance
As a result,recognition morerapidly
deteriorates 0
list
with increasing lengththandoesfrequency discrimination. W
Thedifference sl~ is evident theconvergence cross-
in in and
ing of the curves (providing a contrast to the null effectsof for- ~
~
.10
getting shownin Figure 7). This prediction of differential effects
of list length on recognition and frequency-discrimination ac-
curacy cannot be related directly to existing data, and therefore .00
requiresexperimental test.
OrientingTasks Frequency
. Sev~aI publishedstudies have failedto ~nd~ effeerations), 15-feature curve were only about 1.5 times as large as those
in
contained the present (B)
encoding. would~1ap substantially underlying the 100feature curve which is not great enough to
(A).
with informationfrom pastencodlngs In thiScase, when even . . .' . .. .
of
thereis a highdegree compatibility be~ thetraceB andthe OUtweighthe difference In sl~. Free(nontarget frequency).Thus, if there were no discrimina-
thesetwo techniques, and it is the one that has been explored tion of frequency according to list membership, the intensity
in the presentwork. It should be pointed out that the contextual increment for each unit on the abscissa would be the same as
features that are appr~riate to the target list are not necessarily the separationbetweenadjacent curves. If the ability were per-
the contextual features that would be present during testing. fect, the curves would be well separatedand flat. A simple dis-
Thus, in order for the preactivation schemeto work, the system crimination index can be calculated by taking the ratio of the
mustbe able to usethe instructions to retrieve the discriminat- variance among means accounted for by nontarget frequency
ing featuresfrom memory, and then to add them to the probe. to that accounted for by target frequency. Assuming linear
.
---
536 DOUGLAS L, HINTZMAN
& 1971
4 Hlntzm.. 81ock. versus external generation of the same items (Johnson, Raye,
5 & I in the
Wang, Taylor, 979)-could be modeled exactly same
way, Retrieval in the model is highly context-dependent where
~ 2 as informationis concerned
generic, well asepisodic, (Hintz-
man, 1986b), An important characteristic of MINERVA 2 is its
capacity to determine at the time of retrieval which subset of
traces of a particular item will be strongly activated, Frequency
0
-~~_.~-_.:-~'-,
o 2 5 4
judgments not haveto be prestored, canbe generated
do but
from memory on demand; in the realm of generic memory, the
same holds for concepts (Hintzman, I 986b), Jacoby and
3 Brooks ( 1984) discussed several advantages of viewing memory
~
()
- "~"""'--
, """", 2
in this way.
Recognition and Similarity
W The model has several implications for effects of similarity
C 1 on recognition memory. One is that echo intensity, and there-
: 0 fore the tendency to identify an item as old on a recognition-
~ memory test, should be enhanced if there are items similar to
DI -045 the test item in the list. The phenomenon of false recognition
, (e,g., Anisfeld & Knapp, I 968)-in which lures that are seman-
tically similar to old items are called old more often than are
control items-is consistent with the model. In this regard,
0 1 2 3 4 MINERVA 2 predicts that the tendency to identify a probe as old
. should increase with the number list items that partially match
Nontarget-List Frequency the probe, and this should hold for correct, as ~ll as for false
Figure 10. Discrimination be~n List 1 and List 2 frequenciesby the recognition, There is some evidence for this tendency where
model. (Echointensitiesto List I and List 2 probeshavebeencombined, false recognition is concerned (Hall & Kozloff, 1973), but the
A discrimination index, DI, of 0 indicates perfect list discrimination; I prediction appears not to have been tested in correct recogni-
indicates no discrimination. Inset contains experimental data: mean tion. Both effects are fundamental, as they are expected by sev-
frequencyjudgments.) eral theories of recognition memory (e.g., Bowles & Glanzer,
1983; Gillund & Shiffrin, 1984; Shepard, 1961; Underwood,
1965). Experiment I was designed to help fill this gap.
trends in both cases, the ratio is DI = rN2/rT2, where rN and rT
represent correlations of the means with nontarget and target Experiment I
frequencies, respectively. DI = 0 indicates perfect discrimina-
tion (i,e" no generalization from the nontarget list), and DI = I Method
indicates a complete failure to discriminate. As is shown in Fig- Materials. The experimental words ~ 288 familiar nouns (includ-
ure 10, the simulated data had a DI of .045, whereas DI for the ing prC4'Jernames),6 falling in eachof 48 semanticcategories. The cate-
Hintzman and Block (1971) experiment was .097, gories~ selectedfor high within-category and low ~-category
Briefly, the entire set of simulations showed the following: (a) similarity. Examples are booklet, pamphlet, comic book, periodical.
The more list tags that are used, the better is discrimination magazine, brochure; Scotch,rum, brand}\ vodka, whiskey.gin; Jessica
(e.g., a simulation identical to that in Figure 10 but using just Lange. Sissy S~~k. Van~ssa RedK!aw, Meryl Streep. Sally.Field,
four list tags yielded DI = .207). (b) The higher the learning D~bra Winger; mlnlstel; priest, rabbi, pasto~ 'preacher, parson;jacket,
t th bett . d ' , , b' ( I b. ,de .cal tha , shirt. coat, s~atel; blouse,dress;mouse,prairie dog.groundhog,MJod-
ra e, e er IS lSCf1mma on a Slmu a on I nb to t "'" .
f F ' I0 b . L - 50 ' Ided 0 - 06 chuck, gopher, chipmunk,' and Indiana, Wisconsin, Minnesota. IllinoIs,
0 Igure , ut USing -. yie I -, 5), (c) Con- Michigan I~.
structing probes using excitation alone is not as effective as The 48 'categories ~ diviOOdi~to four setsof 12, and each setwas
those using both excitation and inhibition (a simulation identi- a
assigned presentationfrequencyof 0, 1, 3, or 5, indicating the number
cal to that of Figure 10, but without inhibition, yielded DI = of different categorymembersto appearinti1e list. Words~ ~
.095). It is interesting that there appears to be no way to com- randomly in the list, with the constraint that ~ membersof the same
pletely eliminate generalization from the nontarget to the target An
categorycould not appear in closesuccession. additional 92 unre-
Jist by manipulating these parameters; even if nontarget echo lated filler no~ns ~ppeareda~~m lis.tpositions. The 200 ~ in
intensities are made negative by using inhibition and designat- the presentatIon list ~e pnnted In a ~ngle .colu~n, ex~ over
ing a high pr~rtion of features as list tags the curves relating four .pages,To the left. of each noun. was Iisted.lts seriaI.POSItIon?~ to
, , " , the nght was a blank line for the subject to use In recording an onentIng-
Intensity to nontarget frequency always have some poSItIve task response,A single recognition test list was constructed, listing 96
sl~.. '" , , words, randomly ordered and numbered sequentially on the left.
The capacity for selective retrieval IS not restricted to dis- The test list contained ~ words from eachcategory.For categories
criminating among lists. Source-specific frequency judg- having a frequency of 0, both words ~ new; and for those havinga
ments-for example, judgments of the frequency of internal frequency of 1,3, and 5, one word wasold and the other was new,Al-
.
.
. j
MODEL
MULTIPLE-TRACE 537
together,eight presentationlists ~e constructedaccordingto this pat- 100
.gh I. h ed d ed . 1
Experiment
for a category ~
tern. Across interchanged, and each category was rotated
the el t ISts, t e present an non present test Items 80
through the four frequency conditions. In all casesthe test list was the 60
same.
Subjects. There ~e 87 subjects,recruited for course credit from 40
undergraduatepsychologyclasses the University of Oregon. Subjects
~e tested groups.
in Approximately
at
equalnumbers
~e giveneach 20 ~~
/
list.
presentation . 0
Procedure. Each subject was given a booklet containing (a) instruc- =c ~==/--",
tions to rate nouns on an activitY scale ranging from I to 5, (b) the 0 80
presentationlist, (c) instructions for a filler task consistingof a sequence : Old ::
i ~f four paper-and-pencil mazes(~e mazes~e fairly ~fficult.and ~e -g ~~~-
Intended to occupy all of the subjectsfor at least 10 min. Subjects~e = 60 ~'~
told that if they finished the fourth maze before time wascalled by the ~ ;]ij
experimenter they ~e to ~t and not turn the page.),(d) the four ~
mazes,and (e) the recognition test page,which included the instruction ~ 1tt
to circle the number correspondingto each word that had occurred in tV 40
the presentationlist. To allow everysubjectan ~portunitY to finish the
..
~
activitY ratings and to provide a short additional retention interval, the ~
experimenter~ted 20 min betweenhanding out the booklets and tell- ~ 20
ing subjectsto st~ ~king the mazesand turn to the recognition test. c..
Results 0
: The data in the main panel of Figure 11are from a simulation 0 1 2 3 4 5
j that will be describedfollowing the presentationof Experiment . .
! 2. Theinsetof thefigure the and
shows hit rates falsealarm Category Members In List
I rates from Experi.ment 1. As was ex~ed, ~t rates and false hit
Figure11.Simulated andfalse for and
alarmrates related unrelated
alarm rates both Increasedmonotonically WIth the number of (Inset:
testitems. data I.)
corresponding from Experiment
same-category items in the list. For purposesof statistical analy-
sis,the data for each presentationlist were combined over sub-
j jects and were treated as a macrosubject. Although the linear .. E . O .
her
Ii
Ii
ed
48
'
h
.
tren s own y t rates was sm , It wasSlgnl cant, , - . f h . the fr .
tau freq
teg . .
d
h
b
hi
all
.
.
.
fi
£(1
7)
-
given
In
xpertment
I.
nelt
orm~
st
palTS,eac
conSlst-
. . Ing0 nouns omca ones aVing same presen on uency In
19.9,p 0; thus, Var[IA - IB] is smaller when probesA and
ing number of category members should increaseecho-inten- B are similar than when they are from different categories.
sity variability. The effect of related versusunrelated test pairs Although it was not Statisticallysignificant, there was a hint
was also anticipated, partly on theoretical grounds and partly of an interaction in the data of Experiment 2 that wasnot dupli-
becauseof similar findings reported earlier in the literature. catedby the model (seeFigure 12).There may havebeena ceil-
These matters will receive further discussionafter considering ing effect in the human data. There is little reasonto doubt that
simulations of the two experiments. the differencebetweenrelated and unrelated conditions is pres-
ent evenwhen only one categorymember wasoriginally stored.
Simulation of Experiments I and 2 lW'ving (1981) noted a consistentdifference betweensemanti-
cally related and unrelated conditions in severalpublishedstud-
The program used for the basic frequency-osed search mechanism would lead one
frequency-judgment than under recognition-memory instruc- expect.
tions, but this result has repeatedly failed to replicate (Begg et A multiple-trace model for both absolute judgments of f
al., 1986; Harris et al., 1980; Malmi, 1977), and so will not be quency and frequency discrimination was prq>osed by SchmJ
considered here. (1978). Much as in MINERVA 2, frequency estimates are bas
Another argument against the continuity of recency and fre- on the number of traces that the test item retrieves. The mo O} (recognItIon) versus Pr{Judgment = tal traces is negligible, but even list traces of low similarity tc I
2 }, with one point for each frequency and recency combination, probe will be activated by the probe to some degree, so a ma:
revealed functions that were different for frequency = I and 2. source of difficulty is in determining whether the retrieved j
When the presentation frequency was 2, the two values were tensity reflects activation of target-item traces or only of nont:
monotonically related in the predicted way, but when presenta- get-item traces from the list (cf. Gillund & Shiffrin, 1984; R;
tiop frequency was I, Pr{judgment = 2} was constant at about cliff, 1978).
.10, whereas Pr{judgment > O} varied over a range of .80 to A model devel~ by Ratcliff(1978) has been applied or
nearly 1.00. It is ~ though s.ubjects were reluctant to give ajudg- to recognition and not to memory for frequency, but it bel
ment of 2 to ~ I!em haVl~g freq~ency = 1, no matter how similarities to MINERVA 2. Repetition is assumed to give n
strong the familiarIty of the Item might be. to multiple traces, and a memory probe contacts all traces
W~lls's (1974) results are not what the pr~sent model would parallel, with each resonating according to its relatedness
are rc:asons t>.ecautIous
~redlct,but ther.e t:-vo to abo~taccept- timesarepredicted a cc
similarityto the probe.Decision by
mg her con~l.uslon. First, m .a runmng .frequenCY-Jud~ent tinuous diffusion process, similar to the discrete, random-wa I
task, prq>osltIonal representatIons regarding frequency will be mechanism described earlier (the primary focus of Ratclifl
en.c~ during pre.sentation and ma: play some role in deter- article is on recognition decision times). One difference t
.especl3l!y shortlags.,
mlmng ~ubsequent
Ju.dgments, at Second, the is is
tween two models that relatednessa primitiveconstru :
recency IS a confounding
fPr{ . d O} ".
factor m Wells s results. That Is, values
h . fr . f 2
. Rat I .ff '
m CI sm
odel alth
, ou
gh .t derives fr
I
-'
omuvt;llapplngle
. " -
0 JU gment > lor Items aVlng equenCles oland . .
t ures m MINE RV A. 2 A no ther IS that Rat c liff assumes th a t tI
were equated only when the former were tested at much shorter f h tr d . .ts . divi
'
dual
decisi
.
th t odel ...:'- t '
I
h
h
1
S
b
.
.
ood
disc
...
resonance
0
eac
ace
nves
I
own
m
on
pro
ags t an t e atter. u ~ects are quite g at nmmatIng h .
.. .. . cess w ereas m e presen m resonances or a"u.a Ion va
recenCles over Intervals of seconds to minutes (e.g., Hmnchs & ' led
Buschke, 1968). If Wells's subjects discriminated the recencies ues are poo . . ..
of items that exceeded c. , they could have adjusted C2as seemed In ~e ~rrent literature, the model that handles r~tI(
appr~riate for each item's recency. This possible explanation m~t .Slmilarly to ~INERVA 2 appears to be the Gillund ar
again underscores the need for an accounting of how criteria Shiffrin (1984) version of SAM. Devel~ment of the SAM mod
are set. has proceeded in different directions than the present wor
dealing with free and cued recall as well as with recogniti(
memory. To date, SAM has not been applied to schema abstra.
Comparison With Other Models tion or memory for frequency, although the latter has been SUI
gested as a natural extension of the recognition model (Gillun
Two other multiple-trace models of memory for frequency & Shiffrin, 1984). In SAM, as in MINER VA 2, traces are activate
have been prq>osed. Estes (1976; Whitlow & Estes, 1979) pos- according to similarity to the retrieving probe, and activatio
tulated a limited-capacity memory in which the total number of each trace is a positively accelerated function of the degrf
of traces is constrained by the number of contextual elements of overlap of the trace and the probe. MINERVA 2 and SAM ar
to which they can become attached. Frequency discrimination alike in summing the activation of all traces to make what Gi
is done by searching this memory for both alternatives, A and lund and Shiffrin call a global recognition decision. but SA1
B, and responding with the one that is discovered first. As it has retrieves and processes individual images in order to do recal
been devel~, the model does not apply to numerical fre- Despite the similarity in the two models' assumptions regardin
quency judgments, and so it is limited in ~. Moreover, it the basic recognition process, there are some differences in spe
appears to be inconsistent with several findings from the cific applications. The difference in approaches to the mirro
frequency-discrimination task (Hintzman et al., 1981): Sub- effect were noted earlier, and in SAM, forgetting is caused b
jects' performance often exceeds the strict maxima that the retroactive interference rather than by information loss. SAM'
model predicts; data do not show as much retroactive interfer- recognition performance is not affected by a shift in context
ence as the limited-capacity assumption implies; and the pri- because a matching context simply multiplies the activation 0
mary determinant of response latencies is the difference be- target traces and nontarget traces by the same amount, so tha
tween the frequencies of A and B, not their absolute frequen- signal-to-noise ratios remain the same. This was the resul
.
, .",
~.
~,.
MULTIPLE-TRACEMODEL 549
to noted earlier for the presentmodel when the exponentis deleted ceases once mastery is reached. A primary reason for taking
from Equation 2. a multiple-trace approach to effects of repetition has been to
re- McDowd and Murdock ( 1986)recently compared MINER A v explain this otherwiseanomaloussetof results.
dt 2 with Murdock's model, TODAM,in their fit to data from an
ed experiment by Avant and Bevan (1968). The experiment con- Relerences
ed th ffi .. f .. ~
leI cern e e ect on recogmtlon memory 0 varIation among
of training stimuli. There were 20 categoriesof nonsense shapes, J. J. R.
Anderson, A., Silverstein, W., Ritz, S. A., & Jones, S. (1977).
in and each subject saw four stimuli from each category.For one Distinctive categorical
features, and
perception, probability learning:
Ite group of subjects,all four stimuli were the category prototype; Some of
applications a neuralmodel. Ps~h%gicalRevi~ 84.413-
es other groups sawthe prototype 3, 2, or I time(s), supplemented 451.
:1- by from 1 to 3 distortions of the prototype to bring the total to M.,
Anisfeld, & Knapp,M. (1968). Association, and
synonymity, direc-
s- 4. Following presentation of the patterns, which the subjects tionalityin false JournalofExperimental
recognition. Ps~h%gy,77,
11- were simply told to learn, a recognition-memory test wasgiven 171-179.
that included the 20 rotot and 20 new distractors. Re- F.
Attneave, (1953). Psychological as
p~obability a functionof ex peri-
a . p ypes .. enced frequency.Journalof Experimental Ps~h%gy,46,81-86.
)r ported hIt rates showeda comple.xpattern In whIch the group L.
Avant, L., & Bevan, (1968).
W. of
Recognition a stimulus classmem-
}- that had seeneach prototype 4 times performed worst (71%), of
beraftertrainingwith variednumbers cases class. per Journalof
r- and the group that had seeneachprototype 3 times and a distor- General Ps~h%gy,78,241-246.
t- tion of each prototype 1 time performed best (84%). The other M.
Bain,J. D., & Humphreys, S.(in press). context:Inde-
Relational
two groups fell in between. pendent or In
cues,meanings, configurations? G. Davies D. M. &
Y Although neither model correctly predicted that the group ~omson (Eds.), Memoryin context: Contextin memo~ London:
OS that had studied only the prototypes would perform most Wiley. . .
e v
poorly, TODAMfit the data better than did MINER A 2. Indeed, ~.
Banks, P.( 1970~. theory
Slgnal-cessl~ and
Overview clOSIng com-
.t . I -
I n general, I IS d.ffiCult t 0 see hOWa cI0 sedI~ modeI (e.g., ments. In L. S. Cermak & F. I. M. Cralk (Eds.), Levels ofprocessing
. h
h . f .. .. odel In uman memory (pp. 447-461) .Hillsdal e,. . Erlba
NJ urn.
t eversIon 0 TODAM In question or a c~nnect1omstm Craik,F.I. M., & Lockhart,R. S.(1972). of A
Levels processing: frame-
basedon the delta rule) could handle expenments on memory M>rkfor memoryresearch. Journalof Verbal Learningand Verbal
for frequency,such as those that havebeen of central concern 11,
Behavio1; 671-684.
here. Judged frequency continues to increase with repetition, Eich, J. E. (1980).The cue-ciations memory for serial position. Journal of Experimenta,
Graf, P., Squire, L. R., & Mandler, G. (1984). The information that 97,
Ps}{;hology. 220-229.
amnesicpatients do not fo~. Journal of Experimental Ps}{;hology: Hintzman, D. L., & Gold, E. (1983). A congruity effectin the discrimi-
Learning, Memo'}: and Cognition, 10, 164-178. nation of presentationfrequencies:Somedata and a model. Bulletin
Greene, R. L. (1984). Incidental learning of event frequency.Memory of the Ps}{;honomic Societ)! 21, 11-14.
& Cognition, 12,90-95. Hintzman, D. L., Grandy, C. A., & Gold, E. (1981). Memory for &e-
Groninger,L. D. (1976). Predicting recognition during storage:The ca- quency: A comparison of~ multiple-trace theories.Journal of Ex-
pacity of the memory systemto evaluateitself. Bulletin of the PS}{;ho- perimental PS}{;hology: Human Learning and Memo'}: 7, 231-240.
nomicSociet)! 7,425-428. Hintzman, D. L., Nozawa,G., & Irmscher, M. (1982). Frequencyas a
Hall, J. F. (1979). Recognition as a function of word frequency.Ameri- nonpr~tional attribute of memory. Journal of VerbalLearning
can JournalofPs}{;hology. 92,497-505. and VerbalBehaviol; 21, 127-141.
Hall, J. F., & Kozloff, E. E. (1970). Falserecognitions as a function of Hintzman, D. L., & Stern, L. D. (1978). Contextual variability and
number of presentations.American Journal of Ps}{;hology. 272-83, memory for frequency.Journal of Experimental Ps}{;hology: Human
279. Learning and Memo'}: 4, 539-549.
Hall, J. F., & KozlolI; E. E. (1973). Falserecognitions of a.w>ciates of Hintzman, D. L., & Stern, L. D. (1984). A comparison of forgetting
convergingversusrepeatedwords. American Journal of Ps}{;hology. rates in frequency discriminatiQn and recognition. Bulletin of the
86, 133-139. Ps}{;honomicSociet)!22, 409-412.
Harris, G., 8egg, I., & Mitterer, J. (1980). On the relation bet\veenfre- Hintzman, D. L., Summers,J. J., & Block, R. A. (1975). What causes
quency estimatesand recognition memory. Memory & Cognition, 8, the spacingeffect?Some effectsof repetition, duration, and spacing
99-104. on memory for pictures. Memory & Cognition, 3, 287-294.
Hart, J. T. (1965). Memory and the feeling-of-knowing experience. Hock, H. S., Malcus, L., & Hasher,L. (1986). Frequencydiscrimina-
56,
Journal of Educational Ps}{;hology. 208-216. tion: Assessing global-leveland element-levelunits in memory.Jour-
of
Hasher,L., & Chromiak, W. (1977). The processing frequencyinfor- naI of Experimental PS}{;h%gy: Learning, Memo'}: and Cognition,
mation: An automatic mechanism?Journal of VerbalLearning and 12, 232-240.
VerbalBehaviol; 16, 173-184. Hockley, W. E. (1984). Retrieval of item frequency information in a
Hasher, L., & Zacks, R. T. (1979). Automatic and effortful processes continuous memory task. Memory &; Cognition, 12, 229-242.
in memory. Journal of Experimental Ps}{;hology: General,108, 356- H~ll, W. C. (1973). Storageof eventsand event frequencies:A com-
388. parison of ~ paradigms in memory. Journal of Experimental Psy--
Hasher,L., & Zacks,R. T. (1984). Automatic processing offundamental chology.98, 260-263.
information: The caseof frequencyof occurrence.American Ps}{;hol- Jacoby,L. L., & Brooks, L. R. (1984). NonanaIytic cognition: Memory,
ogist, 39, 1372-1388. perception,and concept learning. In G. H. ~ (Ed.), Theps}{;hol-
Hinrichs, J. V., & Buschke, H. (1968). Judgment of recency under ogy of learning and motimtion (Vol. 18, pp. 1-47). New York: Aca-
steady-stateconditions. Journal of Experimental Ps}{;hology. 78, demic Press.
574-579. Johnson, M. K., Ra)\:, C. L., Wang,A. Y., & Taylor,T. H. (1979). Fact
Hintzman, D. L. (1969). Apparent frequencyasa function offrequency and fantasy:The roles of accuracyand variability in confusingimagi-
and the spacingof repetitions. Journal of Experimental Ps}{;hology. nations with perceptual experiences.Journal of Experimental Psy--
80,139-145. ch%gy: Human Learning and Memo'}: 5, 229-240.
Hintzman, D. L. (1974). Theoretical implications of the spacingeffect. Kausler, D. H., Lichty, W., & Hakami, M. K. (1984). Frequencyjudg-
In R. L. Solso (Ed.), Theories in cognitive PS}{;h%gy: The Loyola mentsfor distractor items in a short-term memory task: Instructional
Symposium(pp. 77-99). Potomac,MD: Erlbaum. wriation and adult agedifferences.Journal of VerbalLearning and
Hintzman, D. L. (1976). Repetition and memory. In G. H. ~ (Ed.), VerbalBehaviol; 23, 660-668.
ThePS}{;h%gyofiearningand motivation (Vol. 10,pp. 47-91). New Koflka, K. (1935). PrincipiesofGestaltps}{;holOg}( New York: Harcourt
York: Academic Press. BraceJavanovich.
.